A classification identification method and system based on industrial big data

Abstract

This invention discloses a classification and recognition method and system based on industrial big data, belonging to the field of image recognition technology. It includes a multimodal data acquisition and annotation module for collecting multimodal data of industrial waste and performing dual annotation, dividing the dataset, merging it, and uploading it to a hybrid storage architecture; a multidimensional data preprocessing module for performing basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality; and a multimodal feature fusion and dynamic extraction model training module for constructing a feature extraction architecture and completing multimodal feature fusion, extraction, and dynamic model optimization training. This invention addresses the problems of low classification accuracy, poor adaptability, insufficient practicality, and difficulty in improving accuracy due to the lack of cyclical optimization in data processing by expanding the dimensions of multimodal data, constructing a dynamically optimized hybrid model, adding a closed-loop function throughout the entire process, strengthening anti-interference processing, and designing a cyclically optimized data processing sequence and dynamic threshold mechanism.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to a classification and recognition method and system based on industrial big data. Background Technology

[0002] Chinese invention patent discloses a classification and identification method and system based on big data, publication number CN116503663A, including: collecting shape and state data of various industrial wastes, classifying them, and uploading them to a cloud-based big data platform; preprocessing, obtaining the original dataset for training a feature extraction model through random sampling and noise reduction; establishing a feature extraction model, establishing mapping relationships, constructing a model parameter optimization problem based on unsupervised learning, and solving for the optimal solution of the model parameters; training the feature dataset of the feature classification model; establishing a feature multi-classifier, establishing mapping relationships, constructing a model parameter optimization problem based on supervised learning, and training the feature multi-classifier; collecting shape and state data of industrial wastes in real time, combining the trained feature extraction model and feature multi-classifier, using the feature extraction model to extract high-order features from the real-time data, and then using the feature multi-classifier to calculate the probability of occurrence of various types of industrial wastes, thereby realizing the identification and classification of industrial wastes.

[0003] The existing big data-based industrial waste sorting technology relies solely on basic physical characteristics such as shape, color, weight, and wetness. This single data dimension makes it difficult to distinguish between industrial wastes with similar compositions but vastly different processing methods (e.g., heavy metal-containing plastics versus ordinary plastics, toxic sludge versus ordinary wet waste). Furthermore, its fixed unsupervised and supervised learning architecture, with parameters fixed after model training, fails to adapt to dynamic scenarios where industrial waste composition changes with production process upgrades, leading to a decline in sorting accuracy over time. In addition, the sorting results only output the waste type without linking it to subsequent processing solutions, resulting in insufficient practicality and failing to consider the impact of environmental interference (such as light and dust) on data quality during data collection.

[0004] Furthermore, in existing technologies, the data processing flow adopts a linear sequential execution method, with each module running independently and without a loop optimization mechanism. During the data processing process, the processing order cannot be dynamically adjusted according to the real-time processing effect, making it difficult to further improve the data processing accuracy and failing to meet the core requirement of high precision for the refined classification of complex industrial waste.

[0005] Therefore, existing technologies suffer from limitations in data dimensions, poor model adaptability, limited functionality, weak anti-interference capabilities, lack of cyclical optimization in data processing, and insufficient accuracy. An improved solution that can overcome these bottlenecks is urgently needed. Summary of the Invention

[0006] To address the problems of existing technologies, this invention provides a classification and identification method and system based on industrial big data. By expanding the dimensions of multimodal data, constructing a dynamically optimized hybrid model, adding a closed-loop function for the entire process, and strengthening anti-interference processing, while designing a cyclically optimized data processing sequence and a dynamic threshold mechanism, this invention solves the problems of low classification accuracy, poor adaptability, and insufficient practicality of existing technologies, as well as the lack of cyclic optimization in data processing and the difficulty in improving accuracy.

[0007] This invention is implemented as follows: a classification and identification system based on industrial big data, comprising: The multimodal data acquisition and annotation module is used to collect multimodal data of industrial waste and perform dual annotation, divide the dataset, merge and upload it to a hybrid storage architecture; The multi-dimensional data preprocessing module is used to perform basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality. The multimodal feature fusion and dynamic extraction model training module is used to build a feature extraction architecture and complete multimodal feature fusion, extraction, and dynamic model optimization training. The hierarchical multi-classifier training and dynamic optimization module is used to build a hierarchical classification architecture and complete classifier training, optimization, and preliminary classification result output. The real-time classification and dynamic update module is used to complete the real-time classification of industrial waste, calculate the classification confidence level, and provide feedback on the classification results, thereby achieving bidirectional optimization of data and models.

[0008] As a preferred embodiment of the present invention, it further includes a loop processing module, which is based on five core data processing modules and executes them in a preset order until a stopping condition is met. Specifically, it includes: S1. Initialize the system, complete the first multimodal data acquisition and annotation of the multimodal data acquisition and annotation module, obtain the initial raw data, set the initial number of loops k=1 (k is a positive integer, k≥1), and set the relevant parameters of the evaluation parameter calculation formula and the dynamic threshold calculation formula. S2. This loop is executed in the following order: multimodal data acquisition and annotation module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module. Each module is linked in turn, and the output of the previous module is used as the input of the next module. After the hierarchical multi-classifier training and dynamic optimization module outputs the preliminary classification results, the evaluation parameters P(k) and dynamic threshold T(k) of this loop are calculated according to the preset formula. S3. Compare the evaluation parameter P(k) with the dynamic threshold T(k) to determine whether the stopping condition is met. If it is met, stop the loop and output the final classification data. If it is not met, update the data and the number of loops, and return to step S2 to repeat the loop.

[0009] As a preferred embodiment of the present invention, the stopping conditions in step S3 include two situations: one is that the evaluation parameter P(k) ≥ the dynamic threshold T(k); the other is that the number of loops reaches the preset maximum limit (which can be set according to the actual scenario to avoid infinite loops), and even if P(k) < T(k), the loop stops and the current optimal classification data is output, and a reminder signal is issued at the same time.

[0010] As a preferred embodiment of the present invention, in step S3, when the stopping condition is not met, the data update method is as follows: the output results of the hierarchical multi-classifier training and dynamic optimization module and the feedback data of the real-time classification and dynamic update module are sent back to the multimodal data acquisition and labeling module to update the original data and labeling information of the multimodal data acquisition and labeling module, and the loop count is updated to k=k+1.

[0011] As a preferred embodiment of the present invention, the evaluation parameter P(k) is calculated using a fractional formula, specifically: , where A(k) is the classification accuracy of the k-th cycle, D(k) is the feature discrimination of the k-th cycle, C(k) is the data confidence of the k-th cycle, and α, β, γ are weighting coefficients and α + β + γ = 1.

[0012] Preferably, the feature discrimination index D(k) is calculated by the ratio of intra-class variance to inter-class variance, specifically as follows: The value range is [0,1]; the weighting coefficients are assumed to be α=0.5, β=0.3, and γ=0.2 by default, and can be adjusted according to the actual industrial classification scenario.

[0013] As a preferred embodiment of the present invention, the formula for calculating the dynamic threshold T(k) is: T(k) = T0 + (T max -T0)×[1-e^(-k / λ)], where T0 is the initial threshold, T max λ is the maximum threshold, λ is the decay coefficient, and k is the current loop count; T0 defaults to 0.7; T max The default value is 0.95; the default value of λ is 5. Both can be adjusted according to the actual scenario, and in complex scenarios, T0 can be adaptively reduced to 0.68.

[0014] As a preferred embodiment of the present invention, the multimodal data acquisition and labeling module acquires physical characteristic data (shape, color, weight, wetness and dryness) and compositional characteristic data (elemental composition acquired by a portable spectrometer and volatile organic compound content acquired by a gas sensor) of industrial waste, as well as environmental auxiliary data (light intensity and dust concentration at the time of acquisition). The labeling method adopts a combination of "manual labeling and machine pre-labeling". The labeling content includes waste type, hazard level (0, 3) and recommended treatment method (recycling, harmless incineration, safe landfill, special treatment). The labeled data is divided into training set, validation set and test set and uploaded to a hybrid storage architecture of edge computing node and cloud big data platform.

[0015] As a preferred embodiment of the present invention, the multimodal feature fusion and dynamic extraction model training module adopts a feature extraction architecture of "branch extraction and cross-modal attention fusion", which includes three feature extraction branches to extract physical features, component features and environmental auxiliary features respectively. The features of the three branches are weighted and fused through a cross-modal attention mechanism. The weights are dynamically adjusted according to the confidence of each modality data, and the fusion weights and feature extraction parameters are updated in real time in combination with an online learning mechanism. At the same time, a dynamic optimization objective function with the dual objectives of maximizing feature discrimination and minimizing model complexity is constructed.

[0016] A classification and identification method based on industrial big data includes the following steps: Multimodal data of industrial waste was collected and double-labeled, and the dataset was divided, merged, and uploaded to a hybrid storage architecture; the multimodal data included at least image data. Perform basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality; Construct a feature extraction architecture to complete multimodal feature fusion, extraction, and dynamic model optimization training; Construct a hierarchical classification architecture, complete classifier training, optimization, and output preliminary classification results; The system enables real-time classification of industrial waste, calculation of classification confidence, and feedback of classification results, achieving bidirectional optimization of data and models.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: Compared to existing technologies, this invention breaks through the technical bottlenecks of traditional industrial waste sorting and identification, especially in terms of innovative improvements in data processing flow and accuracy optimization. It possesses outstanding creativity and practicality, and its core functional advantages and creativity are manifested as follows: the various improvements work synergistically to form a complete technological innovation system, rather than isolated technical adjustments: 1. The iterative processing flow breaks through linear limitations and significantly improves classification accuracy: The design employs a iterative processing sequence consisting of a multimodal data acquisition and labeling module, a multimodal feature fusion and dynamic extraction model training module, a real-time classification and dynamic update module, a multi-dimensional data preprocessing module, and a hierarchical multi-classifier training and dynamic optimization module. This breaks the limitations of existing technologies' linear processing and independent operation of each module, achieving a closed-loop linkage of "data acquisition, feature fusion, real-time feedback, data correction, and classification optimization." Each iteration, based on the previous processing results, corrects the original data in the multimodal data acquisition and labeling module using feedback data from the real-time classification and dynamic update module, optimizes the preprocessing parameters in the multi-dimensional data preprocessing module, and adjusts the model parameters in the multimodal feature fusion and dynamic extraction model training module and the hierarchical multi-classifier training and dynamic optimization module. This continuously improves data quality and classification accuracy, perfectly aligning with the core objective of "improving accuracy." The iterative processing logic is not simply repetitive execution but a closed-loop design with dynamic optimization, representing a fundamental technological innovation.

[0018] 2. Dynamic Threshold and Evaluation Parameter Design: Achieving a Balance Between Accuracy and Resource Efficiency. The innovative design incorporates quantitative evaluation parameters P(k) and a dynamic threshold T(k), along with a clear calculation formula. This addresses the issues of lacking a clear accuracy evaluation standard and vague loop stopping conditions in existing technologies. The evaluation parameters comprehensively consider classification accuracy, feature discrimination, and data confidence, fully reflecting classification accuracy. The dynamic threshold increases dynamically with the number of loops, ensuring normal progress in the initial stage of the loop while simultaneously driving continuous improvement in accuracy in later stages. This avoids insufficient accuracy or resource waste caused by fixed thresholds, achieving a dual balance between accuracy optimization and system resource efficiency. The calculation formula is designed to align with loop optimization logic and adapt to the actual needs of industrial scenarios.

[0019] 3. Multimodal Collaborative Fusion Architecture Breaks Through the Limitations of Single Data Dimensions: Overcoming the limitations of existing technologies that rely solely on physical features, this architecture innovatively introduces component features and environmental auxiliary features to form a multimodal data system. It designs a dedicated architecture for branch extraction and cross-modal attention fusion, and through an adaptive weight dynamic allocation mechanism, allows different modal data to autonomously adapt to classification scenarios based on confidence levels. This fundamentally solves the industry pain point of difficulty in distinguishing industrial waste with similar components, providing data and feature-level support for improved accuracy. Compared to existing fixed feature extraction methods, this represents a fundamental technological improvement, effectively compensating for the limitations of single-data-dimensional classification.

[0020] 4. Dynamic optimization mechanism enables adaptive model development throughout its entire lifecycle: A dual dynamic optimization system of online learning and incremental training is constructed. Combined with a cyclic processing flow, it breaks through the bottleneck of fixed model parameters in existing technologies. The model parameters and fusion weights can be updated in real time based on changes in industrial waste composition and cyclic feedback data, without the need to retrain the entire model. This ensures classification adaptability and significantly reduces model maintenance costs. The dynamic adaptation logic, combined with cyclic processing and threshold control, enables adaptive optimization throughout the model's entire lifecycle, further enhancing the stability of accuracy.

[0021] 5. Combining anti-interference design with iterative optimization enhances accuracy and stability in complex scenarios. Addressing interference factors such as light and dust in industrial settings, an innovative combined anti-interference scheme of "environmental data correction and multi-dimensional noise reduction" is adopted, forming a complete anti-interference logic encompassing data acquisition, correction, and noise reduction. This anti-interference design, combined with iterative processing, utilizes feedback data from real-time classification and dynamic updating modules to specifically optimize anti-interference parameters, ensuring the model remains stable in complex industrial environments. This adapts to harsh sorting scenarios that existing technologies cannot cover, expanding the application scope of the technology while ensuring the stability of accuracy during iterative optimization.

[0022] 6. The hybrid edge and cloud architecture balances efficiency and reliability, supporting innovative loop processing. It employs a hybrid storage and processing architecture combining edge computing nodes and a cloud platform. High-frequency operations such as real-time preprocessing, feature extraction, and loop judgment are deployed on edge nodes, achieving low-latency response. The cloud handles big data storage and global model optimization, achieving high-speed data synchronization through dual-mode transmission technology. This architecture provides efficient and reliable hardware and transmission support for loop processing, ensuring efficient execution of data backhaul, parameter updates, and threshold and evaluation parameter calculations during the loop process. This guarantees both the efficiency of real-time classification and the reliability of loop optimization, achieving a dual improvement in efficiency and reliability.

[0023] 7. A complete functional closed loop adapts to actual industrial needs. It constructs a complete functional closed loop of classification, hazard assessment, treatment suggestions, and feedback optimization, deeply binding the classification results with the actual treatment needs of industrial waste. The newly added hazard level assessment and personalized treatment suggestions fill the gap of the disconnect between classification and treatment in the existing technology. At the same time, the design of cyclic treatment, threshold control, and assessment parameter calculation can adjust relevant parameters (such as weighting coefficients, initial thresholds, maximum thresholds, etc.) according to the actual needs of industrial scenarios, adapting to different types of industrial waste classification scenarios with different accuracy requirements, making it highly practical. Attached Figure Description

[0024] Figure 1 This is a schematic block diagram of the classification and recognition system based on industrial big data provided in an embodiment of the present invention; Figure 2This is a flowchart of the classification and identification method based on industrial big data provided in the embodiments of the present invention. Detailed Implementation

[0025] To further understand the invention's content, features, and effects, the following embodiments are provided, and detailed descriptions are given in conjunction with the accompanying drawings.

[0026] The structure of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] like Figures 1 to 2 As shown in the figure, an embodiment of the present invention provides a classification and recognition system based on industrial big data, including a multimodal data acquisition and annotation module, used to acquire multimodal data of industrial waste and perform dual annotation, divide the dataset, merge and upload it to a hybrid storage architecture; wherein, the multimodal data includes at least image data; The multi-dimensional data preprocessing module is used to perform basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality. The multimodal feature fusion and dynamic extraction model training module is used to build a feature extraction architecture and complete multimodal feature fusion, extraction, and dynamic model optimization training. The hierarchical multi-classifier training and dynamic optimization module is used to build a hierarchical classification architecture and complete classifier training, optimization, and preliminary classification result output. The real-time classification and dynamic update module is used to complete the real-time classification of industrial waste, calculate the classification confidence level, and provide feedback on the classification results, thereby achieving bidirectional optimization of data and models.

[0028] It also includes a loop processing module, which is based on five core data processing modules and executes them in a preset order until a stopping condition is met. Specifically, it includes: S1. Initialize the system, complete the first multimodal data acquisition and annotation of the multimodal data acquisition and annotation module, obtain the initial raw data, set the initial number of loops k=1, k is a positive integer, k≥1, and set the relevant parameters of the evaluation parameter calculation formula and the dynamic threshold calculation formula. S2. This loop is executed in the following order: multimodal data acquisition and annotation module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module. Each module is linked in turn, and the output of the previous module is used as the input of the next module. After the hierarchical multi-classifier training and dynamic optimization module outputs the preliminary classification results, the evaluation parameters P(k) and dynamic threshold T(k) of this loop are calculated according to the preset formula. S3. Compare the evaluation parameter P(k) with the dynamic threshold T(k) to determine whether the stopping condition is met. If it is met, stop the loop and output the final classification data. If it is not met, update the data and the number of loops, and return to step S2 to repeat the loop.

[0029] (I) Specific Implementation of Core Modules Multimodal data acquisition and annotation module (Multimodal data acquisition and annotation) Collect basic physical characteristics data (shape, color, weight, dryness and wetness) and compositional characteristics data (elemental composition collected by a portable spectrometer and volatile organic compound content collected by a gas sensor) of industrial waste, as well as environmental auxiliary data (light intensity and dust concentration at the time of collection). The data is double-labeled using a combination of manual labeling and machine pre-labeling. The labeling content includes the type of waste (refined into 12 categories: plastics containing heavy metals, ordinary plastics, toxic sludge, ordinary wet waste, waste metal, waste glass, waste rubber, waste textiles, hazardous chemical residues, construction waste, electronic waste, and other industrial waste), the hazard level (0-3), and the recommended treatment method (recycling, harmless incineration, secure landfill, and special treatment). The labeled data is divided into training, validation, and test sets, and uploaded to a hybrid storage architecture of edge computing nodes and cloud big data platforms to provide raw data support for subsequent iterative processing.

[0030] Multi-dimensional data preprocessing module (Multi-dimensional data preprocessing) Basic preprocessing: missing data is filled using a classification imputation method (numeric data is filled with the median, and categorical data is filled with the mode), duplicate data is cleaned, and outliers are removed using a combination of the "3σ principle and the Local Outlier (LOF) algorithm". Feature enhancement processing: Wavelet transform and smoothing filtering are used to reduce noise in the component feature data, and normalization and feature correction are used in the environmental auxiliary data (color data is corrected according to light intensity, and shape recognition results are corrected according to dust concentration). Data standardization: RankGauss normalization is used for physical feature data, Z-score normalization is used for component feature data, and Min-Max normalization is used for environmental data to unify the data scale and provide high-quality data for feature extraction and classification training.

[0031] Multimodal Feature Fusion and Dynamic Extraction Model Training Module (Multimodal Feature Fusion and Dynamic Extraction Model Training) A feature extraction architecture of "branch extraction and cross-modal attention fusion" is constructed: Branch 1 (physical feature branch) uses an improved sparse filtering algorithm (introducing adaptive regularization parameters) to extract local and global features of physical features; Branch 2 (component feature branch) uses a deep autoencoder (DAE) to extract high-order nonlinear features of component data; Branch 3 (environmental auxiliary branch) uses principal component analysis (PCA) to extract key influencing factors of environmental data. A cross-modal attention mechanism is introduced to perform weighted fusion of features from the three branches. The weights are dynamically adjusted based on the confidence of each modality (the confidence is calculated inversely from the classification accuracy of the validation set). Construct a dynamic optimization objective function: with the dual objectives of maximizing feature discriminative power and minimizing model complexity, combined with an online learning mechanism, the model can update the fusion weights and feature extraction parameters in real time when new data is input; The dynamic feature extraction model is trained using the training set, and the hyperparameters are adjusted using the validation set to ensure the stability of feature extraction under different garbage types, and the fused high-order feature data is output.

[0032] Hierarchical Multi-Classifier Training and Dynamic Optimization Module A hierarchical classification architecture of "first-level coarse classification, second-level fine classification, and third-level attribute classification" is constructed: the first-level coarse classifier adopts an improved SoftMax logistic regression to divide waste into 4 major categories based on fused features; the second-level fine classifier uses a dedicated model for each of the 4 major categories to refine them into 12 subcategories; the third-level attribute classifier adopts gradient boosting decision tree (GBDT) to output the hazard level and recommended treatment method. The training process introduces a "dynamic learning rate adjustment and incremental training mechanism". The cosine annealing algorithm is used to adjust the learning rate. When incremental training is triggered, only some parameters of the model are updated to ensure training efficiency. The optimization objective is to minimize the weighted sum of multi-level classification error rates. The model performance is verified through a test set, and the classification model parameters and preliminary classification results are output.

[0033] Real-time classification and dynamic update module: Collects multimodal data of industrial waste in real time, and quickly completes preprocessing and feature extraction through edge computing nodes to ensure efficient real-time classification response; calls the trained hierarchical multi-classifier to output the detailed categories of waste, hazard levels, recommended treatment methods, and calculates classification confidence; establishes a classification result feedback mechanism: transmits the correct data after manual verification and the actual treatment effect data back to the cloud to provide feedback data for model optimization and data correction in subsequent cyclic processing, realizing two-way optimization of data and model.

[0034] (II) Circular Processing Flow One of the core innovations of this invention lies in breaking the linear processing logic and using a cyclical sequential processing method for data. The specific process is as follows: Initialization: Start the system, complete the first multimodal data acquisition and annotation of the multimodal data acquisition and annotation module, obtain the initial raw data, and set the initial number of loops k=1 (k is a positive integer, k≥1). The first loop executes in the following order: multimodal data acquisition and annotation module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module. During the execution of each module, the output of the previous module is used as the input of the next module. After the hierarchical multi-classifier training and dynamic optimization module outputs the preliminary classification results, the evaluation parameter P(k) for this loop is calculated, and the dynamic threshold T(k) is calculated based on the current loop number k. Loop Check: Compare the current evaluation parameter P(k) with the dynamic threshold T(k). If P(k) ≥ T(k), the stopping condition is met, the loop stops, and the final classification data (including subcategories, hazard levels, and recommended treatment methods) is output. If P(k) < T(k), the stopping condition is not met, and the output results of the hierarchical multi-classifier training and dynamic optimization module and the feedback data of the real-time classification and dynamic update module are sent back to the multimodal data acquisition and annotation module to update the original data and annotation information. The loop count is k = k + 1, and the next loop begins. Repeat steps 2 and 3 until the evaluated parameter P(k) is greater than or equal to the dynamic threshold T(k) in a certain loop. Stop the loop and output the final data. If the number of loops reaches the preset maximum limit (which can be set according to the actual scenario to avoid infinite loops), stop the loop and output the current best classification data even if the threshold is not reached, and issue a reminder signal.

[0035] (III) Evaluation Parameters and Calculation Formulas The evaluation parameter P(k) is the core quantitative indicator for measuring the classification accuracy of the k-th iteration. It comprehensively considers three key dimensions: classification accuracy, feature discrimination, and data confidence. It is calculated using a weighted summation method to ensure that the parameter can comprehensively and objectively reflect the accuracy effect of the iterative processing. The specific calculation formula is as follows: Calculation formula: Parameter descriptions: P(k): The evaluation parameter for the kth iteration, with a value range of [0,1]. The larger P(k) is, the higher the classification accuracy of this iteration. A(k): The classification accuracy in the kth iteration, which is the ratio of the number of correctly classified garbage samples to the total number of samples in the classification results output by the hierarchical multi-classifier training and dynamic optimization module. The value range is [0,1]. D(k): The feature discrimination in the k-th iteration, i.e., the degree of separation of different categories of industrial waste features in the fused features output by the multimodal feature fusion and dynamic extraction model training module, calculated by the ratio of intra-class variance to inter-class variance (which can be expressed as...). The value range is [0,1]. The larger D(k) is, the higher the feature discrimination and the easier it is to classify accurately. C(k): The data confidence score in the kth iteration, which is the average classification confidence score output by the real-time classification and dynamic update module. The value range is [0,1]. The larger C(k) is, the higher the reliability of the data and classification results. α, β, γ: Weighting coefficients, used as weighting terms in the numerator of the fraction, with the denominator α + β + γ = 1 (ensuring the calculation result is consistent with the original logic, and the value range remains [0,1]). These can be adjusted according to the actual industrial waste classification scenario. The default values ​​are: α = 0.5 (classification accuracy is the core weight), β = 0.3 (feature discrimination is the secondary weight), and γ = 0.2 (data confidence is the auxiliary weight), ensuring that the evaluation parameters are prominent and comprehensive. The fractional design can further standardize the weighted calculation logic, avoid weight superposition deviation, and meet the requirements of the fractional form.

[0036] (iv) Dynamic threshold and calculation formula The dynamic threshold T(k) is the core criterion for determining whether the loop should stop. Its value changes dynamically with the number of loops k, avoiding the problem that a fixed threshold cannot adapt to the loop optimization process, leading to insufficient accuracy or wasted resources. In the early stages of the loop, the threshold is set relatively low to ensure that the loop can proceed normally and complete the initial optimization of data and model. In the later stages of the loop, the threshold gradually increases, forcing the classification accuracy to continuously improve. The specific calculation formula is as follows: Calculation formula: T(k) = T0 + (T max -T0)×[1-e^(-k / λ)] Parameter descriptions: T(k): The dynamic threshold for the k-th iteration, with a value ranging from [T0, Tk]. max The value of ] corresponds to the evaluation parameter P(k), and its range is [0,1]. T0: Initial threshold, i.e. the threshold when k=1, with a default value of 0.7, representing the minimum accuracy requirement in the early stage of the loop, which can be adjusted according to the actual scenario; T max: Maximum threshold, i.e. the upper limit of the threshold, with a default value of 0.95, representing the target accuracy of industrial waste classification, which can be adjusted according to the actual scenario (e.g., the classification of high-risk industrial waste can be increased to 0.98). k: Current loop count, k≥1 (positive integer); λ: Attenuation coefficient, which controls the rate at which the threshold rises. The default value is 5. The smaller λ is, the faster the threshold rises, which forces the accuracy to improve faster. The larger λ is, the slower the threshold rises, and the smoother the loop optimization. It can be adjusted according to the actual scenario.

[0037] Threshold change logic: When k=1, T(1)=T0, the threshold is low, ensuring that the basic processing can be completed in the first loop; as k increases, T(k) gradually approaches T max As the required precision increases, T(k) approaches infinity, and T(k) becomes infinitely close to T. max To achieve the ultimate optimization of accuracy.

[0038] Data acquisition equipment selection: Select specialized acquisition equipment adapted to industrial scenarios, covering multi-dimensional acquisition needs such as images, physical parameters, composition, and environment. The equipment has stable performance and strong anti-interference capabilities, and can meet the needs of continuous and accurate data acquisition in complex industrial environments, providing support for the cyclical data update of the multimodal data acquisition and annotation module.

[0039] Model training environment: Build a collaborative model training environment between the edge and the cloud, configure hardware and software systems adapted to multimodal data processing and model training, ensure the efficiency and stability of model training, support dynamic updates and optimization of model parameters during the loop, and support the rapid calculation of evaluation parameters and dynamic thresholds.

[0040] Initial settings for loop parameters: Based on the scenario of refined classification of ordinary industrial waste, the initial parameters are set as follows: evaluation parameter weighting coefficients α=0.5, β=0.3, γ=0.2; dynamic threshold parameters T0=0.7, T... max =0.95, λ=5; the maximum loop count is set to 20 (which can be adjusted according to the actual scenario).

[0041] Implementation steps details: Initialization: Start the system, loop count k=1, complete the first multimodal data acquisition and annotation through the multimodal data acquisition and annotation module, and upload it to the hybrid storage architecture; set the initial calculation parameters for evaluation parameters and dynamic thresholds, and start the loop control module.

[0042] Loop execution: The modules are executed in the following order: multimodal data acquisition and labeling module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module. After the hierarchical multi-classifier training and dynamic optimization module outputs the preliminary classification results, P(1) and T(1) are calculated according to the formula. If P(1) ≥ T(1), the loop stops and the results are output. If P(1) < T(1), the feedback data of the real-time classification and dynamic update module is sent back to the multimodal data acquisition and labeling module to update the original data. k=2 and the loop continues.

[0043] Optimization by loop: Repeat the above loop process. As k increases, T(k) gradually increases and P(k) continues to increase in the loop optimization. When k=8, assume P(8)=0.95 and T(8)=0.95, satisfying P(k)≥T(k), stop the loop, and output the final classification data (including sub-categories, hazard levels, and recommended processing methods).

[0044] Abnormal handling: If the number of loops reaches 20 and P(20) still does not reach T(20), the loop will stop, the current best classification data will be output, and a reminder signal will be issued to prompt the staff to check the parameter settings or equipment status. The loop can be restarted after optimization.

[0045] Implementation Results: Through the above-mentioned cyclical processing flow, combined with technologies such as multimodal fusion and dynamic optimization, the accuracy of industrial waste sorting has been significantly improved compared with existing technologies. It can effectively distinguish industrial waste with similar components, adapt to complex industrial scenarios, and avoid resource waste. It achieves dual optimization of accuracy and efficiency, and fully meets the actual needs of refined industrial waste sorting.

[0046] To further verify the feasibility of the evaluation parameter formula, dynamic threshold formula, and cyclic processing technology of this invention, two specific data verification examples from different industrial scenarios are designed below. By combining actual data calculation and cyclic judgment, the rationality of the formula logic and the solution is verified.

[0047] Verification prerequisites: Both examples use the following initial parameters (set to fit the actual industrial scenario and consistent with the parameters mentioned above): evaluation parameter weighting coefficients α=0.5, β=0.3, γ=0.2 (satisfying α+β+γ=1); dynamic threshold parameters T0=0.7, T… max =0.95, λ=5; maximum loop count = 20 times; stop looping when the evaluation parameter P(k) ≥ T(k) and output the classification result.

[0048] Example 1: Typical industrial waste scenario (simple composition, minimal interference, such as a mixture of ordinary plastics, scrap metal, and construction waste) The industrial waste in this scenario has a simple composition and no significant interference, making recycling optimization relatively easy. The accuracy of the calculation formula and the rationality of the recycling process are verified using the following data and verification process: Initialization: k=1 (first loop), collect 1000 labeled samples (300 ordinary plastics, 400 scrap metals, and 300 construction wastes) through the multimodal data acquisition and labeling module, and execute the loop in sequence.

[0049] First iteration (k=1): Calculate the basic parameters: classification accuracy A(1) = 0.75 (750 samples were classified correctly); feature discrimination D(1) = inter-class variance / intra-class variance = 0.7 (the feature separation between ordinary plastics and waste metals is good); data confidence C(1) = 0.72 (the average classification confidence output by the real-time classification and dynamic update module).

[0050] Calculation of evaluation parameter P(1) (substituting into the fractional formula): .

[0051] Calculation of dynamic threshold T(1): .

[0052] Loop judgment: P(1)=0.729<T(1)=0.745, the stopping condition is not met; the 250 misclassified samples (mainly construction waste and ordinary plastics confused) fed back by the real-time classification and dynamic update module are sent back to the multimodal data acquisition and labeling module to update the labeling data, k=2 to enter the next loop.

[0053] Second cycle (k=2): Updated basic parameters: A(2) = 0.82 (820 samples were correctly classified); D(2) = 0.78 (the multimodal feature fusion and dynamic extraction model training module optimized the feature fusion weights, improving feature discrimination); C(2) = 0.80 (the classification confidence was improved).

[0054] Calculate P(2): .

[0055] T(2) calculation: .

[0056] Loop condition: P(2) = 0.804 ≥ T(2) = 0.782, the stopping condition is met, stop the loop, and output the final classification result.

[0057] Verification conclusion: In this scenario, optimization was completed through two iterations. The calculation logic of P(k) and T(k) was correct, and the fractional formula calculation was without deviation. The iterative process closely followed the order of the multimodal data acquisition and annotation module, the multimodal feature fusion and dynamic extraction model training module, the real-time classification and dynamic update module, the multidimensional data preprocessing module, and the hierarchical multi-classifier training and dynamic optimization module. The feedback optimization mechanism was effective, and the classification accuracy improved from 75% to 82%, verifying the feasibility of the formula and technical solution.

[0058] Example 2: Complex industrial waste scenarios (complex composition, significant interference, such as mixed scenarios containing heavy metal plastics, toxic sludge, and ordinary wet waste, with light interference). The industrial waste in this scenario has a similar composition (e.g., plastics containing heavy metals versus ordinary plastics), and significant environmental disturbances, making recycling optimization quite challenging. This study verifies the adaptability of the formula to complex scenarios and the solution's resistance to interference and optimization capabilities. Specific data and the verification process are as follows: Initialization: k=1 (first loop). Collect and label 1200 samples (including 300 heavy metal plastics, 400 toxic sludge, and 500 ordinary wet waste) through the multimodal data acquisition and labeling module. There is slight light interference. The loop is executed in the following order: multimodal data acquisition and labeling module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module.

[0059] First iteration (k=1): Calculate the basic parameters: classification accuracy A(1) = 0.71 (852 samples were classified correctly, the main error was confusion between plastics containing heavy metals and ordinary plastics); feature discrimination D(1) = inter-class variance / intra-class variance = 0.65 (similar composition leads to low discrimination); data confidence C(1) = 0.69 (light interference leads to low confidence).

[0060] Evaluation parameter P(1) is calculated as follows: .

[0061] The dynamic threshold T(1) is calculated as follows: consistent with Example 1, T(1) = 0.745.

[0062] Loop judgment: P(1) = 0.688 < T(1) = 0.745, the stopping condition is not met; the 348 misclassified samples and illumination interference data fed back by the real-time classification and dynamic update module are sent back to the multimodal data acquisition and annotation module (update annotation) and the multidimensional data preprocessing module (optimize illumination correction parameters), k=2 to enter the next loop.

[0063] The 3rd iteration (k=3, intermediate iterations omitted, only key optimization nodes are shown): Updated basic parameters: After two iterations of optimization (k=2, k=3), the multi-dimensional data preprocessing module completed the correction of illumination interference, and the multi-modal feature fusion and dynamic extraction model training module optimized the cross-modal fusion weights. A(3)=0.83 (996 samples were correctly classified); D(3)=0.81 (the component feature discrimination was significantly improved); C(3)=0.82 (the confidence was improved).

[0064] Calculate P(3): .

[0065] T(3) calculation: .

[0066] Loop judgment: P(3) = 0.822 ≥ T(3) = 0.813, the stopping condition is met, stop the loop, and output the final classification result (including sub-categories, hazard level, and recommended treatment method).

[0067] Verification conclusion: In this complex scenario, optimization was completed through 3 iterations. The calculation logic of P(k) and T(k) formulas is consistent and the data is reasonable. The iterative processing effectively corrected interference factors, improved the differentiation of similar waste components, and improved the classification accuracy. This verified the adaptability of the formula in complex scenarios and proved the feasibility of the iterative scheme and anti-interference design.

[0068] Formula and Scheme Optimization Explanation: Through two examples, no logical deviations in the formula were found. The calculation result of the fractional formula for evaluating parameter P(k) can accurately reflect the classification accuracy. The dynamic threshold T(k) formula can dynamically increase the threshold according to the number of iterations, which meets the core requirements of loop optimization. The loop process can be continuously optimized through feedback data to adapt to different complexity scenarios. The only point that can be optimized is: In example 2, when k=1, P(1)=0.688<T(1)=0.745, which is slightly larger than expected. The efficiency of the initial loop optimization can be accelerated by adjusting the adaptation logic of the initial threshold T0 (T0 can be lowered to 0.68 in complex scenarios). The optimization will not affect the overall formula logic. After optimization, T0 can be adaptively adjusted according to the scenario type (T0=0.7 in normal scenarios and T0=0.68 in complex scenarios).

[0069] Working principle of the invention: The core working principle of this invention is modular cyclic linkage, dynamic threshold control, and multi-dimensional precise optimization, with the core objective of improving the accuracy of industrial waste classification. It combines multimodal fusion, dynamic model optimization, and other technologies to form a complete working logic, as detailed below: Initialization Phase: The multimodal data acquisition and annotation module completes the acquisition and annotation of raw multimodal data, providing the data foundation for the entire system. Simultaneously, the loop count k=1 is initialized, and the relevant parameters (T0, T...) of the threshold calculation formula are set.max , λ) and the weighting coefficients (α, β, γ) of the evaluation parameters; Loop processing phase: Each module is executed sequentially in a loop, forming a closed loop of "data acquisition, feature fusion, real-time feedback, data correction, and classification optimization". The multimodal data acquisition and annotation module provides raw data, while the multimodal feature fusion and dynamic extraction model training module completes multimodal feature fusion and extraction based on the raw data, providing high-quality feature support for classification. The real-time classification and dynamic update module completes real-time classification based on the extracted features, and at the same time provides feedback on classification confidence and abnormal data, providing correction direction for the data preprocessing of the multi-dimensional data preprocessing module (such as optimizing the noise reduction and correction parameters of the preprocessing based on the low confidence data). The multi-dimensional data preprocessing module performs precise preprocessing on the raw data based on feedback from the real-time classification and dynamic update module, removing interference and correcting biases, providing higher quality data for the hierarchical multi-classifier training and dynamic optimization module's classifier training. The hierarchical multi-classifier training and dynamic optimization module completes classifier training and optimization based on the optimized data, outputs more accurate classification results, and provides the data classification accuracy A(k) and feature discrimination D(k) required for calculating the evaluation parameters P(k) in this loop.

[0070] Dynamic Judgment Phase: After each loop, the evaluation parameter P(k) and dynamic threshold T(k) for this loop are calculated according to the preset formula. The comparison between the two determines whether to stop the loop. If P(k) does not reach T(k), the feedback data is sent back to the multimodal data acquisition and annotation module to update the data and enter the next loop to continuously optimize data quality, model parameters and classification effect. If P(k) reaches T(k), it means that the classification accuracy meets the requirements of the current scenario, the loop stops and the final result is output. Core logic: Through iterative processing, bidirectional linkage and dynamic optimization of each module are achieved, breaking the limitations of linear processing and forming a closed loop of data processing, feature extraction, and classification training. Each iteration can be corrected based on the processing effect of the previous iteration. At the same time, the design of dynamic thresholds ensures that the iterative optimization has a clear objective and can adapt to the needs of different iteration stages. While ensuring accuracy, it avoids the waste of resources caused by infinite loops, ultimately achieving a significant improvement in the accuracy of industrial waste classification.

[0071] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A classification and identification system based on industrial big data, characterized in that, include: The multimodal data acquisition and annotation module is used to collect multimodal data of industrial waste and perform dual annotation, divide the dataset, merge and upload it to a hybrid storage architecture; the multimodal data includes at least image data; The multi-dimensional data preprocessing module is used to perform basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality. The multimodal feature fusion and dynamic extraction model training module is used to build a feature extraction architecture and complete multimodal feature fusion, extraction, and dynamic model optimization training. The hierarchical multi-classifier training and dynamic optimization module is used to build a hierarchical classification architecture and complete classifier training, optimization, and preliminary classification result output. The real-time classification and dynamic update module is used to complete the real-time classification of industrial waste, calculate the classification confidence level, and provide feedback on the classification results, thereby achieving bidirectional optimization of data and models.

2. The classification and identification system based on industrial big data according to claim 1, characterized in that, It also includes a loop processing module, which is based on five core data processing modules and executes them in a preset order until a stopping condition is met. Specifically, it includes: S1. Initialize the system, complete the first multimodal data acquisition and annotation of the multimodal data acquisition and annotation module, obtain the initial raw data, set the initial number of loops k=1, k is a positive integer, k≥1, and set the relevant parameters of the evaluation parameter calculation formula and the dynamic threshold calculation formula. S2. This loop is executed in the following order: multimodal data acquisition and annotation module, multimodal feature fusion and dynamic extraction model training module, real-time classification and dynamic update module, multidimensional data preprocessing module, and hierarchical multi-classifier training and dynamic optimization module. Each module is linked in turn, and the output of the previous module is used as the input of the next module. After the hierarchical multi-classifier training and dynamic optimization module outputs the preliminary classification results, the evaluation parameters P(k) and dynamic threshold T(k) of this loop are calculated according to the preset formula. S3. Compare the evaluation parameter P(k) with the dynamic threshold T(k) to determine whether the stopping condition is met. If it is met, stop the loop and output the final classification data. If it is not met, update the data and the number of loops, and return to step S2 to repeat the loop.

3. The classification and identification system based on industrial big data according to claim 2, characterized in that, In step S3, the stopping conditions include two situations: one is that the evaluation parameter P(k) ≥ the dynamic threshold T(k); the other is that the number of loops reaches the preset maximum limit, and even if P(k) < T(k), the loop stops and the current best classification data is output, and a reminder signal is issued at the same time.

4. The classification and identification system based on industrial big data according to claim 2, characterized in that, In step S3, if the stopping condition is not met, the data update method is as follows: the output results of the hierarchical multi-classifier training and dynamic optimization module and the feedback data of the real-time classification and dynamic update module are sent back to the multimodal data acquisition and labeling module to update the original data and labeling information of the multimodal data acquisition and labeling module, and the loop count is updated to k=k+1.

5. A classification and identification system based on industrial big data according to claim 2, characterized in that, The evaluation parameter P(k) is calculated using a fractional formula, specifically: , where A(k) is the classification accuracy of the k-th cycle, D(k) is the feature discrimination of the k-th cycle, C(k) is the data confidence of the k-th cycle, and α, β, γ are weighting coefficients and α + β + γ = 1.

6. A classification and identification system based on industrial big data according to claim 5, characterized in that, The feature discrimination index D(k) is calculated as the ratio of intra-class variance to inter-class variance, specifically as follows: The value range is [0,1]; the weighting coefficients are assumed to be α=0.5, β=0.3, and γ=0.2 by default, and can be adjusted according to the actual industrial classification scenario.

7. A classification and identification system based on industrial big data according to claim 2, characterized in that, The formula for calculating the dynamic threshold T(k) is: T(k) = T0 + (T max -T0)×[1-e^(-k / λ)], where T0 is the initial threshold, T max λ is the maximum threshold, λ is the decay coefficient, and k is the current loop count; T0 defaults to 0.7; T max The default value is 0.95; the default value of λ is 5. Both can be adjusted according to the actual scenario, and T0 will be adaptively reduced to 0.68 in complex scenarios.

8. A classification and identification system based on industrial big data according to claim 1, characterized in that, The multimodal data acquisition and labeling module acquires physical characteristic data, compositional characteristic data, and environmental auxiliary data of industrial waste. The labeling method combines manual labeling and machine pre-labeling. The labeling content includes waste type, hazard level, and recommended treatment method. The labeled data is divided into training set, validation set, and test set and uploaded to a hybrid storage architecture of edge computing node and cloud big data platform.

9. A classification and identification system based on industrial big data according to claim 1, characterized in that, The multimodal feature fusion and dynamic extraction model training module adopts a feature extraction architecture of branch extraction and cross-modal attention fusion, including three feature extraction branches, which extract physical features, component features and environmental auxiliary features respectively. The features of the three branches are weighted and fused through a cross-modal attention mechanism. The weights are dynamically adjusted according to the confidence of each modality data, and the fusion weights and feature extraction parameters are updated in real time in combination with an online learning mechanism. At the same time, a dynamic optimization objective function with the dual objectives of maximizing feature discrimination and minimizing model complexity is constructed.

10. A classification and identification method based on industrial big data, characterized in that, Includes the following steps: Collect multimodal data of industrial waste and perform dual annotation, then divide the dataset, merge and upload it to a hybrid storage architecture; Perform basic preprocessing, feature enhancement, and data standardization on the collected multimodal data to optimize data quality; Construct a feature extraction architecture to complete multimodal feature fusion, extraction, and dynamic model optimization training; Construct a hierarchical classification architecture, complete classifier training, optimization, and output preliminary classification results; The system enables real-time classification of industrial waste, calculation of classification confidence, and feedback of classification results, achieving bidirectional optimization of data and models.

Images