A multi-working-condition online monitoring and regulation system and method suitable for waste plastic resourceization production line

Abstract

The present application relates to the technical field of waste plastic treatment, and discloses a multi-working-condition online monitoring and regulation system and method suitable for waste plastic resource production lines, which comprises the following steps: real-time evaluation of material composition, moisture content, particle size distribution and impurity content through multi-sensor fusion of near-infrared spectroscopy, machine vision and microwave humidity; automatic discovery of working condition categories based on density peak clustering, detection of working condition migration and updating of a knowledge base; full-process simulation and parameter optimization driven by digital twinning, construction of a multi-physical field coupling digital twinning model, acceleration of multi-objective optimization through a proxy model; multi-process collaborative regulation based on deep reinforcement learning, and multi-objective collaborative optimization of yield, quality and energy consumption through multi-agent reinforcement learning. The present application can improve the efficiency of waste plastic resource production lines and improve product quality and operating efficiency.

Description

Technical Field

[0001] This invention relates to the field of waste plastic treatment technology, specifically to a multi-condition online monitoring and control system and method adapted to waste plastic resource recovery production lines. Background Technology

[0002] With the rapid development of the plastics industry and the widespread use of plastic products, the amount of waste plastics generated continues to grow. The resource utilization of waste plastics has become an important way to achieve a circular economy and sustainable development. Waste plastic resource utilization production lines typically include multiple processes such as crushing, washing, sorting, extrusion, and granulation, using physical or chemical methods to transform waste plastics into recycled plastic pellets or other recycled products.

[0003] However, existing waste plastic recycling production lines face the following technical problems in actual operation: Waste plastic raw materials have complex sources and variable compositions. Traditional production lines mainly rely on manual sampling and offline testing, which cannot monitor key parameters such as material composition, moisture content, particle size distribution, and impurity content in real time. This results in process adjustments lagging behind material changes and significant fluctuations in product quality. Although some production lines have introduced single sensors (such as metal detectors and weight sensors), these sensors can only detect one aspect of information and cannot comprehensively characterize the material's state, let alone achieve effective fusion of multi-source information.

[0004] The operating conditions of waste plastic production lines are affected by various factors such as raw material type, moisture content, impurity content, and equipment condition, resulting in numerous and dynamically changing operating conditions. Most existing production lines use fixed operating condition classifications and preset parameter schemes, which cannot automatically identify the current operating condition type, let alone detect changes in operating conditions in a timely manner. When operating conditions change, continuing to use the original parameter scheme leads to decreased production efficiency and unstable product quality.

[0005] The process parameter settings of existing production lines mainly rely on the experience of operators, lacking simulation optimization methods based on physical models. Since the waste plastic recycling process involves the coupling of multiple physical fields such as heat, force, fluid, and chemistry, there are complex nonlinear relationships between the parameters of each process. It is difficult to find the globally optimal parameter combination based solely on human experience, resulting in low production line operating efficiency, high energy consumption, and unstable product quality.

[0006] Waste plastic recycling production lines involve multiple processes such as crushing, washing, sorting, extrusion, and granulation, with strong coupling relationships between these processes. Most existing production lines adopt an independent control approach for each process, lacking a global collaborative optimization mechanism. Parameter adjustments in upstream processes often fail to consider the impact on downstream processes, resulting in poor overall operational performance and difficulty in achieving collaborative optimization of multiple objectives such as output, quality, and energy consumption.

[0007] Existing production line anomaly detection mainly relies on threshold alarms and manual inspections, which can only detect obvious anomalies and struggle to identify early, subtle anomalies. When an anomaly occurs, fault diagnosis depends on the experience of maintenance personnel, resulting in low efficiency, poor accuracy, and a lack of automatic self-healing capabilities. This leads to long anomaly handling cycles, significant unplanned downtime, and substantial production losses. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-condition online monitoring and control system and method adapted to waste plastic resource recovery production lines, thereby solving the problems in the prior art.

[0009] The present invention adopts the following technical solution: A multi-condition online monitoring and control method adapted to waste plastic recycling production lines includes the following steps: S1. Collect multimodal data of waste plastic materials through multiple sensors, and evaluate the composition, moisture content, particle size distribution and impurity content of the materials in real time after multimodal feature fusion to generate a comprehensive characterization of the material state. S2. Based on the joint feature space of material status and operating parameters, the current working condition type of the production line is automatically identified through unsupervised clustering, and working condition migration events are detected to dynamically update the working condition classification system. S3. Construct a multi-physics coupled digital twin model of the production line, perform full-process simulation based on the current working conditions and material status, and obtain the optimal combination of process parameters for each process through simulation optimization. S4. The optimization of multi-process parameters is modeled as a multi-agent reinforcement learning problem. The agents of each process make collaborative decisions to generate the optimal control strategy under the constraints of multiple objectives such as output, quality and energy consumption. S5. Real-time monitoring of production line operation status, using graph neural networks and knowledge graphs to trace the root causes of anomalies, assess the severity of anomalies, and generate self-healing control strategies. S6. Construct a soft measurement model for product quality based on process parameters, predict quality indicators in real time, and achieve rapid correction of quality deviations through feedforward-feedback composite control; S7 performs real-time data processing and rapid adjustment at the edge, performs big data analysis and model training in the cloud, and achieves edge-cloud collaborative optimization through model distillation and dynamic distribution.

[0010] Preferably, the data acquisition in step S1 specifically includes: S11. Multimodal data acquisition: Material spectral data is acquired through a near-infrared spectral sensor, material visual images are acquired through an industrial camera, moisture content data is acquired through a microwave humidity sensor, flow rate data is acquired through a weighing sensor, and temperature data is acquired through a temperature sensor. S12. Single-mode feature extraction: Feature extraction is performed on each mode of data. For spectral data, features such as absorption peak position, peak height, and peak area are extracted. For image data, features such as color, texture, and shape are extracted. For waveform data, time-domain and frequency-domain features are extracted. S13. Multimodal feature fusion: A multimodal fusion network based on attention mechanism is adopted to adaptively learn the weights of each modality feature and fuse the multimodal features into a unified material state feature vector. S14. Quantitative assessment of material state: Based on the fused feature vector, the material is quantitatively assessed through a regression model, including key indicators such as the proportion of plastic types, moisture content, particle size distribution, and impurity content, to generate a comprehensive material quality score.

[0011] Preferably, the initial screening method in step S2 specifically includes: S21. Construction of Joint Feature Space: The material state features are combined with the production line operating parameter features to construct a high-dimensional joint feature space; S22. Density Peak Clustering for Working Condition Discovery: The density peak clustering algorithm is used to automatically discover working condition categories in the joint feature space without pre-specifying the number of categories. The optimal number of clusters is automatically determined through the decision graph. S23. Working condition migration detection: Based on statistical process control methods, it monitors changes in feature distribution. When the distribution deviation exceeds the threshold, it is determined to be a working condition migration, triggering a reclassification of the working condition. S24. Working Condition Knowledge Base Update: Newly identified working conditions and their corresponding optimal parameter combinations are stored in the working condition knowledge base, which is continuously updated and expanded through online learning.

[0012] Preferably, step S3 specifically includes: S31. Multiphysics Coupled Modeling: Construct a multiphysics coupled digital twin model that includes thermal field, flow field, force field, and chemical reaction field to describe the physicochemical processes of each step, such as crushing, washing, extrusion, and granulation. S32. Hybrid simulation calculation: A hybrid simulation strategy combining the finite element method and the discrete element method is adopted. The finite element method is used to solve the continuous medium, and the discrete element method is used to solve the particulate material, thereby improving the simulation accuracy and efficiency. S33, Proxy Model Acceleration: Based on simulation samples, train the Kriging proxy model and use the proxy model to replace part of the time-consuming full-size simulation, thereby accelerating the optimization iteration process; S34. Multi-objective parameter optimization: With the objectives of maximizing output, optimizing quality, and minimizing energy consumption, a multi-objective evolutionary algorithm is used to search for the optimal combination of process parameters in the simulation environment to generate a Pareto optimal solution set.

[0013] Preferably, step S4 specifically includes: S41. Multi-agent modeling: Each process of the production line is modeled as an independent reinforcement learning agent, with each agent responsible for parameter control decisions for its own process; S42. Centralized training and distributed execution: A centralized training and distributed execution architecture is adopted. During training, global information is used for centralized value evaluation, and during execution, each agent makes distributed decisions based only on local observations. S43. Multi-objective reward design: Design a multi-objective reward function that includes output reward, quality reward, and energy consumption penalty, and adjust the importance of each objective through weighting coefficients; S44. Strategy Transfer and Fine-tuning: Transfer the strategy trained in the simulation environment to the actual production line, and use the actual operating data for online fine-tuning to quickly adapt to the real environment.

[0014] Preferably, step S5 specifically includes: S51. Anomaly Detection and Localization: Construct a probabilistic model of normal operating conditions based on a variational autoencoder, detect anomalies by reconstructing errors, and locate the process and parameters where the anomaly occurs through an attention mechanism. S52. Graph Neural Network Root Cause Analysis: Model each process and parameter of the production line as a graph structure, use graph neural networks to propagate abnormal information, and trace the root cause of the abnormality. S53. Knowledge Graph Fault Reasoning: Construct a knowledge graph for fault mode and impact analysis, and diagnose fault types and severity based on graph matching and rule reasoning; S54. Self-healing regulation strategy generation: Based on the diagnostic results, the optimal self-healing regulation strategy is matched from the strategy library, or a regulation strategy is generated in real time through reinforcement learning, and the regulation action is automatically executed.

[0015] Preferably, step S6 specifically includes: S61. Soft Measurement Model Construction: A product quality soft measurement model is constructed using a deep belief network, with process parameters that can be measured online as input and quality indicators that are difficult to measure online as output. S62. Quality Deviation Prediction: Based on the current process parameters and material status, predict the trend of quality index changes in the future period and calculate the estimated quality deviation. S63. Feedforward control: Based on the estimated quality deviation, the adjustment amount of each parameter is calculated based on sensitivity analysis, and the quality deviation is corrected in advance through feedforward control. S64. Feedback Correction: Based on the actual quality inspection results, feedback correction is performed through model predictive control to form a feedforward-feedback composite closed-loop control.

[0016] Preferably, it also includes intelligent energy consumption optimization steps: S81. Real-time monitoring and decomposition of energy consumption: Real-time collection of energy consumption data for each process, decomposition of total energy consumption into energy consumption components for each piece of equipment and each link, and establishment of an energy consumption decomposition model. S82. Dynamic adjustment of energy consumption baseline: Based on the current operating condition, dynamically adjust the energy consumption baseline and calculate the deviation between the actual energy consumption and the baseline. S83. Energy-saving optimization strategy generation: Based on the operating condition-energy consumption correlation model, generate energy-saving optimization strategies for the current operating condition, including parameter adjustment, equipment start-up and shutdown optimization, load allocation, etc. S84. Energy Efficiency Assessment and Feedback: Regularly assess the effectiveness of energy efficiency improvements, feed the energy-saving effects back to the optimization model, and continuously improve energy-saving strategies.

[0017] Preferably, it also includes virtual sample generation and data augmentation steps: S91, Virtual Samples for Mechanism Simulation: Virtual operating samples under different working conditions and parameter combinations are generated based on physical mechanism models. S92. Enhanced Generative Adversarial Network: Conditional Generative Adversarial Networks are used to learn the distribution of real-world data and generate high-quality virtual samples. S93. Sample quality assessment: Evaluate the quality of virtual samples from dimensions such as distribution consistency, physical rationality, and diversity, and screen high-quality samples; S94, Hybrid Training Enhancement: Train the model using a mix of real and virtual samples to improve the model's generalization ability under small sample sizes or rare conditions.

[0018] Preferably, a multi-condition online monitoring and control system adapted to a waste plastic recycling production line includes: The system comprises several modules: a multi-source sensing module integrating near-infrared spectroscopy, machine vision, microwave humidity, weighing, and temperature sensors for collecting multimodal data on materials and production line operation data; a material state assessment module for feature fusion of multi-source data to quantitatively assess material composition, moisture content, particle size distribution, and impurity content; a working condition identification module for automatically identifying the current working condition type of the production line, detecting working condition migration, and maintaining a working condition knowledge base; a digital twin simulation module for constructing a multi-physics coupled digital twin model of the production line for full-process simulation and parameter optimization; a collaborative control module for achieving multi-process collaborative control based on deep reinforcement learning to optimize output, quality, and energy consumption; an anomaly diagnosis and self-healing module for anomaly detection, root cause tracing, and generation of self-healing control strategies; a quality closed-loop control module for soft measurement of product quality and feedforward-feedback composite control; and an edge-cloud collaboration module for optimizing real-time processing at the edge and big data analysis in the cloud.

[0019] The beneficial effects of this invention are as follows: By leveraging the collaborative work of multiple sensors, including near-infrared spectroscopy, machine vision, microwave humidity, weighing, and temperature sensors, combined with attention-based multimodal feature fusion technology, it can comprehensively, in real-time, and quantitatively assess key state parameters of materials, such as composition, moisture content, particle size distribution, and impurity content. This improves the accuracy of material composition identification and solves the problems of traditional monitoring methods being single, lagging, and unable to comprehensively characterize the material state. The density peak clustering algorithm automatically identifies operating condition categories without requiring pre-specifying the number of categories, adapting to the variable characteristics of waste plastic raw materials. The operating condition migration detection based on statistical process control can promptly detect changes in operating conditions with a low false alarm rate. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the control system architecture of the present invention.

[0021] Figure 2 This is a schematic diagram of the control method of the present invention.

[0022] Figure 3 This is a flowchart of the online material status sensing method of the present invention. Detailed Implementation

[0023] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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 are within the scope of protection of the present invention.

[0024] Example 1: refer to Figure 1-3 As shown, a multi-condition online monitoring and control method adapted to waste plastic recycling production lines is proposed. This method achieves refined online perception of material status through multi-source heterogeneous data fusion, adaptive identification and dynamic classification of operating conditions based on density peak clustering and statistical process control, full-process simulation and parameter optimization using a multi-physics coupled digital twin model, collaborative control of multiple processes using multi-agent deep reinforcement learning, intelligent diagnosis and self-healing control of abnormal operating conditions by combining graph neural networks and knowledge graphs, a feedforward-feedback composite closed-loop quality control system constructed through deep belief network soft measurement and model predictive control, and distributed intelligent control achieved through an edge-cloud collaborative architecture. Furthermore, it incorporates intelligent energy consumption optimization and virtual sample data enhancement mechanisms to comprehensively improve the operating efficiency, product quality, and intelligence level of waste plastic recycling production lines under multiple operating conditions.

[0025] In the online sensing of material status through multi-source heterogeneous data fusion, multiple sensors deployed at key nodes of the production line collaboratively collect multimodal data of waste plastic materials. The near-infrared spectral sensor acquires continuous spectral data of materials in the 900-2500nm near-infrared band using diffuse reflection, with a spectral resolution of no less than 10nm, a sampling interval of 2nm, and at least 32 scans to improve the signal-to-noise ratio. The industrial camera uses a linear or area array CCD / CMOS sensor to simultaneously acquire RGB color and depth images of the materials, with an image resolution of no less than 1920×1080 and a frame rate of no less than 30fps. It is equipped with a ring LED light source to eliminate the influence of shadows. The microwave humidity sensor inverts the moisture content by measuring the attenuation and phase shift of microwaves in the material, with a measurement range covering 0-50%, a measurement accuracy of ±1%, and a response time of less than 500ms. The weighing sensor adopts a high-precision strain gauge structure to acquire the instantaneous flow rate and cumulative weight of the material in real time, with an accuracy of ±0.5%FS. The infrared temperature sensor adopts a non-contact temperature measurement method, with a measurement range of 0-200℃ and an accuracy of ±2℃, and is used to monitor the surface temperature distribution of materials. In addition, vibration sensors, current sensors, pressure sensors and other sensors are used to collect equipment operating status data to form a multi-dimensional material-equipment joint sensing system.

[0026] Before entering the fusion network, each modal data requires separate single-modal feature extraction. Spectral data undergoes standard normal transformation to eliminate scattering effects, multivariate scattering correction for baseline drift, and Savitzky-Golay polynomial smoothing to remove random noise. After preprocessing, a 48-dimensional spectral feature vector is extracted, including parameters such as the position, peak height, peak area, full width at half maximum (FWHM), and peak shape asymmetry coefficient of each absorption peak, as well as normalized reflectance features after continuum removal. Image data undergoes median filtering for denoising, histogram equalization enhancement, and particle segmentation based on thresholding and morphological operations, resulting in a 68-dimensional visual feature vector. This vector covers color histograms, gray-level co-occurrence matrix texture features, shape features, and particle size distribution statistics. Waveform data undergoes temporal and frequency domain feature extraction, totaling 25 dimensions of waveform features.

[0027] Multimodal feature fusion employs an attention-based multimodal fusion network architecture, comprising three parallel feature encoding channels: a spectral feature branch, an image feature branch, and a waveform feature branch. The spectral branch consists of three one-dimensional convolutional layers and two fully connected layers, with kernel sizes of 7, 5, and 3, and channel numbers of 64, 128, and 256, respectively, used to extract local absorption pattern features of the spectrum. The image branch uses a lightweight convolutional neural network (MobileNetV3 architecture) to encode the input image into a 128-dimensional deep visual feature vector. The waveform branch consists of two fully connected layers, mapping 25-dimensional waveform features to 64-dimensional abstract features. The output features of the three branches are adaptively weighted and fused through a cross-modal attention module, which uses a scaled dot product attention mechanism, with attention weights satisfying the formula: Where Q is the query vector (using spectral features as the query), K is the key vector (using image features as the key), V is the value vector (using waveform features as the value), and d_k is the dimension of the key vector. Dividing by √d_k is used to scale the dot product value to avoid gradient vanishing. The attention weights are normalized using the softmax function, and finally, the weighted multimodal features are concatenated into a 256-dimensional unified material state feature vector. The introduction of the attention mechanism enables the network to automatically adjust the contribution of each modality according to the material type and operating conditions. For example, for the moisture content detection task, it automatically increases the weight of microwave humidity and spectral features, and for the impurity identification task, it automatically increases the weight of visual features.

[0028] The quantitative assessment of material condition is based on a 256-dimensional fused feature vector. A multi-output regression model is constructed using the Gradient Boosting Regression Tree (GBDT) algorithm to quantitatively assess various key indicators of the material. The model output includes 28 indicators across six categories: plastic type percentage (mass fractions of 8 major plastics including PET, HDPE, LDPE, PP, PS, PVC, ABS, and PC), moisture content (mass percentage), particle size distribution (D10, D50, D90, uniformity coefficient, and curvature coefficient), impurity content (mass fractions of 4 impurities including metal, paper, sediment, and organic matter), material temperature (average surface temperature and temperature standard deviation), and a comprehensive material quality score (0-100 points). Model training employs a hybrid approach using real and virtual samples. Real samples are derived from historical production line testing data, while virtual samples are generated based on mechanism simulation. The overall material quality score is calculated using a weighted summation method, satisfying the formula: Score = 0.4 × S_purity + 0.2 × S_moisture + 0.2 × S_size + 0.2 × S_impurity, where S_purity is the purity score, S_moisture is the moisture content score, S_size is the particle size uniformity score, and S_impurity is the impurity content score. Each score is normalized to the range of 0-100. The score result also outputs a confidence interval; when the confidence level falls below a threshold, a manual review mechanism is triggered.

[0029] In the multi-condition adaptive identification and dynamic classification process, a high-dimensional joint feature space is first constructed as the foundation for condition identification. The aforementioned 256-dimensional material state feature vector is concatenated with the production line operating parameter feature vector. The operating parameter features cover key process parameters and equipment state parameters for each stage, including temperature, pressure, speed, current, vibration, and liquid level, totaling approximately 128 parameter features. The concatenation of material features and operating parameter features forms a 384-dimensional high-dimensional joint feature space. To reduce computational complexity and eliminate feature redundancy, principal component analysis (PCA) is used to reduce the feature space to 64 dimensions, retaining over 95% of the variance information. During dimensionality reduction, the contribution rate and cumulative contribution rate of each principal component, as well as the loadings of each original feature on the principal components, are calculated simultaneously for subsequent feature importance analysis and condition interpretation.

[0030] The automatic working condition detection employs a density peak clustering algorithm to perform unsupervised clustering in the dimensionality-reduced 64-dimensional feature space. The algorithm first calculates the local density ρ_i for each sample point using a Gaussian kernel: ρ_i = Σ_{j≠ Where d_ij is the distance between sample point i and sample point j, and d_c is the cutoff distance (the 2nd quantile of all sample point distances). The Gaussian kernel ensures that samples that are closer together contribute more to the density. The distance metric uses Mahalanobis distance to eliminate the influence of differences in dimensions. Mahalanobis distance satisfies the formula: d_ √ Σ is the sample covariance matrix. Mahalanobis distance is invariant to the distribution shape of the data and can effectively handle the correlation between features. Then, the relative distance δ_i of each sample point is calculated, which is the minimum distance from that point to all sample points with higher density; for the point with the highest density, δ_i is the maximum distance to all other points. A decision graph is plotted with ρ as the horizontal axis and δ as the vertical axis. The point located in the upper right corner of the decision graph (which has both high density and large relative distance) is determined as the cluster center. The number of cluster centers is automatically determined by a threshold γ_i = ρ_i × δ_i, where the threshold is the inflection point of the γ value sequence. The remaining sample points are assigned to the corresponding clusters according to the principle of "higher density and closest distance". Each cluster corresponds to a working condition type. To improve the stability of clustering, an incremental clustering strategy based on a sliding window is adopted. The window size is 1000 samples, and the sliding step size is 100 samples. After each sliding, the clustering results are recalculated, and the clusters of the previous and next windows are matched using the Hungarian algorithm to maintain the consistency of the working condition labels.

[0031] The operating condition migration detection method is based on the Statistical Process Control (SPC) approach, employing a Multivariate Cumulative Sum (MCUSUM) control chart to monitor changes in the characteristic distribution. Using the characteristic mean vector μ_0 and covariance matrix Σ_0 of the current normal operating condition as a benchmark, the Mahalanobis distance between the characteristic mean vector and the benchmark mean vector is calculated within a sliding time window (window size of 50 samples, sampling interval of 1 minute). The MCUSUM statistic satisfies the formula: C_t Where C_t is the cumulative sum statistic at time t, d_ Let C_t be the Mahalanobis distance between the sample and the baseline mean at time t, and k be a reference value (usually half of the allowable offset; in this invention, k = 0.5σ). When C_t exceeds the control limit h, it is determined as a working condition transition event, triggering a reclassification of the working condition. The control limit h is determined by the kernel density estimation method, using the 99th percentile of normal fluctuations in historical data as the initial control limit, with the false alarm rate controlled below 1%. After detecting a working condition transition, the system automatically initiates the working condition re-identification process, adding the new data after the transition to cluster analysis to determine the new working condition type and its corresponding parameter range. Simultaneously, a Hidden Markov Model (HMM) is used to model the working condition transition pattern, treating each working condition type as a hidden state and the observed features as the observed state. The state transition probability matrix and observation probability matrix are learned using the Baum-Welch algorithm, and the Viterbi algorithm is used to predict the most likely working condition transition path in the future, providing a basis for feedforward regulation.

[0032] The operating condition knowledge base is constructed and dynamically updated using a case-based reasoning (CBR) mechanism. The knowledge base stores information such as feature descriptions, typical parameter ranges, optimal operating parameter combinations, historical operating records, failure rates, quality performance, and energy consumption levels for each operating condition type in a structured format. Each operating condition case includes fields such as operating condition identifier, feature center vector, feature covariance matrix, optimal parameter set, performance indicators, and applicable conditions. When a new operating condition is identified, the system first searches the knowledge base for the most similar historical operating condition case using k-nearest neighbor retrieval, with the similarity satisfying the formula: Where d_M is the Mahalanobis distance and σ_s is the similarity bandwidth parameter. The optimal parameters of the k most similar cases are weighted and averaged according to similarity as the initial recommendation parameters; then, they are adjusted online based on actual operating results, and the adjusted results are stored as new cases in the knowledge base. The knowledge base is updated using an incremental learning strategy, with a full update every 24 hours, or updated immediately when a new operating condition is detected. During the update process, the case base is maintained simultaneously, including merging redundant cases, eliminating outdated cases, and adjusting case weights, to maintain the simplicity and effectiveness of the knowledge base. The confidence of the operating condition-parameter mapping relationship in the knowledge base is dynamically calculated based on the number of cases and verification results, and mapping relationships with high confidence are given priority for control decisions.

[0033] The full-process simulation and parameter optimization driven by digital twins first constructs a multi-physics coupled digital twin model. The model covers all key processes of the production line, with each process employing a corresponding modeling method based on its physical mechanism. For the crushing process, a discrete element method (DEM) is used to establish a particle crushing model, simulating the motion, collision, and crushing process of particles within the crusher. The model considers the particle's shape, size, hardness, brittleness, and other physical properties, as well as the crusher's speed, clearance, and liner shape, among other structural parameters. A bonded particle model (BPM) is used to simulate particle crushing behavior, characterizing particle crushing through the fracture of parallel bonds. Particle contact is handled using the Hertz-Mindlin contact model, with the normal contact force satisfying the formula: ), where E* is the equivalent elastic modulus, R* is the equivalent radius, and δ_n is the normal overlap. The cleaning process employs a computational fluid dynamics-discrete element method (CFD-DEM), and the fluid phase is described by the Navier-Stokes equations, satisfying the formula: ∂ (Continuity equation), ∂ (Momentum equation), where α_f is the fluid volume fraction, ρ_f is the fluid density, u_f is the fluid velocity, p is the pressure, τ_f is the viscous stress tensor, and F_interaction is the fluid-structure interaction force; the particulate phase is described by Newton's laws of motion, and momentum exchange occurs between the two phases through drag, lift, and pressure gradient forces, simulating the interaction between water flow and plastic particles, impurity particles, and the flotation separation process. The extrusion process uses a thermo-fluid coupling model, based on the non-Newtonian fluid constitutive equation (Cross model) to describe the rheological behavior of the molten plastic. Cross model: Where η is viscosity, Let η_0 be the shear rate, η_∞ be the zero-shear viscosity, η_∞ be the infinite-shear viscosity, K be the time constant, and n be the power-law exponent. Coupled energy equations are used to solve for the temperature and velocity field distributions, while considering viscous dissipation and heat conduction effects. The granulation process employs a thermo-mechanical coupling model to simulate the entire process of molten plastic being extruded through a die, cut by a pelletizer, and cooled and shaped into granules, considering physical processes such as melt stretching, cooling and solidification, and granule deformation. The models for each process are coupled through material flow and energy flow interfaces to achieve coherent simulation of the entire production line.

[0034] Hybrid simulation employs a strategy combining the Finite Element Method (FEM) and the Discrete Element Method (DEM), selecting appropriate numerical methods based on the characteristics of each physical domain. For continuous medium domains (such as molten plastic, water flow, and air), the FEM or Finite Volume Method is used to solve partial differential equations, employing structured or unstructured meshes, and explicit or implicit time discretization schemes. For discrete particle domains (such as plastic fragments, particles, and impurities), the DEM is used to solve Newton's equations of motion, with a soft sphere model used for particle contact, and contact forces including normal elastic force, normal damping force, tangential friction force, and tangential damping force. At the interface between the continuous medium and discrete particles, a coupling algorithm is used to achieve information exchange, including the fluid's force on the particles and the particles' reaction force on the fluid (momentum source term). The coupling method adopts an Eulerian-Lagrange framework, with the fluid phase solved on an Eulerian mesh and the particle phase tracked in a Lagrange coordinate system. To improve computational efficiency, an adaptive time step strategy is adopted, which automatically adjusts the time step based on CFL conditions and particle collision frequency. At the same time, parallel computing technology is adopted, which divides the computational domain into multiple subdomains based on the domain decomposition method, and solves them in parallel on different processors. Data exchange between subdomains is carried out through message passing interface (MPI).

[0035] The surrogate model acceleration mechanism is built based on the Kriging method to replace the time-consuming high-precision simulation model and accelerate the optimization process. Sample points are selected using the Latin hypercube sampling (LHS) method, with sampling points evenly distributed within the design space, and an initial sample size of 50-100. The performance indicators (yield, quality, energy consumption, etc.) corresponding to each sample point are obtained through high-precision simulation calculations. The Kriging model takes process parameters as input and performance indicators as output, employing a Gaussian process regression framework. The predicted mean of the Kriging model satisfies the formula: Where μ is the process mean, r(x) is the correlation vector between the point to be predicted x and the sample points, R is the correlation matrix between the sample points, Y is the output value vector of the sample points, and 1 represents an all-1 vector. The prediction variance satisfies the formula: , where σ 2 Let Variance be the process variance. The model's predicted mean is used to estimate the performance index value, and the predicted variance is used to estimate the prediction uncertainty. To progressively improve the accuracy of the surrogate model, a sequential sampling strategy is adopted, adaptively selecting the next sampling point based on the expected improvement (EI) criterion. The expected improvement satisfies the formula: Where f_min is the current optimal value, Φ is the standard normal cumulative function, and φ is the standard normal probability density function. This means supplementing sampling points in regions with high prediction uncertainty but potential for improvement. The Kriging model is retrained after each supplementary sampling point until the model's cross-validation error is below 5% or the maximum sampling budget is reached. For multi-output problems, a multi-output Kriging model is used, or multiple single-output Kriging models are independently constructed. The correlation between output variables is considered through the covariance function between the models.

[0036] Multi-objective parameter optimization aims to maximize output, optimize quality, and minimize energy consumption. It employs a Non-Dominated Ranking Genetic Algorithm (NSGA-II) with an elitist strategy to perform optimization search on a surrogate model. Optimization variables include key adjustable parameters for each process, such as crushing speed, washing water temperature, flotation density, extrusion temperature, screw speed, and pelletizing speed, totaling 12 parameters, each with upper and lower limits. The objective function includes output, quality, and energy consumption. The core of the NSGA-II algorithm is fast non-dominated ranking and crowding distance calculation. Non-dominated ranking stratifies the population according to Pareto dominance relationships. Solutions not dominated by any other solution constitute the first layer of non-dominated fronts (Pareto fronts), and so on, obtaining each subsequent layer of non-dominated fronts. The crowding distance satisfies the formula: Where M is the number of objectives, and f_m(i+1) and f_m(i-1) are the values ​​of the i-th solution's immediate and next neighbors on the m-th objective. and These represent the maximum and minimum values ​​for the m-th objective, respectively. A larger crowding distance indicates a more uniform distribution of solutions on that front. The algorithm population size is set to 100, and the number of generations is set to 200. Offspring are generated through binary tournament selection, simulated binary crossover, and polynomial mutation. After merging parent and offspring generations, non-dominated sorting and crowding calculation are performed, and the top N individuals are selected for the next generation. After optimization, a Pareto optimal solution set is generated, containing approximately 100 combinations of non-dominated parameters. Decision-makers can select the most suitable parameter combination from the Pareto solution set based on actual production needs (such as order priority, raw material costs, electricity prices, etc.). Simultaneously, the system provides multi-criteria decision support, ranking and recommending Pareto solutions based on the TOPSIS method or the Analytic Hierarchy Process (AHP).

[0037] Based on deep reinforcement learning, the multi-process collaborative control mechanism first models each process of the production line as an independent reinforcement learning agent. A typical waste plastic recycling production line usually includes 5-8 key processes such as crushing, washing, sorting, extrusion, and granulation, with each process corresponding to one agent. The multi-process collaborative control problem is modeled as a Markov Decision Process (MDP), described by a quintuple (S, A, P, R, γ) consisting of a state space S, an action space A, state transition probabilities P, a reward function R, and a discount factor γ. Each agent's observation space includes the operating parameters of its current process, the state parameters of the materials entering its current process, and the key parameters of adjacent processes, with an observation dimension of approximately 30-50 dimensions. The action space includes adjustable parameters of its current process, such as speed setpoints, temperature setpoints, pressure setpoints, and feed rate, with each action dimension being a continuous value, and an action dimension of approximately 5-10 dimensions. The actions of each agent are transmitted to a PLC or DCS system via actuators to achieve real-time control of the production line.

[0038] Multi-agent collaboration employs a Centralized Training and Distributed Execution (CTDE) architecture, specifically based on the Multi-Agent Deep Deterministic Policy Gradient Algorithm (MADDPG). During training, a centralized critic network is used. This network takes global state information (observations and actions of all agents) as input, evaluates the value of joint actions, and provides gradient signals from a global perspective to each agent. During execution, a distributed actuator network is used, where each agent makes decisions based solely on its own local observations, eliminating the need for communication and meeting real-time requirements. The Q-value of the evaluation by the centralized critic network satisfies the formula: ~P} Here, x represents the global state, a_i represents the action of the i-th agent, N represents the number of agents, μ represents the policy network, μ´ represents the target policy network, r_i represents the reward of the i-th agent, and γ represents the discount factor. The centralized commentator network consists of 3 fully connected layers with a hidden layer dimension of 256, uses ReLU activation function, and outputs Q-values. The distributed executor network consists of 2 fully connected layers with a hidden layer dimension of 128, and uses the tanh activation function to map actions to the interval [-1,1] in the output layer, and then maps them to the actual parameter range through a linear transformation. To improve training stability, a soft update strategy for the target network and an experience replay mechanism are adopted, with a replay buffer size of 1 million records and a batch size of 256 samples per sampling.

[0039] The multi-objective reward function design comprehensively considers three optimization objectives: output, quality, and energy consumption. The reward function satisfies the formula: r_t=w_1r_prod(t)+w_2r_qual(t)-w_3r_energy(t)-w_4r_constraint(t)+w_5r_smooth(t), where r_prod is the output reward, r_qual is the quality reward, r_energy is the energy consumption penalty, r_constraint is the constraint penalty, r_smooth is the smoothness reward, and w_1~w_5 are the weight coefficients of each reward item. The output reward is directly proportional to the actual output and is normalized. The maximum output reward is obtained when the output reaches the designed capacity. The quality reward is directly proportional to the product quality score, which is calculated based on online detection or soft measurement results. The higher the score, the greater the reward. The energy consumption penalty is inversely proportional to the actual energy consumption. The higher the energy consumption, the greater the penalty. The weighting coefficients of the three objectives are dynamically adjusted based on the current operating conditions and production targets. During normal production, the weighting is 0.5 for quality, 0.3 for output, and 0.2 for energy consumption. When orders are urgent, the weighting of output increases to 0.5, and the weighting of quality is adjusted to 0.3. During periods of energy shortage or peak electricity prices, the weighting of energy consumption increases to 0.4. A constraint penalty term is also added to the reward function, providing a larger negative reward when parameters exceed the safe range or an alarm is triggered, ensuring that the agent learns a strategy that meets the safety constraints. Furthermore, a smoothness reward term is added to penalize abrupt changes in actions, ensuring the stability of the production line operation.

[0040] A policy transfer and fine-tuning mechanism addresses the transfer problem from simulation to reality. First, large-scale pre-training (1 million steps) is performed in a digital twin simulation environment to achieve good initial performance of the policy. During pre-training, domain randomization is employed to apply random perturbations to the simulation model's parameters (such as friction coefficient, density, and viscosity), enabling the policy to learn more robust behavior and improve its adaptability to the real environment. The parameter perturbation range of domain randomization satisfies the formula: θ_sim = θ_real × , where ε~U For random perturbation, δ represents the perturbation amplitude (typically 0.1-0.3). Parameters are randomly sampled during training to adapt the policy to different environmental dynamics. After pre-training, the policy network parameters are transferred to the actual production line for online fine-tuning using real-world data. During fine-tuning, importance sampling is used to correct for differences in data distribution between the simulation and real environments. The importance weights satisfy the formula: w(s,a)=π_real(s,a) / π_sim(s,a), where π_real is the real environment policy and π_sim is the simulation environment policy. Weight correction reduces the bias caused by distribution shifts. A relatively small learning rate, approximately 1 / 10 of the pre-training learning rate, is used during fine-tuning to avoid destroying learned useful knowledge. Policy performance is continuously monitored during fine-tuning. Fine-tuning is considered complete when performance indicators consistently improve over multiple cycles and meet requirements. To ensure safety, a safety constraint mechanism is used during the fine-tuning phase to limit the magnitude of action changes within a safe range, and a manual control interface is provided, allowing switching to manual control when necessary.

[0041] In the intelligent diagnosis and self-healing control of abnormal operating conditions, the first step is anomaly detection and localization. A probabilistic model of normal operating conditions is constructed based on a variational autoencoder (VAE). The encoder consists of three fully connected layers, mapping high-dimensional observation data (approximately 64 features) to a low-dimensional latent space (16 dimensions), where the latent variables follow a standard normal distribution. The decoder consists of three fully connected layers, reconstructing the observation data from the latent space. The optimization objective of the VAE is the lower bound of evidence (ELBO), satisfying the formula: L(θ,φ;x)=E_{q_φ(z|x)} _ The first term is the reconstruction likelihood expectation, and the second term is the KL divergence between the latent variable posterior distribution qφ(z|x) and the prior distribution p(z)=N(0,I). Anomalies are detected by calculating the negative log-likelihood of the reconstruction error; anomalies are identified when their scores exceed a threshold. The threshold is determined through kernel density estimation, using the 99th percentile of the reconstruction error from normal data as the initial threshold, with the false alarm rate controlled below 1%. Anomaly localization is achieved by calculating the contribution of each input dimension to the reconstruction error; the dimensions with the highest contributions correspond to the parameters where the anomaly occurred. Simultaneously, the process topology is combined to locate the process where the anomaly occurred, achieving a localization accuracy of no less than 90%. To adapt to changing operating conditions, the VAE model employs an incremental update strategy, fine-tuning the model parameters every 24 hours using the latest normal data.

[0042] The root cause analysis of anomalies employs a graph neural network approach, modeling each process and parameter of the production line as a graph structure. Nodes in the graph represent processes or parameters, with approximately 50-80 nodes; edges represent physical connections, material transfer relationships, or causal relationships, with approximately 100-150 edges. Anomaly information is propagated using a graph attention network (GAT), which consists of three graph attention layers, each with four attention heads, aggregating neighborhood information through a multi-head attention mechanism. The attention coefficient satisfies the formula: α_{ij} ∈ , where e_{ij}=Leaky For attention score, 'a' is the attention vector, 'W' is the learnable weight matrix, 'h_i' is the feature vector of node i, 'N_i' is the set of neighbors of node i, and '||' is the concatenation operation. During anomaly tracing, the anomaly signal is propagated backward along the edges of the graph from the anomaly node, and the anomaly contribution of each node is calculated. The node with the highest contribution is the root cause of the anomaly. To improve the accuracy of tracing, a causal graph is constructed based on prior knowledge. The direction of the edges represents the causal relationship. During tracing, propagation is only performed backward along the causal edges to avoid interference from false associations. The graph neural network is trained using a semi-supervised approach, using labeled data from known fault cases for supervised training, and simultaneously using normal operation data for unsupervised pre-training.

[0043] Fault reasoning is implemented based on a Fault Mode and Effects Analysis (FMEA) knowledge graph. The knowledge graph contains entities and their relationships, including fault modes, fault causes, fault effects, remedial measures, detection methods, severity, frequency of occurrence, and detectability. The number of entities is approximately 200-300, and the number of relationships is approximately 400-600. A graph matching algorithm is used to match the current anomaly pattern with fault modes in the knowledge graph, calculating a similarity score. The top N fault modes with the highest similarity are selected as candidate diagnostic results. The matching process comprehensively considers multiple dimensions of information, including anomaly parameters, anomaly severity, anomaly development trends, and process location. Rule-based reasoning is used to further diagnose the fault type and severity. Rules include IF-THEN production rules and fault tree-based logical reasoning rules. Severity assessment uses the Risk Priority Number (RPN) method, satisfying the formula: RPN = S × O × D, where S is the severity score (1-10 points), O is the frequency of occurrence score (1-10 points), and D is the detectability score (1-10 points). A higher RPN value indicates a greater risk. The fault diagnosis accuracy rate is no less than 85%, and for common fault types, the accuracy rate can reach over 95%. The knowledge graph adopts an incremental update mechanism; newly discovered fault modes and their handling experience are added to the knowledge graph after expert review.

[0044] The self-healing control strategy generation and execution matches the optimal self-healing control strategy from the strategy library based on the diagnostic results. The strategy library contains preset control strategies for common faults, with multiple alternative strategies corresponding to each fault, ordered by priority. The strategy takes the form of a parameter adjustment scheme, including which parameters to adjust, the direction of adjustment, the adjustment magnitude, and the adjustment rate. For new fault types not in the strategy library, control strategies are generated in real-time through reinforcement learning. A model-based reinforcement learning method is used, leveraging a digital twin model to quickly evaluate the effectiveness of different strategies and select the optimal strategy for execution. Strategy evaluation employs a multi-objective utility function, comprehensively considering multiple objectives such as changes in quality, output, energy consumption, recovery time, and risk level. Before execution, the generated strategy undergoes safety verification, including checking whether parameters are within safe ranges, whether adjustments will trigger other faults, and whether the impact on product quality is acceptable. Strategies that fail verification are automatically discarded and regenerated. Strategies that pass verification are automatically sent to the control system for execution. During execution, the effect is continuously monitored; if anomalies are not eliminated or new anomalies occur, strategy regeneration is triggered. For severe faults, the system will issue an alarm to the operator while implementing the self-healing strategy, and will automatically shut down the equipment if necessary to protect it. The response time of the self-healing control is no more than 30 seconds, and the self-healing success rate for common faults is no less than 80%.

[0045] In the product quality soft measurement and feedforward-feedback closed-loop control process, a product quality soft measurement model based on a deep belief network (DBN) is first constructed. The model takes online measurable process parameters as input, including approximately 50 parameters such as temperature, pressure, rotational speed, current, vibration, and liquid level at various stages, and approximately 10 parameters related to material state characteristics (composition, moisture content, particle size distribution, etc.), for a total of approximately 60 input variables. It outputs quality indicators that are difficult to measure online, including approximately 8-10 indicators such as melt flow index, density, tensile strength, impact strength, ash content, melt flow rate, molecular weight distribution, and color. The deep belief network is composed of multiple stacked restricted Boltzmann machines (RBMs), containing 3 RBM layers with 200, 100, and 50 hidden layer nodes respectively. Model training employs a two-stage strategy of layer-by-layer pre-training followed by global fine-tuning: the pre-training stage uses the contrastive divergence (CD) algorithm to train each RBM layer by layer; the fine-tuning stage uses the backpropagation algorithm to perform supervised fine-tuning of the entire network. To prevent overfitting, Dropout regularization and L2 regularization are employed. The prediction accuracy of the soft sensor model is measured by the root mean square error (RMSE) and the coefficient of determination (R²). 2 Evaluation, R of key quality indicators 2 Not less than 0.95.

[0046] Quality deviation prediction is achieved based on a time series forecasting model, predicting the trend of quality indicators over a future period. The prediction model employs a Long Short-Term Memory (LSTM) network. The input sequence consists of process parameters and soft-sensor quality results from the past period, with a sequence length of 60 time steps (corresponding to data from the past 30 minutes, with a sampling interval of 30 seconds). The output sequence is the predicted value of the quality indicator for the future period, with a sequence length of 60 time steps (corresponding to the prediction for the next 30 minutes). The LSTM network consists of two LSTM layers with a hidden layer dimension of 128, followed by a fully connected layer to output the prediction result. To improve prediction accuracy, an attention mechanism is used to assign different weights to different time steps in the input sequence, enabling the model to automatically focus on historical data with a greater impact on the future. The mean absolute percentage error of the prediction model is controlled within 8%, providing a warning of quality deviation 15-30 minutes in advance, allowing sufficient adjustment time for feedforward control. The prediction result also outputs a confidence interval; manual verification is triggered when the prediction confidence is low.

[0047] Feedforward control optimizes parameter adjustment based on sensitivity analysis and the pseudo-inverse method. Sensitivity analysis is obtained through a perturbation method, where a small perturbation (typically 1% of the parameter range) is applied to each input parameter, the change in output quality index is observed, and the sensitivity coefficient is calculated to form a sensitivity matrix. The elements S_ij of the sensitivity matrix satisfy the formula: S_ij=∂y_i / ∂x_j≈Δy_i / Δx_j, representing the change in the i-th quality index when the j-th parameter changes by a unit amount, where y_i is the i-th quality index and x_j is the j-th process parameter. The sensitivity matrix is ​​updated periodically, either weekly or immediately when significant changes occur in operating conditions. Based on the predicted quality deviation and the sensitivity matrix, the optimal parameter adjustment is calculated using the pseudo-inverse method (Moore-Penrose pseudo-inverse), which is the parameter adjustment scheme that minimizes the quality deviation while satisfying parameter constraints. The parameter adjustment satisfies the formula: ,in Let Δy be the Moore-Penrose pseudoinverse of the sensitivity matrix S, Δy be the quality deviation vector, and Δx be the parameter adjustment vector. When the predicted quality deviation exceeds the allowable range, the feedforward control system automatically calculates the parameter adjustment and executes it to correct the quality deviation in advance and avoid quality problems. The response time of the feedforward control is no more than 10 seconds, enabling parameter adjustment to be completed before the actual quality deviation occurs. For multivariable coupled systems, the feedforward control adopts a decoupling strategy to reduce mutual interference between parameter adjustments.

[0048] Feedback correction is implemented based on Model Predictive Control (MPC), using actual quality inspection results for closed-loop correction. Actual quality inspection results come from offline laboratory data or online quality inspection instruments, with inspection cycles typically ranging from 15 minutes to 1 hour. Model Predictive Control aims to minimize quality deviation, considering both parameter constraints (upper and lower limits) and rate of change constraints (parameter adjustment rate limits). In each control cycle, it solves a finite-time optimization problem to obtain the optimal control sequence, executes the first control action, and then solves again in the next cycle. The control cycle is 1 minute, the prediction time domain is 30 steps, and the control time domain is 10 steps. The optimization objective function of MPC satisfies the formula: Where N_p is the prediction time domain and N_c is the control time domain. Let y_ref(k) be the predicted output, Δu(k) be the reference trajectory, Δu(k) be the control increment, Q be the output error weight matrix, and R be the control increment weight matrix. Constraints include upper and lower limits for parameters and rate of change limits. The optimization solution employs quadratic programming (QP). Feedback correction and feedforward control work in tandem; feedforward control is responsible for suppressing predictable disturbances in advance, while feedback control is responsible for eliminating unpredictable disturbances and model errors, forming a feedforward-feedback composite closed-loop control system, significantly improving the accuracy and stability of quality control.

[0049] The edge-cloud collaborative distributed intelligent control architecture is divided into two layers: the edge side and the cloud side. Through collaborative optimization, it achieves a balance between real-time performance and global capabilities. At the edge side, edge computing nodes are deployed at each process stage of the production line. These edge nodes are equipped with industrial-grade AI processors with a computing power of at least 10 TOPS, supporting local inference for various AI models. The edge side is responsible for tasks with high real-time requirements, such as real-time data acquisition and preprocessing, feature extraction, rapid anomaly detection, and local control execution. The data acquisition frequency is determined based on the parameter type: 100Hz for rapidly changing parameters (such as current and vibration) and 1Hz for slowly changing parameters (such as temperature and pressure). Feature extraction is performed locally at the edge nodes, reducing the amount of data uploaded to the cloud. Rapid anomaly detection uses lightweight models with a response time of no more than 100 milliseconds, enabling timely detection and triggering of local protection mechanisms in the early stages of anomalies. Local control execution includes conventional PID control, feedforward control, and simple logic control, ensuring the production line can maintain basic operation even during network outages. Edge nodes are also responsible for data caching, buffering data during network interruptions and retransmitting it after network recovery.

[0050] A cloud-based big data analytics platform and AI training platform are deployed to handle computationally intensive tasks such as data storage and analysis across the entire production line, model training and optimization, global parameter optimization, and knowledge base updates. The cloud platform employs a distributed computing architecture, containerized based on Kubernetes, supporting parallel processing and elastic scaling of massive datasets. Data storage utilizes a time-series database for time-series data, a relational database for structured configuration data, and object storage for unstructured data (such as images and spectral data). Model training is conducted weekly, using newly accumulated data to incrementally train the model and continuously improve its accuracy. Global optimization is performed daily, optimizing global parameters based on a digital twin model and historical data to generate recommended operating parameters for the following day. The knowledge base is updated every 24 hours, adding newly discovered operating conditions, fault cases, and control strategies. The cloud platform also provides a visual monitoring interface, report statistics, alarm management, and remote diagnostics to facilitate monitoring and decision-making by management personnel.

[0051] Model distillation and lightweighting techniques address the challenge of deploying large cloud-based models to edge computing. The large cloud-trained model (teacher model) is compressed into a smaller, edge-deployable model (student model) through knowledge distillation. The teacher model has approximately 10-20 million parameters and is trained and run on a cloud GPU server; the student model has approximately 800,000-1.2 million parameters and can run in real-time on an edge CPU or lightweight GPU. During distillation, the student model learns not only the true labels (hard labels) but also the soft outputs (soft labels, i.e., probability distributions of each category) of the teacher model to preserve the teacher model's generalization ability and hidden knowledge. The distillation loss function satisfies the formula: L_distill=αL_Hard+(1 L_Soft), where L_Hard is the cross-entropy loss between the student model and the hard labels, L_Soft is the KL divergence loss between the soft outputs of the student model and the teacher model, and α is the balance coefficient (usually taken as 0.1-0.3). After distillation, the model accuracy decreases by no more than 3%, the inference speed is improved by more than 5 times, and the memory usage is reduced by more than 80%, meeting the resource constraints of edge computing. For different hardware platforms, further optimization techniques such as quantization and pruning are adopted to quantize the model from FP32 to INT8, controlling the accuracy loss to within 1% and further improving the inference speed by 2-3 times.

[0052] A dynamic delivery and version management mechanism enables OTA updates and multi-version management of models. Lightweight models trained in the cloud are dynamically delivered to edge nodes via OTA, supporting incremental updates that transmit only the changed model parameters, reducing network bandwidth consumption. Multi-version model management is supported; each version includes model files, version number, release date, accuracy metrics, applicable operating conditions, test reports, etc., allowing switching between different versions as needed. Models undergo thorough testing before delivery, including functional, performance, security, and stability testing, ensuring model security and stability. Canary releases are supported, allowing testing on select edge nodes or production lines before full rollout; if issues arise, a quick rollback to the previous stable version is possible. The model delivery process employs breakpoint resumption and verification mechanisms to ensure transmission integrity and correctness. Upon receiving a new model, edge nodes first load and test it in a backup environment; only after verification is the model switched to the production environment, achieving seamless upgrades.

[0053] Intelligent energy consumption optimization, as an auxiliary functional module, further improves the energy efficiency of the production line. Real-time energy consumption monitoring and decomposition are achieved by installing smart meters on key equipment in each process, collecting real-time electrical energy data such as voltage, current, power, power factor, and harmonics at a sampling frequency of no less than 1Hz. Total energy consumption is decomposed into energy consumption components for each piece of equipment and each stage, establishing an energy consumption decomposition model. Energy consumption decomposition employs non-intrusive load decomposition technology, identifying the energy consumption contribution of each piece of equipment based on current waveform characteristics, achieving fine-grained energy consumption monitoring without the need to install sensors on each piece of equipment. The decomposition accuracy is no less than 90%, capable of identifying the start-up and shutdown status and energy consumption changes of major energy-consuming equipment. Energy consumption data is statistically analyzed according to dimensions such as equipment, process, shift, and product type, generating energy consumption reports and energy efficiency analysis reports.

[0054] The energy consumption baseline is dynamically adjusted, automatically switching baseline values ​​based on the current operating condition. The baseline is determined based on the best operating records from historical data. Each operating condition corresponds to a set of baseline values, including unit product energy consumption, energy consumption percentage of each piece of equipment, and energy efficiency indicators. The baseline is calculated using quantile regression, taking the lowest 10% quantile of energy consumption in historical data as the baseline, representing the optimal energy efficiency level under that operating condition. The objective function of quantile regression satisfies the formula: Where τ is the quantile (τ=0.1 in this invention), β is the regression coefficient, x is the input feature, and y is the energy consumption value. When a shift in operating conditions is detected, the system automatically switches to the baseline of the corresponding operating condition. The deviation between the actual energy consumption and the baseline is calculated, and when the deviation exceeds a threshold (e.g., 5%), an energy consumption optimization reminder or automatic optimization is triggered. The baseline is updated regularly, once a month, incorporating new operating data to ensure that the baseline reflects the latest status and best practices of the equipment. At the same time, an energy consumption benchmark library is established to collect energy efficiency data from similar production lines in the same industry as an external benchmarking reference.

[0055] Energy-saving optimization strategies are generated based on a working condition-energy consumption correlation model, producing strategies tailored to the current working conditions. These strategies encompass multiple aspects: process parameter adjustments, such as reducing extrusion temperature, adjusting screw speed, and optimizing cooling water temperature, to reduce energy consumption while ensuring product quality; equipment start-up and shutdown optimization, such as starting and stopping auxiliary equipment (fans, pumps, heaters, etc.) on demand to avoid idling; load distribution optimization, such as balancing load distribution when multiple extruders are running in parallel to improve overall efficiency; and standby mode management, such as entering a low-power standby state instead of complete shutdown during short-term shutdowns. The generation of energy-saving strategies employs a multi-objective optimization method, aiming to minimize energy consumption while constraining product quality and output. The effectiveness of different strategies is evaluated based on a digital twin model. Energy-saving strategies are implemented after simulation verification to ensure no impact on product quality and production schedule. The execution of energy-saving optimizations includes both automatic and manual confirmation modes. Adjustments with low risk are executed automatically, while adjustments with significant impact require operator confirmation before execution.

[0056] Energy efficiency assessment and feedback regularly evaluate the effectiveness of energy efficiency improvements, calculating energy savings and energy efficiency rates. Assessment cycles include daily, weekly, and monthly assessments, analyzing energy efficiency changes across different time scales. Energy savings are calculated using a baseline method, using a benchmark energy consumption as a reference to calculate the actual energy savings relative to the benchmark. Energy savings are fed back to the optimization model to continuously improve energy-saving strategies, forming a closed-loop optimization. An energy efficiency ranking system is established to rank the energy efficiency performance of different shifts, operating conditions, and operators, incentivizing operators to optimize operating parameters. Simultaneously, an energy consumption anomaly alarm mechanism is established; when energy consumption suddenly increases and cannot be explained by changes in operating conditions, an alarm is triggered, indicating potential equipment failure or operational abnormalities. Energy efficiency data is analyzed in conjunction with production and quality data to uncover the correlation between energy efficiency, quality, and output, providing a basis for comprehensive optimization.

[0057] Virtual sample generation and data augmentation mechanisms address the issues of data scarcity and sample imbalance in industrial scenarios. Mechanism simulation virtual samples generate virtual operating samples under different operating conditions and parameter combinations based on physical mechanism models. The simulation model includes the aforementioned digital twin model, capable of simulating various situations of normal and abnormal operation. By changing input parameters (material composition, equipment parameters, environmental conditions, etc.), diverse virtual samples are systematically generated within the parameter space. Sample generation employs experimental design methods, such as full factorial design, orthogonal design, and Latin hypercube sampling, to ensure uniform distribution and representativeness of samples within the parameter space. For normal operating condition samples, scenarios covering different material types, production capacity levels, and quality requirements are included; for abnormal operating condition samples, scenarios covering different fault types, fault degrees, and development stages are included. The number of virtual samples is determined as needed, typically 3-10 times that of real samples.

[0058] Generative Adversarial Network (GAN) augmentation employs Conditional Generative Adversarial Networks (CGANs) to learn the distribution of real-world operational data and generate high-quality virtual samples. The generator takes operating condition labels and random noise as input to produce operational data samples that conform to the distribution of that operating condition; the discriminator distinguishes between real and generated samples. Through adversarial training, the samples generated by the generator tend to be more consistent with real samples in terms of distribution. Conditional labels control attributes such as operating condition type, fault type, and severity of the generated samples, achieving targeted data augmentation. Training uses the Wasserstein GAN with Gradient Penalty (WGAN-GP) framework to improve training stability and generation quality. The objective function of WGAN-GP satisfies the formula: Where P_r is the actual data distribution and P_g is the generated data distribution. Let λ be the interpolation distribution between real and generated samples, and λ be the gradient penalty coefficient (usually taken as 10). The diversity of generated samples is guaranteed by the randomness of noise, and the authenticity of generated samples is guaranteed by adversarial training of the discriminator.

[0059] Sample quality assessment evaluates the quality of virtual samples from multiple dimensions, selecting high-quality samples for model training. Distribution consistency is measured by the maximum mean difference (MMD), which measures the difference between virtual and real samples in the feature space distribution. MMD satisfies the formula: Where X and Y represent the real and virtual sample sets, respectively, φ is the feature mapping function, and H is the reproducing kernel Hilbert space. A smaller MMD value indicates a more consistent distribution. Physical rationality is verified through physical constraints, confirming whether the virtual samples meet basic physical laws, such as energy conservation, mass conservation, material balance, and parameter upper and lower limits. Samples violating physical constraints are marked as invalid. Diversity is measured by sample coverage, evaluating the coverage of virtual samples in the feature space; higher coverage indicates better diversity. A sample quality score is calculated based on these three dimensions, and virtual samples with quality scores above a threshold are selected for model training. The threshold is dynamically adjusted according to task requirements; a higher threshold is set for tasks with high accuracy requirements, while a lower threshold is appropriately set for tasks with extremely scarce data.

[0060] Hybrid training enhances the model's generalization ability by training it with a mix of real and virtual samples in a specific ratio. The ratio of real to virtual samples is dynamically adjusted based on task requirements, typically between 1:1 and 1:3, meaning one real sample is paired with one to three virtual samples. Adjustment strategies include: reducing the virtual sample ratio when real samples are plentiful to avoid introducing excessive bias; and increasing the virtual sample ratio when real samples are scarce to expand the training data. A learning-by-course strategy is employed: pre-training on virtual samples allows the model to learn basic physical laws and data patterns; fine-tuning on real samples then adapts the model to the distribution and characteristics of real data. A larger learning rate is used during pre-training to quickly converge to a reasonably correct solution; a smaller learning rate is used during fine-tuning to precisely adjust model parameters. Hybrid training significantly improves the model's generalization ability under small sample sizes or rare conditions, maintaining similar model accuracy even with a 50% reduction in sample size.

[0061] This invention also provides a multi-condition online monitoring and control system adapted to waste plastic resource recovery production lines, including a multi-source sensing module that integrates various sensors such as near-infrared spectral sensors, industrial cameras, microwave humidity sensors, weighing sensors, temperature sensors, current sensors, and vibration sensors to collect multi-modal data of materials and production line operation data; a material state assessment module for extracting single-modal features and fusing multi-modal features from multi-source data to quantitatively assess material composition, moisture content, particle size distribution, and impurity content, generating a comprehensive material quality score; a condition identification module for constructing a joint feature space, automatically identifying condition types through density peak clustering, detecting condition migration, and maintaining and updating the condition knowledge base; and a digital twin simulation module for constructing a multi-physics coupled digital twin model of the production line for simulation. The system includes: a full-process simulation calculation module for accelerating multi-objective parameter optimization based on a surrogate model; a collaborative control module for achieving multi-process collaborative control based on deep reinforcement learning to optimize multiple objectives such as output, quality, and energy consumption; an anomaly diagnosis and self-healing module for anomaly detection and localization, root cause tracing, fault reasoning, and generation and execution of self-healing control strategies; a quality closed-loop control module for soft measurement of product quality, quality deviation prediction, feedforward control, and feedback correction; an edge-cloud collaboration module for achieving collaborative optimization of real-time processing at the edge and big data analysis in the cloud, including model distillation, dynamic distribution, and version management; an energy consumption optimization module for real-time energy consumption monitoring and decomposition, dynamic adjustment of baselines, generation of energy-saving strategies, and energy efficiency assessment; and a data augmentation module for virtual sample generation and quality assessment to achieve data augmentation for model training.

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-condition online monitoring and control method adapted to a waste plastic resource recovery production line, characterized in that, Includes the following steps: S1. Collect multimodal data of waste plastic materials through multiple sensors, and evaluate the composition, moisture content, particle size distribution and impurity content of the materials in real time after multimodal feature fusion to generate a comprehensive characterization of the material state. S2. Based on the joint feature space of material status and operating parameters, the current working condition type of the production line is automatically identified through unsupervised clustering, and working condition migration events are detected to dynamically update the working condition classification system. S3. Construct a multi-physics coupled digital twin model of the production line, perform full-process simulation based on the current working conditions and material status, and obtain the optimal combination of process parameters for each process through simulation optimization. S4. The optimization of multi-process parameters is modeled as a multi-agent reinforcement learning problem. The agents of each process make collaborative decisions to generate the optimal control strategy under the constraints of multiple objectives such as output, quality and energy consumption. S5. Real-time monitoring of production line operation status, using graph neural networks and knowledge graphs to trace the root causes of anomalies, assess the severity of anomalies, and generate self-healing control strategies. S6. Construct a soft measurement model for product quality based on process parameters, predict quality indicators in real time, and achieve rapid correction of quality deviations through feedforward-feedback composite control; S7 performs real-time data processing and rapid adjustment at the edge, performs big data analysis and model training in the cloud, and achieves edge-cloud collaborative optimization through model distillation and dynamic distribution.

2. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, The data acquisition in step S1 specifically includes: S11. Multimodal data acquisition: Material spectral data is acquired through a near-infrared spectral sensor, material visual images are acquired through an industrial camera, moisture content data is acquired through a microwave humidity sensor, flow rate data is acquired through a weighing sensor, and temperature data is acquired through a temperature sensor. S12, Single-mode feature extraction: Feature extraction is performed on each mode of data. For spectral data, features such as absorption peak position, peak height, and peak area are extracted. For image data, features such as color, texture, and shape are extracted. For waveform data, time-domain and frequency-domain features are extracted. S13. Multimodal feature fusion: A multimodal fusion network based on attention mechanism is adopted to adaptively learn the weights of each modality feature and fuse the multimodal features into a unified material state feature vector. S14. Quantitative assessment of material state: Based on the fused feature vector, the material is quantitatively assessed through a regression model, including key indicators such as the proportion of plastic types, moisture content, particle size distribution, and impurity content, to generate a comprehensive material quality score.

3. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, The initial screening method in step S2 specifically includes: S21. Construction of Joint Feature Space: The material state features are combined with the production line operating parameter features to construct a high-dimensional joint feature space; S22. Density Peak Clustering for Working Condition Discovery: The density peak clustering algorithm is used to automatically discover working condition categories in the joint feature space without pre-specifying the number of categories. The optimal number of clusters is automatically determined through the decision graph. S23. Working condition migration detection: Based on statistical process control methods, it monitors changes in feature distribution. When the distribution deviation exceeds the threshold, it is determined to be a working condition migration, triggering a reclassification of the working condition. S24. Working Condition Knowledge Base Update: Newly identified working conditions and their corresponding optimal parameter combinations are stored in the working condition knowledge base, which is continuously updated and expanded through online learning.

4. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, Step S3 specifically includes: S31. Multiphysics Coupled Modeling: Construct a multiphysics coupled digital twin model that includes thermal field, flow field, force field, and chemical reaction field to describe the physicochemical processes of each step, such as crushing, washing, extrusion, and granulation. S32. Hybrid simulation calculation: A hybrid simulation strategy combining the finite element method and the discrete element method is adopted. The finite element method is used to solve the continuous medium, and the discrete element method is used to solve the particulate material, thereby improving the simulation accuracy and efficiency. S33, Proxy Model Acceleration: Based on simulation samples, train the Kriging proxy model and use the proxy model to replace part of the time-consuming full-size simulation, thereby accelerating the optimization iteration process; S34. Multi-objective parameter optimization: With the objectives of maximizing output, optimizing quality, and minimizing energy consumption, a multi-objective evolutionary algorithm is used to search for the optimal combination of process parameters in the simulation environment to generate a Pareto optimal solution set.

5. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, Step S4 specifically includes: S41. Multi-agent modeling: Each process of the production line is modeled as an independent reinforcement learning agent, with each agent responsible for parameter control decisions for its own process; S42. Centralized training and distributed execution: A centralized training and distributed execution architecture is adopted. During training, global information is used for centralized value evaluation, and during execution, each agent makes distributed decisions based only on local observations. S43. Multi-objective reward design: Design a multi-objective reward function that includes output reward, quality reward, and energy consumption penalty, and adjust the importance of each objective through weighting coefficients; S44. Strategy Transfer and Fine-tuning: Transfer the strategy trained in the simulation environment to the actual production line, and use the actual operating data for online fine-tuning to quickly adapt to the real environment.

6. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, Step S5 specifically includes: S51. Anomaly Detection and Localization: Construct a probabilistic model of normal operating conditions based on a variational autoencoder, detect anomalies by reconstructing errors, and locate the process and parameters where the anomaly occurs through an attention mechanism. S52. Graph Neural Network Root Cause Analysis: Model each process and parameter of the production line as a graph structure, use graph neural networks to propagate abnormal information, and trace the root cause of the abnormality. S53. Knowledge Graph Fault Reasoning: Construct a knowledge graph for fault mode and impact analysis, and diagnose fault types and severity based on graph matching and rule reasoning; S54. Self-healing regulation strategy generation: Based on the diagnostic results, the optimal self-healing regulation strategy is matched from the strategy library, or a regulation strategy is generated in real time through reinforcement learning, and the regulation action is automatically executed.

7. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, Step S6 specifically includes: S61. Soft Measurement Model Construction: A product quality soft measurement model is constructed using a deep belief network, with process parameters that can be measured online as input and quality indicators that are difficult to measure online as output. S62. Quality Deviation Prediction: Based on the current process parameters and material status, predict the trend of quality index changes in the future period and calculate the estimated quality deviation. S63. Feedforward control: Based on the estimated quality deviation, the adjustment amount of each parameter is calculated based on sensitivity analysis, and the quality deviation is corrected in advance through feedforward control. S64. Feedback Correction: Based on the actual quality inspection results, feedback correction is performed through model predictive control to form a feedforward-feedback composite closed-loop control.

8. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, It also includes intelligent energy consumption optimization steps: S81. Real-time monitoring and decomposition of energy consumption: Real-time collection of energy consumption data for each process, decomposition of total energy consumption into energy consumption components for each piece of equipment and each link, and establishment of an energy consumption decomposition model. S82. Dynamic adjustment of energy consumption baseline: Based on the current operating condition, dynamically adjust the energy consumption baseline and calculate the deviation between the actual energy consumption and the baseline. S83. Energy-saving optimization strategy generation: Based on the operating condition-energy consumption correlation model, generate energy-saving optimization strategies for the current operating condition, including parameter adjustment, equipment start-up and shutdown optimization, load allocation, etc. S84. Energy Efficiency Assessment and Feedback: Regularly assess the effectiveness of energy efficiency improvements, feed the energy-saving effects back to the optimization model, and continuously improve energy-saving strategies.

9. The multi-condition online monitoring and control method for a waste plastic resource recovery production line according to claim 1, characterized in that, It also includes virtual sample generation and data augmentation steps: S91, Virtual Samples for Mechanism Simulation: Virtual operating samples under different working conditions and parameter combinations are generated based on physical mechanism models. S92. Enhanced Generative Adversarial Network: Conditional Generative Adversarial Networks are used to learn the distribution of real-world data and generate high-quality virtual samples. S93. Sample quality assessment: Evaluate the quality of virtual samples from dimensions such as distribution consistency, physical rationality, and diversity, and screen high-quality samples; S94, Hybrid Training Enhancement: Train the model using a mix of real and virtual samples to improve the model's generalization ability under small sample sizes or rare conditions.

10. A multi-condition online monitoring and control system adapted to a waste plastic resource recovery production line, employing the control method described in any one of claims 1-9, characterized in that, include: The multi-source sensing module integrates various sensors such as near-infrared spectroscopy, machine vision, microwave humidity, weighing, and temperature to collect multimodal data of materials and production line operation data; The material state assessment module is used to perform feature fusion on multi-source data and quantitatively assess material composition, moisture content, particle size distribution and impurity content. The working condition identification module is used to automatically identify the current working condition type of the production line, detect working condition migration, and maintain the working condition knowledge base; The digital twin simulation module is used to build a multi-physics coupled digital twin model of the production line for full-process simulation and parameter optimization; collaborative... The regulation module is used to achieve multi-process collaborative regulation based on deep reinforcement learning, and optimize output, quality and energy consumption; the anomaly diagnosis and self-healing module is used for anomaly detection, root cause tracing and generation of self-healing regulation strategies. The quality closed-loop control module is used for soft measurement and feedforward-feedback composite control of product quality. The edge-cloud collaboration module is used to optimize real-time processing at the edge and big data analysis in the cloud.

Images