Accurate forming method for floor support plate

Abstract

The invention belongs to the technical field of floor support plate forming, and discloses a floor support plate accurate forming method, which integrates dynamic parameter adaptation and hidden defect pre-judgment in the whole process from design drawing analysis to real-time production adjustment, effectively avoids hidden dangers such as stress concentration and local buckling, enables core indexes such as ultimate bearing capacity and midspan deflection of a floor support plate to stably reach the standard, and improves the production efficiency. The reject ratio of products is reduced, and the precision requirements of high-end buildings are met; real-time linkage of virtual equipment and physical equipment is realized through a 1: 1 digital twin system, and a federated learning architecture not only ensures multi-node data security, but also can realize collaborative optimization; the multi-working-condition simulation scene library and the parameterized model can flexibly adapt to different production requirements, intelligent parameter optimization, dynamic deviation correction and model iteration can be completed without manual intervention in the whole process, the process debugging period is shortened, meanwhile, full-amount tracing of production data is achieved, the forming process is changed from experience driving to data driving, and the production efficiency is improved. And the production controllability and the overall efficiency are improved.

Description

Technical Field

[0001] This invention belongs to the field of floor decking forming technology, specifically a method for precise forming of floor decking. Background Technology

[0002] As a core load-bearing component of prefabricated buildings, the forming precision of floor decking directly affects the safety, stability, and construction efficiency of the building structure. Currently, floor decking forming mostly adopts traditional mechanical rolling processes, relying on pre-set fixed molds and process parameters to complete the forming, which presents the following technical problems.

[0003] Traditional process parameter design relies heavily on manual experience and fails to fully consider the impact of variables such as material property fluctuations and environmental temperature changes on molding accuracy. This can easily lead to quality problems such as rib height deviation and uneven wave pitch in the product, resulting in large conventional errors that are difficult to meet the stringent millimeter-level precision requirements of super high-rise and large-span buildings.

[0004] With the deepening of industrialization in the construction industry, digital technologies such as BIM and finite element simulation have been gradually applied in the construction field. However, their deep integration in the field of floor decking molding still has significant shortcomings. Most existing technologies only apply digital technologies to the early design and modeling stage, failing to achieve closed-loop management of the entire process from design and simulation to production execution and quality feedback. Furthermore, finite element simulation models are mostly developed for single-specification floor decking, without forming a reusable and iterative parametric model library. This results in significant deviations between simulation results and actual molding, making it difficult to effectively guide actual production work. Summary of the Invention

[0005] The purpose of this invention is to provide a method for precise forming of floor decking, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for precise forming of floor decking, comprising the following specific steps: Preferably, the multi-field modeling stage constructs a multi-physics parameterized model of the floor decking that integrates dynamic feature perception. The multi-physics field specifically includes a structural mechanical field (describing the stress and strain distribution during the substrate molding process) and a thermodynamic field (describing the influence of temperature changes on material properties and molding accuracy). The two are integrated through parameter coupling, that is, the temperature data of the thermodynamic field is used as the input parameter of the structural mechanical field to dynamically correct the material mechanical property parameters, ensuring that the model fits the multi-field coupling effect in actual molding. The specific coupling logic is as follows: when the ambient temperature or the substrate temperature changes by 5°C during the molding process, the elastic modulus, Poisson's ratio and other parameters in the structural mechanical field are synchronously corrected through the material property dynamic prediction sub-model. The correction magnitude is linearly related to the temperature change, ensuring real-time linkage of multiple field data and conforming to the parameter coupling effect in actual molding.

[0007] An attention mechanism is used to achieve adaptive focusing of key parameters. BIM technology is used to analyze the floor deck design drawings, extract core geometric parameters such as rib height, wave pitch, and coverage width, and build a basic three-dimensional model. Simultaneously, the stress and strain curves, temperature sensitivity coefficient, Poisson's ratio and other mechanical parameters of the substrate material are collected. Combined with real-time environmental data such as temperature and humidity and atmospheric pressure, a standardized multidimensional parameter dataset is formed. A sub-model for dynamic prediction of material properties is constructed based on deep learning. The model is trained using historical material data, and the inputs are environmental parameters and basic material parameters. The output is the mechanical parameters that change dynamically with the working conditions. This sub-model employs a 3-layer fully connected neural network architecture. The number of neurons in the input layer is the total dimension of environmental parameters (temperature, humidity, atmospheric pressure) and basic material parameters (density, initial elastic modulus). There are two hidden layers: the first layer has 128 neurons, and the second layer has 64 neurons. The output layer contains three mechanical parameters, including dynamic elastic modulus, dynamic Poisson's ratio, and dynamic temperature sensitivity coefficient, corresponding to three neurons. The activation function is ReLU, the loss function is mean squared error, the optimizer is Adam, the learning rate is set to 0.001, the number of training iterations is 500, and the batch size is 32. An early stopping strategy is adopted during training, with patience=20. Training stops when the validation set loss does not decrease for 20 consecutive iterations to ensure the model's generalization ability.

[0008] A parametric framework is built using the ANSYS parametric design language, and an attention weight allocation module is embedded to dynamically assign weights to geometric parameters, material dynamic parameters, and environmental parameters. The correlation weights of key parameters are emphasized to generate a dynamic iterative multiphysics fusion model. By employing a multi-scale meshing strategy, a differentiated setting is adopted, using fine meshes in the core stress area and coarse meshes in non-core areas. Combined with the dynamic parameters output by the sub-model, the model is constructed, thereby improving the matching degree between the simulation results and the actual forming process.

[0009] The mesh size of the core stress-bearing areas, such as the ribs and the contact area between the crests and troughs of the floor deck, is set to 0.5mm×0.5mm. The mesh size of the non-core areas, such as the non-stress-bearing areas at the edge of the substrate, is set to 2mm×2mm. The mesh transition area adopts a gradual size, gradually transitioning from 0.5mm to 2mm to avoid simulation errors caused by abrupt mesh changes.

[0010] Preferably, the simulation prediction stage is based on the multi-physics field fusion model output by the multi-field modeling stage, constructs a multi-condition simulation scenario library, includes different rolling speed, pressure, and ambient temperature combination conditions, adopts an adaptive mesh generation algorithm that coordinates shell elements and solid elements, optimizes the mesh density according to the parameter weights output by the attention module, and improves the mesh accuracy in the region affected by key parameters. The multi-condition simulation scenario library contains 120 typical conditions, including rolling speeds ranging from 0.5m / s, 1.0m / s, 1.5m / s, to 2.0m / s (4 gradients), pressure ranging from 100MPa, 150MPa, 200MPa, 250MPa, to 300MPa (5 gradients), and ambient temperature ranging from -10℃, 0℃, 10℃, 20℃, 30℃, to 40℃ (6 gradients). Through the combination of all factors, a standardized condition library is formed, covering more than 95% of the actual production conditions.

[0011] Meanwhile, physical information neural networks are used to enhance simulation accuracy. Material constitutive equations and thermodynamic equilibrium equations are embedded as hard constraints into the finite element simulation model to achieve a dual fitting of physical laws and data-driven processes in the forming process. The physical information neural network adopts an encoder-decoder architecture. The encoder is a 4-layer convolutional neural network with a kernel size of 3×3, a stride of 1, padding=1, and output channels of 16, 32, 64, and 128 respectively. The decoder is a 3-layer deconvolutional network with a physical constraint layer embedded in the middle. The hard constraint is embedded by using the material constitutive equation (… ), thermodynamic equilibrium equation ( The loss function is incorporated, and the total loss function is... (in For data fitting loss, For physical constraint loss, =0.5 is the balance coefficient); the model is trained using the SGD optimizer with a learning rate of 0.0005, 300 iterations, and a batch size of 16. The simulation system comprehensively analyzes key performance indicators such as elastic buckling moment, ultimate bearing capacity, and mid-span deflection. It uses anomaly detection algorithms to identify hidden defects such as stress concentration, local buckling, and dimensional deviations. The output includes a multidimensional dataset containing working condition parameters, simulation results, defect risk level, and defect location coordinates.

[0012] The anomaly detection algorithm adopts the isolated forest algorithm, with 100 trees and 256 samples per tree. The anomaly judgment threshold is set to 0.3. When the sample anomaly score is ≥0.3, it is judged to have a latent defect. Through unsupervised learning of feature data such as stress distribution and size deviation during the simulation process, the algorithm can accurately identify small latent defects.

[0013] Preferably, the parameter optimization stage relies on the multi-dimensional simulation dataset generated in the simulation prediction stage to construct an attention mechanism-driven genetic algorithm neural network hybrid optimization model to achieve intelligent optimization of forming parameters. The simulation dataset is randomly divided into training set and test set in an 8:2 ratio. The core optimization objectives are clearly defined as precise control of rib height deviation, improvement of wave pitch uniformity, and achievement of bearing capacity. Rolling speed, pressure, die gap, and temperature compensation are selected as key optimization variables. The input parameters are adapted to different scenarios: rolling speed 0.5-2m / s, pressure 100-300MPa, die gap 0.1-0.5mm, and temperature compensation -5-5℃, set according to the different thermal expansion coefficients of materials such as SGCC galvanized steel and G350AZ aluminized zinc steel; the output results are correlated with quality standards: rib height deviation output value ≤±0.1mm is considered compliant, wave pitch uniformity output value ≥98% is considered compliant, and load-bearing capacity output value ≥1.1 times the design value is considered compliant. The model prediction results are directly mapped to the production parameter adjustment threshold.

[0014] A self-attention selection operator is embedded in the genetic algorithm. The correlation between each parameter and the optimization objective is determined by calculating the Pearson correlation coefficient, and the crossover and mutation probabilities are adaptively adjusted. The computational logic of the self-attention selection operator is as follows: First, the four key optimization variables—rolling speed, pressure, die clearance, and temperature compensation—are normalized. Then, the attention weights are calculated by the dot product of the query vector (optimization target feature vector), the key vector (variable feature vector), and the value vector (original variable value). The weight calculation formula is as follows: ,in The dimension of the key vector is set to 4; The crossover probability is adjusted positively based on the correlation: 0.8 when the correlation is ≥0.7 and 0.5 when the correlation is <0.7. The mutation probability is adjusted negatively: 0.05 when the correlation is ≥0.7 and 0.1 when the correlation is <0.7.

[0015] The parameters optimized by the genetic algorithm are used as input to the neural network to construct a multilayer perceptron model containing an input layer, a hidden layer, and an output layer. The hidden layer uses the ReLU activation function and outputs the molding quality prediction result. The multilayer perceptron has 4 input layer neurons, corresponding to 4 key optimization variables. There are 2 hidden layers, with 64 neurons in the first layer and 32 neurons in the second layer. The output layer has 3 core optimization objectives, including rib height deviation, wave distance uniformity, and bearing capacity compliance rate, corresponding to 3 neurons. The alternating training steps of genetic algorithms and neural networks are as follows: ① The genetic algorithm iterates for 30 generations to obtain the initial set of optimized parameters; ② Input the parameter set into the neural network and train it for 100 rounds; ③ The fitness function of the genetic algorithm is corrected by using the prediction error of the neural network; ④ Repeat steps ①-③ 5 times to complete the co-optimization. The convergence conditions are prediction error ≤ 0.01 mm (rib height deviation) and wave distance variation coefficient ≤ 0.02.

[0016] The genetic algorithm and neural network are optimized through alternating training mode. The model accuracy is verified by mean square error index using test set. When the prediction deviation meets the preset accuracy requirements, the optimal molding parameter combination and standardized parameter configuration file are output. At the same time, a parameter sensitivity analysis report is generated to clarify the weight ranking of the influence of each parameter on molding quality.

[0017] The report includes the following core dimensions: ① The sensitivity coefficients of each key optimization variable (rolling speed, pressure, die clearance, temperature compensation), with values ​​ranging from 0 to 1. The larger the coefficient, the more significant the impact. ② The influence curve trend of single parameter changes on each optimization objective (rib height deviation, wave pitch uniformity, bearing capacity); ③ The degree of coupling influence of parameter interaction on molding quality is divided into three levels: strong, medium, and weak; ④ Recommended parameter adjustment priority sorting.

[0018] Preferably, the twin linkage stage uses the optimal parameter configuration file output from the parameter optimization stage as the core input to construct a digital twin system for the floor decking forming production line under the federated learning architecture. The digital twin model replicates the core components such as rolling equipment, sensors, transmission mechanisms, and detection devices at a 1:1 ratio with the physical production line, and reproduces the geometric structure, motion relationship, and physical characteristics of each component. The optimal parameter configuration file is imported into the twin system. A low-latency real-time communication link between the virtual model and the physical equipment is established through industrial Ethernet. The MQTT industrial internet protocol is used to synchronize equipment operating status data such as motor speed, mold position, real-time pressure, and energy consumption. The edge and cloud aggregation strategy of federated learning is adopted to divide the production line into multiple edge nodes according to function. Each edge node builds a data security isolation zone locally and trains a local twin model based on local data. Only the model parameters are encrypted and uploaded to the cloud server. Encryption employs an asymmetric encryption algorithm. Each edge node generates its own public and private keys. The public key is uploaded to the cloud, where it is used to encrypt the optimization parameters and control strategies sent to the edge nodes. The private key is stored locally to decrypt the optimization parameters and control strategies sent from the cloud, ensuring data security during model parameter transmission and preventing information leakage.

[0019] The cloud uses a weighted average algorithm to aggregate local models to generate a globally optimized twin model, and then encrypts and distributes the optimized parameters and control strategies to each edge node.

[0020] The weights of the weighted average consist of two parts: first, the weight of the production data volume of the edge nodes, accounting for 60%, with higher weights for larger data volumes, calculated according to the proportion of each node's data volume to the total data volume; and second, the weight of the local model's validation accuracy, accounting for 40%, with higher weights for smaller validation set errors, calculated as "1 - normalized validation error". After the weights are normalized, the parameters of each local model are weighted and summed to obtain the global model parameters.

[0021] Preferably, the data preprocessing stage is based on the real-time communication link and sensor deployment foundation built by the twin linkage stage. Laser displacement sensors, pressure sensors, and infrared temperature sensors are installed at key forming positions of the rolling die, corners of the transmission channel, and core areas of the finished product inspection station. The sensor sampling frequency is set to 100Hz to collect key dimensional data such as rib height and wave distance, as well as process data such as rolling pressure, substrate temperature, and transmission speed. A lightweight data preprocessing module is deployed at the edge node, integrating wavelet denoising algorithm and isolated forest outlier removal algorithm to perform real-time denoising and outlier filtering on the collected raw data, and remove invalid data caused by factors such as equipment vibration and electromagnetic interference. The wavelet denoising algorithm uses the db4 wavelet basis function, with a decomposition layer of 3. It employs a soft threshold denoising method, with the threshold calculation formula determined based on the absolute deviation of the data median: threshold = 1.4826 × absolute deviation of the data median. This value is used to estimate the data standard deviation, ensuring denoising accuracy and effectively filtering low-frequency vibration interference below 50Hz and high-frequency electromagnetic noise.

[0022] The min-max normalization algorithm is used to convert heterogeneous data from multiple sources into standardized data in a unified format, generating a real-time data stream that meets the interaction requirements of the digital twin model; at the same time, the data freshness index is calculated and the data transmission latency is monitored in real time.

[0023] Preferably, the closed-loop control stage utilizes the real-time data stream output from the data preprocessing stage to construct a deviation intelligent identification and dynamic correction module. This module compares the preprocessed real-time production data from the edge nodes with the simulation prediction data output from the digital twin model, parameter by parameter. A cosine similarity algorithm is used to calculate the similarity between the real-time production data and the simulation prediction data. Simultaneously, a difference calculation method is used to accurately obtain specific key indicators such as size deviation and pressure deviation. When the similarity is below a preset threshold (0.95) and the size deviation exceeds ±0.1mm, a double confirmation of deviation exceeding the standard is established. The preset similarity threshold is set to 0.95. That is, when the cosine similarity between real-time production data and simulation prediction data is <0.95, the result of determining whether the size deviation exceeds ±0.1mm is combined simultaneously to double confirm whether the deviation exceeds the standard, thus avoiding false triggering caused by a single judgment standard.

[0024] The preset deviation threshold is determined based on historical production data and quality standards. When the deviation value exceeds the threshold, the closed-loop control process is automatically triggered. The parameter correction amount is deduced in reverse through the digital twin model. Combined with the optimization direction output by the federated learning global model, parameter correction instructions are generated to adjust key parameters such as die gap, rolling speed, and temperature compensation value of the rolling equipment in real time. The corrected production data is fed back to each edge node in real time, updating the local twin model and the multiphysics parameterized model, while the correction record is encrypted and uploaded to the cloud.

[0025] The response time of closed-loop control (from the determination of deviation exceeding the standard to the completion of equipment parameter adjustment) does not exceed 500ms. This is achieved by optimizing the transmission efficiency of the communication protocol and the response speed of the equipment actuator, ensuring timely correction of deviations and avoiding quality problems in batch products.

[0026] Preferably, the model iteration stage integrates the correction records uploaded in the closed-loop control stage and the full data accumulated in each stage to construct a knowledge graph and distributed knowledge base for floor decking forming. The knowledge base adopts a federated learning architecture and is distributed and stored on edge nodes and the cloud. The edge nodes are responsible for storing the full data of each production batch locally, including design parameters, optimization parameters, real-time production data, quality inspection data, and defect handling solutions. The cloud stores the global common data and standardized knowledge graph shared by each edge node. The knowledge graph is stored in the Neo4j graph database format. Entities are stored as nodes, which include unique identifiers (IDs) and key-value pairs of core attributes. Relationships are stored as directed edges, which include relationship type and association strength values. It supports batch import and fast query, and is adapted to the needs of multi-node data synchronization and association mining.

[0027] The knowledge graph uses parameters, working conditions, defects, and solutions as core association dimensions to construct a structured knowledge network with multiple entities and relationships. Through graph neural networks, it deeply mines the implicit relationships between data and extracts the forming parameter patterns and quality control points of floor decking of different specifications and materials. The core entities and their attributes are as follows: ① Parameter entity, attributes: parameter name, value range, unit, associated material; ② Working condition entity, attributes: working condition type, rolling speed, pressure, ambient temperature, applicable specifications; ③ Defective entity, attributes: defect type, risk level, location characteristics, and scope of impact; ④ Solution entity, attributes: correction parameters, adjustment range, implementation effect, applicable scenarios; The core relationship types include: parameter-condition ("adaptation" relationship), condition-defect ("inducement" relationship), defect-solution ("correspondence" relationship), and parameter-solution ("adjustment" relationship).

[0028] The graph neural network uses a graph attention network, which consists of one input layer, two attention mechanism layers, and one output layer. The input layer is a node feature vector: parameter nodes are 12-dimensional, working condition nodes are 8-dimensional, defect nodes are 6-dimensional, and solution nodes are 10-dimensional. The first attention layer has an output dimension of 64, and the second layer has 32. The output layer is a node association weight vector. During training, the entities such as parameters and working conditions in the floor decking forming knowledge graph are encoded as feature vectors. Parameter nodes use min-max normalization encoding, working condition nodes use one-hot encoding, defect nodes use label encoding, and solution nodes use word embedding encoding. The optimizer is AdamW with a learning rate of 0.002, 200 iterations, and a batch size of 64. The model output is the association strength between entities, ranging from [0, 1]. An association strength ≥ 0.6 is considered a strong association, which serves as the basis for extracting forming patterns.

[0029] Based on the knowledge graph, an intelligent model iteration triggering mechanism is constructed. A data difference threshold is preset. When a new batch of data is entered into the database, the graph neural network algorithm automatically analyzes the difference between the new data and the existing knowledge. When the difference exceeds the preset threshold, the federated learning model iteration process is automatically triggered. The parameter weights of the machine learning optimization model and the finite element simulation model are updated using incremental learning. At the same time, the newly extracted formation rules and knowledge are added to the knowledge graph.

[0030] The preset data difference threshold is set to 0.2. The difference is calculated based on the feature similarity between the new batch of data and the same type of data in the existing knowledge graph. The average similarity of key parameters such as rolling speed, pressure, and rib height deviation is calculated by cosine similarity. When the average similarity is ≤0.8, that is, difference = 1 - average similarity ≥0.2, the model iteration process is triggered.

[0031] The knowledge graph update rules are as follows: new entities (parameters, working conditions, defects, solutions) must be associated with more than 3 valid production data as support, and new relationships must be verified by the association strength calculated by the graph neural network ≥ 0.6; an automatic review mechanism is set up, new knowledge is first stored in a temporary database, and after cross-verification with data from 3 different production batches, it is officially written into the knowledge graph to ensure the accuracy and reliability of the knowledge.

[0032] The beneficial effects of this invention are as follows: 1. This invention constructs a multi-physics parameterized model that integrates dynamic feature perception, uses an attention mechanism to focus on key parameters, and improves simulation accuracy by combining multi-scale meshing and physical information neural networks. Then, it achieves ±0.1mm-level deviation control through closed-loop control. From design drawing analysis to real-time production adjustment, the entire process incorporates dynamic parameter adaptation and hidden defect prediction, effectively avoiding hidden dangers such as stress concentration and local buckling. This ensures that the core indicators of the floor deck, such as ultimate bearing capacity and mid-span deflection, consistently meet the standards, reduces the product defect rate, and meets the precision requirements of high-end buildings.

[0033] 2. This invention achieves real-time linkage between virtual and physical devices through a 1:1 digital twin system. The federated learning architecture not only ensures data security across multiple nodes but also enables collaborative optimization. The multi-condition simulation scenario library and parameterized models can flexibly adapt to different production needs. The entire process can complete intelligent parameter optimization, dynamic deviation correction, and model iteration without manual intervention, shortening the process debugging cycle, reducing ineffective consumption, and simultaneously enabling full traceability of production data. This allows the molding process to shift from experience-driven to data-driven, improving production controllability and overall efficiency.

[0034] 3. This invention constructs a knowledge graph and distributed knowledge base for floor decking forming, utilizes graph neural networks to mine implicit relationships between data, and extracts the forming rules of floor decking of different specifications and materials; through an intelligent iterative triggering mechanism with a preset data difference threshold, incremental learning is used to update model parameter weights, realizing the collaborative upgrading of the model and knowledge base; multi-source sensor network and lightweight preprocessing module ensure the real-time performance and reliability of data, and the attention mechanism-driven hybrid optimization model can accurately adapt to various materials such as SGCC galvanized steel and G350AZ aluminized zinc steel, and can quickly switch production schemes without reconstructing the production system, reducing production changeover costs and manpower dependence, and achieving long-term production cost reduction and flexible improvement of production capacity through continuous optimization. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the multi-field modeling stage of the present invention; Figure 3 This is a flowchart of the closed-loop control stage of the present invention. Detailed Implementation

[0036] 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.

[0037] like Figures 1 to 3 As shown, this embodiment of the invention provides a method for precise forming of floor decking, including the following specific steps: In the multi-field modeling stage, a multi-physics parameterized model of the floor decking is constructed, which integrates dynamic feature perception. An attention mechanism is used to achieve adaptive focusing of key parameters. The floor decking design drawings are analyzed with high precision using BIM technology to accurately extract core geometric parameters such as rib height, wave pitch, and coverage width and to construct a basic three-dimensional model. Simultaneously, complete mechanical parameters such as stress and strain curves, temperature sensitivity coefficient, and Poisson's ratio of the substrate material are collected. Combined with real-time environmental data such as temperature and humidity and atmospheric pressure, a standardized multi-dimensional parameter dataset is formed. The substrate materials include SGCC galvanized steel, G350AZ aluminized zinc steel, etc.

[0038] A dynamic prediction sub-model for material properties is constructed based on deep learning. The model is trained with a large amount of historical material data. By inputting environmental parameters and basic material parameters, it can output accurate mechanical parameters that dynamically change with working conditions. A parameterized framework is built using the ANSYS parametric design language. An attention weight allocation module is embedded to dynamically assign weights to geometric parameters, material dynamic parameters, and environmental parameters. The correlation weights of key parameters that play a decisive role in forming accuracy, such as rib height and rolling temperature, are strengthened. A dynamic iterative multiphysics fusion model is generated to achieve intelligent adaptation of parameter correlation.

[0039] By employing a multi-scale meshing strategy, a differentiated setting is adopted, using fine meshes in the core stress area and coarse meshes in non-core areas. Combined with the dynamic parameters output by the sub-model, the model is refined and the matching degree between the simulation results and the actual forming process is improved.

[0040] The simulation prediction stage is based on the high-precision multi-physics fusion model output from the multi-field modeling stage. It constructs a multi-condition simulation scenario library, which includes different combinations of rolling speed, pressure, and ambient temperature. It adopts an adaptive mesh generation algorithm that combines shell elements and solid elements, optimizes the mesh density according to the parameter weights output by the attention module, and improves the mesh accuracy in the region affected by key parameters to balance computational efficiency and simulation accuracy.

[0041] For critical influence areas with parameter weights ≥ 0.7, the mesh density is increased by 50% on top of the basic fine mesh (0.5mm × 0.5mm); for general influence areas with parameter weights 0.3-0.7, the basic fine mesh is used; and for minor influence areas with parameter weights < 0.3, a coarse mesh (2mm × 2mm) is used. The algorithm dynamically adjusts the mesh generation scheme in real time according to the parameter weights, balancing simulation accuracy and computational efficiency.

[0042] Meanwhile, physical information neural networks are used to enhance simulation accuracy. Material constitutive equations and thermodynamic equilibrium equations are embedded as hard constraints into the finite element simulation model to achieve a dual fitting of physical laws and data-driven processes in the forming process. The training data for the physical information neural network consists of two parts: 70% is simulation data generated by a multi-condition simulation scenario library, and 30% is high-quality data from historical actual production. After anomaly removal and standardization, the data is shuffled according to time series and used for model training, balancing training efficiency and model adaptability to real-world scenarios.

[0043] The simulation system comprehensively analyzes key performance indicators such as elastic buckling moment, ultimate bearing capacity, and mid-span deflection, and uses anomaly detection algorithms to accurately identify latent defects such as stress concentration, local buckling, and dimensional deviations. The output is a multidimensional dataset containing operating parameters, simulation results, defect risk levels, and defect location coordinates.

[0044] Defect risk levels are divided into three levels: high, medium, and low, and the determination is based on the degree of impact and probability of occurrence of the defect. High risk: Defects causing a decrease in load-bearing capacity of ≥10% or dimensional deviations exceeding ±0.3mm, with an occurrence probability of ≥30%; Medium risk: Defects may cause a 5%-10% decrease in load-bearing capacity or a dimensional deviation of ±0.15mm-±0.3mm, with an occurrence probability of 10%-30%. Low risk: Defects causing a decrease in load-bearing capacity of <5% or a dimensional deviation of <±0.15mm, with an occurrence probability of <10%. The risk level is directly related to the priority of subsequent parameter adjustments.

[0045] In the parameter optimization stage, the multi-dimensional simulation dataset generated in the simulation prediction stage is used to construct an attention mechanism-driven genetic algorithm neural network hybrid optimization model to achieve intelligent optimization of forming parameters. The simulation dataset is randomly divided into training set and test set in an 8:2 ratio. The core optimization objectives are defined as precise control of rib height deviation, improvement of wave pitch uniformity, and achievement of bearing capacity. Rolling speed, pressure, die gap, and temperature compensation are selected as key optimization variables.

[0046] A self-attention selection operator is embedded in the genetic algorithm. The correlation between each parameter and the optimization target is determined by calculating the Pearson correlation coefficient. Based on this, the crossover and mutation probabilities are adaptively adjusted to improve the optimization efficiency and accuracy of key parameters that have a significant impact on molding quality. The parameters optimized by the genetic algorithm are used as the input of the neural network to construct a multilayer perceptron model with an input layer, a hidden layer, and an output layer. The ReLU activation function is used in the hidden layer to improve the nonlinear fitting ability of the model and output the molding quality prediction result.

[0047] The genetic algorithm and neural network are optimized through alternating training mode. The model accuracy is verified by mean square error index using test set. When the prediction deviation meets the preset accuracy requirements, the optimal molding parameter combination and standardized parameter configuration file are output. At the same time, a parameter sensitivity analysis report is generated to clarify the weight ranking of the influence of each parameter on molding quality.

[0048] The configuration file uses JSON format and contains the following core modules: ①Basic information, including document number, generation time, and applicable floor decking specifications / materials); ② The core parameter group, including the specific values ​​and accuracy of rolling speed, pressure, die clearance, and temperature compensation, are retained to two decimal places; ③ Applicable operating conditions, including rolling ambient temperature range and substrate thickness range; ④ Constraints, including the upper limit of equipment operating power and the limit of parameter adjustment step size; ⑤ Validity period: 6 months by default. Automatic re-optimization will be triggered upon expiration.

[0049] The twin linkage stage takes the optimal parameter configuration file output from the parameter optimization stage as the core input to construct a digital twin system for the floor decking forming production line under the federated learning architecture. This system achieves data privacy protection and collaborative optimization for multiple devices. The digital twin model strictly replicates the core components such as rolling equipment, sensors, transmission mechanisms, and detection devices at a 1:1 ratio according to the physical production line, accurately reproducing the geometric structure, motion relationship, and physical characteristics of each component.

[0050] The replicated physical properties specifically include the stiffness, damping coefficient, thermal conductivity, and wear resistance parameters of the components. The stiffness and damping coefficient are corrected based on the equipment's factory parameters and historical operating loss data, while the thermal conductivity is dynamically adapted in combination with the substrate material characteristics to ensure the consistency of the physical response between the virtual model and the physical equipment.

[0051] The optimal parameter configuration file is imported into the twin system, and a low-latency real-time communication link between the virtual model and the physical equipment is established through industrial Ethernet. The MQTT industrial internet protocol is used to synchronize equipment operating status data such as motor speed, mold position, real-time pressure, and energy consumption data. The equipment operating status data includes motor speed, mold position, real-time pressure, and energy consumption data. The edge and cloud aggregation strategy of federated learning is adopted to divide the production line into multiple edge nodes such as rolling unit, detection unit, and transmission unit according to function. Each edge node builds a data security isolation zone locally and trains a local twin model based on local data. Only the model parameters are encrypted and uploaded to the cloud server to avoid leakage of original production data.

[0052] The edge nodes are divided into 3 by default, and their specific functions are as follows: ① Edge nodes of rolling units: responsible for collecting rolling speed, pressure, and temperature data, and training a twin model of the local rolling process; ② Edge nodes of the detection unit: responsible for collecting quality inspection data such as size and load-bearing capacity, and optimizing the model parameters related to defect identification; ③ Edge nodes of the transmission unit: responsible for collecting transmission speed and tension data to ensure parameter adaptation during transmission; the number of nodes can be increased or decreased according to the scale of the production line, with a maximum of 5 nodes.

[0053] The real-time synchronization frequency is set to 10Hz, which means 10 synchronizations per second, ensuring that the deviation between the virtual model and the physical device's operating status is ≤50ms, meeting the real-time control requirements of virtual-physical linkage, and avoiding parameter adjustment lag caused by synchronization delay.

[0054] The cloud aggregates local models using a weighted average algorithm to generate a globally optimized twin model. The optimized parameters and control strategies are then encrypted and distributed to each edge node to achieve distributed collaborative optimization that links the virtual and real worlds.

[0055] The data preprocessing stage is based on the real-time communication link and sensor deployment foundation built by the twin linkage stage. A multi-source sensor network is deployed, and high-precision laser displacement sensors, pressure sensors, and infrared temperature sensors are installed at key forming positions of the rolling die, corners of the transmission channel, and core areas of the finished product inspection station. The sensor sampling frequency is set to 100Hz to ensure the real-time nature of data acquisition, and to accurately collect key dimensional data such as rib height and wave pitch, as well as process data such as rolling pressure, substrate temperature, and transmission speed.

[0056] The laser displacement sensor has a measurement range of 0-500mm and a measurement accuracy of ±0.01mm; the pressure sensor has a range of 0-500MPa and an accuracy class of 0.1; the infrared temperature sensor has a measurement range of -20℃-500℃ and a measurement accuracy of ±0.5℃, ensuring the accuracy of key data acquisition.

[0057] A lightweight data preprocessing module is deployed at the edge node, integrating wavelet denoising algorithm and isolated forest outlier removal algorithm to perform real-time denoising and outlier filtering on the collected raw data, removing invalid data caused by factors such as equipment vibration and electromagnetic interference; the min-max normalization algorithm is used to convert multi-source heterogeneous data into standardized data in a unified format, generating a real-time data stream that meets the interaction requirements of the digital twin model.

[0058] After standardization, all data is mapped to the range [0, 1] and stored in CSV format. The file fields are "data acquisition timestamp, sensor number, parameter name, original value, standardized value, and data status (valid / invalid)" to facilitate rapid reading and parsing by the digital twin model.

[0059] Simultaneously, data freshness indicators are calculated, and data transmission latency is monitored in real time to ensure that the data input into the digital twin model is highly timely.

[0060] The data freshness index is defined as "the total time taken from data collection to transmission to the digital twin model and completion of preprocessing". The preset threshold is 100ms. When the total time exceeds 100ms, the data is judged to be invalid, triggering the sensor network to re-collect data to ensure that the timeliness of the data input to the model meets the real-time control requirements.

[0061] In the closed-loop control stage, the high-quality real-time data stream output from the data preprocessing stage is used to construct a deviation intelligent identification and dynamic correction module. The real-time production data after edge node preprocessing is compared with the simulation prediction data output by the digital twin model parameter by parameter, and the cosine similarity algorithm is used to accurately calculate key difference indicators such as size deviation and pressure deviation.

[0062] The preset deviation threshold (±0.1mm) is determined based on a large amount of historical production data and quality standards. When the deviation value exceeds this threshold, the closed-loop control process is automatically triggered. The parameter correction amount is deduced in reverse through the digital twin model. Combined with the optimization direction output by the federated learning global model, the targeted and highly accurate parameter correction instructions are generated to adjust key parameters such as die gap, rolling speed, and temperature compensation value of the rolling equipment in real time.

[0063] The corrected production data is fed back to each edge node in real time, updating the local twin model and multiphysics parameterized model to complete the single-round closed loop, while the correction record is encrypted and uploaded to the cloud.

[0064] The parameter adjustment adopts a step-by-step fine-tuning strategy with the following step size limits: rolling speed adjustment step size ≤ 0.1m / s, pressure adjustment step size ≤ 10MPa, die gap adjustment step size ≤ 0.01mm, and temperature compensation value adjustment step size ≤ 0.5℃, to avoid fluctuations in molding quality due to sudden parameter changes.

[0065] In the model iteration phase, the correction records uploaded in the closed-loop control phase and the full data accumulated in each phase are integrated to construct a knowledge graph and a distributed knowledge base for floor decking forming. The knowledge base adopts a federated learning architecture and is distributed and stored on edge nodes and in the cloud. The edge nodes are responsible for storing the full data of each production batch locally, including design parameters, optimization parameters, real-time production data, quality inspection data, defect handling solutions, etc. The cloud stores the global common data and standardized knowledge graph shared by each edge node.

[0066] The specific quality inspection data includes: the actual measured value of the rib height of a single floor deck (the average value of 5 test points), the measured value of the wave pitch uniformity, the test results of the ultimate bearing capacity, the test value of the mid-span deflection, and the number of surface defects (such as scratches, deformation, etc.). All test data are retained to two decimal places and are stored in correspondence with the design standard values.

[0067] The knowledge graph uses parameters, working conditions, defects, and solutions as core relational dimensions to construct a structured knowledge network with multiple entities and relationships. Through graph neural networks, it deeply mines the implicit relationships between data and extracts the forming parameter patterns and quality control points of floor decking of different specifications and materials.

[0068] Based on a knowledge graph, an intelligent model iteration triggering mechanism is constructed. A preset data difference threshold is used. When a new batch of data is entered into the database, the graph neural network algorithm automatically analyzes the difference between the new data and existing knowledge. When the difference exceeds the preset threshold, the federated learning model iteration process is automatically triggered. The parameter weights of the machine learning optimization model and the finite element simulation model are updated using incremental learning. At the same time, the newly extracted molding rules and knowledge are added to the knowledge graph to achieve collaborative iterative upgrades of the model and the knowledge base. This improves the optimization efficiency and accuracy of subsequent molding processes and adapts to the production needs of more specifications of floor decking.

[0069] The incremental learning process is as follows: ① Freeze the underlying basic parameters of the model, accounting for 70%, to ensure the stability of the model's core logic; ② Only the parameters at the top level of the model that are related to the new data are updated, accounting for 30%; ③ Use 10% of the new batch of data as a validation set to monitor the prediction accuracy of the updated model; ④ If the verification accuracy improves by ≥5%, the updated parameters are retained; if the accuracy decreases, the model parameters before the update are automatically rolled back. ⑤ The update frequency of incremental learning is synchronized with the production batch. After each production batch is completed, the difference between the new batch data and the existing knowledge is automatically analyzed. When the difference exceeds the preset threshold, the incremental learning update process is triggered.

[0070] 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.

[0071] 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 precise forming of floor decking, characterized in that, The specific steps include the following: Multi-field modeling stage: Construct a multi-physics parameterized model of the floor decking that integrates dynamic feature perception, analyze the design drawings to extract geometric parameters, collect material mechanical parameters and environmental data to form a standardized dataset, embed an attention mechanism to realize dynamic weight assignment of parameters, and complete the model construction through multi-scale mesh division; Simulation prediction stage: Based on the multi-physics field fusion model, a multi-condition simulation scenario library is built to enhance simulation accuracy, identify latent defects in the formed system, and output multidimensional datasets; Parameter optimization stage: Based on the simulation dataset, a hybrid optimization model of genetic algorithm and neural network driven by attention mechanism is constructed. The optimization objectives and variables are set. After verifying the accuracy through collaborative optimization, the optimal parameter combination, configuration file and parameter sensitivity report are output. Twin linkage stage: Construct a digital twin system for the floor decking production line with a federated learning architecture, replicate the core components in proportion, establish a real-time communication link between the virtual model and the physical equipment, train and optimize the twin model through edge and cloud aggregation strategies, and achieve distributed collaborative optimization of virtual and physical linkage; Data preprocessing stage: Based on the multi-source sensor network deployed in the twin linkage stage, key dimension data and process data of the forming are collected. The edge node preprocessing module completes data noise reduction, anomaly removal and standardization transformation to generate real-time data stream; Closed-loop control phase: Construct a deviation identification and dynamic correction module, compare real-time production data with simulation prediction data to calculate deviation, trigger closed-loop control to adjust equipment parameters when the threshold is exceeded, feed back data to update the model and upload to the cloud; Model iteration phase: Construct a knowledge graph and distributed knowledge base to store all production data, mine data correlations and extract established patterns, establish a model iteration trigger mechanism, and use incremental learning to update and optimize the model and simulation model.

2. The method for precise forming of floor decking according to claim 1, characterized in that, In the multi-field modeling stage, BIM technology is used to analyze the floor decking design drawings to extract core geometric parameters and construct a basic three-dimensional model. Simultaneously collect mechanical parameters of the substrate material and multi-dimensional environmental data, combine historical material data to train a deep learning material property dynamic prediction sub-model, and output mechanical parameters that dynamically change with working conditions; adopt a parameterized framework to embed an attention weight allocation module to dynamically assign weights to geometric parameters, material dynamic parameters, and environmental parameters, and complete the model construction by setting differentiated mesh division between core stress areas and non-core areas and combining dynamic parameters.

3. The method for precise forming of floor decking according to claim 2, characterized in that, In the simulation prediction stage, the multi-condition simulation scenario library contains combinations of different rolling speeds, pressures, and ambient temperatures; an adaptive mesh generation algorithm is adopted and the mesh density is optimized according to the parameter weights output by the attention module. At the same time, the material constitutive equation and thermodynamic equilibrium equation are embedded as hard constraints into the finite element simulation model; key performance indicators are analyzed through the simulation system, hidden defects are identified using anomaly detection algorithms, and a multidimensional dataset containing operating parameters, simulation results, defect risk levels, and defect location coordinates is output.

4. The method for precise forming of floor decking according to claim 3, characterized in that, In the parameter optimization stage, the simulation dataset is divided into a training set and a test set. The core optimization objectives are precise control of rib height deviation, improvement of wave pitch uniformity, and achievement of load-bearing capacity. Rolling speed, pressure, die clearance, and temperature compensation are selected as key optimization variables. A self-attention selection operator is embedded in the genetic algorithm. The crossover and mutation probabilities are adaptively adjusted based on the correlation between the parameters and the optimization objectives. The optimized parameters are input into the multilayer perceptron model to output the forming quality prediction results. Collaborative optimization is achieved through alternating training. After the accuracy is verified by the test set, the optimal forming parameter combination, standardized parameter configuration file, and parameter sensitivity analysis report are output.

5. The method for precise forming of floor decking according to claim 4, characterized in that, In the twin linkage stage, the digital twin model replicates the core components of the production line at a 1:1 scale, including their physical characteristics, motion characteristics, and sensing characteristics. After importing the optimal parameter configuration file, a low-latency real-time communication link is established via industrial Ethernet to synchronize the equipment's operating status data. The production line is divided into multiple edge nodes according to function. Each edge node builds a local data security isolation zone to train a local twin model, and only the model parameters are encrypted and uploaded to the cloud. The cloud aggregates the local models using a weighted average algorithm to generate a globally optimized twin model, and then encrypts and distributes the optimized parameters and control strategies to each edge node.

6. The method for precise forming of floor decking according to claim 5, characterized in that, In the data preprocessing stage, based on the multi-source sensor network deployed in the twin linkage stage, key dimension data and process data of molding are collected in the key molding area; the preprocessing module of the edge node integrates noise reduction algorithm and outlier removal algorithm, filters invalid data, and then converts multi-source heterogeneous data into standardized data in a unified format through standardization algorithm, generating a real-time data stream that meets the interaction requirements and monitoring data transmission latency.

7. The method for precise forming of floor decking according to claim 6, characterized in that, In the closed-loop control phase, a cosine similarity algorithm is used to calculate the similarity between real-time production data and simulation prediction data, and a preset similarity threshold is used to determine whether the deviation exceeds the standard. At the same time, the difference calculation method is used to obtain specific key indicators, including size deviation and pressure deviation. The preset deviation threshold is determined based on historical production data and quality standards. When the deviation exceeds the threshold, the parameter correction amount is deduced by reverse engineering through the digital twin model, and a correction instruction is generated by combining the optimization direction output by the federated learning global model to adjust the parameters of key equipment in real time. The corrected production data is fed back to each edge node to update the model, and the correction record is encrypted and uploaded to the cloud.

8. The method for precise forming of floor decking according to claim 7, characterized in that, During the model iteration phase, the distributed knowledge base adopts a federated learning architecture, with data distributed across edge nodes and the cloud. Edge nodes store all local production data, while the cloud stores global common data and standardized knowledge graphs. The knowledge graph constructs a structured knowledge network with parameters, operating conditions, defects, and solutions as core association dimensions. It uses graph neural networks to mine implicit associations in the data and extract established patterns. A model iteration triggering mechanism is built based on the knowledge graph. After a new batch of data is added to the database, its difference from existing knowledge is analyzed. When the difference exceeds a preset threshold, the federated learning model iteration process is triggered. Incremental learning is used to update the model parameter weights and supplement new knowledge into the knowledge graph.

Images