Intelligent welding forming method and system for steel heating radiator for green building

Abstract

The invention discloses an intelligent welding forming method and system for a green building steel heating radiator, and the method comprises the following steps: carrying out the surface defect recognition of a steel heating radiator base material based on an AI visual inspection system, recognizing a qualified base material, and automatically matching the type of a welding material from a material database according to the material and thickness parameters of the qualified base material. Welding parameters are intelligently matched through AI visual inspection and a neural network model, a laser and friction stir hybrid welding process is combined, traditional manual operation is replaced, the welding efficiency and precision are improved, and the problems of uneven welding seams and the like are solved; welding data are analyzed in real time through a multi-mode AI model, parameters are dynamically adjusted, intelligent defect recognition and repair welding are achieved in cooperation with 3D visual inspection, and the quality stability is improved through whole-process monitoring; smoke dust is treated through an environment-friendly process, acid pickling is replaced with mechanical rust removal, efficient recycling of materials is achieved through waste recycling, a green manufacturing system is constructed, and the sustainable development requirement of green buildings is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of steel radiator manufacturing, in particular to a green building steel radiator intelligent welding forming method and system. BACKGROUND

[0002] As the core heat exchange component of the green building heating system, the welding forming quality of the steel radiator directly determines the structural stability, heat dissipation performance and corrosion resistance of the product. Precise welding process controls the sealing of the weld to avoid water leakage during operation; uniform welds can reduce stress concentration and prolong service life; low defect forming can reduce contact thermal resistance, improve the thermal efficiency of the radiator, and directly affect the energy consumption and economy of the building heating system, which meets the demand of green building for high durability and high energy efficiency equipment. The existing welding process mainly relies on manual operation or semi-automatic equipment to realize the connection of steel components through argon arc welding, resistance welding and other technologies, which to some extent meets the demand of large-scale production. However, there are still some problems: Firstly, the traditional welding process relying on manual operation is not only low in efficiency, but also easy to cause uneven welds and large heat-affected zone; Secondly, although the existing semi-automatic equipment can improve efficiency, it lacks real-time quality monitoring and parameter optimization mechanism, and it is difficult to meet the requirements of green building for environmental protection and energy saving; Thirdly, the smoke, waste gas and material loss generated during the welding process have not been effectively controlled, which does not meet the sustainable development goal of green building; Therefore, a green building steel radiator intelligent welding forming method and system are proposed. SUMMARY

[0003] The purpose of the present application is to provide a green building steel radiator intelligent welding forming method and system to solve one of the problems in the background art.

[0004] To solve the above technical problems, one technical solution adopted by the present application is: a green building steel radiator intelligent welding forming method, comprising the following steps: Step one, based on the AI visual detection system, the surface defects of the steel radiator base material are identified, the qualified base material is identified, and the welding material type is automatically matched from the material database according to the material quality and thickness parameters of the qualified base material; Step two, the surface oxide layer of the qualified base material is treated by mechanical derusting process, and the structure parameters and environmental temperature data of the treated base material are obtained by three-dimensional scanning; Step three, a welding parameter prediction model is constructed and trained based on neural network, the welding material type, base material structure parameters and environmental temperature data are input, and the welding parameters are output; Step four, execute the welding program according to the welding parameters, collect welding data in real time, and analyze the welding data through a multi-modal AI model to dynamically adjust the welding parameters; Step five, perform welding using a combined welding process of laser welding and friction stir welding according to the adjusted welding parameters; Step six, perform comprehensive scanning of the weld based on a 3D visual inspection system, identify weld defects using a deep learning algorithm, and automatically trigger a laser repair welding mechanism according to the weld defect type; Step seven, process welding fumes based on plasma catalytic purification technology, and classify and recycle welding spatters and corner materials for regeneration treatment through a waste recycling module.

[0005] As a further optimization of the technical solution: in step three, the method of constructing and training a welding parameter prediction model based on a neural network includes the following steps: Step 301, obtain historical process data, perform noise reduction and normalization processing on the historical process data, and divide them into a training set, a validation set, and a test set according to a proportion; the historical process data includes welding material type, base material structure parameters, environmental temperature data, and corresponding welding parameters; Step 302, build a welding parameter prediction model including an input layer, a hidden layer, and an output layer; the input layer corresponds to welding material type, base material structure parameters, and environmental temperature data, and the output layer corresponds to welding parameters; Step 303, use the training set to iteratively train the welding parameter prediction model, adjust the weights of the welding parameter prediction model through the Adam algorithm, and use the validation set to evaluate the generalization ability of the welding parameter prediction model; Step 304, use the test set to evaluate the final performance of the welding parameter prediction model, so that the welding parameter prediction accuracy meets the preset standard; Step 305, deploy the tested welding parameter prediction model to a welding parameter intelligent decision-making module, receive welding material type, base material structure parameters, and environmental temperature data in real time, and output welding parameters.

[0006] As a further optimization of the technical solution: in step four, the welding parameters include welding current, welding voltage, welding speed, and protective gas ratio; the welding data includes arc signal, molten pool image, and temperature data; The fusion algorithm of the multi-modal AI model includes: Perform wavelet transform on the arc spectrum signal to extract 5-15 kHz frequency band features; Perform threshold segmentation on the molten pool infrared image to calculate the area and perimeter of the molten pool; Use a Kalman filter to fuse multi-sensor data and output a welding quality evaluation value.

[0007] As a further preferred of the technical solution: in step five, the laser welding adopts a pulsed fiber laser, the pulse frequency of the pulsed fiber laser is 50-200 Hz, and the average power is 1-3 kW; the friction stir welding adopts a tungsten carbide stir head, and the rotation speed and pressure parameters are dynamically matched according to the thickness of the base material.

[0008] As a further preferred of the technical solution: in step one, the AI visual detection system is equipped with a line array CCD camera and a deep learning defect recognition model.

[0009] As a further preferred of the technical solution: in step two, the mechanical derusting process adopts a sand blasting or shot blasting process, the sand blasting medium is quartz sand or steel shot, and the derusting level reaches Sa2.5 level; the three-dimensional scanning adopts a laser three-dimensional scanner to obtain the base material structure size, geometric shape parameters and environmental temperature data with an accuracy of ±0.1 mm.

[0010] As a further preferred of the technical solution: in step six, the 3D visual detection system generates three-dimensional point cloud data of the weld with an accuracy of ±0.05 mm based on the principle of structured light projection, and intelligently classifies and identifies through a defect feature library.

[0011] To solve the above technical problems, another technical solution adopted by the present application is: an intelligent welding and forming system for steel heating radiators for green buildings, comprising a base material screening module, an environmental protection treatment module, a parameter prediction module, a parameter control module, a composite welding execution module, an intelligent detection and repair module, and a smoke and waste treatment module. The base material screening module is configured to identify the surface defects of the steel heating radiator base material based on an AI visual detection system, identify qualified base materials, and automatically match welding material types from a material database according to the material quality and thickness parameters of the qualified base materials. The environmental protection treatment module is configured to use a mechanical derusting process to treat the surface oxide layer of the qualified base material, and obtain the structure parameters and environmental temperature data of the treated base material through three-dimensional scanning. The parameter prediction module is configured to build and train a welding parameter prediction model based on a neural network, input the welding material type, base material structure parameters and environmental temperature data, and output the welding parameters. The parameter control module is configured to execute the welding program according to the welding parameters, collect welding data in real time, analyze the welding data through a multi-modal AI model, and dynamically adjust the welding parameters. The composite welding execution module is configured to use a composite welding process combining laser welding and friction stir welding to perform welding according to the adjusted welding parameters. The intelligent detection and repair module is configured to perform omnidirectional scanning on the weld by a 3D visual detection system, recognize weld defects by using a deep learning algorithm, and automatically trigger a laser repair welding mechanism according to the weld defect type. The smoke and waste treatment module is configured to treat welding smoke based on a plasma catalytic purification technology, and simultaneously classify and recycle and regenerate welding spatters and corner materials.

[0012] As a further preferred aspect of the technical solution: the parameter prediction module includes a data preprocessing unit, a model training unit and a parameter output unit; The data preprocessing unit is used for denoising, normalizing and dividing training set, validation set and test set of historical process data; The model training unit is used for building a welding parameter prediction model and iteratively training by an Adam algorithm; The parameter output unit is used for deploying the welding parameter prediction model that passes the test and outputting welding parameters in real time.

[0013] As a further preferred aspect of the technical solution: the parameter control module includes a sensor acquisition unit and an AI fusion decision unit; The sensor acquisition unit integrates a spectrometer, an infrared thermal imager and a K-type thermocouple, and is used for real-time acquisition of arc signal, molten pool dynamics and temperature field data; The AI fusion decision unit is used for wavelet transform on the arc spectrum signal, extraction of 5-15kHz frequency band features, threshold segmentation on the molten pool infrared image, calculation of the molten pool area and circumference, fusion of multi-sensor data by using a Kalman filter, and output of a welding quality evaluation value; if the welding quality evaluation value is less than 0.8, the parameter adjustment is automatically triggered.

[0014] Advantages of the present application: 1. The present application matches welding parameters intelligently by AI visual detection and neural network model, replaces traditional manual operation by combining laser and friction stir complex welding process, significantly improves welding efficiency, realizes high-precision control of weld uniformity, and solves the problems of uneven weld and excessive heat affected zone; 2. The present application realizes real-time collection and analysis of welding data and dynamic adjustment of parameters by a multi-modal AI model, realizes intelligent recognition and repair welding of weld defects by 3D visual detection, realizes full-process intelligent monitoring and dynamic parameter adjustment of the welding process, and effectively improves the stability of welding quality; 3. The present application treats welding smoke by using an environmentally friendly process, replaces pickling with mechanical derusting to avoid chemical pollution, and realizes efficient classification and recycling of welding spatters and corner materials by a waste recycling module, thereby constructing a green manufacturing recycling system and meeting the sustainable development requirements of green building. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0016] Figure 1 This is a schematic diagram of a process for an intelligent welding method for a steel radiator for a green building according to the present invention; Figure 2 Schematic diagram of the process of constructing and training a welding parameter prediction model based on a neural network according to the present invention; Figure 3 This is a schematic diagram of the functional modules of an intelligent welding and forming system for steel radiators for green buildings according to the present invention. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] Example Figure 1 This is a flow chart of an intelligent welding method for steel radiators for green buildings according to an embodiment of the present invention. It should be noted that if there are substantially the same results, the method of this application is not based on Figure 1 The process sequence shown is limited. Figures 1-2 As shown: A smart welding forming method for steel radiators for green buildings, comprising the following steps: Step 1: Use the AI ​​visual inspection system to identify surface defects on the steel radiator base material, identify qualified base materials, and automatically match the welding material type from the material database based on the material and thickness parameters of the qualified base materials; Specifically, first, an AI visual inspection system is deployed to identify surface defects on the parent material. The AI ​​visual inspection system uses an industrial-grade linear array CCD camera as its core, combined with a ring light source component, to scan the steel parent material line by line (with a scanning accuracy of 0.05mm / pixel). The specific identification method is as follows: Images captured by the camera are transmitted to the edge computing unit via the Gigabit network and analyzed in real time using a pre-trained deep learning model (such as YOLOv5s). This model can identify surface defects such as cracks with a width of ≥0.1mm and pinholes with a diameter of ≥0.5mm. For suspected defect areas, the system automatically triggers a secondary review mechanism - by rotating the mechanical arm to adjust the shooting angle, using structured light three-dimensional reconstruction technology to generate local point cloud data (accuracy ± 0.02 mm), and combining with the defect feature library (containing 2000+ labeled samples) for stereo shape comparison, to finally determine whether the base material is qualified; The entire detection process is fully automated, and the detection time of a single base material is ≤15 seconds, which is more than 10 times more efficient than manual visual inspection; Then, the material and thickness parameters of the qualified base material are accurately collected, as follows: Using laser-induced breakdown spectroscopy (LIBS) technology, high-energy laser pulses are used to excite the surface of the base material to generate plasma, and the spectrometer captures the characteristic spectrum of 350-750 nm wavelength. After neural network algorithm analysis, the C, Mn, Si element content (accuracy ± 0.02%) can be obtained within 3 seconds, and the base material grade (such as Q235B, 304L) can be automatically matched; Based on the principle of laser triangulation, the laser thickness head is driven by a servo motor to move along the longitudinal direction of the base material (speed 50 mm / s), and 2000 thickness data points are collected every second to construct a base material thickness distribution heat map. The system has built-in threshold judgment logic, which automatically marks "thickness abnormal area" and generates a warning report when the local thickness deviation exceeds ± 0.1 mm; Finally, based on multi-dimensional parameters, intelligent matching of welding materials is realized. The material database uses a distributed architecture and stores 1000+ groups of "base material attribute-welding material" mapping relationships, including low-carbon steel ER50-6 welding wire, stainless steel ER308L welding wire, and other environmentally friendly welding material information. The matching algorithm is executed in two steps, as follows: According to the base material quality (such as carbon equivalent CE ≤ 0.4% to determine low alloy steel) and thickness (t ≤ 3 mm for thin-walled parts), 3-5 candidate welding materials are filtered out; Introduce a fuzzy logic controller, input environmental temperature, welding position (flat / vertical), and other auxiliary parameters, calculate the comprehensive adaptation degree of each candidate welding material (weight distribution: material matching 40%, thickness matching 30%, process adaptability 20%, environmental indicator 10%), and finally output the optimal welding material type and diameter specification (such as recommending φ1.2mm ER50-G welding wire for Q235B steel, thickness 4mm base material). The matching results are synchronized to the welding parameter intelligent decision-making module to provide basic data for subsequent process planning.

[0019] Step two, use mechanical derusting process to treat the surface oxide layer of the qualified base material, and obtain the structure parameters and environmental temperature data of the treated base material through three-dimensional scanning; Specifically, first, according to the base material quality (low carbon steel or stainless steel) and the surface rust degree, the sand blasting or shot blasting process is automatically selected; wherein, the sand blasting process uses 0.5-1.5mm quartz sand or steel shot, which is sprayed to the surface of the base material by 0.5-0.7MPa compressed air to remove the oxide skin, rust and dirt; the shot blasting process uses high-speed rotating impeller (speed 2000-3000rpm) to throw 0.8-1.2mm steel shot to achieve efficient rust removal; both processes are equipped with a dust recovery system, which collects metal dust through a magnetic filter device (filtering accuracy ≥99.9% @0.3μm) to avoid environmental pollution; after rust removal, the surface roughness of the base material reaches Ra≤25μm, the rust removal level meets the ISO8501-1 standard Sa2.5 level, and the requirements of laser welding for surface cleanliness are met; Then, the base material after rust removal is scanned in all directions by deploying a laser three-dimensional scanner (such as FARO Edge, accuracy ±0.1mm); the scanner emits a laser beam (wavelength 635nm) to project onto the surface of the base material, and after receiving the reflected light, it generates three-dimensional point cloud data through triangulation method, and extracts structure parameters and environmental parameters after algorithm processing; wherein, the structure parameters include the length, width, pipe diameter (accuracy ±0.1mm), bending angle (accuracy ±0.5°), bevel size (angle error ±1°, blunt edge thickness ±0.2mm) and other geometric characteristics; the environmental temperature (measurement range-10℃~40℃, accuracy ±0.5℃) is collected by the built-in temperature sensor in real time, which is used for subsequent welding parameter compensation calculation; Finally, the scanning data is transmitted in real time to the central control system through industrial Ethernet, and the system automatically compares the design drawing parameters, if the structure size deviation exceeds ±0.3mm (such as pipe diameter out of tolerance, bending angle deviation too large), an alarm is triggered and the process is returned to the previous step; the qualified data and environmental temperature value together constitute the base material feature vector, which is input into the welding parameter prediction model to provide key basis for subsequent process parameter calculation; In addition, the rust removal equipment and the three-dimensional scanner are linked through PLC to realize linkage control, when it is detected that the surface roughness of the base material does not meet the standard (such as Ra>25μm), the sand blasting / shot blasting time is automatically extended by 10-30 seconds, and the rust removal quality and data acquisition accuracy are double controlled.

[0020] Step three, build and train a welding parameter prediction model based on neural network, input welding material type, base material structure parameters and environmental temperature data, output welding parameters; Specifically, first, build a neural network model architecture, as follows: The mixed deep neural network includes a 1D convolution layer to extract time sequence features (such as changes in welding speed), a graph neural network (GNN) to model parameter relationships (such as the matching degree of material and welding material), and a multi-layer fully connected output layer; the model integrates an attention mechanism to dynamically allocate different feature weights (such as thickness weight 0.4, material weight 0.3), and designs an environment perception module to adaptively adjust the parameter prediction strategy according to temperature data (such as increasing the current weight by 0.15 in a low temperature environment); Then, historical welding data (≥10,000 groups) is collected, including base material parameters, welding material types, environmental temperatures, and corresponding welding parameters and quality indicators; after cleaning, standardizing, and enhancing the data, K-fold cross-validation (K=5) is used to divide the data set; the model is trained with a multi-objective loss function (60% parameter prediction error + 30% penetration prediction error + 10% defect probability), using the Adam optimizer (initial learning rate 0.001, decay 0.9 every 10 rounds), and the generalization ability is improved through early stopping and model fusion; Finally, the trained model is lightened (reducing the parameter quantity by 60%) and deployed to an edge computing node to achieve a single-sample prediction delay of <10ms; in actual application, the model outputs the optimal welding parameter combination (such as current 220±5A) based on the input base material parameters, welding materials, and environmental temperature, and verifies the prediction stability through digital twin technology; the system performs incremental learning every 100 times of actual welding data collection, and triggers manual intervention when the prediction error exceeds the threshold (such as penetration deviation >±0.3mm), forming a closed-loop iterative mechanism of "prediction-verification-optimization" to continuously improve model accuracy and adaptability.

[0021] Step four, execute the welding program according to the welding parameters, real-time collect welding data, and analyze the welding data through a multi-modal AI model to dynamically adjust the welding parameters; Specifically, first, based on the welding parameters output in step three (such as current 220A, voltage 25V, welding speed 300mm / min), the industrial robot automatically calls the corresponding welding process package to control the laser-arc hybrid welding head to perform the welding task; during welding, the current sensor (sampling frequency 10kHz, accuracy ±0.5A) and voltage sensor (accuracy ±0.1V) arranged on the welding torch are used to collect real-time electrical parameters, while a high-speed camera (frame rate 500fps, resolution 1280×720) captures dynamic images of the molten pool, and an infrared thermal imager (temperature measurement range 0-1500℃, accuracy ±2℃) monitors the temperature distribution of the weld; Then, the collected electrical parameters, images, and temperature data are preprocessed by the edge computing unit and input into the multi-modal AI model. The model uses a dual-flow neural network architecture, where the time series analysis branch uses an LSTM network to process current / voltage fluctuation data, identifying features such as short-circuit transition frequency and droplet transition stability. The visual analysis branch uses YOLOv8 to detect molten pool morphology (e.g., melt width, melt depth) and spatter rate (accuracy ≥98%) in real-time. The thermal field analysis branch uses U-Net semantic segmentation model to identify overheated areas (>1200℃) based on infrared data. Next, after the model fuses multiple source features, it calculates parameter adjustment strategies (e.g., when the molten pool width deviation is >0.5mm, dynamically adjust the welding speed ±5mm / min) through reinforcement learning algorithms. Finally, the parameter adjustment instructions (e.g., current fine-tuning to 225A) output by the AI model are transmitted in real-time to the welding power supply through the PLC control system, with a response time delay <20ms. At the same time, the system establishes a quality prediction mechanism, which automatically triggers an early warning and executes a three-level response when it detects a sudden increase in spatter rate (>5%) or molten pool fluctuation exceeding the threshold value (±0.3mm). Specifically, it includes: First-level adjustment: fine-tune the welding speed (±10mm / min) or protective gas flow (±1L / min); Second-level intervention: pause welding and start laser scanning to detect weld quality; Third-level correction: call the repair welding program to accurately repair defects based on 3D reconstruction data; The entire process forms a closed-loop control of "execution-monitoring-analysis-adjustment", ensuring that the welding parameters are always within the optimal range. According to actual measurements, the dynamic adjustment mechanism can control the weld width deviation within ±0.2mm, and the porosity rate is reduced to below 0.8%, improving the quality stability by 27% compared to traditional fixed-parameter welding processes.

[0022] Step five, according to the adjusted welding parameters, use the combined welding process of laser welding and friction stir welding to weld; Specifically, first, according to the dynamically adjusted welding parameters (e.g., laser power 1.5kW, stirring head speed 1200rpm), automatically match the welding area through process switching logic; for thin-walled components or precision butt welds with a plate thickness ≤3mm, preferentially use pulsed fiber laser welding (pulse frequency 100Hz, spot diameter 0.2mm) to achieve 0.1mm-level precision welding with a heat-affected zone width ≤0.5mm; for structural components or large-area splicing welds with a plate thickness >3mm, use friction stir welding (tungsten carbide stirring head diameter 8mm, welding pressure 10kN) to achieve solid-state connection under the action of upsetting force, avoiding gas hole and crack defects in traditional melting welding; Then, the six-axis industrial robot (repeat positioning accuracy ± 0.05mm) equipped with a double welding head adjusts the welding gun position in real time according to the welding trajectory planning through a laser tracking sensor (accuracy ± 0.02mm); during the laser welding stage, the galvanometer scanning system fills the path at a speed of 2m / s, and the coaxial argon blowing protection (flow rate 15L / min) is used to prevent weld oxidation; when switching to friction stir welding, the robot automatically adjusts the posture to make the stir head form a 2.5°inclination angle with the base material surface to optimize the plastic metal flow state, and at the same time, the water cooling system controls the welding area temperature ≤200℃ to reduce thermal deformation; Finally, a linkage model of laser energy and friction stir parameters is established, when the laser power is increased by 10%, the stir head speed is automatically reduced by 5% and the welding pressure is increased by 8% to balance the heat input and plastic deformation; real-time monitoring of torque feedback (accuracy ± 0.1N·m) and laser power stability (fluctuation ≤±2%) during welding, when the stir head wear causes abnormal torque (>15% of the set value), the automatic tool changing program is triggered (tool changing time ≤30s) to ensure the continuity and consistency of the composite welding process; the process makes the welding efficiency increase by 35% compared with single laser welding, and the joint strength reaches more than 95% of the base material, meeting the stringent requirements of green buildings on the pressure performance of heating radiators (≥1.6MPa).

[0023] Step six, based on the 3D visual detection system, the weld is scanned in all directions, the weld defect is identified by deep learning algorithm, and the laser repair welding mechanism is automatically triggered according to the weld defect type; Specifically, first, a 3D visual detection system using structured light projection principle (projecting 64 sinusoidal fringes) is used to scan the weld, a laser line projector (wavelength 450nm) and an industrial camera (resolution 2448×2048 pixels) cooperate to collect weld surface point cloud data at a speed of 50 frames per second, the scanning accuracy reaches ±0.05mm, and the maximum coverage range can reach 500mm×500mm; the system automatically builds a three-dimensional model of the weld, detects geometric defects such as concave, uneven reinforcement, etc. on the weld surface through curvature analysis algorithm, and identifies macroscopic morphology abnormalities such as undercut and weld tumor; Then, after converting the point cloud data into a depth image, it is input into the pre-trained YOLOv6s defect detection model (training samples include 100,000+ weld images, labeled with 20 types of defects), which can identify microscopic defects such as pores (diameter ≥0.2mm), cracks (length ≥0.5mm), and incomplete fusion, with an identification accuracy of ≥99%; for suspected defect areas, the system automatically calls the CT scanning module (resolution 100μm) for tomographic scanning, quantifies the defect size and depth through threshold segmentation and morphological analysis algorithm, and generates a detection report containing the defect position, type, and severity; Finally, according to the defect detection results, the intelligent planning of the repair welding path and process parameters; for surface porosity, local remelting is performed using pulsed laser (power 300 W, pulse width 0.5 ms), and a smooth weld is formed after the molten pool cools down; for deep cracks, a five-axis linkage laser head carried by the robot is used to perform layer-by-layer build-up welding along the crack direction (welding speed 2 mm / s, wire feeding rate 0.8 g / min), and the molten pool temperature (controlled at 1100-1300°C) and wire feeding stability are monitored in real time during the repair welding process; After the repair welding is completed, the system automatically performs a second scan on the repair welding area, and if residual defects are detected (such as the porosity rate after repair welding >1%), the repair welding process is repeated until it is qualified, achieving full closed-loop automation of defect repair; this mechanism improves the first-time weld pass rate from 89% to 98.5%, and the repair welding efficiency is improved by 70% compared to manual operation, while avoiding the thermal damage to the surrounding base material caused by traditional repair welding processes.

[0024] Step seven, the welding fumes are treated based on plasma catalytic purification technology, and the welding spatters and leftover materials are classified and recycled for regeneration treatment through the waste recycling module; Specifically, first, the plasma catalytic purification technology is used, high-frequency high-voltage electric field (voltage 15-20 kV) is used to charge the particulate matter (such as Fe2O3, MnO2) in the welding fumes, and the particles with a particle size of ≥0.1 μm are efficiently captured in the electrostatic precipitation chamber (electric field strength ≥3000 V / cm), with a purification efficiency of ≥99.5%; at the same time, the TiO2-based catalyst is filled in the catalytic reactor, and under ultraviolet light irradiation, harmful components such as nitrogen oxides (NOx) and carbon monoxide (CO) in the welding exhaust gas are decomposed, ultimately achieving a particulate matter emission concentration of ≤10 mg / m³ and a harmful gas removal rate of ≥95%, meeting the first-level standard of the "Integrated Emission Standard of Air Pollutants" (GB16297-1996); Then, the metal particles in the welding spatters are separated by a magnetic conveying belt (magnetic field strength 1.2 T) (recovery rate ≥98%), and after classification by a vibrating screen device (screen hole size 0.5 mm), the spatters with a particle size of >1 mm are fed into a smelting furnace (temperature 1600°C) for regeneration smelting to produce steel billets for welding wires; the metal dust with a particle size of ≤1 mm is conveyed by airflow to a nanometer material preparation system and is made into 3D printing metal powder (particle size distribution 20-50 μm) by an atomization deposition process (atomization speed 300 m / s); for leftover materials generated by cutting and grinding, the material sorting (accuracy rate ≥99%) is performed using a laser thickness gauge and a visual recognition system, the low-carbon steel leftover materials are processed by a shearing machine (shearing force 500 kN) and then remelted, and the stainless steel leftover materials are made into standard parts (such as bolts and nuts) by precision machining, achieving a material recycling rate of ≥95%; Finally, the high-temperature waste gas (temperature 800-1000℃) generated in the smelting process is recovered by the waste heat boiler to meet 30% of the heating demand of the workshop; the plasma purification system and the recovery equipment are powered by photovoltaic power (installed capacity 500kW, annual power generation 600,000 kWh), and the intelligent energy management system (accuracy ± 5%) dynamically regulates and controls the equipment power, which reduces the energy consumption by more than 40% compared with the traditional processing technology; the whole process builds a green manufacturing closed loop of "pollution control-resource recovery-energy circulation", significantly reduces the environmental load of welding production, and meets the sustainable development requirements of green building.

[0025] In this embodiment, specifically: in step three, the method for constructing and training the welding parameter prediction model based on the neural network comprises the following steps: Step 301, obtain historical process data, and after denoising and normalization processing, divide the historical process data into a training set, a validation set and a test set according to a proportion; the historical process data includes welding material types, base material structure parameters, environmental temperature data and corresponding welding parameters; Specifically, first, extract past production records from the manufacturing execution system (MES), covering different batches of welding material types (such as ER50-6 welding wire, E308 welding rod, etc.), base material structure parameters (including material, thickness, groove angle, etc.), environmental temperature data (workshop real-time temperature and humidity) and corresponding welding parameters (current, voltage, welding speed, protective gas ratio, etc.), and the cumulative data amount needs to reach the level of ten thousand groups to ensure the sufficiency of model training; Then, for time series data such as current and voltage, sliding average filtering (window size 5) is used to remove high-frequency noise, and isolated forest algorithm is used to detect and eliminate outliers (such as records with environmental temperature exceeding the range of-10℃-50℃); for continuous features (such as base material thickness, environmental temperature, welding current), Z-score standardization is performed to scale the values to the interval with mean value 0 and standard deviation 1; for classification features (such as welding material type, base material quality), one-hot encoding is implemented to convert them into binary vectors to adapt to the input requirements of neural networks; Finally, the processed data set is divided into a training set, a validation set and a test set according to the proportion of 8:1:1; among them, the training set is used for iterative optimization of model parameters, the validation set is used to monitor the overfitting in the training process and adjust the hyperparameters, and the test set is used to independently evaluate the final generalization ability of the model to ensure that the data distribution of each subset is uniform and contains various process scenarios, avoiding model prediction distortion caused by data bias.

[0026] Step 302, build a welding parameter prediction model including an input layer, a hidden layer and an output layer, the input layer corresponds to welding material types, base material structure parameters and environmental temperature data, and the output layer corresponds to welding parameters; Specifically, first, the model input layer is constructed to map the pre-processed multi-dimensional features to the neural network input nodes; among them, the welding material type is formed into an 8-dimensional vector after one-hot encoding (corresponding to 8 typical welding materials), the base material structure parameters are divided into thickness (1-dimensional), material code (6-dimensional), bevel angle (1-dimensional), and surface roughness (1-dimensional) for a total of 9 dimensions, and the environmental temperature data is 1-dimensional, totaling 18 input features; a batch normalization layer (Batch Normalization) is used to standardize the input data in real time, accelerate model convergence, and enhance stability; Then, the hidden layer architecture is designed, a deep residual network (ResNet) structure is used, 4 residual blocks are stacked (each residual block contains 3 fully connected layers with 128 neurons each), and the gradient vanishing problem is alleviated through jump connection; a multi-head attention mechanism (Multi-head Attention) is introduced, 8 attention heads are set to automatically capture the nonlinear relationship between input features (such as the coupling relationship between material and welding current); a Dropout layer (dropout rate 0.2) is added after each fully connected layer to prevent model overfitting; Finally, the hidden layer features are mapped to a 4-dimensional output vector through a 3-layer fully connected network (with 64, 32, and 4 neurons respectively), corresponding to the welding current, voltage, speed, and protective gas flow 4 key parameters; for parameters such as current and voltage that are more continuous, a linear activation function is used; for parameters such as welding speed that have an upper physical limit, a ReLU6 activation function is used for constraint; the overall model parameter size is about 1.5 million, which is compressed to 1 / 4 of the original size through parameter quantization technology (8-bit integer) to adapt to the deployment requirements of industrial edge computing devices.

[0027] Step 303, iteratively train the welding parameter prediction model using the training set, adjust the weights of the welding parameter prediction model through the Adam algorithm, and evaluate the generalization ability of the welding parameter prediction model using the validation set; Specifically, first, the adaptive moment estimation (Adam) optimization algorithm is used, the initial learning rate is set to 0.001, β1=0.9, and β2=0.999 to balance the first and second moments of gradient estimation; batch normalization (Batch Normalization) layer is used to accelerate convergence, and 64 samples are processed per batch; the loss function is designed as weighted mean square error (MSE), and the current, voltage, welding speed and gas flow are respectively assigned weights of 0.4, 0.3, 0.2 and 0.1 to highlight the prediction accuracy of key parameters; Then, the model is iteratively trained on the training set for 100 epochs, with performance evaluated on the validation set after each epoch. An early stopping mechanism is introduced to prevent overfitting, terminating training when the validation set loss has not improved for 5 consecutive epochs. Meanwhile, a learning rate decay strategy is adopted, reducing the learning rate to 0.7 times every 10 epochs to ensure stable convergence of the model. During training, gradient clipping is used to limit the gradient norm to within 5 to avoid the problem of gradient explosion. Finally, the validation set is used to calculate multiple performance indicators, including mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). For current prediction, MAE ≤ 5A and R² ≥ 0.95 are required; for voltage prediction, MAE ≤ 1V and R² ≥ 0.93 are required. The Monte Carlo Dropout technique is used to quantify the prediction uncertainty, ensuring that the 95% confidence interval covers the true parameter range. In addition, special evaluations are conducted on abnormal working conditions (such as low temperature environment and special material welding) in the validation set. If the error of a specific scene exceeds the threshold, additional samples of that type are added for incremental training to continuously optimize the model's generalization ability.

[0028] Step 304, using the test set to evaluate the final performance of the welding parameter prediction model, so that the welding parameter prediction accuracy meets the preset standard; Specifically, first, an independent test set (1,000 samples) is used to evaluate the model performance from multiple dimensions, calculating the prediction error indicators for current, voltage, welding speed, and gas flow, as follows: Absolute error indicators: mean absolute error (MAE), root mean square error (RMSE); Relative error indicators: mean absolute percentage error (MAPE); Goodness of fit indicators: coefficient of determination (R²); The preset standards are: current MAE ≤ 5A (MAPE ≤ 3%), voltage MAE ≤ 1V (MAPE ≤ 5%), welding speed MAE ≤ 15 mm / min (MAPE ≤ 5%), gas flow MAE ≤ 1.5 L / min (MAPE ≤ 8%), and R² ≥ 0.95 for each; Then, the Monte Carlo Dropout (executed 50 times forward propagation) generates a prediction distribution, and the 95% confidence interval width is calculated. The current prediction interval width is required to be ≤ 15A, and the voltage interval width is required to be ≤ 3V, with the proportion of samples exceeding this range not exceeding 5%. Meanwhile, quantile regression technology is used to verify the coverage rate of the prediction interval, ensuring that the probability of the actual parameter falling within the prediction interval is ≥ 90%; Finally, special conditions (such as low temperature environment [-5℃~5℃], large thickness base material [>8mm]) in the test set are evaluated; if the error exceeds 1.2 times of the preset standard, the model iteration process is triggered, as follows: Supplement such working condition samples from historical data; Fine-tune the model using knowledge distillation technology; Re-test until all conditions meet the standard; When the average error of all test samples and the error of special conditions meet the preset standard, and the single sample prediction delay of the model on the edge computing device is ≤10ms, the performance verification is passed, and the production deployment phase is entered.

[0029] Step 305, deploy the welding parameter prediction model that passes the test to the welding parameter intelligent decision-making module, receive welding material type, base material structure parameters and environmental temperature data in real time, and output welding parameters; Specifically, first, deploy the trained model (TensorFlow SavedModel format) to an industrial-grade edge computing unit (such as UNO-3272G from Advantech, equipped with Intel i7 processor and NVIDIA Jetson GPU), and ensure environmental consistency through Docker containerization packaging; develop RESTful API interface to support receiving real-time data in JSON format and returning prediction results, with response delay controlled within 10ms; Then, get welding material type, base material thickness and other parameters from the workshop PLC system through OPC UA protocol, and receive environmental temperature sensor data through MQTT protocol; and perform online standardization (reuse the mean and standard deviation of the training stage) on the input data, and detect data jumps through a sliding window (size 5), and abnormal values automatically trigger the manual review process; at the same time, input the preprocessed data into the model, output the optimal welding parameter combination, and add the prediction confidence; Finally, deploy 3 independently trained models for integrated prediction, use weighted voting mechanism (weights based on validation set performance) to output the final parameters, and reduce the bias risk of a single model; After completing 100 welds, compare the actual parameters with the predicted values, if the deviation exceeds the threshold (such as current deviation >±8A) for 5 consecutive times, trigger the online fine-tuning of the model; when the prediction confidence <0.85, the system automatically pushes the warning to the process engineer terminal, and provides historical similar cases as a decision-making reference.

[0030] In this embodiment, specifically: in step four, the welding parameters include welding current, welding voltage, welding speed and protective gas ratio; Among them, the welding current (unit: A) is the core parameter of determining the arc energy density, which directly affects the penetration and weld width; too large current is easy to cause burn-through and increase spatter; too small current will cause incomplete fusion defects; typical range: 100-150 A for thin plate (2-3 mm) welding, 180-250 A for medium-thick plate (4-8 mm) welding; Voltage (unit: V) cooperates with current to control the arc length and stability; too high voltage will make the arc drift and increase the fusion width; too low voltage will cause insufficient arc rigidity and easy to produce slag inclusion; generally, voltage and current have a linear relationship (such as in MAG welding, voltage ≈ 14 + 0.05 × current); Welding speed (unit: mm / min) affects the weld formation and heat input; too fast speed will result in shallow penetration and undercut; too slow speed will expand the heat-affected zone and easily produce deformation; typical range: 150-300 mm / min for manual welding, 300-500 mm / min for automatic welding; Gas mixture ratio (unit: %) is used to protect the weld from oxidation, and different gas mixtures are suitable for different materials; for example: Carbon steel welding: 80% Ar + 20% CO2 (reduce spatter); Stainless steel welding: 98% Ar + 2% O2 (improve molten pool fluidity); Aluminum alloy welding: pure Ar (prevent alloy element burnout).

[0031] Welding data includes arc signal, molten pool image, and temperature data; Arc signal includes current / voltage waveform and arc spectrum; Among them, current / voltage waveform is collected by Rogowski coil (bandwidth 10 MHz) and Hall sensor, with a sampling frequency ≥ 10 kHz; waveform features (such as short circuit frequency, peak current duration) reflect the droplet transfer mode (such as short circuit transfer, spray transfer), which is used to monitor the welding stability; Arc spectrum is the spectral intensity of specific wavelengths (such as FeⅠ 386.0 nm corresponding to iron ionization), which can analyze the molten pool composition and protection effect, and the spectrometer resolution needs to reach 0.1 nm; Molten pool image includes visible light image and infrared image; Among them, the visible light image is the molten pool dynamics captured by a high-speed camera (frame rate ≥ 500 fps) with a narrow-band filter (such as 532 nm), which is used to measure the molten pool length, width (accuracy ± 0.1 mm), and oscillation frequency (normal range 20-50 Hz); Infrared image refers to the thermal imager (temperature measurement range 0-1500℃) monitoring the temperature distribution of the molten pool, and the overheated area (> 1200℃) indicates the presence of pores, and the temperature gradient (≥ 100℃ / mm) affects the weld crystal direction; Temperature data is collected through contact and non-contact temperature measurement: Contact temperature measurement uses a K-type thermocouple (response time <10ms) to measure the weld cooling rate (critical value: 300-500℃ / s, affecting the hardness of the heat-affected zone); The non-contact temperature measurement uses a two-color infrared thermometer to measure the center temperature of the molten pool (error ± 2%), which is used to adjust the heat input parameters in real time.

[0032] The fusion algorithms of the multimodal AI model include: Perform wavelet transform on the arc spectrum signal to extract the 5-15kHz frequency band features; Specifically, during the welding process, the arc spectrum signal contains key information about the welding status. The 5-15kHz frequency band is closely related to droplet transfer. When the droplet transfer is unstable, the signal in this frequency band will show significant changes. Wavelet transform is like using a "precision comb" to comb through the arc spectrum signal, and the signal in the 5-15kHz frequency band can be "picked out" separately. The features extracted from it can reflect whether the droplet transfer is smooth and the welding process is stable. For example, if the signal in this frequency band shows abnormal fluctuations, it means that there is a problem with the droplet transfer, and spatter is likely to occur during welding. In this case, the welding parameters need to be adjusted. Perform threshold segmentation on the molten pool infrared image and calculate the molten pool area and perimeter; Specifically, the shape and size of the molten pool directly affect the quality of the weld, and the molten pool infrared image records the state of the molten pool. Threshold segmentation of the molten pool infrared image is to clearly separate the molten pool area from the background. Specifically, it is like setting a "screening standard" for the image. Pixels that meet the conditions are identified as belonging to the molten pool, and those that do not meet the conditions are classified as background. After the segmentation is completed, the area and perimeter of the molten pool are calculated. A molten pool area that is too large or too small indicates a problem with the welding heat input. For example, a large area means too high a current, which can easily burn through the parent material. Changes in the perimeter can reflect the shape of the molten pool edge. If the perimeter suddenly increases, it indicates that the molten pool oscillation has intensified, which will cause defects such as undercuts. These geometric features provide an intuitive basis for judging weld quality and adjusting the welding process. Use Kalman filter to fuse multi-sensor data and output welding quality assessment value; Specifically, during the welding process, the data of a single sensor has errors or limitations, such as the arc spectrum signal being affected by electromagnetic interference, and the molten pool image being affected by light; the Kalman filter is like a "data integration master" that can analyze and process the data collected by other sensors, such as the characteristics extracted from the arc spectrum signal, the geometric parameters calculated from the molten pool infrared image, and the temperature; it can predict and optimize the welding quality state according to the historical trend of the data and the current measurement value, eliminate noise and uncertainty in the data; finally, a reliable welding quality evaluation value is output, which intuitively represents the quality of the current welding; if the evaluation value is low, the system will issue a warning in time and adjust the welding parameters according to the preset rules to ensure stable welding quality.

[0033] In this embodiment, specifically: in step five, the laser welding uses a pulsed fiber laser, the pulse frequency of the pulsed fiber laser is 50-200Hz, and the average power is 1-3kW; by adjusting the pulse frequency of the pulsed fiber laser within the range of 50-200Hz and the average power within the range of 1-3kW, the pulsed fiber laser can accurately control the energy input, and then realize instantaneous high-energy output through high-frequency pulse, which can effectively reduce the range of heat-affected zone and avoid deformation of the base material due to overheating, especially suitable for welding of thin-walled parts or precise butt welds, to ensure that the weld has high strength and aesthetics.

[0034] Friction stir welding uses a tungsten carbide stirrer, and dynamically matches the speed and pressure parameters according to the thickness of the base material; the tungsten carbide stirrer has the characteristics of high hardness and strong wear resistance; during the welding process, the speed and pressure parameters are dynamically matched according to the thickness of the base material; for thinner base materials (such as less than 3mm thick), the stirrer rotates at a higher speed (800-1200rpm) to quickly rotate with smaller pressure (5-8kN) to make the base material locally plasticized to complete the welding, preventing the plate from being crushed; for thicker base materials (more than 3mm thick), the speed is reduced to 400-800rpm, and the pressure is increased to 8-12kN to ensure that the deep material is fully stirred and fused to form a dense welded joint; through this dynamic matching, high-quality welding results can be obtained for base materials of different thicknesses.

[0035] In this embodiment, specifically: in step one, the AI visual detection system is equipped with a line array CCD camera and a deep learning defect recognition model; The line array CCD camera can perform high-speed and high-precision scanning imaging of the weld through linearly arranged photosensitive elements, and can capture micron-level details (such as small cracks and pores on the surface of the weld), with a scanning speed of dozens of meters per minute, suitable for real-time detection requirements of industrialized assembly lines; The deep learning defect recognition model is trained based on a large number of labeled weld image data, and can automatically learn defect features (such as the edge shape of undercut and the gray difference of incomplete fusion); after the line array CCD camera collects the weld image, the model extracts image features layer by layer through a convolutional neural network (CNN), combines a YOLO, a Faster R-CNN and other target detection algorithms, realizes rapid recognition and positioning of various defects such as pores, cracks and welding bumps, and the detection accuracy can reach more than 99%, and the single image processing time is less than 50 milliseconds, which provides efficient technical support for real-time judgment of welding quality.

[0036] In this embodiment, specifically: in step two, the mechanical derusting process adopts a sand blasting or shot blasting process, the sand blasting medium is quartz sand or steel shot, and the derusting level reaches Sa2.5 level; Among them, the sand blasting process uses quartz sand as the medium (particle size 0.5-2mm), which is sprayed at high speed to the surface of the base material through compressed air to form a uniform rough surface; the shot blasting process uses steel shot (diameter 1-3mm), which is impacted on the base material by high-speed rotating impeller to achieve a stronger impact derusting effect; both processes need to ensure that the derusting level reaches the Sa2.5 level standard, that is, more than 95% of the oxide scale, rust and pollutants on the surface of the base material are removed, and a uniform metal luster is presented, providing a clean surface for subsequent welding and avoiding defects such as pores and cracks in the weld caused by impurities.

[0037] The three-dimensional scanning adopts a laser three-dimensional scanner to obtain the base material structure size, geometric shape parameters and environmental temperature data with an accuracy of ±0.1mm; Specifically, the laser three-dimensional scanner is used to scan the base material in all directions, and the device constructs a three-dimensional point cloud model of the base material by emitting a laser beam and receiving a reflected signal; this technology can accurately obtain the base material structure size (such as length, thickness, bevel angle) and geometric shape parameters (such as curved surface radian, edge flatness) with an accuracy of ±0.1mm, and simultaneously collect environmental temperature data (accuracy ±0.5℃); after processing the scanning data, it is used to verify the base material machining accuracy, plan the welding path, and serve as the input basis for the welding parameter prediction model, to ensure that the welding process and the base material characteristics are accurately matched, and to improve the welding quality stability.

[0038] In this embodiment, specifically: in step six, the 3D visual detection system generates weld three-dimensional point cloud data with an accuracy of ±0.05mm based on the principle of structured light projection, and intelligently classifies and identifies through a defect feature library; Specifically, the system projects a specific pattern of structured light (such as sinusoidal stripes, grid patterns) onto the weld surface through a projector. When the structured light encounters uneven weld surfaces, the stripes will deform. The industrial camera mounted on the system captures the deformed light stripe image at sub-millisecond speed, calculates the pixel-level disparity through the principle of triangulation, and generates three-dimensional point cloud data with an accuracy of ±0.05mm, which can clearly restore the geometric characteristics of the weld such as height, width, and curvature (such as whether the excess height is uniform and whether the weld leg size meets the standard). The generated point cloud data is processed by noise reduction and splicing, and compared with the built-in defect feature library (containing three-dimensional feature templates of 20 typical defects such as pores, undercut, and incomplete fusion); through the point cloud segmentation algorithm based on deep learning (such as PointNet++), the defect type is automatically identified and the specific position in the weld is located; for example, when the local area point cloud density is abnormally reduced, the system will determine it as a pore defect; if the curvature of the weld edge point cloud is found to be suddenly changed, it will be classified as an undercut defect; the whole detection process can be completed within 5 seconds, with a detection accuracy of ≥99%, providing accurate defect position and type information for subsequent laser repair welding, and realizing the whole-process closed-loop control of welding quality.

[0039] In summary, the intelligent welding and forming method for steel heating radiators for green buildings provided by the embodiment of the present application realizes the identification of base material defects and the intelligent matching of welding materials through an AI vision detection system, completes surface treatment and data acquisition through mechanical derusting and three-dimensional scanning, realizes intelligent planning of process parameters based on a neural network trained welding parameter prediction model, dynamically adjusts welding parameters with the help of a multi-modal AI model and improves welding precision by combining laser and friction stir hybrid welding processes, realizes defect closed-loop repair through 3D vision detection and laser repair welding, and finally adopts plasma catalytic purification and waste recycling technology to build a green manufacturing system, which improves welding efficiency, quality stability and environmental benefits throughout the whole process, and meets the sustainable development needs of green buildings.

[0040] Figure 3 is a functional module schematic diagram of an intelligent welding and forming system for steel heating radiators for green buildings according to an embodiment of the present application, as Figure 3 shown, an intelligent welding and forming system for steel heating radiators for green buildings, comprising a base material screening module, an environmental protection treatment module, a parameter prediction module, a parameter control module, a hybrid welding execution module, an intelligent detection and repair module, and a smoke and waste treatment module. The base material screening module is configured to identify the surface defects of the steel heating radiator base material based on an AI vision detection system, identify qualified base materials, and automatically match the type of welding material from the material database according to the material quality and thickness parameters of the qualified base materials. The environmental treatment module is configured to adopt a mechanical rust removal process to treat the surface oxide layer of the qualified base material, and obtain the structure parameters of the treated base material and the environmental temperature data through three-dimensional scanning; The parameter prediction module is configured to construct and train a welding parameter prediction model based on a neural network, input the welding material type, base material structure parameters and environmental temperature data, and output the welding parameters; The parameter control module is configured to execute a welding program according to the welding parameters, collect welding data in real time, analyze the welding data through a multi-modal AI model, and dynamically adjust the welding parameters; The composite welding execution module is configured to adopt a composite welding process combining laser welding and friction stir welding to perform welding according to the adjusted welding parameters; The intelligent detection and repair module is configured to perform omnidirectional scanning on the weld based on a 3D vision detection system, identify weld defects using a deep learning algorithm, and automatically trigger a laser repair welding mechanism according to the weld defect type; The smoke and waste treatment module is configured to process welding smoke based on plasma catalytic purification technology, and simultaneously classify and recycle welding spatters and corner scraps for regeneration treatment.

[0041] In this embodiment, specifically, the parameter prediction module includes a data preprocessing unit, a model training unit, and a parameter output unit; The data preprocessing unit is used for denoising, normalizing, and dividing the training set, validation set, and test set of historical process data; The model training unit is used for building a welding parameter prediction model and iteratively training it through the Adam algorithm; The parameter output unit is used for deploying the tested welding parameter prediction model and outputting welding parameters in real time.

[0042] In this embodiment, specifically, the parameter control module includes a sensor acquisition unit and an AI fusion decision unit; The sensor acquisition unit integrates a spectrometer, an infrared thermal imager, and a K-type thermocouple, and is used for collecting arc signals, molten pool dynamics, and temperature field data in real time; The AI fusion decision unit is used for wavelet transform of the arc spectrum signal, extraction of 5-15 kHz frequency band features, threshold segmentation of the molten pool infrared image, calculation of the molten pool area and perimeter, fusion of multi-sensor data using a Kalman filter, and output of a welding quality evaluation value; if the welding quality evaluation value is <0.8, the parameter adjustment is automatically triggered.

[0043] To sum up, the green building steel heating radiator intelligent welding forming system provided by the embodiment of the application realizes parent material defect identification and intelligent matching of welding materials through the parent material screening module, completes surface rust removal and three-dimensional data acquisition by using the environmental protection processing module, realizes intelligent planning of welding parameters by means of the neural network model of the parameter prediction module, dynamically optimizes the welding process by the multi-modal AI model of the parameter control module, realizes efficient combination of laser and friction stir welding by using the composite welding execution module, completes intelligent identification and closed-loop repair of weld defects by relying on the intelligent detection and repair module, and finally realizes welding smoke purification and waste regeneration treatment by the smoke and waste treatment module, so that an intelligent and green manufacturing system is constructed in the whole process, welding efficiency, quality stability and resource recycling rate are significantly improved, and the demand of green building for efficient and environmentally friendly manufacturing processes is met.

[0044] For other details of the technical solutions of each module in the green building steel heating radiator intelligent welding forming system described above, refer to the description of the green building steel heating radiator intelligent welding forming method in the above embodiment, which will not be repeated here.

[0045] It should be noted that each embodiment in the present specification adopts a progressive manner for description, and each embodiment focuses on the difference from other embodiments, and the same and similar parts between each embodiment can be referred to. For system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and the relevant parts can be referred to the part of the method embodiment.

[0046] The above only describes the preferred embodiments of the application and should not be used to limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application should be included in the protection scope of the application.

Claims

1. An intelligent welding forming method for steel radiators for green buildings, characterized in that: The following steps are involved: The AI-based visual inspection system identifies surface defects on steel radiator base materials, identifies qualified base materials, and automatically matches welding material types from the material database based on the material and thickness parameters of the qualified base materials. The surface oxide layer of qualified parent materials is treated by mechanical rust removal process, and the structural parameters and ambient temperature data of the treated parent materials are obtained through 3D scanning; Build and train a welding parameter prediction model based on a neural network, input welding material type, parent material structural parameters and ambient temperature data, and output welding parameters; Execute welding procedures according to welding parameters, collect welding data in real time, analyze welding data through multimodal AI models, and dynamically adjust welding parameters; According to the adjusted welding parameters, a hybrid welding process combining laser welding and friction stir welding is used for welding; The weld seam is scanned in all directions based on a 3D visual inspection system, weld defects are identified using a deep learning algorithm, and the laser repair welding mechanism is automatically triggered based on the type of weld defect. Welding fume is treated based on plasma catalytic purification technology, while welding spatter and scraps are sorted, recycled and regenerated through the waste recovery module.

2. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The method for constructing and training a welding parameter prediction model based on a neural network comprises the following steps: Obtain historical process data, perform noise reduction and normalization on the historical process data, and then divide the data into a training set, a validation set, and a test set in proportion; the historical process data includes welding material type, base material structural parameters, ambient temperature data, and corresponding welding parameters; Build a welding parameter prediction model including an input layer, a hidden layer, and an output layer, wherein the input layer corresponds to the welding material type, parent material structural parameters, and ambient temperature data, and the output layer corresponds to the welding parameters; The welding parameter prediction model was iteratively trained using the training set, the weight of the welding parameter prediction model was adjusted using the Adam algorithm, and the generalization ability of the welding parameter prediction model was evaluated using the validation set. The final performance of the welding parameter prediction model is evaluated using the test set to ensure that the welding parameter prediction accuracy meets the preset standards; The tested welding parameter prediction model is deployed to the welding parameter intelligent decision-making module, which receives welding material type, parent material structural parameters and ambient temperature data in real time and outputs welding parameters.

3. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The welding parameters include welding current, welding voltage, welding speed and shielding gas ratio; the welding data include arc signal, molten pool image and temperature data; The fusion algorithm of the multimodal AI model includes: Perform wavelet transform on the arc spectrum signal to extract the 5-15kHz frequency band features; Perform threshold segmentation on the molten pool infrared image and calculate the molten pool area and perimeter; The Kalman filter is used to fuse multi-sensor data and output welding quality assessment values.

4. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The laser welding uses a pulsed fiber laser with a pulse frequency of 50-200 Hz and an average power of 1-3 kW; the friction stir welding uses a tungsten carbide stirring head and dynamically matches the rotation speed and pressure parameters according to the thickness of the base material.

5. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The AI ​​visual inspection system is equipped with a linear array CCD camera and a deep learning defect recognition model.

6. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The mechanical rust removal process adopts sandblasting or shot blasting, the sandblasting medium is quartz sand or steel shot, and the rust removal level reaches Sa2.5 level; the three-dimensional scanning adopts a laser three-dimensional scanner to obtain the parent material structure size, geometric shape parameters and ambient temperature data with an accuracy of ±0.1mm.

7. The intelligent welding forming method for steel radiators for green buildings according to claim 1, characterized in that: The 3D visual inspection system is based on the principle of structured light projection, generates three-dimensional point cloud data of welds with an accuracy of ±0.05mm, and performs intelligent classification and identification through a defect feature library.

8. An intelligent welding and forming system for steel radiators for green buildings, applied to an intelligent welding and forming method for steel radiators for green buildings according to any one of claims 1 to 7, characterized in that: It includes a base material screening module, an environmental protection treatment module, a parameter prediction module, a parameter control module, a composite welding execution module, an intelligent detection and repair module, and a smoke and waste treatment module; The parent material screening module is configured to identify surface defects of the steel radiator parent material based on the AI ​​visual inspection system, identify qualified parent materials, and automatically match the welding material type from the material database based on the material and thickness parameters of the qualified parent materials; The environmental protection treatment module is configured to use a mechanical rust removal process to treat the surface oxide layer of the qualified parent material, and obtain the structural parameters and ambient temperature data of the treated parent material through three-dimensional scanning; The parameter prediction module is configured to construct and train a welding parameter prediction model based on a neural network, input welding material type, parent material structural parameters and ambient temperature data, and output welding parameters; The parameter control module is configured to execute the welding program according to the welding parameters, collect welding data in real time, analyze the welding data through a multimodal AI model, and dynamically adjust the welding parameters; The hybrid welding execution module is configured to perform welding using a hybrid welding process combining laser welding and friction stir welding according to the adjusted welding parameters; The intelligent inspection and repair module is configured to perform an all-round scan of the weld based on a 3D visual inspection system, identify weld defects using a deep learning algorithm, and automatically trigger a laser repair welding mechanism based on the type of weld defect; The fume and waste treatment module is configured to treat welding fume based on plasma catalytic purification technology, and at the same time classify, recycle and regenerate welding spatter and scraps.

9. The intelligent welding and forming system for steel radiators for green buildings according to claim 8, characterized in that: The parameter prediction module includes a data preprocessing unit, a model training unit and a parameter output unit; The data preprocessing unit is used to perform noise reduction, normalization processing and division of historical process data into training set, verification set and test set; The model training unit is used to build a welding parameter prediction model and iteratively train it through the Adam algorithm; The parameter output unit is used to deploy the tested welding parameter prediction model and output the welding parameters in real time.

10. The intelligent welding and forming system for steel radiators for green buildings according to claim 8, characterized in that: The parameter control module includes a sensor acquisition unit and an AI fusion decision unit; The sensor acquisition unit integrates a spectrometer, an infrared thermal imager and a K-type thermocouple, and is used to collect arc signals, molten pool dynamics and temperature field data in real time; The AI ​​fusion decision-making unit is used to perform wavelet transform on the arc spectrum signal, extract the 5-15kHz frequency band features, perform threshold segmentation on the molten pool infrared image, calculate the molten pool area and perimeter, and use a Kalman filter to fuse multi-sensor data to output a welding quality assessment value; if the welding quality assessment value is less than 0.8, parameter adjustment is automatically triggered.

Images