A method and system for online monitoring and tracing the production quality of color steel profiled sheets and aluminum-manganese-magnesium roof panels

Abstract

The application belongs to the technical field of intelligent manufacturing, and discloses a production quality online monitoring and tracing method and system for color steel profiled plates and aluminum-magnesium-manganese roof plates. Special sensing equipment is deployed in the whole link of raw material entering, profile forming and coating curing. Terahertz spectrum sensors, laser composite sensing units, infrared thermal imagers and other devices work together, and are matched with adaptive adjustment and self-calibration technologies to ensure the accuracy of process parameters and defect information collection. Cross-modal feature extraction and double-branch models are used for defect identification, which can accurately identify subtle surface defects, accurately locate internal micro-cracks and plate shape deviations, and realize full-process online monitoring. A hybrid trusted architecture combining blockchain, NFT and RFID is used to configure a unique identifier and digital certificate for each batch of plates, embed core quality information, and implant a special chip on the surface of the plate to support fast reading and writing in all scenarios.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, specifically a method and system for online monitoring and traceability of production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels. Background Technology

[0002] Corrugated steel sheets and aluminum-magnesium-manganese roofing panels possess advantages such as lightweight, high strength, corrosion resistance, and convenient construction, and have been widely used in various building scenarios such as industrial plants, large stadiums, and logistics warehouses. Currently, quality monitoring during the production process of corrugated steel sheets and aluminum-magnesium-manganese roofing panels still mainly relies on offline sampling and manual visual inspection, which presents the following technical problems: First, the detection timeliness is insufficient, making it difficult to grasp the dynamic quality changes in the production process in real time. Problems such as plate shape deviation, uneven coating thickness, and surface scratches in the molding process often cannot be detected in time, which can easily lead to batches of unqualified products and significantly increase production losses and costs. Secondly, the detection accuracy is limited. The human visual inspection is not capable of identifying minor defects and is easily affected by subjective factors such as the experience and fatigue of the inspectors, making it difficult to meet the refined quality requirements of high-end buildings for roofing panels.

[0003] Existing traceability technologies largely rely on traditional database records, which commonly suffer from issues such as data tampering and incomplete traceability chains. In practical applications, existing data typically only records basic information such as raw material batches and production dates, failing to cover the entire process from raw material arrival, production and processing, quality inspection and warehousing to installation and use. Furthermore, centralized data storage models are prone to data loss or tampering due to human error or system failures. Once quality disputes arise, it is difficult to quickly and accurately pinpoint the root cause of the problem and effectively clarify the boundaries of responsibility. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for online monitoring and traceability of the production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels, comprising: Preferably, in the raw material entry stage, a coating thickness detector, a substrate composition analyzer, and a terahertz spectral sensor are deployed. The microscopic defects of the substrate are identified through terahertz penetration detection. Simultaneously, an environmental temperature and humidity sensing unit is configured to establish a reference library that associates material parameters with environmental parameters. The reference library can be specifically adapted according to the characteristics of roof panels of different materials. In the molding section, a composite sensing unit integrating laser and terahertz sensors is deployed to simultaneously acquire data on the plate's contour and internal stress distribution. An industrial camera equipped with a polarized light imaging module uses adaptive exposure adjustment technology to ensure stable identification of minor surface scratches and creases under both strong and low light conditions. Simultaneously, pressure and temperature sensors are configured, with built-in self-calibration algorithms to correct detection deviations in real time, ensuring the accuracy of process parameter acquisition. The sensor sampling interval dynamically adapts to the production speed, with a maximum acquisition frequency of 100Hz.

[0006] The adaptive exposure adjustment technology uses the industrial camera's built-in light sensor to collect ambient light intensity in real time, with the sampling frequency matching the camera's frame rate at 60fps. When the light intensity is ≥5000 lux, it automatically reduces the exposure time and increases the sensitivity; when the light intensity is ≤1000 lux, it automatically extends the exposure time and activates noise reduction mode. Simultaneously, it maintains a shutter speed that matches the production speed, specifically: shutter speed = production speed (m / min) × 0.015s / m. For example, when the production speed is 10m / min, the shutter speed is set to 0.15s to avoid motion blur and ensure consistent image sharpness for defects under different lighting conditions.

[0007] In the coating curing section, an infrared thermal imager, a coating adhesion sensor, a coating porosity sensor, and a VOCs concentration monitoring unit are deployed to simultaneously collect curing temperature, coating bonding strength, coating porosity, and other parameters related to green production. The sensing devices achieve bidirectional communication with edge computing nodes via industrial 5G, dynamically adjusting the acquisition frequency and detection accuracy according to production conditions, and reducing transmission redundancy through local data preprocessing at the edge nodes.

[0008] Preferably, the data fusion stage adopts a two-level anomaly removal mechanism that combines the improved 3σ criterion with the isolated forest algorithm. First, the 3σ criterion is used to filter explicit abnormal data, and then the isolated forest algorithm is used to perform in-depth screening of the remaining data to identify latent anomalies such as sensor drift. The isolated forest algorithm sets up a forest model with 100 isolated trees, and each tree has 256 samples, with a feature sampling ratio of 70%. The latent anomaly detection rule is: if the average path length of the data in the forest is less than 30% of the average path length of all data in the current batch to be screened (the current batch to be screened has 500 data sets, consistent with the improved 3σ criterion sliding window), it is considered a latent anomaly. The algorithm retrains the forest model every 10,000 data sets to adapt to the dynamic changes in production data.

[0009] The improved 3σ criterion is based on the normal distribution characteristics of roof panel production data. The confidence interval of the traditional 3σ is adjusted to 2.8σ. At the same time, a sliding window mechanism is added, with the window size set to 500 sets of data. The mean and standard deviation of the data are updated in real time to avoid the insufficient adaptation of static 3σ to dynamic production data. Explicit anomalies are defined as single data points that exceed the 2.8σ interval. After being removed, the valid data are retained for subsequent screening of implicit anomalies.

[0010] For multimodal data such as numerical process parameters, image-based surface defects, and waveform-based stress data, an adaptive standardization strategy is designed. Through data normalization and feature mapping, different types of data are converted into feature vectors of a unified dimension. At the same time, a horizontal federated learning framework is introduced to achieve collaborative training of data features among multiple plants based on encrypted shared parameters. The specific steps are as follows: ① Numerical data such as temperature and pressure are normalized using min-max and mapped to the [0, 1] interval, with the extreme values ​​of historical production data as the normalization benchmark; ②Image data such as surface defect images are normalized by dividing the pixel value by 255, and then edge and texture features are extracted and mapped into vectors through convolution kernels; ③ For waveform data such as terahertz waveforms, 500 sampling points before and after the peak are extracted, arranged in time series, normalized, and then the frequency features are extracted and mapped into vectors through Fourier transform. Finally, all types of data are mapped into 256-dimensional feature vectors.

[0011] The encrypted shared parameters adopt a homomorphic encryption algorithm, which only encrypts the feature vectors of process parameters and does not involve the original production data transmission. The collaborative training process is as follows: each factory trains the feature layer parameters of the model locally, encrypts them and uploads them to the federated learning center node. The center node aggregates and averages the encrypted parameters and then sends the aggregated parameters back to each factory to update the local model. The number of training iterations is uniformly set to 200 rounds, and the parameter synchronization interval of each round is ≤5 seconds to ensure the alignment of data features across multiple factories.

[0012] A multimodal fusion algorithm combining graph learning and attention mechanisms is set up. A sensor data association network is constructed through graph learning to uncover hidden associations between different sensor data. Then, the fusion weights of each data dimension are dynamically adjusted using the attention mechanism, and higher weights are given to key quality-related data. The fused data matrix is ​​output, and the fused data is hashed and timestamped through blockchain.

[0013] Preferably, the defect identification stage is designed with a cross-modal shared feature extraction layer, which extracts common features of laser point cloud, polarization image and terahertz waveform through deep convolutional neural network, and then connects them to two special identification branches to realize the identification of different types of defects; This deep convolutional neural network consists of 5 convolutional layers (with Same Padding) and 3 max pooling layers, using ReLU as the activation function. Laser point cloud data, polarization image data, and terahertz waveform data are uniformly converted into 256×64×64 feature maps through their respective adaptation layers (point cloud voxelization layer, image downsampling layer, and waveform feature flattening layer) before being input. The shared feature extraction layer finally outputs a 512-dimensional common feature vector.

[0014] Branch 1 is based on the improved YOLOv8 algorithm, which introduces deformable convolution to enhance the adaptability to irregular defects and combines coordinate attention mechanism to strengthen the feature response of defect area and identify subtle defects; Branch 2 is based on the improved U-Net network, which incorporates cross-modal features of laser point cloud and terahertz data, and improves the segmentation accuracy of defect boundary through multi-scale feature fusion to achieve semantic segmentation of internal microcracks and plate shape waviness deviation. The deformable convolution kernel size is 3×3, and the dilation rate is set to 2. The channel compression ratio of the coordinate attention mechanism is 16, and the feature response weight coefficients are normalized by the Sigmoid function. The improved YOLOv8 retains the original three detectors and adds a cross-scale feature fusion layer, fusing 8x, 16x, and 32x downsampled features. During training, an industrial roof panel defect dataset containing 100,000+ samples covering 20 types of defects is used. The initial learning rate is 0.01, and a cosine annealing decay strategy (Tmax=100) is adopted. The batch size is 32, the number of iterations is 500, and the loss function is CIoU+FocalLoss, αF=0.25, and γ=2.0.

[0015] The improved U-Net encoder has 4 layers (each layer contains 2 convolutional layers + 1 pooling layer), and the decoder has 4 layers (each layer contains 2 convolutional layers + 1 upsampling layer). The skip connections adopt the method of "feature concatenation + 1×1 convolution dimensionality reduction". Multi-scale feature fusion covers 2x, 4x, and 8x downsampled features, and the fusion weights are dynamically allocated through Softmax. The training dataset is consistent with branch one, with an initial learning rate of 0.001, an optimizer of Adam, β1=0.9, β2=0.999, batch size=16, 300 iterations, and a loss function of DiceLoss + cross-entropy loss.

[0016] A four-dimensional hierarchical model based on defect type, size, location, and working condition is established. The hierarchical threshold is dynamically adjusted according to the quality requirements of different application scenarios. The model also supports online self-updating driven by federated learning and continuously optimizes algorithm parameters by accumulating defect data from multiple factories.

[0017] The tiered threshold adjustment rule is as follows: Threshold = Basic threshold × Scene weight, with an industrial plant scene weight of 1.0 and a large venue scene weight of 0.8. The basic threshold is obtained by fitting historical qualified data through a normal distribution. Federated Learning Online Self-Update Steps: 1) Train model parameters locally in each factory and encrypt them; 2) Aggregate parameters using the federated average algorithm; aggregation weight = factory data volume / total data volume. 3) After aggregation, the parameters are sent back to each factory to update the local model. The update is triggered when the number of newly added defect samples is ≥1000. Model input format: defect type, size, location, working condition; output is a graded result, divided into four levels: A, B, C, and D, corresponding to qualified, minor defect, moderate defect, and severe defect; working conditions include molding speed, curing temperature, and ambient humidity.

[0018] Preferably, the closed-loop control stage is based on the construction of a digital twin model of the entire roof panel production process based on fused data, which includes multi-dimensional information such as raw material characteristics, equipment operating status, and process parameter changes, to achieve real-time virtual-real correspondence of each link such as raw material entry, molding, and coating curing, and to simulate product quality under different combinations of process parameters through the digital twin model; A hybrid optimization strategy combining BP neural network and deep reinforcement learning is adopted. First, the BP neural network is trained on historical production data to build an initial correspondence model between process parameters and quality indicators. Then, the deep reinforcement learning algorithm is used to construct a reward function with the dual objectives of optimal quality and minimum energy consumption to dynamically optimize process parameters such as forming speed, roll surface temperature, and pressure. The BP neural network contains 3 hidden layers. The input layer consists of 6-dimensional process parameters, including forming speed, roller surface temperature, pressure, coating thickness, curing time, and ambient temperature and humidity. The output layer consists of 3-dimensional quality indicators, including plate shape deviation, coating adhesion, and porosity. The activation function is ReLU, the optimizer is SGD, the learning rate is 0.005, the momentum is 0.9, the training data size is ≥50,000 records, and the iteration stopping condition is that the validation set loss is ≤0.001. In deep reinforcement learning, the agent is a process parameter adjustment unit with a state space of (current process parameters + quality index + energy consumption value) and an action space of parameter adjustment (forming speed ±0.5m / min, roller surface temperature ±2℃, pressure ±0.1MPa). The reward function R = 0.6 × quality score + 0.4 × (1 - energy consumption normalization value). The exploration rate ε is initially 0.9, decays by 0.05 every 100 steps, and has a minimum of 0.1. The training cycle is 5000 steps.

[0019] When the quality indicators exceed the threshold detected online, the effect of parameter adjustment is first predicted through a digital twin model, and then the optimized parameter instructions are fed back to the production equipment control system in real time. At the same time, the quality data and energy consumption data before and after the parameter adjustment are recorded simultaneously.

[0020] Preferably, the traceability and evidence storage stage assigns a unique blockchain address and an NFT digital certificate as dual identifiers to each batch of boards. The NFT digital certificate uses non-homogeneous encryption encoding technology and contains core information such as raw material composition, production standards, and quality requirements. At the same time, a high-temperature resistant and corrosion-resistant passive RFID chip is embedded on the surface of the board. The chip stores the associated identifier of the NFT certificate, realizing the binding of physical boards and digital certificates and supporting reading and writing in the entire life cycle scenario. The NFT non-fungible encryption code uses the SHA-256 hash algorithm to generate a unique identifier with a code length of 64 bits. The built-in information is stored in JSON format, and key fields are protected by AES-128 encryption. The passive RFID chip operates in the UHF band at 860-960MHz, has a storage capacity of ≥512 bytes, a reading distance of 0.5-3m, an operating temperature range of -40℃ to 125℃, and an IP68 protection rating, which can withstand stamping, high-temperature curing and other processes in production and processing.

[0021] It adopts a hybrid blockchain architecture that combines consortium blockchains and public blockchains. The consortium blockchain is jointly maintained by core entities such as suppliers, manufacturers, construction parties, and supervisors. It stores detailed data on the entire process, including raw material entry, production and processing, quality inspection and warehousing, installation and maintenance. Key node data is synchronously anchored to the industry public blockchain through cross-chain technology. The decentralized nature of the public blockchain ensures that the data is tamper-proof, while the consortium blockchain enables data sharing among the various entities. The system introduces smart contracts to automatically execute the traceability process, pre-setting data recording trigger conditions for different stages. This includes automatically triggering production stage data recording after raw material quality inspection is qualified, automatically updating the traceability status after installation and acceptance, and locating the entire chain of related data through NFT credentials when quality disputes occur, tracing the stage and cause of the problem, and providing a basis for liability determination.

[0022] The specific trigger thresholds and recorded information are as follows: ① Raw material quality inspection qualification trigger conditions: coating thickness deviation ≤ ±0.02mm, substrate composition qualification rate ≥ 99.5%. After triggering, record the raw material batch number, testing equipment number, and testing personnel employee number; ② Data recording content in the production process: six core parameters including molding speed, roller surface temperature, pressure, curing temperature, coating thickness, and inspection timestamp; ③ Installation acceptance completion trigger conditions: installation position deviation ≤ ±5mm, fastener qualification rate 100%. After triggering, update the installation completion status and record the signature information of the construction party and the supervision party.

[0023] Preferably, the visual management stage supports access from multiple terminals, including PC, mobile, and AR glasses. By scanning RFID chips or NFT QR codes with AR, the full life cycle data of the board material can be retrieved and displayed in real time. At the installation site, AR glasses can overlay the board material installation standards and quality requirements with the actual installation scene to assist construction personnel in aligning the installation. Based on the digital twin model, the quality change trend of the board material is simulated under different usage environments. Combined with the material aging mechanism, early warnings are given for possible problems such as coating aging and board deformation, and the warning information is pushed to relevant management terminals. The platform supports multi-dimensional queries and statistics by batch, time, defect type, application scenario, etc., and automatically generates quality analysis reports and process optimization suggestions. At the same time, it is connected to the enterprise ERP system to realize the linkage management of quality data with production planning and cost accounting.

[0024] Preferably, the process optimization stage constructs a generative AI process optimization model based on a diffusion model, inputs constraints such as raw material parameters, production equipment status, and quality targets, and automatically generates multiple feasible process parameter adaptation schemes through the model's generation capabilities, while simultaneously predicting the quality effect and energy consumption level corresponding to each scheme; a dynamic decision-making mechanism for quality priority, energy consumption priority, and cost priority is established, and the weight of each priority is adjusted according to the different production needs of the enterprise, and the optimal solution is selected from the generated schemes; The diffusion model adopts the DDIM architecture with 1000 diffusion steps and a linear scheduler for noise. The input data format includes raw material parameters, equipment status, and quality targets, all standardized to the [0, 1] interval. Raw material parameters include substrate thickness, coating composition, material density, tensile strength, and elongation. Equipment status includes equipment running time, roller wear, and sensor accuracy. Quality targets include plate shape deviation threshold, coating adhesion threshold, and porosity threshold. The output consists of 10 process schemes and corresponding predicted values. Each process scheme includes 6 process parameters: molding speed 5-15m / min, roller surface temperature 80-150℃, pressure 1-5MPa, coating thickness 0.1-0.5mm, curing time 30-120s, and ambient humidity 30%-60%. The corresponding predicted values ​​include quality score, energy consumption value, and cost value. Dynamic decision weight formula: w_quality = αW × scene importance; w_energy consumption = βW × (1 - scenario importance); cost = 1 - αW - βW; αW and βW are adjustable coefficients from 0 to 1 and satisfy αW+βW≤1. For high-end venues, αW=0.6 and βW=0.3 (αW+βW=0.9≤1), and for ordinary factory buildings, αW=0.3 and βW=0.5 (αW+βW=0.8≤1). When αW+βW>1, it is automatically normalized proportionally, w_mass'=w_mass / (w_mass+w_energy_consumption), w_energy_consumption'=w_energy_consumption / (w_mass+w_energy_consumption), and w_cost=0, ensuring the self-consistency of the weight logic. The model training used historical process data from 10 factories, with a learning rate of 0.0001, a batch size of 64, and 800 training rounds. The selection criteria for generated solutions were: comprehensive score = w quality × quality score + w energy consumption × (1 - energy consumption normalization value) + w cost × (1 - cost normalization value), and the top 3 optimal solutions were selected.

[0025] To address the process differences between color steel profiled sheets and aluminum-magnesium-manganese roofing panels, specialized optimization sub-models were constructed for each. The optimized process parameters were applied to actual production, and the parameters of the generative AI model were iteratively optimized by comparing and verifying the production data with the predicted results. At the same time, a process optimization knowledge graph was constructed to collect the optimal process solutions under different working conditions.

[0026] When the actual quality effect of a certain process scheme deviates from the predicted value by more than ±8%, or the actual energy consumption deviates from the predicted value by more than ±10%, or when the same scheme is used in three consecutive batches of production and two or more defects of grade B or above occur, the model iteration is triggered. During the iteration, only the feature layer parameters related to the deviation are adjusted, and the adjustment range is ≤10%. After the iteration, it is verified by a small batch of 50-100 pieces. After the standard is met, it is updated to the process optimization knowledge graph.

[0027] This invention also provides an online monitoring and traceability system for the production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels. Based on the above method, it consists of a multi-source sensor acquisition module, a data fusion processing module, a defect identification and classification module, a closed-loop control module, a traceability and evidence storage module, a visualization management module, a process optimization module, and a core control unit. Each module achieves data interaction and linkage through industrial 5G communication and bus technology. The multi-source sensor acquisition module is deployed throughout the entire process of raw material entry, molding, and coating curing. It integrates terahertz spectral sensors, laser sensing units, industrial cameras, environmental sensing units, and other equipment to complete the real-time acquisition and preprocessing of material parameters, process data, and defect information. The data fusion processing module is equipped with an anomaly removal algorithm and a multimodal fusion algorithm. It uses a federated learning framework to achieve collaborative training of data from multiple factories, outputs fused data, and completes blockchain notarization. The defect identification and grading module constructs a cross-modal feature extraction and dual-branch identification architecture to realize the identification and four-dimensional grading of surface, internal and plate shape defects, and outputs standardized identification results. The closed-loop control module dynamically adjusts production process parameters based on a digital twin model and a hybrid optimization strategy. The traceability and evidence storage module adopts an architecture composed of blockchain, NFT and RFID to achieve reliable data storage and traceability throughout the entire process; The visualization management module supports multi-terminal access, enabling data visualization, quality trend warnings, and multi-dimensional statistical analysis, and linking with the enterprise ERP system to optimize resource allocation. The process optimization module is based on a generative AI model to generate process solutions that adapt to different needs and iteratively optimize them. The core control unit schedules the collaborative work of each module to achieve full-link data flow and overall system operation control.

[0028] The beneficial effects of this invention are as follows: 1. This invention deploys dedicated sensing equipment throughout the entire process of raw material entry, molding, and coating curing. Terahertz spectral sensors, laser composite sensing units, and infrared thermal imagers work together, combined with adaptive adjustment and self-calibration technologies, to ensure the accuracy of process parameters and defect information acquisition. Defect identification employs cross-modal feature extraction and a dual-branch model, which can accurately identify minute surface defects and precisely locate internal microcracks and plate shape deviations, enabling full-process online monitoring. It can detect plate shape deviations, uneven coatings, and other problems in real time, avoiding batch quality risks and reducing production losses and costs.

[0029] 2. This invention adopts a hybrid trusted architecture combining blockchain, NFT, and RFID, assigning a unique identifier and digital certificate to each batch of boards, embedding core quality information, and embedding a dedicated chip on the surface of the boards to support fast reading and writing in all scenarios; the consortium blockchain and public blockchain collaboratively store data, with multiple entities such as suppliers and manufacturers jointly maintaining the consortium blockchain, and key node data anchored to the public blockchain to ensure immutability, while also enabling cross-entity data sharing; smart contracts automatically trigger data recording at each stage, and when quality disputes occur, the entire chain of raw materials, production, and installation data can be quickly traced through exclusive certificates to locate the problematic link and cause.

[0030] 3. This invention relies on integrated data to construct a digital twin model of the entire production process, including multi-dimensional information such as raw material characteristics, equipment status, and process parameters. This enables real-time virtual-real correspondence between each production stage and can simulate the quality effects of different combinations of process parameters, providing a low-cost virtual test scenario for parameter optimization. A hybrid optimization strategy is adopted, aiming for optimal quality and minimum energy consumption, dynamically optimizing key process parameters such as forming speed and roller surface temperature. When quality indicators are abnormal, the digital twin model predicts the adjustment effect and feeds it back to the equipment control system, ensuring production stability. In the process optimization stage, a generative process optimization model is constructed, combined with a dynamic decision-making mechanism, to generate the optimal process solution based on different production needs. Specialized sub-models are built for the process differences of different types of roof panels, improving the targeting of optimization. By accumulating process knowledge, rapid adaptation to new production lines is achieved, improving product quality stability, reducing energy consumption and material loss, and linking with the enterprise ERP system to optimize resource allocation, thereby improving production efficiency and overall economic benefits. Attached Figure Description

[0031] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the multi-source sensing acquisition process of the present invention; Figure 3 This is a flowchart of the defect identification and classification process of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] like Figures 1 to 3 As shown, this embodiment of the invention provides an online monitoring and traceability method for the production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels, including: In the raw material entry stage, a coating thickness detector, a substrate composition analyzer, and a terahertz spectral sensor are deployed. The terahertz penetrating detection is used to identify micro-defects in the substrate, such as microcracks and impurity inclusions. At the same time, an environmental temperature and humidity sensing unit is configured to establish a reference library that links material parameters with environmental parameters. The reference library can be specially adapted according to the characteristics of roof panels of different materials. The core parameters of the benchmark library include: substrate thickness (0.5-2.0mm), coating thickness (5-20μm), substrate composition (percentage of elements such as iron, aluminum, magnesium, and manganese), and ambient temperature and humidity (temperature 5-35℃, humidity 30%-70%). For color-coated steel profiled sheets, the focus is on the correlation between coating thickness and ambient humidity. For every 10% increase in humidity, the coating thickness detection accuracy is compensated by 0.1μm. For aluminum-magnesium-manganese roofing panels, the focus is on the correlation between substrate composition and ambient temperature. For every 5℃ decrease in temperature, the composition detection sampling time is extended by 0.2 seconds. The benchmark library supports manually adding adaptation rules according to new material types.

[0034] In the molding section, a composite sensing unit integrating laser and terahertz is deployed to simultaneously collect data on the plate shape contour and internal stress distribution. The industrial camera is equipped with a polarized light imaging module, which uses adaptive exposure adjustment technology to ensure stable identification of minor scratches and creases on the surface under strong or weak light conditions. At the same time, pressure sensors and temperature sensors are configured, and a built-in self-calibration algorithm corrects detection deviations in real time to ensure the accuracy of process parameter acquisition. The sensor sampling interval is dynamically adapted according to the production speed. The self-calibration algorithm is triggered when 1000 sets of data are collected or when the deviation between the detection data and the reference library exceeds 0.01 mm. The calibration process is as follows: first, the standard data in the parameter association reference library is called, then the plate-shaped standard sample data collected by the laser sensing unit is used as a reference to calculate the current detection deviation value, and finally the output parameters of the pressure and temperature sensors are linearly corrected. After correction, the deviation is controlled within ±0.005 mm.

[0035] In the coating curing section, an infrared thermal imager, a coating adhesion sensor, a coating porosity sensor, and a VOCs concentration monitoring unit are deployed to simultaneously collect curing temperature, coating bonding strength, coating porosity, and green production-related parameters. The sensing devices achieve bidirectional communication with edge computing nodes via industrial 5G and dynamically adjust the acquisition frequency (up to 100Hz) and detection accuracy according to production conditions. Local data preprocessing at the edge nodes reduces transmission redundancy.

[0036] The VOCs concentration correlation rule is as follows: The safe threshold for VOCs is set at 80 mg / m³. 3 When the monitored value exceeds the threshold, the curing temperature is automatically adjusted accordingly (for every 10 mg / m² exceeding the threshold). 3 The curing temperature was increased by 5°C and the curing time was extended by 10 seconds. The VOC concentration and coating performance data before and after the adjustment were recorded. The green production parameters also include energy consumption indicators. The energy consumption for curing a unit board is ≤0.5kWh / kg. When the energy consumption exceeds the indicator, the pressing speed parameter is reduced by 0.5m / min.

[0037] Local data preprocessing specifically includes: preliminary filtering of outliers for collected numerical data such as temperature and pressure (removing data exceeding the equipment's measurement range), and data smoothing and noise reduction (using the moving average method with a window size of 5 data sets); size normalization (uniform scaling to 512×512 pixels) and grayscale correction for image data such as surface defect images; baseline correction and peak extraction for waveform data such as terahertz waveforms, retaining only valid feature data after preprocessing.

[0038] The data fusion stage employs a two-level anomaly removal mechanism that combines the improved 3σ criterion with the isolated forest algorithm. First, significant abnormal data is filtered out using the 3σ criterion, and then the isolated forest algorithm is used to perform in-depth screening of the remaining data to identify latent anomalies, such as gradual deviations caused by sensor drift, thus ensuring the reliability of the original data. For multimodal data such as numerical process parameters, image-based surface defects, and waveform-based stress data, an adaptive standardization strategy is designed. Through data normalization and feature mapping, different types of data are converted into feature vectors of a unified dimension. At the same time, a horizontal federated learning framework is introduced to achieve collaborative training of data features across multiple factories based on encrypted shared parameters. This avoids the privacy leakage risk caused by cross-factory transmission of raw data and integrates data from multiple scenarios, thereby improving the generalization ability of data fusion. A multimodal fusion algorithm combining graph learning and attention mechanisms is set up. Graph learning is used to build a network of sensor data associations and to uncover hidden relationships between different sensor data, such as the nonlinear relationship between molding pressure and plate shape deviation. Then, the attention mechanism is used to dynamically adjust the fusion weights of each data dimension, giving higher weights to key quality-related data and outputting a fused data matrix. The fused data is then hashed and timestamped through blockchain.

[0039] When constructing a relational network using graph learning, sensor devices are used as nodes and data correlations are used as edges. The Pearson correlation coefficient is calculated to quantify the correlation strength between different sensor data. A correlation coefficient with an absolute value ≥ 0.6 is considered a strong correlation. The attention mechanism calculates the mutual information value between each data dimension and quality indicator, and assigns weights according to the proportion of mutual information value. The higher the mutual information value, the greater the weight proportion. The weight proportion of key quality-related data such as coating adhesion and plate shape deviation is not less than 30%.

[0040] In the defect identification stage, a cross-modal shared feature extraction layer is designed. A deep convolutional neural network is used to extract common features of laser point cloud (plate shape), polarization image (surface), and terahertz waveform (internal), reducing feature redundancy and improving feature extraction efficiency. Then, it is connected to two specialized identification branches to realize the identification of different types of defects. Branch 1 (Surface Defect Recognition) is based on an improved YOLOv8 algorithm, introducing deformable convolution to enhance adaptability to irregular defects. It also combines coordinate attention mechanism to strengthen the feature response of defect areas, improving the positioning accuracy of irregular defects such as coating peeling and wavy edges. This branch has a detection speed of 60fps and can accurately identify surface micro-defects smaller than 0.5mm, meeting the real-time detection requirements of high-speed production lines. Branch 2 (Internal or Plate Shape Defect Recognition) is based on an improved U-Net network, incorporating cross-modal features of laser point cloud and terahertz data. It improves the segmentation accuracy of defect boundaries through multi-scale feature fusion, achieving semantic segmentation of internal micro-cracks and plate shape deviations. The defect area positioning error does not exceed 0.1mm. A four-dimensional hierarchical model based on defect type, size, location, and working condition is established. The hierarchical threshold is dynamically adjusted according to the quality requirements of different application scenarios (industrial plants, large venues). The model also supports online self-updating driven by federated learning. By accumulating defect data from multiple plants, the algorithm parameters are continuously optimized to improve generalization ability.

[0041] The specific grading standards are as follows: ① Grade A (Qualified): Defect size < 0.3mm, located in non-stressed area, and operating parameters within the standard range; ② Grade B (Minor Defects): Defect size ≤ 0.3mm < 0.8mm, located in a secondary stress area, with operating parameters fluctuating ≤ ±5%; ③ Grade C (Moderate Defect): Defect size ≤ 0.8mm < 1.5mm, located in the main stress area, or the fluctuation of operating parameters > ±5% and ≤ ±10%; ④ Grade D (Severe Defect): Defect size ≥ 1.5mm, or located in a critical stress area, or fluctuation of operating parameters > ±10%; the classification result is directly related to the intensity of subsequent process adjustments. Grade D triggers emergency shutdown adjustment, and Grade C triggers parameter fine-tuning.

[0042] The closed-loop control stage is based on the construction of a digital twin model of the entire roof panel production process based on fused data. It includes multi-dimensional information such as raw material characteristics, equipment operating status, and process parameter changes, so as to realize real-time virtual-real correspondence of raw material entry, molding, coating curing and other links. The virtual-real correspondence error does not exceed 0.5%. The digital twin model simulates the product quality under different process parameter combinations, providing a low-cost and high-efficiency virtual test scenario for parameter optimization and reducing material loss caused by physical tests. The model building process is as follows: ① Recreate the three-dimensional structure of equipment such as roller press and curing oven at a 1:1 scale according to the production equipment, and establish digital prototypes of different models of roof panels according to the specifications of the panels; ② Associate parameters such as raw material density, tensile strength, and coating adhesion from the fused data with the digital prototype to establish the correspondence between material parameters and physical properties; ③ Receive fused data synchronously at a frequency of 100ms / time, and dynamically adjust the equipment operating parameters and board defect status in the model; ④ Select three sets of quality inspection data from actual production boards daily, and use them to reverse-correct the physical property mapping coefficients of the model to ensure that the error between virtual and real correspondence is stable within ±0.5%.

[0043] When the deviation between virtual and real data in a certain link is greater than ±0.5% and two sets of data appear consecutively, model calibration is triggered. During the calibration process, the parameters of other links are frozen first, and only the mapping coefficient of the link with excessive deviation is corrected. After correction, it is verified by a new set of production data. If the deviation is ≤ ±0.5%, the calibration is completed. Otherwise, the correction is repeated, up to 3 times. If it still does not meet the standard, an equipment failure warning is triggered.

[0044] A hybrid optimization strategy combining BP neural network and deep reinforcement learning is adopted. First, the BP neural network is trained on historical production data to build an initial correspondence model between process parameters and quality indicators. Then, the deep reinforcement learning algorithm is used to construct a reward function with the dual objectives of optimal quality and minimum energy consumption to dynamically optimize process parameters such as forming speed, roll surface temperature, and pressure. The optimization response time does not exceed 50ms, ensuring the real-time performance of adjustment instructions. When online monitoring detects that quality indicators exceed the threshold, the effect of parameter adjustment is first predicted through a digital twin model to avoid quality fluctuations caused by blind adjustments. Then, the optimized parameter instructions are fed back to the production equipment control system in real time. At the same time, the quality data and energy consumption data before and after parameter adjustment are recorded simultaneously to ensure stable and controllable production quality.

[0045] The quality indicator thresholds are set dynamically, based on historical statistical data of qualified products and determined in conjunction with the requirements of the application scenario: the threshold for ordinary industrial plant scenarios is the 95th percentile of the statistical data, and the threshold for large venues and high-end scenarios is the 99th percentile. The thresholds are updated every 7 days based on the latest production data. During the update, abnormal production batch data, such as batches caused by equipment failure or unqualified raw materials, are removed to ensure that the thresholds are adapted to the actual production.

[0046] In the traceability and evidence storage stage, each batch of boards is assigned a unique blockchain address and an NFT digital certificate as dual identifiers. The NFT digital certificate adopts non-homogeneous encryption coding technology and contains core information such as raw material composition, production standards, and quality requirements, realizing irreplaceable traceability with one certificate for each item. At the same time, a high-temperature resistant and corrosion-resistant passive RFID chip is embedded on the surface of the board. The chip stores the NFT certificate association information and supports rapid reading and writing in the entire life cycle scenarios such as production, logistics, and installation. It adopts a hybrid blockchain architecture that combines consortium blockchains and public blockchains. The consortium blockchain is jointly maintained by core entities such as suppliers, manufacturers, construction parties, and supervisors. It stores detailed data on the entire process, including raw material entry, production and processing, quality inspection and warehousing, installation and maintenance. Key node data such as quality inspection results and defect records are synchronously anchored to the industry public blockchain through cross-chain technology. The decentralized nature of the public blockchain ensures that the data is tamper-proof, while the consortium blockchain enables data sharing among the various entities. The cross-chain technology employs a hash-locking mechanism combined with a relay chain. The relay chain uses dedicated cross-chain nodes from consortium blockchains to connect the consortium blockchain with industry public blockchains. The anchoring process is as follows: ① Consortium blockchain nodes calculate hash values ​​for key data such as quality inspection reports and defect records; ② Synchronize the hash value and data timestamp to the public chain via the relay chain; ③ The public chain records the hash value and the corresponding blockchain address to form an immutable anchor credential; the cross-chain synchronization delay is ≤3 seconds to ensure data real-time performance.

[0047] By introducing smart contracts to automatically execute the traceability process, data recording trigger conditions are preset for different stages. For example, data recording in the production stage is automatically triggered after the raw materials pass quality inspection, and the traceability status is automatically updated after installation and acceptance. When quality disputes occur, the NFT certificate is used to locate the relevant data across the entire chain, trace the stage and cause of the problem, and provide a basis for liability determination.

[0048] The visual management stage supports access from multiple terminals, including PC, mobile, and AR glasses. By scanning RFID chips or NFT QR codes with AR, it can retrieve and display the entire life cycle data of the board material in real time, including raw material information, production process, quality inspection report, and installation location. At the installation site, AR glasses can overlay the board material installation standards and quality requirements with the actual installation scene to assist construction personnel in aligning the installation and reduce human error. Supported AR device models include HoloLens 2, Magic Leap 2 and above; effective scanning distance is 0.5-3m; data transmission latency is ≤200ms; successful scanning feedback is provided by a "visual pop-up (displaying the unique identifier of the board) + voice prompt (data retrieval successful)"; when the scanning distance is outside the range or the signal is weak, a pop-up will prompt "Please adjust the distance to 0.5-3m and try again", and the signal reception power will be automatically increased.

[0049] Based on the digital twin model, the quality change trend of the board material is simulated under different usage environments (temperature, humidity, load). Combined with the material aging mechanism, early warnings are given for possible problems such as coating aging and board deformation. The early warning lead time is no less than 3 months, and the early warning information is pushed to relevant management terminals. The platform supports multi-dimensional queries and statistics by batch, time, defect type, application scenario, etc., and automatically generates quality analysis reports and process optimization suggestions. At the same time, it is connected to the enterprise ERP system to realize the linkage management of quality data with production planning and cost accounting, which helps enterprises optimize the allocation of production resources and improve overall operational efficiency.

[0050] The process optimization stage is based on a diffusion model to construct a generative AI process optimization model. It takes raw material parameters, production equipment status, quality targets and other constraints as input, and automatically generates a variety of feasible process parameter adaptation schemes through the model’s generation capabilities. It also predicts the quality effect and energy consumption level of each scheme at the same time. A dynamic decision-making mechanism for quality priority, energy consumption priority and cost priority is established. The weight of each priority is adjusted according to the different production needs of the enterprise. For example, the quality of high-end venue panels is given priority, while the cost of ordinary factory panels is given priority. The optimal solution is selected from the generated schemes. To address the process differences between color steel profiled sheets and aluminum-magnesium-manganese roofing panels, specialized optimization sub-models were constructed to enhance the targeting of optimization. The optimized process parameters were applied to actual production, and the parameters of the generative AI model were iteratively optimized by comparing and verifying production data with predicted results. At the same time, a process optimization knowledge graph was constructed to collect the optimal process solutions under different working conditions, enabling rapid adaptation and process replication for new production lines.

[0051] The optimization focus of the sub-model for color steel profiled sheets is on matching the forming speed with the roller pressure, with a forming speed range of 5-12 m / min and a roller pressure range of 2-4 MPa, emphasizing the optimization of sheet flatness. The optimization focus of the sub-model for aluminum-magnesium-manganese roofing panels is on matching the curing temperature with the coating thickness, with a curing temperature range of 100-150℃ and a coating thickness range of 0.2-0.5 mm, emphasizing the optimization of coating adhesion and corrosion resistance. The two sub-models adapt to the differences in material characteristics through independent feature extraction layers and share the main network structure of the model.

[0052] The core dimensions of the knowledge graph include material type (color steel / aluminum-magnesium-manganese), sheet thickness (0.5-2.0mm), equipment runtime (0-10000h), quality target (normal / high-end), and optimal process parameter set; the data storage structure adopts the form of "key-value" pairs, where the key is "material type + quality target + sheet thickness", and the value is the corresponding complete process parameters such as forming speed and roller surface temperature; The calling logic is as follows: After the basic parameters of the new production line are input, the system automatically retrieves the optimal solution with a similarity of ≥95% in the knowledge graph and outputs the solution adaptability assessment simultaneously, including equipment adaptability requirements and energy consumption estimates.

[0053] This invention also provides an online monitoring and traceability system for the production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels. Based on the above method, it consists of a multi-source sensor acquisition module, a data fusion processing module, a defect identification and classification module, a closed-loop control module, a traceability and evidence storage module, a visualization management module, a process optimization module, and a core control unit. Each module achieves data interaction and linkage through industrial 5G communication and bus technology. The multi-source sensor acquisition module is deployed throughout the entire process of raw material entry, molding, and coating curing. It integrates terahertz spectral sensors, laser sensing units, industrial cameras, environmental sensing units, and other equipment to complete the real-time acquisition and preprocessing of material parameters, process data, and defect information. The data fusion processing module is equipped with an anomaly removal algorithm and a multimodal fusion algorithm. It uses a federated learning framework to achieve collaborative training of data from multiple factories, outputs fused data, and completes blockchain notarization. The defect identification and grading module constructs a cross-modal feature extraction and dual-branch identification architecture to realize the identification and four-dimensional grading of surface, internal and plate shape defects, and outputs standardized identification results. The closed-loop control module dynamically adjusts production process parameters based on a digital twin model and a hybrid optimization strategy. The traceability and evidence storage module adopts an architecture composed of blockchain, NFT and RFID to achieve reliable data storage and traceability throughout the entire process; The visualization management module supports multi-terminal access, enabling data visualization, quality trend warnings, and multi-dimensional statistical analysis, and linking with the enterprise ERP system to optimize resource allocation. The process optimization module is based on a generative AI model to generate process solutions that adapt to different needs and iteratively optimize them. The core control unit schedules the collaborative work of each module to achieve full-link data flow and overall system operation control.

[0054] 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.

[0055] 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 method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels, characterized in that, include: Sensing deployment phase: Construct a multi-type sensor collaborative acquisition network to synchronously collect material properties, process parameters, environmental conditions and quality-related data, establish a parameter correlation benchmark library and complete local data preprocessing to achieve dynamic adaptation of acquisition frequency and accuracy; Data fusion phase: Establish a multi-level abnormal data screening mechanism to ensure the reliability of the original data, adopt an adaptive standardization strategy to unify the dimensions of multimodal data, mine the inherent correlation of data through collaborative fusion algorithms and dynamically allocate weights to generate fused data, and simultaneously complete the credible storage and traceability marking of fused data; Defect identification stage: Construct an integrated architecture for cross-modal feature extraction and specialized identification to identify and locate defects on the surface, inside and in the shape of the roof panel. Establish a four-dimensional defect classification model based on defect type, size, location and working condition, and support online self-updating optimization based on multi-scenario data. Output standardized defect identification and classification results. Closed-loop control stage: Based on the fusion data, a digital twin model of the entire roof panel production process is built to realize the simulation of the production process and the prediction of parameter optimization. Intelligent optimization strategies are adopted to dynamically adjust key process parameters and synchronously transfer all process and quality data to the traceability and evidence storage stage. Traceability and Evidence Preservation Stage: Configure a unique traceable identifier for the roof panel, build a multi-entity collaborative hybrid trusted storage architecture, store data from the entire production process and ensure that the data is tamper-proof and can be shared across entities, and automatically execute the traceability process through smart contracts to locate the root cause of quality problems; Visual management phase: Build a multi-terminal visual management platform to realize the retrieval of data throughout the entire life cycle of roof panels, quality trend early warning and multi-dimensional statistical analysis, link with the enterprise ERP system to optimize resource allocation, and output quality analysis reports and optimization suggestions; Process optimization stage: Establish an intelligent process optimization system with multi-objective dynamic decision-making, generate process solutions that adapt to different production needs and iteratively optimize them, accumulate the optimal process solution, and guide the setting of front-end production parameters in reverse.

2. The method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 1, characterized in that, In the sensor deployment phase, dedicated sensor units are deployed for the raw material entry section, molding section, and coating curing section respectively. Each sensor unit communicates bidirectionally with the edge computing node through industrial 5G. The acquisition frequency and detection accuracy are dynamically adjusted according to the production conditions, and local data preprocessing is completed simultaneously to reduce transmission redundancy. Among them, the raw material entry section is equipped with a coating thickness detector, a substrate composition analyzer and a terahertz spectral sensor, and an environmental temperature and humidity sensing unit is configured simultaneously to establish a reference library for the correlation between material and environmental parameters that is adapted to different material properties. The molding section is equipped with a composite sensing unit that integrates laser and terahertz sensors, an industrial camera, and a process parameter sensing unit. It has a built-in self-calibration algorithm to correct detection deviations, and the sensor sampling interval is dynamically adapted according to the production speed. The coating curing section is equipped with an infrared thermal imager, a coating performance sensor, and a VOCs concentration monitoring unit to simultaneously collect curing temperature, coating performance, and green production-related parameters.

3. The method for online monitoring and traceability of production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 2, characterized in that, The data fusion stage employs a two-level anomaly removal mechanism that combines the improved 3σ criterion with the isolated forest algorithm to filter explicit and implicit anomaly data in sequence. An adaptive standardization strategy is designed for multimodal data. Data is normalized and feature mapping is used to convert it into a unified dimension feature vector. A horizontal federated learning framework is introduced to achieve collaborative training of data features among multiple factories based on encrypted shared parameters. A multimodal fusion algorithm combining graph learning and attention mechanisms is adopted to mine the correlation of sensor data and dynamically adjust the fusion weight of each dimension. Key quality-related data are given higher weights, and a fusion data matrix is ​​output. The fusion data is hashed and timestamped through blockchain.

4. The method for online monitoring and traceability of production quality of color steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 3, characterized in that, The defect identification stage extracts common features from multiple types of sensor data through a cross-modal shared feature extraction layer, and then connects to two specialized identification branches to identify different types of defects. Branch 1 is based on the improved YOLOv8 algorithm to identify minute surface defects, and Branch 2 is based on the improved U-Net network to achieve semantic segmentation of internal microcracks and plate shape waviness deviations. A four-dimensional hierarchical model is established, which includes defect type, size, location, and working condition. The hierarchical threshold is dynamically adjusted according to the quality requirements of the application scenario. The model supports online self-updating driven by federated learning.

5. The method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 4, characterized in that, The closed-loop control stage is based on the construction of a digital twin model of the entire production process using fused data, which enables real-time virtual-real correspondence and quality simulation under multiple process parameter combinations in each link; A hybrid optimization strategy combining BP neural network and deep reinforcement learning is adopted to construct an initial correspondence model between process parameters and quality indicators, and to dynamically optimize key process parameters with the dual objectives of optimal quality and minimum energy consumption. When quality indicators exceed the threshold, the effect of parameter adjustment is first predicted through a digital twin model, and then the optimization parameter instructions are fed back to the equipment control system, while simultaneously recording the quality and energy consumption data before and after the adjustment.

6. The method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 5, characterized in that, The traceability and evidence storage stage assigns a unique blockchain address and NFT digital certificate dual identifier to each batch of boards, and embeds a high-temperature and corrosion-resistant passive RFID chip on the surface of the boards to store NFT-related information. It adopts a hybrid blockchain architecture that combines consortium blockchains and public blockchains. Multiple entities jointly maintain the consortium blockchain to store detailed data throughout the entire process, while key node data is synchronously anchored to the public blockchain through cross-chain technology to ensure immutability. By introducing smart contracts to pre-set data recording trigger conditions, the traceability process is automatically executed; when quality disputes occur, the entire chain of related data is located through NFT credentials to trace the problematic link and cause.

7. The method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 6, characterized in that, The visual management phase supports multi-terminal access and uses AR scanning technology to retrieve the full life cycle data of the board material in real time, and overlays the installation standards with the actual scene at the installation site. Based on the simulation of quality change trends using a digital twin model, early warnings of potential quality problems are issued and warning information is pushed out. It supports multi-dimensional query and statistics, automatically generates quality analysis reports and optimization suggestions, and integrates with enterprise ERP systems to achieve linked management of quality data, production plans, and cost accounting.

8. The method for online monitoring and traceability of production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels according to claim 7, characterized in that, The process optimization stage is based on a diffusion model to build a generative AI process optimization model. After inputting constraints, it automatically generates multiple process parameter adaptation schemes and predicts quality and energy consumption effects. Establish a multi-objective dynamic decision-making mechanism to adjust priority weights and select the optimal solution based on the enterprise's production needs; To address the process differences between the two types of roof panels, a specific optimization sub-model was constructed. The model parameters were iteratively optimized through comparison and verification, and a process optimization knowledge graph was built to accumulate the optimal process solution.

9. An online monitoring and traceability system for the production quality of color-coated steel profiled sheets and aluminum-magnesium-manganese roofing panels, based on the method described in claim 8, characterized in that, It consists of a multi-source sensor acquisition module, a data fusion and processing module, a defect identification and classification module, a closed-loop control module, a traceability and evidence storage module, a visualization management module, a process optimization module, and a core control unit. Each module achieves data interaction and linkage through industrial 5G communication and bus technology. Multi-source sensor acquisition modules are deployed throughout the entire production process to complete the real-time acquisition and preprocessing of various types of data. The data fusion processing module enables anomaly removal, multimodal data fusion, and blockchain-based evidence storage. The defect identification and grading module realizes defect identification and four-dimensional grading; the closed-loop control module dynamically adjusts process parameters based on a digital twin model and a hybrid optimization strategy. The traceability and evidence storage module adopts a blockchain, NFT and RFID integrated architecture to achieve full-process trusted traceability; the visualization management module realizes multi-terminal data visualization and linkage management; The process optimization module generates and iteratively optimizes process solutions based on a generative AI model; the core control unit schedules the collaborative work of each module to achieve full-link data flow and overall system control.

Images