Forming evaluation method for metal curtain wall plate

Abstract

The invention belongs to the technical field of curtain wall plate forming evaluation, and discloses a forming evaluation method for a metal curtain wall plate, and the method comprises the steps: constructing an improved Swin Transformer model fusing structural similarity and gradient features, capturing defects of different scales through introducing a multi-scale feature fusion module, and optimizing a training process through matching with a double-branch loss function, thereby achieving the forming evaluation of the metal curtain wall plate. Multiple types of defects can be accurately identified, and the capturing capability on tiny defects is enhanced; the method comprises the following steps: realizing double coverage of defect parameter quantification and macroscopic size detection by combining three-dimensional scanning full-coverage acquisition and a professional size measurement algorithm, and ensuring weight objectivity by improving an analytic hierarchy process based on a multi-dimensional evaluation index system so as to comprehensively reflect the quality of a plate through comprehensive scores; based on a distributed federal learning data acquisition framework, dynamic process parameters in a forming process and three-dimensional shape data after forming are acquired in real time, and the dynamic process parameters and the three-dimensional shape data are quickly processed by an edge calculation module and then input into a reinforcement learning agent model.

Description

A molding evaluation method for metal curtain wall panels Technical Field

[0001] This invention belongs to the field of curtain wall panel forming evaluation technology, specifically a forming evaluation method for metal curtain wall panels. Background Technology

[0002] The quality of metal curtain wall panel molding directly affects the building's facade appearance, structural safety, and service life. As the construction industry's demand for customized and high-precision molding of curtain wall panels continues to increase, traditional molding quality evaluation methods are increasingly inadequate for actual production needs. The main technical problems are as follows: Traditional metal curtain wall panel molding evaluation largely relies on manual sampling inspection of product dimensions. This can only determine a few macroscopic dimensional indicators such as panel length, width, and thickness, failing to comprehensively cover microscopic defects such as surface micro-cracks and internal stress concentration areas generated during the molding process, and also making it difficult to capture dynamic parameter deviations during molding. This single-dimensional evaluation method is prone to omissions and misjudgments, failing to accurately reflect the actual molding quality of the panels, thus affecting subsequent installation accuracy and building structural stability.

[0003] Existing evaluation methods are mostly offline inspection modes, requiring sampling inspection after the metal curtain wall panels are fully formed, making real-time monitoring and evaluation of the forming process impossible. Once quality problems are discovered, a large number of defective products have often already been produced, resulting in waste of raw materials and reduced production efficiency. Furthermore, while existing technologies include solutions for multi-source data fusion and defect identification, most lack dedicated algorithm design for metal curtain wall panel forming scenarios. They also lack integrated designs for data privacy protection, improved model generalization capabilities, and dynamic optimization of process parameters, exhibiting significant shortcomings and failing to meet the high-precision and high-efficiency requirements for forming evaluation in the context of intelligent manufacturing. Summary of the Invention

[0004] The purpose of this invention is to provide a molding evaluation method for metal curtain wall panels to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a molding evaluation method for metal curtain wall panels, comprising a data acquisition and processing stage, a defect feature extraction stage, a quality prediction and correlation stage, a three-dimensional detection and evaluation stage, a federated model optimization stage, and a report feedback closed-loop stage; preferably, the data acquisition and processing stage establishes a distributed federated learning data acquisition framework, with each production node deploying an edge computing module with edge computing and local caching functions, capable of receiving and temporarily storing acquired data in real time, synchronously acquiring dynamic process parameters during the molding process and three-dimensional morphological data after molding, with each node only uploading model parameter increments through an encrypted channel, and achieving collaborative data utilization through the parameter aggregation mechanism of the federated learning central server; the parameter aggregation process is as follows: after each node completes 10 rounds of local training, it uploads model parameter increments, the central server verifies the validity of the received parameter increments, removes abnormal parameters that exceed the normal fluctuation range, and then generates global parameters by weighted average according to the node sample size ratio, and after aggregation, the global parameters are returned to each node for initialization of the next round of local training, ensuring synchronous optimization of model parameters of each node.

[0006] The edge computing module has a single-node data processing capacity of no less than 1GB / s, a local cache capacity of 500GB, and supports data interruption and resumption. When the network is interrupted, it can temporarily store the collected data within 24 hours and automatically synchronize it to the encrypted channel after the network is restored.

[0007] Simultaneously, an adaptive multimodal data preprocessing algorithm was designed. The process parameter preprocessing adopted an improved wavelet threshold denoising algorithm to remove high-frequency noise, combined with the federated K-nearest neighbor algorithm to complete missing value filling, and Z-score standardization to achieve cross-node data scale uniformity. The number of neighbors in the federated K-nearest neighbor algorithm was set to 15, and a cross-node neighbor screening strategy was adopted to prioritize the process parameter data of the same material and specification of the board as the neighbor sample to ensure the pertinence and accuracy of missing value filling.

[0008] A point cloud filtering algorithm based on an attention mechanism is proposed for 3D morphological data preprocessing. This algorithm can identify and filter redundant point clouds caused by ambient light interference and equipment vibration. The coordinate system unification and density simplification of multi-node 3D point cloud data are achieved through a federated feature alignment module. Finally, a standardized federated shared process parameter dataset and a 3D point cloud dataset are generated.

[0009] Preferably, the defect feature extraction stage, based on the standardized 3D point cloud dataset generated in the acquisition and processing stage, constructs an improved Swin Transformer defect feature extraction model that integrates structural similarity and gradient features. The preprocessed 3D point cloud data is labeled with defects, including surface microcracks, depressions, protrusions, edge warping, and internal stress concentration areas. The 3D point cloud data is then converted into a sequence feature format suitable for Transformer processing through voxel encoding, preserving the spatial topological relationships of the point cloud during encoding to generate a fixed-length feature sequence. This improved Swin Transformer model contains four feature extraction modules, each consisting of two SwinTransformer blocks with head numbers of 4, 8, 16, and 16 respectively. The Transformer block employs a window attention mechanism with a window size of 7×7. It contains three sub-modules: layer normalization, multi-head self-attention, and a feedforward neural network. The hidden layer dimension of the feedforward neural network is four times the feature dimension, and the activation function is GELU to ensure the non-linear expressive power of feature extraction. The feature map size is downsampled to 1 / 2, 1 / 4, 1 / 8, and 1 / 16 sequentially through a window shift operation with a stride of 2. The multi-scale feature fusion module adopts a horizontal connection + top-down upsampling fusion structure. The output features of the four stages are passed through a 1×1 convolution kernel to unify the number of channels, then upsampled to the original feature map size through bilinear interpolation, and finally multi-scale feature fusion is achieved by element-wise addition. The fusion weights of the features at each stage are dynamically allocated by the attention mechanism.

[0010] Subsequently, a multi-scale feature fusion module was introduced into the Swin Transformer architecture. This module can simultaneously extract defect features at different scales, enhancing the ability to capture features of minute defects. A dual-branch loss function was designed to calculate the reconstruction loss based on structural similarity and the defect classification loss based on gradient magnitude similarity deviation, respectively. The model training was optimized through a dynamic weighted summation mechanism. The initial learning rate of the model training was set to 1e-4, and a cosine annealing learning rate scheduling strategy was adopted. The number of iterations was 200, and the batch size was set to 32. In the dual-branch loss function, the dynamic calculation formulas for the reconstruction loss weight α and the classification loss weight β are α=0.3+0.2×exp(-epoch / 100) and β=1-α, where epoch is the current training epoch. This formula enables dynamic optimization by emphasizing classification loss in the early stage of training and reconstruction loss in the later stage.

[0011] The system employs a federated prototype similarity distillation technique, where a central server aggregates model parameters from each node to generate a globally shared defect feature extraction model. Each node then fine-tunes the global model based on its local dedicated dataset. Cross-node knowledge distillation is achieved through prototype feature similarity calculation, integrating defect identification experience from different production scenarios, and outputting the extracted defect feature vector and the trained model.

[0012] The specific steps of federated prototype similarity distillation are as follows: ① Each node calculates prototype feature vectors for 5 types of defects based on its local dataset, and takes the feature mean of 100 samples for each type of defect; ② The central server collects the prototype feature vectors of each node and calculates the global prototype feature mean as the distillation target; ③ When each node performs local fine-tuning, a prototype feature similarity loss L is introduced. proto =||Local Prototype - Global Prototype||2 2 , and the weighted summation of the bi-branch loss function, L proto The weight is set to 0.1; the central server uses the FedAvg algorithm to aggregate parameters, and the aggregation weight of each node is the proportion of the sample size of the dataset of that node to the total sample size, with an aggregation interval of 10 training rounds.

[0013] Preferably, in the quality prediction correlation stage, reinforcement learning and multimodal data fusion classification techniques are integrated to construct a dynamic correlation and real-time prediction model between process parameters and molding quality. The preprocessed process parameter features from the acquisition and processing stage are fused with the defect feature vector output by the defect feature extraction model at a federated level. An attention weight allocation module is introduced during the fusion process. This module dynamically adjusts the weights of each feature dimension according to the degree of influence of different features on quality evaluation, generating a multi-dimensional fused feature vector that represents the nonlinear correlation between process parameters and defects. A deep reinforcement learning-based intelligent agent model is constructed, with training data derived from a federated shared process parameter dataset and defect feature vector set, covering at least 5000 groups of different materials (stainless steel, aluminum alloy, etc.) and different specifications (0.5...). The model uses forming data for sheet metal ranging from m×0.3m to 3m×1.5m. Each set of data includes a complete sequence of process parameter changes, defect features, and quality results to ensure the model is adaptable to training needs in multiple scenarios. Real-time process parameters and defect features are used as state inputs, while the quantitative value of defect severity and macroscopic dimensional deviation are used as output targets. A multi-objective reward function is designed to simulate the quality change patterns under different combinations of process parameters through a state transition model. The model learns the nonlinear correlation between process parameters and forming quality, enabling real-time prediction of forming quality and dynamic optimization suggestions for process parameters. The state space is 32-dimensional, with dynamic process parameters such as stamping pressure, bending angle, forming temperature, and feed speed occupying 4 dimensions, all normalized to the [0, 1] interval. The defect feature vector occupies 28 dimensions and is output by an improved Swin Transformer. The action space is a 4-dimensional continuous space, corresponding to the adjustment ranges of 4 types of process parameters: stamping pressure ±5MPa, bending angle ±0.5°, forming temperature ±3℃, and feed speed ±0.1mm / s, with an adjustment step size of 0.1 units.

[0014] The state transition model is adapted to the metal curtain wall panel forming scenario. When the adjustment of process parameters exceeds 5% or the change in defect severity exceeds 2 minutes, the state transition is triggered. During the transition, the previous 3 historical state data are retained to correct prediction deviations and ensure the model's adaptability to dynamic changes in the forming process.

[0015] The process parameters, which are collected and preprocessed in real time, are input into the reinforcement learning model through the edge computing module. The reinforcement learning model outputs the prediction results of defect type, severity and size deviation in real time. The dynamic thresholds of each quality index are preset. If the prediction result exceeds the threshold, a real-time warning is triggered and parameter adjustment suggestions are pushed to the molding equipment control system, and the prediction result is output.

[0016] Preferably, the three-dimensional inspection and evaluation stage combines the early warning information and key inspection guidance from the quality prediction association stage to perform three-dimensional laser scanning on the surface of the formed metal curtain wall panel. The scanning path adopts a spiral full-coverage planning. The acquired three-dimensional point cloud data is input into the trained improved Swin Transformer model to automatically identify and locate actual defects and quantify key parameters such as defect size, quantity, and distribution location. The scanning path starts from the geometric center of the panel and expands outward in a clockwise spiral direction. The scanning step size is set to 0.5mm, the overlap rate of adjacent spiral trajectories is 10%, and the step size in the edge area is reduced to 0.2mm. Priority is given to covering the defect areas predicted by the model and key parts such as panel corners and joints.

[0017] Simultaneously, a dimension measurement algorithm based on matrix operations is adopted. This algorithm can complete the coordinate transformation and dimension fitting of point cloud data, and calculate macroscopic dimensional indicators such as the length, width, thickness, and diagonal deviation of the plate. The specific operation steps are as follows: First, a local coordinate system is established with the preset reference hole of the plate as the origin. The point cloud data is mapped to this coordinate system through matrix transformation. Then, the point cloud data in the length and width directions are fitted with straight lines, and the thickness direction is fitted with planes. The extreme difference of the fitting results is taken as the corresponding dimension indicator. The diagonal deviation is calculated using the distance formula between two points.

[0018] Combining the evaluation experience of each node in federated learning with the industry-standard metal curtain wall panel forming quality, a multi-dimensional evaluation index system including appearance defects, dimensional accuracy, and structural integrity is established. An improved analytic hierarchy process is used to determine the weight of each index. The comprehensive quality score is calculated by weighted summation to achieve forming quality evaluation. The measured evaluation results and quality data are then output to the federated model optimization stage.

[0019] Preferably, the federated model optimization stage is based on the measured data and evaluation results output from the 3D inspection and evaluation stage, and constructs a federated model verification and iterative optimization mechanism. Each production node uses contact measuring instruments to measure the key dimensions and defect parameters of the board. The measured points are selected to cover the key areas of the board and the defect areas predicted by the model. The measured data is compared with the local model evaluation results point by point, and the error value is calculated. Each node encrypts the error data and uploads it to the central server using an asymmetric encryption algorithm. Based on the error data of each node, the central server uses federated model distillation technology to collaboratively calibrate the defect feature extraction model and the reinforcement learning prediction optimization model. For nodes whose errors exceed the allowable range, the model parameters are fine-tuned using the local gradient descent algorithm to achieve local optimization. At the same time, the evaluation data of each node is included in the federated learning training set, and a periodic iterative training cycle is set. The optimized model will be fed back to the aforementioned acquisition and processing stage, defect feature extraction stage, quality prediction correlation stage, and 3D inspection and evaluation stage.

[0020] In addition to monthly regular iterations, when the cumulative average error of each node as counted by the central server exceeds 30% of the allowable error of ±0.02mm (i.e. ±0.006mm), or when the number of newly added production nodes reaches 5, a dynamic iteration process is triggered. During dynamic iteration, the measured data of nodes with higher errors and the scenario data of newly added nodes are used first for model training, and the iteration rounds are adjusted to 100 rounds to accelerate model adaptation.

[0021] Preferably, the report feedback closed-loop stage connects the evaluation model output from the federated model optimization stage with the calibrated data. A multi-dimensional molding quality evaluation report is automatically generated using data processing report generation technology. The report uses a combination of text and graphics, and includes basic information about the sheet metal, molding process parameter curves, defect distribution maps, dimensional deviation data, comprehensive quality scores, and pass / fail determination results. For non-conforming sheets, a reinforcement learning model is used to trace the trajectory of process parameter changes, deduce the key process parameter influencing factors that lead to quality problems, generate process parameter adjustment schemes, and feed them back to the molding equipment control system. The tracing logic is as follows: priority is given to tracing process parameters strongly correlated with the defect type; for example, for dent defects, stamping pressure is prioritized, and for edge warping, bending angle is prioritized. The tracing time window is set to the process parameter change data within 30 seconds before the defect prediction result appears. By comparing the parameter fluctuation range of normal sheets at the same stage, parameters exceeding the fluctuation threshold are identified as key influencing factors.

[0022] Meanwhile, the reported data is processed through a federated encryption module for hierarchical encryption and then stored in a distributed database. The database uses a high-frequency backup mechanism to ensure data security, and the stored data will be used to supplement the training set in the federated model optimization phase.

[0023] The distributed database adopts a master-slave architecture, with one master node and three backup nodes. Data is stored in shards according to the production batch of the board. The high-frequency backup frequency is a full backup every hour plus an incremental backup every 15 minutes. In case of master node failure, it will automatically switch to the backup node within 30 seconds to ensure data continuity.

[0024] The beneficial effects of this invention are as follows: 1. This invention constructs an improved Swin Transformer model that integrates structural similarity and gradient features. By introducing a multi-scale feature fusion module to capture defects at different scales, and by optimizing the training process with a dual-branch loss function, it can accurately identify multiple types of defects such as surface microcracks and internal stress concentration, thereby enhancing the ability to capture minute defects. Combined with three-dimensional scanning full-coverage acquisition and professional dimensional measurement algorithms, it achieves dual coverage of defect parameter quantification and macroscopic dimension detection. Furthermore, relying on a multi-dimensional evaluation index system, it ensures the objectivity of weights through an improved hierarchical analysis method, and comprehensively reflects the quality of the board with a comprehensive score.

[0025] 2. This invention relies on a distributed federated learning data acquisition framework to collect dynamic process parameters during the molding process and three-dimensional morphological data after molding in real time. After rapid processing by the edge computing module, the data is input into the reinforcement learning agent model. Through the attention weight allocation module, multi-feature federated fusion is achieved to accurately characterize the nonlinear correlation between process parameters and quality, and to output defect prediction results and early warning information in real time. Once the prediction result exceeds the preset dynamic threshold, a real-time early warning is immediately triggered and a parameter adjustment plan is pushed to correct deviations in a timely manner during the molding process, reduce the generation of defective parts, reduce raw material waste and rework costs, and avoid the time loss of offline detection.

[0026] 3. This invention adopts a federated learning architecture, where each node only uploads model parameters rather than raw data. Combined with asymmetric and hierarchical encryption technologies, it establishes a robust privacy and security defense from data transmission to storage. Through federated prototype similarity distillation technology, it integrates evaluation experience from different production scenarios, and continuously optimizes model parameters through regular iterative training, improving the model's generalization ability and adapting to personalized customization and multi-condition production needs. The report feedback loop feeds back measured data and adjustment plans to model training and equipment control, forming a complete closed-loop process of collection, evaluation, optimization, and feedback, driving the continuous intelligent upgrading of the evaluation system. Attached Figure Description

[0027] Figure 1 is an overall flowchart of the method of the present invention; Figure 2 is a flowchart of the adaptive multimodal data preprocessing of the present invention; Figure 3 is a flowchart of the improved Swin Transformer defect feature extraction of the present invention. Detailed Implementation

[0028] 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.

[0029] As shown in Figures 1 to 3, this embodiment of the invention provides a molding evaluation method for metal curtain wall panels, including a data acquisition and processing stage, a defect feature extraction stage, a quality prediction and correlation stage, a three-dimensional detection and evaluation stage, a federated model optimization stage, and a report feedback closed-loop stage. The specific implementation of each stage is as follows: In the data acquisition and processing stage, a distributed federated learning data acquisition framework is built. Each production node deploys an edge computing module with edge computing and local caching functions, which can receive and temporarily store the acquired data in real time, and synchronously acquire dynamic process parameters during the molding process and three-dimensional morphological data after molding. Each node only uploads model parameters through an encrypted channel instead of raw data. Data collaborative utilization is achieved through the parameter aggregation mechanism of the federated learning central server to avoid data privacy leakage and data loss during transmission. The encrypted channel is built using the SSL / TLS 1.3 protocol, which supports two-way authentication and transmission encryption during data transmission. The communication port is uniformly configured as 4433. Data transmission adopts a fragmented transmission and verification code verification mechanism to ensure the integrity and anti-tampering capability of model parameter transmission.

[0030] Dynamic process parameters include stamping pressure, bending angle, forming temperature, and feed speed, with a collection frequency of 100Hz; the three-dimensional morphological data after forming is collected by laser scanning with an accuracy of 0.01mm, covering the entire surface and edge areas, including key areas prone to defects such as plate corners and seams.

[0031] An adaptive multimodal data preprocessing algorithm is designed. The process parameter preprocessing adopts an improved wavelet threshold denoising algorithm to remove high-frequency noise, combined with the federated K-nearest neighbor algorithm to complete missing value filling, and Z-score standardization to achieve cross-node data scale uniformity. The improvement of the improved wavelet threshold denoising algorithm is to introduce an adaptive threshold adjustment function based on the data fluctuation amplitude, which can dynamically adjust the threshold according to the characteristics of different process parameters.

[0032] The adjustment is based on the fluctuation range of process parameters: the fluctuation thresholds for stamping pressure and forming temperature are set to ±5%, and the fluctuation thresholds for bending angle and feed speed are set to ±3%. When the parameter fluctuations exceed the corresponding range, the thresholds are automatically increased by 1.2 times to retain valid fluctuation information.

[0033] The three-dimensional morphological data preprocessing proposes a point cloud filtering algorithm based on an attention mechanism, which can accurately identify and filter redundant point clouds caused by ambient light interference and equipment vibration. Through the federated feature alignment module, the coordinate system of multi-node three-dimensional point cloud data is unified and the density is simplified, and finally a standardized federated shared process parameter dataset and a three-dimensional point cloud dataset are generated.

[0034] The algorithm calculates the curvature and neighborhood point density of each point cloud, assigns high attention weight to effective surface points and low attention weight to redundant points, sets a weight threshold of 0.6, and filters point clouds with weights below the threshold. At the same time, it retains point clouds of key areas such as plate edges and seams, and retains 30% of them even if the weights are below the threshold to avoid feature loss.

[0035] The coordinate system uses the reference coordinate system of each node's production equipment as a reference, and establishes a transformation matrix through federated shared standard calibration board data to map the point cloud data of all nodes to a unified world coordinate system; density simplification adopts the voxel sampling method, setting the target point cloud density to 50 points / mm. 2 We retain the point within the voxel that is closest to the center, ensuring that the simplified point cloud can still fully represent the surface morphology of the plate.

[0036] The defect feature extraction stage is based on the standardized 3D point cloud dataset generated in the acquisition and processing stage. An improved Swin Transformer defect feature extraction model that integrates structural similarity and gradient features is constructed. The preprocessed 3D point cloud data is labeled with defects, including surface microcracks, depressions, protrusions, edge warping, and internal stress concentration areas. The stress distribution during forming is simulated by finite element simulation to help locate stress concentration areas and complete the labeling. The labeling accuracy reaches 0.005mm. The 3D point cloud data is then converted into a sequence feature format suitable for Transformer processing through voxel encoding. The spatial topological relationship of the point cloud is preserved during the encoding process, generating a fixed-length feature sequence.

[0037] The voxelized encoding generates a fixed-length feature sequence of 1024, which is divided into several 7×7 feature blocks according to the surface area of ​​the board. Each feature block corresponds to a local area of ​​the 3D point cloud. The window attention mechanism calculates the attention weight for each region of the feature block to ensure the balanced extraction of defect features in different regions of the board. The topological relationship between regions is preserved by splicing edge features between feature blocks.

[0038] The voxel size for voxel encoding is set to 0.05mm×0.05mm×0.05mm. After voxelizing the 3D point cloud data of the metal curtain wall panel, only voxels containing valid points are retained. The feature vector of each voxel is extracted by mean pooling and then arranged in spatial coordinate order to generate a feature sequence of fixed length 1024, ensuring that the sequence corresponds one-to-one with the spatial topology of the original point cloud.

[0039] A multi-scale feature fusion module is introduced into the Swin Transformer architecture. This module can extract defect features at different scales simultaneously, enhancing the ability to capture features of minute defects. A dual-branch loss function is designed to calculate the reconstruction loss based on structural similarity and the defect classification loss based on gradient magnitude similarity deviation, respectively. The model training is optimized through a dynamic weighted summation mechanism to improve the accuracy of defect identification and localization, and reduce the missed detection rate of minute defects.

[0040] The system employs a federated prototype similarity distillation technique, where a central server aggregates model parameters from each node to generate a globally shared defect feature extraction model. Each node fine-tunes the global model based on its local dataset. Cross-node knowledge distillation is achieved through prototype feature similarity calculation, integrating defect identification experience from different production scenarios to improve the model's generalization ability. The extracted defect feature vectors and the trained model are then output to the quality prediction association stage.

[0041] The quality prediction correlation stage connects with the output of the defect feature extraction stage, and integrates reinforcement learning and multimodal data fusion classification technology to construct a dynamic correlation and real-time prediction model between process parameters and molding quality. The preprocessed process parameter features from the acquisition and processing stage are fused with the defect feature vector output by the defect feature extraction model at the federated level. During the fusion process, an attention weight allocation module is introduced. The attention weight allocation module can dynamically adjust the weight of each feature dimension according to the degree of influence of different features on quality evaluation, and generate a multi-dimensional fused feature vector to represent the nonlinear correlation between process parameters and defects.

[0042] Weight adjustment is achieved through feature importance scoring. First, the Pearson correlation coefficient between each feature and the quality evaluation result is calculated. The correlation coefficient is normalized and used as the initial weight. Then, the gradient contribution value of the feature during model training is combined for iterative correction. The weight allocation scheme is updated every 10 rounds of training.

[0043] A deep reinforcement learning-based intelligent agent model is constructed, with real-time process parameters and defect features as state inputs and the quantitative value of defect severity and macroscopic dimensional deviation as output targets. A multi-objective reward function is designed, with defect control effect, parameter adjustment cost, and production efficiency as core indicators. The weights of each indicator can be dynamically configured according to production needs. The quality change law under different combinations of process parameters is simulated through a state transition model. The nonlinear correlation between process parameters and molding quality is continuously learned, realizing real-time prediction of molding quality and dynamic optimization suggestions for process parameters.

[0044] The multi-objective reward function formula is R=w1×(1-defect severity quantification value / 10)+w2×(1-|parameter adjustment amount| / maximum adjustment amount)+w3×production efficiency coefficient, where w1=0.5, w2=0.3, w3=0.2, which can be manually adjusted according to production scenario requirements. The defect severity quantification value ranges from 0 to 10 points. The production efficiency coefficient = actual production speed / standard production speed (standard production speed is set to 10mm / s). The initial exploration rate ε of reinforcement learning training is 0.9, which decays by 0.01 every 100 training steps, down to a minimum of 0.1. The discount factor γ is set to 0.95, and the number of iterations is 50,000 steps.

[0045] Defect severity quantification standards: Surface microcrack length <0.2mm: 1 point, 0.2-0.5mm: 3 points, 0.5-1mm: 5 points, >1mm: 10 points; Depth of indentation / protrusion <0.1mm: 1 point, 0.1-0.3mm: 3 points, >0.3mm: 8 points; Edge warping <0.2mm: 1 point, >0.2mm: 6 points; Area of ​​internal stress concentration region <1cm² 2 2 points, >1cm 2 The score is 7 points; the macroscopic dimensional deviation is calculated as |measured value - design value| / design value × 100%, rounded to two decimal places.

[0046] The process parameters, which are collected and preprocessed in real time, are quickly input into the reinforcement learning model through the edge computing module. The model outputs the predicted results of defect type, severity, and dimensional deviation in real time. Dynamic thresholds for each quality indicator are preset, such as a surface microcrack length threshold of 0.5 mm and a dimensional deviation threshold of ±0.1 mm. These thresholds can be adaptively adjusted according to the type of sheet. If the predicted result exceeds the threshold, a real-time warning is immediately triggered and parameter adjustment suggestions are pushed to the molding equipment control system via industrial Ethernet to ensure timely intervention in the molding process. The predicted results are then output to the three-dimensional inspection and evaluation stage.

[0047] The parameter adjustment suggestions are encapsulated in a standardized industrial control protocol (Modbus TCP) and sent to the equipment control system. After receiving the suggestions, the system first verifies the rationality of the parameters to ensure that the adjustment amount is within the rated operating range of the equipment. After the verification is passed, the adjustment is executed within 1 second. After the adjustment is completed, the execution result is fed back to the reinforcement learning model. The model then corrects the parameter prediction strategy for the next round based on the feedback result.

[0048] The adjustment rules are determined based on the core properties of the sheet metal: the defect threshold for stainless steel sheet metal is reduced by 20%, and for aluminum alloy sheet metal it is increased by 10%; when the sheet metal size is greater than 2m×1m, the dimensional deviation threshold is relaxed to ±0.15mm, and when it is less than 1m×0.5m, it is tightened to ±0.08mm.

[0049] In the three-dimensional detection and evaluation stage, the early warning information and key detection guidance from the quality prediction and correlation stage are combined to perform three-dimensional laser scanning on the surface of the formed metal curtain wall panel. The scanning path adopts a spiral full-coverage planning to ensure no scanning blind spots. The acquired three-dimensional point cloud data is input into the trained improved Swin Transformer model to automatically identify and locate actual defects and quantify key parameters such as defect size, quantity, and distribution location.

[0050] Meanwhile, a dimension measurement algorithm based on matrix operations is adopted. This algorithm can quickly complete the coordinate transformation and dimension fitting of point cloud data, calculate macroscopic dimensional indicators such as plate length, width, thickness and diagonal deviation, and control the measurement error within ±0.01mm.

[0051] Based on the evaluation experience of each node in federated learning and the industry-standard general criteria for metal curtain wall panel forming quality, a multi-dimensional evaluation index system covering appearance defects, dimensional accuracy, and structural integrity is established. Specific scoring criteria are as follows: Appearance defects (40 points): 40 points for no defects; deduct 2 points for each minor defect (microcrack less than 0.2mm in length), deduct 5 points for moderate defects (microcracks 0.2-0.5mm), and deduct 15 points for severe defects (microcracks greater than 0.5mm); Dimensional accuracy (30 points): 30 points for deviation ≤0.05mm, deduct 5 points for deviation 0.05-0.1mm, deduct 10 points for deviation 0.1-0.15mm, and deduct 20 points for deviation >0.15mm; Structural integrity (30 points): 30 points for no stress concentration areas, and stress concentration area <1cm². 2 Deduct 5 points, 1-3cm 2 Deduct 10 points, >3cm 2 20 points deducted.

[0052] An improved analytic hierarchy process (AHP) is used to determine the weights of each indicator. A comprehensive quality score is calculated through weighted summation to achieve final quality evaluation. The generated measured evaluation results and quality data are output to the subsequent federated model optimization stage. The improvement of the AHP is the introduction of entropy weighting to correct subjective weights and enhance the objectivity of weight allocation. The specific steps are: ① Collect indicator evaluation data from each node in the federation and calculate the information entropy of each indicator; ② Calculate the entropy weight based on the information entropy; the smaller the entropy value, the greater the indicator weight; ③ Combine the subjective weights and entropy weights in a 4:6 weighted ratio to obtain the final comprehensive weight; the subjective weights are obtained by pairwise comparison judgment matrices using the AHP.

[0053] The pairwise comparison judgment matrix is ​​constructed based on the production scenario coverage and evaluation data volume of each node in the federation. The weight of nodes covering more than 3 board specifications in the production scenario is set to 1.2, the weight of nodes with more than 1000 evaluation data sets is set to 1.1, and the weight of other nodes is set to 1.0. The pairwise comparison results of each node are integrated by weighted average to form a globally unified judgment matrix.

[0054] The federated model optimization stage is based on the measured data and evaluation results output from the 3D detection and evaluation stage. It constructs a federated model verification and iterative optimization mechanism. Each production node uses contact measuring instruments such as laser trackers and dial indicators to measure the key dimensions and defect parameters of the board. The measured points cover the key areas of the board and the defect areas predicted by the model. The measured data is compared with the local model evaluation results point by point, and the error value is calculated with an allowable error of ±0.02mm. Each node encrypts the error data and uploads it to the central server through an asymmetric encryption algorithm to avoid leakage and tampering of the original measured data.

[0055] The asymmetric encryption algorithm uses RSA-2048. The public key is distributed to each production node by the central server, while the private key is held only by the central server. The encryption process is as follows: after the node performs SHA-256 hash processing on the error data, it uses the public key to encrypt the hash value and the original error data to form ciphertext, which is then uploaded. The encrypted object is clearly defined as "hash value + original error data" to ensure that the data is not tampered with or leaked during the data transmission process.

[0056] Based on the error data of each node, the central server uses federated model distillation technology to collaboratively calibrate the defect feature extraction model and the reinforcement learning prediction optimization model. For nodes whose errors exceed the allowable range, the local gradient descent algorithm is used to fine-tune the model parameters to achieve local optimization, ensuring the consistency of the model evaluation accuracy of each node.

[0057] The collaborative calibration process is as follows: First, the error data of each node is classified and statistically analyzed according to defect type and plate specifications, and the average error under various scenarios is calculated; then, based on the global model, for scenarios with high error, such as edge warping detection of large aluminum alloy plates, high-quality model parameter fragments of the corresponding nodes are extracted and replaced; finally, through 50 rounds of small batch iterative training, the error of the calibrated model under various scenarios is controlled within the allowable range.

[0058] The learning rate is set to 5×10 during fine-tuning. -5 The iteration count is 50 rounds, using batch gradient descent. Each iteration selects 20% of the local dataset as the fine-tuning batch. When the loss function decreases by less than 1×10⁻⁶, the batch is fine-tuned. -6 Stop fine-tuning when the time comes.

[0059] Meanwhile, the evaluation data of each node, such as process parameters, three-dimensional morphological data, evaluation results, and error data, are incorporated into the federated learning training set. A regular iterative training cycle of once a month is set to continuously optimize the model parameters, improve the model's generalization ability and evaluation accuracy, and the optimized model will be fed back to the aforementioned stages such as the data acquisition and processing stage and the defect feature extraction stage.

[0060] The report feedback closed-loop stage connects the evaluation model output from the federated model optimization stage with the calibrated data. Through data processing report generation technology, a multi-dimensional molding quality evaluation report is automatically generated. The report adopts a combination of text and graphics, and the content includes basic information of the sheet metal, molding process parameter curves, defect distribution maps, dimensional deviation data, comprehensive quality scores, and pass / fail judgment results. The basic information of the sheet metal includes model, production batch, production time, etc.; the molding process parameter curves include the changing trends of key parameters; and the defect distribution maps intuitively mark the location and type of defects.

[0061] For non-conforming parts, the process parameter change trajectory is traced back through reinforcement learning model to deduce the key process parameter influencing factors that lead to quality problems, and generate personalized process parameter adjustment schemes, including adjustment range and adjustment steps. These schemes are fed back to the molding equipment control system through industrial bus to achieve closed-loop optimization. The adjustment schemes can also serve as the basis for prioritizing parameter acquisition during the acquisition and processing stages.

[0062] Meanwhile, the reported data is processed through a federated encryption module for hierarchical encryption and then stored in a distributed database. The database uses a high-frequency backup mechanism to ensure data security, and the stored data will be continuously added to the training set during the federated model optimization phase to further improve the intelligence level of the evaluation system.

[0063] The grading standard is divided into three levels based on data sensitivity: Level 1 data, such as core data of board defects and process adjustment plans, is encrypted using AES-256; Level 2 data, such as dimensional deviation data and process parameter curves, is encrypted using AES-128; and Level 3 data, such as basic information of board components and qualification judgment results, is encrypted using DES. The keys are dynamically distributed by the central server according to the node permissions.

[0064] 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.

[0065] 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 evaluating the forming of metal curtain wall panels, characterized in that, The system comprises four phases: data acquisition and processing, defect feature extraction, quality prediction and correlation, 3D inspection and evaluation, federated model optimization, and report feedback closed-loop. The data acquisition and processing phase involves constructing a distributed federated learning data acquisition framework to simultaneously acquire process parameters and 3D morphological data, uploading only model parameters rather than raw data. An adaptive multimodal data preprocessing mechanism is designed to perform targeted preprocessing on both types of data, generating a standardized federated shared dataset. The defect feature extraction phase involves constructing an improved Swin Transformer defect feature extraction model based on the standardized shared dataset. After data defect labeling and format conversion, a globally shared model is generated through a federated collaborative training mechanism to extract defect feature vectors. The quality prediction and correlation phase involves constructing a dynamic correlation and real-time prediction model integrating multiple technologies. A feature fusion mechanism generates multi-dimensional feature vectors to represent the correlation between process parameters and defects. A reinforcement learning model is used to achieve real-time prediction of molding quality and dynamic optimization suggestions for process parameters. Prediction results are output in real-time with early warnings, and adjustment suggestions are pushed to intervene in the molding process. 3D Inspection and Evaluation Stage: Surface 3D scanning and dimensional measurement are performed on the formed sheet metal. Defect parameters are identified and quantified using the trained model. A multi-dimensional evaluation index system is constructed by combining industry-standard practices and evaluation experience from each node of the federated learning process. A comprehensive quality score is calculated, and the measured data and comprehensive score are output to the federated model optimization stage. Federated Model Optimization Stage: A federated verification and optimization mechanism is built based on the measured data. Each node calculates the error by comparing the measured data with the model results and uploads the data in encrypted form. The central server completes collaborative model calibration and local optimization based on the error data, and incorporates the data from each node into the training set for periodic iteration. Report feedback closed-loop stage: Derivation of key process influencing factors for non-conforming parts and generation of adjustment suggestions are fed back to the equipment. The report data is encrypted and stored and then supplemented into the training set of the federated model optimization stage.

2. The forming evaluation method for metal curtain wall panels according to claim 1, characterized in that, In the acquisition and processing phase, the distributed federated learning data acquisition framework deploys edge computing modules on each production node to receive and temporarily store dynamic process parameters during the molding process and three-dimensional morphological data after molding in real time. Each node uploads model parameters through an encrypted channel, and the data is collaboratively utilized through parameter aggregation by the central server. The adaptive multimodal data preprocessing mechanism performs denoising, completion, and standardization processing on process parameters, and filtering and registration processing on three-dimensional morphological data, ultimately generating a standardized federated shared process parameter dataset and a three-dimensional point cloud dataset.

3. The method for evaluating the forming of metal curtain wall panels according to claim 2, characterized in that, The defect feature extraction stage is based on a standardized 3D point cloud dataset to build a model. The defect labeling categories include surface microcracks, depressions, protrusions, edge warping, and internal stress concentration areas. The 3D point cloud data is converted into a sequence feature format suitable for Transformer processing and the spatial topological relationship is preserved. The federated collaborative training adopts federated prototype similarity distillation technology. The central server aggregates the model parameters of each node to generate a globally shared model. Each node is fine-tuned based on its local dataset, and cross-node knowledge distillation is achieved through prototype feature similarity calculation.

4. The forming evaluation method for metal curtain wall panels according to claim 3, characterized in that, The multi-technology fusion in the quality prediction association stage specifically refers to the fusion of reinforcement learning and multimodal data fusion classification technology; The feature fusion process introduces an attention weight allocation module to dynamically adjust the weights of each feature dimension to generate a multi-dimensional fused feature vector. The reinforcement learning model takes real-time process parameters and defect features as state inputs and the quantitative value of defect severity and macroscopic size deviation as output targets. It learns the nonlinear relationship between process parameters and molding quality through a state transition model. When the prediction result exceeds the preset dynamic threshold, it triggers an early warning and pushes parameter adjustment suggestions to the equipment control system.

5. The method for evaluating the forming of metal curtain wall panels according to claim 4, characterized in that, The three-dimensional inspection and evaluation stage uses a spiral full-coverage scanning path to complete the three-dimensional laser scanning of the surface. The improved SwingTransformer model, which has been trained, identifies and locates actual defects and quantifies key parameters. The coordinate transformation and size fitting method is used to calculate macroscopic dimensional indicators, including the length, width, thickness, and diagonal deviation of the board. The multi-dimensional evaluation index system includes appearance defects, dimensional accuracy, and structural integrity. The improved analytic hierarchy process is used to determine the weight of each index and calculate the comprehensive quality score.

6. The method for evaluating the forming of metal curtain wall panels according to claim 5, characterized in that, During the federated model optimization phase, each production node uses contact measuring instruments to conduct actual measurements on key areas of the board and the defect areas predicted by the model. The error data is encrypted and uploaded using an asymmetric encryption algorithm. The central server uses federated model distillation technology to collaboratively calibrate the defect feature extraction model and the reinforcement learning prediction optimization model. For nodes whose errors exceed the allowable range, the parameters are fine-tuned using the local gradient descent algorithm and incorporated into the training set of evaluation data of each node for iterative optimization according to a set period.

7. The method for evaluating the forming of metal curtain wall panels according to claim 6, characterized in that, The report feedback closed-loop stage automatically generates a multi-dimensional forming quality evaluation report combining text and graphics through data processing report generation technology. The report includes basic information of the sheet, process parameter curves, defect distribution map, dimensional deviation data, comprehensive quality score and pass / fail judgment results. For non-conforming parts, the process parameter change trajectory is traced back through reinforcement learning model to deduce key influencing factors and generate adjustment schemes; the report data is stored in a distributed database after hierarchical encryption and is backed up frequently, and the stored data is added to the training set of the federated model optimization stage.

8. The method for evaluating the forming of metal curtain wall panels according to claim 7, characterized in that, The molding quality evaluation report includes a defect distribution map, presented in a combination of text and graphics.

9. The method for evaluating the forming of metal curtain wall panels according to claim 8, characterized in that, The distributed database employs a backup mechanism that combines full backup with incremental backup.

Images