Aluminum profile production quality management and control system and method

Abstract

The invention discloses an aluminum profile production quality management and control system and method, and relates to the technical field of aluminum profile production, the management and control system comprises a raw material detection module, a production process monitoring module, a quality analysis and early warning module, a finished product detection module and a data management module, and comparing with a preset quality standard threshold value to judge whether the aluminum raw material is qualified or not. According to the aluminum profile production quality management and control system and method, in the production operation process, the model is updated in real time according to actual parameters, like a digital twinborn body is created for a production line; whether the temperature fluctuation during smelting or the pressure change in the extrusion process, the system can capture and compare the parameters with standard process parameters in time; once an abnormal trend occurs, the system gives an early warning in advance, and a worker can adjust the process in time according to the early warning, thereby effectively preventing quality problems caused by abnormal equipment parameters, guaranteeing stable operation of production, and reducing loss caused by production interruption.

Description

Technical Field

[0001] The present invention relates to the technical field of aluminum profile production, and in particular to an aluminum profile production quality control system and method. Background Art

[0002] In the field of modern industrial manufacturing, aluminum profiles have become an indispensable basic material for industries such as construction, transportation, aerospace, and electronics and electrical appliances due to their light weight, high strength, corrosion resistance, and strong recyclability. Taking the construction industry as an example, aluminum profiles are widely used in door and window frames and curtain wall structures of high-rise buildings, and their quality directly affects the safety and aesthetics of the buildings. In the manufacturing of new energy vehicles, lightweight aluminum profile body frames help to improve vehicle range, while quality defects may lead to insufficient body strength and endanger driving safety. As various industries continue to place increasing requirements on the performance and precision of aluminum profiles, quality control in the production process has become a key factor in determining a company's competitiveness.

[0003] Traditional aluminum profile manufacturers face many difficulties in quality control. During the raw material acceptance stage, most companies still use manual sampling combined with chemical titration to detect the composition of aluminum ingots. This method is not only inefficient, but also difficult to detect trace element impurities at the part-per-million level. For example, when the trace titanium content in the aluminum raw material exceeds the standard, the extruded profile will lack toughness. In terms of production process monitoring, the model that relies on workers to regularly inspect equipment and instrument data has obvious lags. During this period, since staff are unable to discover and deal with problems in a timely manner, not only will the impact persist, but it will also form an expanding trend, resulting in very large economic losses.

[0004] The finished product inspection process also has limitations; some companies rely solely on visual inspection and simple dimensional measurements, and lack effective detection methods for internal defects such as pores and cracks; in export trade, returns due to internal quality issues are common, seriously damaging the company's reputation; in addition, data between different factories are isolated from each other, making collaborative analysis impossible; taking multiple aluminum profile production bases under the same industrial group as an example, due to the failure to share data in a timely manner, each factory repeatedly invests resources in the development of similar quality improvement plans, which will result in a huge waste of manpower and material resources.

[0005] With the development of Industry 4.0 and intelligent manufacturing technologies, traditional quality control models are no longer able to meet industry demands. The market urgently needs a full-process, intelligent, and precise quality control system to address existing issues such as low detection accuracy, slow response, and data silos, thereby improving the overall quality and efficiency of aluminum profile production. To this end, we propose a quality control system and method for aluminum profile production. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In view of the deficiencies in the prior art, the present invention provides an aluminum profile production quality control system and method to solve the technical problems pointed out in the above background technology.

[0008] (2) Technical solution

[0009] To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0010] An aluminum profile production quality control system, which consists of a raw material detection module, a production process monitoring module, a quality analysis and early warning module, a finished product detection module, and a data management module;

[0011] The raw material detection module uses a quantum dot spectral detection chip and an atomic force microscope to operate. First, a core-shell quantum dot material is coated on the surface of the quantum dot spectral detection chip. This material produces a fluorescent effect when exposed to light of a specific wavelength. During detection, the aluminum raw material is made into a tiny sample and allowed to flow in the microfluidic channel within the chip. The chip is then illuminated with ultraviolet light with a wavelength of 350-400 nanometers. The quantum dot material on the chip produces a fluorescent signal. The strength of the signal is related to the trace element content in the aluminum raw material. The system captures the signal through a photodetector and calculates the content of 15 trace elements through a signal processing algorithm. At the same time, the atomic force microscope probe scans the raw material surface line by line at a spacing of 5 nanometers in tapping mode. By measuring the changes in atomic force between the probe and the raw material surface, it generates a fine image of the raw material surface with nanometer-level accuracy. Finally, the system compares the detected element content data and surface image with the qualified standard data and image templates pre-stored in the database for pixel-level comparison and parameter matching to determine whether the raw material is qualified.

[0012] The production process monitoring module uses finite element analysis to construct a three-dimensional dynamic simulation model based on the design drawings and operating parameters of actual production equipment. This model contains over one million computing units and fully covers all aspects of aluminum profile production, including melting, extrusion, and aging. During the production process, data is collected through fiber grating temperature sensors, piezoresistive pressure sensors, Hall effect velocity sensors, and piezoelectric vibration sensors installed on the equipment. Fiber grating temperature sensors acquire temperature data by detecting the wavelength offset of the reflected light from the grating, piezoresistive pressure sensors measure pressure based on changes in resistance, Hall effect velocity sensors monitor equipment speed using the Hall effect, and piezoelectric vibration sensors detect equipment vibration through changes in the charge of the piezoelectric material. After data is collected, it is de-noised and fused using a Kalman filter algorithm, and the three-dimensional dynamic model is then updated in real time at a frequency of 20 times per second. By comparing the operating status parameters in the model with the standard ranges pre-set in the process parameter database, potential quality issues in the production process can be identified in advance.

[0013] The quality analysis and early warning module deploys edge computing nodes at each production site, running a lightweight data processing program based on convolutional neural networks to perform preliminary analysis of collected data and extract key feature information. Simultaneously, through the federated learning data sharing framework, each plant locally analyzes and processes the data using the stochastic gradient descent algorithm, and only uploads model parameter update data that has undergone homomorphic encryption. During data transmission, the TLS1.3 secure transmission protocol is used, and Laplace noise is added to protect data privacy. The system is configured to activate the blockchain evidence storage function when the deviation of the detected data from the standard data exceeds three times the standard deviation calculated from historical data. The SHA-256 hash algorithm is used to record the occurrence time and specific value of the abnormal data in the distributed ledger, and an early warning is issued through a pre-programmed smart contract program.

[0014] The finished product inspection module combines a terahertz time-domain spectrometer, a microfocus X-ray imager, and a laser-induced breakdown spectrometer (LIBS) in an optically coaxially aligned configuration. The THD transmits terahertz waves in the frequency range of 0.1-10 THz to penetrate the interior of the finished product, detecting porosity and delamination defects based on the time delay and intensity differences between wave reflection and transmission. The MFI scans the finished product eight times at 45-degree intervals, reconstructing the image using a filtered back-projection algorithm, enabling accurate submillimeter dimension measurement. The LIBS emits laser pulses with energies of 50-100 mJ to bombard the finished product surface, exciting the surface material to produce a plasma spectrum. The intensity and position of characteristic spectral lines in the spectrum are analyzed to determine the elemental distribution on the surface. The system utilizes a multimodal data fusion algorithm based on an attention mechanism. Based on the accuracy and reliability of the three inspection instruments, the system assigns weights to different inspection data sets, fusing the three sets of data to generate a comprehensive quality inspection map containing defect locations, dimensional parameters, and elemental concentrations. Finally, the system performs feature matching and numerical comparison analysis with the finished product quality standard map stored in the quality standard library.

[0015] The data management module adopts a blockchain distributed storage architecture and uses a practical Byzantine fault-tolerant consensus algorithm to ensure data consistency in a ring network consisting of 50 storage nodes. By collecting data from the entire production process and using a graph neural network algorithm, data such as raw material characteristics, production process parameters, equipment operating status, and finished product quality are linked in the form of nodes and edges, forming a data network containing more than 2,000 data nodes and more than 5,000 associated edges, enabling intelligent search and in-depth analysis of data.

[0016] Preferably, the quantum dot spectral detection chip in the raw material detection module greatly improves the sensitivity of detecting trace elements by modifying gold nanoparticles on the surface of quantum dots and utilizing the surface plasmon resonance effect, and can detect trace elements at a content of one billionth; during detection, the raw material sample does not require complex chemical pretreatment, and detection is completed directly in the microfluidic channel within the chip. Compared with the traditional inductively coupled plasma mass spectrometry detection method, the speed is three times faster, and one detection only takes 100 to 120 seconds.

[0017] Preferably, the three-dimensional dynamic model of the production process monitoring module sets boundary conditions such as mold shape and extrusion speed based on the Navier-Stokes equations of fluid mechanics and the Hill yield criterion of metal plastic deformation when constructing a model of the aluminum profile extrusion link; during the operation of the model, new data from the equipment sensor is received every 5 seconds to update the model status; after actual verification, the model predicts the profile outlet temperature with an error within plus or minus 3 degrees Celsius, and can issue an early warning 4 hours before the temperature becomes abnormal and may cause quality problems, so as to adjust the coolant flow and temperature settings of the cooling equipment in advance to avoid quality problems.

[0018] Preferably, when performing data sharing analysis, the quality analysis and early warning module uses a hybrid encryption technology based on RSA and AES to encrypt the data, combined with the secret sharing protocol in secure multi-party computing, to ensure that the data uploaded by each factory area will not be cracked during the combined analysis process; after testing, the possibility of data leakage is less than one in a billion; by integrating the historical production data of 5 factories for joint analysis, the accuracy of the quality prediction model is improved.

[0019] Preferably, the data fusion algorithm of the finished product inspection module first performs discrete wavelet transform on the signal obtained by the terahertz time-domain spectrometer to extract frequency domain features, performs three-dimensional voxel processing on the image generated by the microfocus X-ray imager to obtain geometric features, and performs principal component analysis on the spectral data obtained by the laser induced breakdown spectrometer to extract element features; then, through the multi-head attention mechanism, different weights are assigned to the three features according to the defect detection requirements, dimensional measurement accuracy requirements and component analysis requirements, and they are fused into a whole containing a 128-dimensional feature vector; after actual detection and verification, this method can greatly improve the accuracy of defect detection inside and on the surface of aluminum profiles.

[0020] Preferably, the distributed storage system of the data management module adopts a consistent hashing algorithm to determine the data storage nodes; when new quality data is generated, the hash value of the data is first calculated, and then the node where the data should be stored is quickly determined based on the node distribution on the hash ring, and then the data is stored using parallel writing technology; the system utilizes the parallel processing capabilities of multi-core processors, and the data storage confirmation time is less than 0.3 seconds. It can process 1,500 quality data records per second, which fully meets the data storage needs of 220 aluminum profile products produced per minute.

[0021] Preferably, the data network of the data management module is based on the data of the entire production process. The named entity recognition technology in natural language processing is first used to extract key information from the data, and then the node connection weights of the data network are continuously updated through the graph convolutional neural network algorithm. This network establishes a causal relationship between 236 types of data such as raw material purity, melting temperature, and profile strength. In actual use, the accuracy of this network in predicting quality problems is improved, and it can give specific production process improvement suggestions such as "When the magnesium content in the raw material is reduced by 0.1%, it is recommended to reduce the extrusion speed by 5% and increase the aging temperature by 10°C."

[0022] Preferably, the raw material detection module, production process monitoring module, quality analysis and early warning module, and finished product detection module all establish real-time connections through a data transmission middleware and a data management module based on the message queue telemetry transport protocol (MQTT); the data generated by each module is first temporarily stored in the message queue, and then the data management module uses a priority queue algorithm to read and store the data in sequence according to the generation time and importance of the data; when the data of any module changes, the data management module pushes the updated data to other related modules in a timely manner through a publish-subscribe model, thereby realizing data traceability and collaborative analysis of the entire production process from raw material procurement to finished product warehousing.

[0023] Preferably, after the quality analysis and early warning module issues an early warning, it will automatically retrieve the production process data, equipment maintenance records and raw material batch information of the past week from the data management module, and input them into the three-dimensional dynamic model of the production process monitoring module; in a virtual environment, the system simulates 8 different production process schemes, such as adjusting the extrusion temperature up or down 5 degrees Celsius, the stretching speed up or down 10%, and the mold gap up or down 0.1 mm. By comparing the simulation results with the dimensional tolerances, mechanical performance indicators, and surface quality requirements in the quality standards, the multi-objective optimization algorithm is used to select the optimal adjustment scheme; after actual verification, after adjusting the production process in this way, the product qualification rate is improved, and the waste of raw materials and time costs caused by blindly adjusting the process in actual production are greatly reduced.

[0024] A method for controlling the production quality of aluminum profiles, the method comprising:

[0025] Before production, the raw material detection module uses quantum dot spectral detection chips and atomic force microscopes to detect the element content and surface structure of aluminum raw materials, and compares the detection data with the data stored in the raw material quality standard library. Only qualified raw materials can enter the production process. At the same time, the detection data, raw material supplier name, batch number and other information are uploaded to the data management module for storage; during the production process, the production process monitoring module collects equipment operation data 10 times per second, updates the three-dimensional dynamic model after Kalman filter processing, and the quality analysis and early warning module analyzes the data in real time. When the data is abnormal, the blockchain evidence storage function is activated to record the timestamp, equipment number, parameter value and other information of the abnormal data, and issue low, medium and high level early warnings, and adjust the plan according to the simulation verification of the three-dimensional dynamic model. Production process, such as adjusting the power of the heating furnace, replacing worn molds, etc.; after production, the finished product inspection module uses a combined terahertz time-domain spectrometer, a microfocus X-ray imager and a laser-induced breakdown spectrometer to conduct a comprehensive inspection of the finished products. Qualified finished products are put into storage and upload quality data, including dimensional measurements, mechanical properties test results, surface element content, etc. Unqualified finished products are traced through the data network of the data management module to trace information such as raw material batches, production equipment numbers, specific process parameter settings, analyze the reasons for non-compliance and form an improvement report containing improvement measures and responsible persons; every month, the historical data stored in the data management module is used to optimize the parameters of the quality prediction model through the federated learning algorithm, update the standard range of production process parameters, and continuously improve the quality of aluminum profile production.

[0026] (3) Beneficial effects

[0027] 1. During raw material testing, aluminum raw material samples interact with quantum dot materials within the chip's microfluidic channels, generating fluorescence signals upon excitation with specific ultraviolet light, enabling precise analysis of trace element composition. Atomic force microscopes scan the raw material surface with nanometer-level precision, revealing minute defects. This detection method, requiring no complex sample pre-processing, is fast and accurate, eliminating unqualified raw materials from entering the production line at the source, avoiding resource waste and quality risks in subsequent production, and laying a solid foundation for high-quality production. Furthermore, during production operations, the model is updated in real time based on actual parameters, creating a "digital twin" for the production line. Whether it's temperature fluctuations during smelting or pressure changes during extrusion, the system can promptly capture these fluctuations and compare them with standard process parameters. Once an abnormal trend emerges, the system issues an early warning, allowing staff to adjust the process in a timely manner based on the warning, effectively preventing quality issues caused by abnormal equipment parameters, ensuring stable production operations, and minimizing losses caused by production interruptions.

[0028] 2. Through the federated learning framework, each factory achieves collaborative data analysis while ensuring data privacy and security. Edge computing nodes perform preliminary data processing and encrypt transmission of key parameters to avoid the risk of data leakage. When production data shows anomalies, the system quickly initiates blockchain evidence storage and issues warnings based on preset rules. At the same time, it automatically retrieves relevant data, simulates various process adjustment plans in a virtual environment, and selects the optimal strategy, making production adjustments more scientific and efficient, and significantly improving product quality stability.

[0029] 3. By using a fusion algorithm based on the attention mechanism, different test data are intelligently integrated to generate a comprehensive quality inspection map. Compared with a single detection method, this multimodal fusion detection method can more comprehensively and accurately evaluate the quality of finished products, ensuring that only aluminum profiles that meet high standards enter the market, and improving the company's product reputation and market competitiveness. Secondly, blockchain technology is used to ensure that the data is authentic, reliable, and cannot be tampered with, providing a solid basis for quality traceability. The graph neural network mines the potential correlations between data and builds a knowledge network of raw materials, processes, equipment, and finished product quality. Each module interacts with the data management module in real time, and data flows smoothly within the system. Based on the analysis results of the data network, the system can provide practical suggestions for process optimization, regularly optimize the quality prediction model and process standards every month, promote continuous improvement in the production quality of aluminum profiles, and help companies maintain their leading position in market competition. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention with reference to the accompanying drawings.

[0031] Figure 1 is an overall structural diagram of an embodiment of the present invention;

[0032] Figure 2 This is a flow chart of raw material detection in an embodiment of the present invention;

[0033] Figure 3 This is a flowchart of the production process monitoring process in an embodiment of the present invention;

[0034] Figure 4 This is a flowchart of the finished product inspection process in an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The embodiments of the present application provide an aluminum profile production quality control system and method to solve the problem in the prior art that data between different factory areas are independent of each other, forming information islands, making collaborative analysis and shared utilization of data impossible, and limiting the improvement of overall production quality. Through the federated learning framework, each factory area realizes collaborative data analysis under the premise of ensuring data privacy and security; the edge computing node preliminarily processes the data and encrypts the transmission of key parameters to avoid the risk of data leakage; when an abnormality occurs in the production data, the system quickly starts blockchain evidence storage and issues an early warning based on preset rules; at the same time, it automatically retrieves relevant data, simulates various process adjustment plans in a virtual environment, and screens out the optimal strategy, making production adjustments more scientific and efficient, and significantly improving product quality stability.

[0036] Aluminum profiles are widely used in many fields such as construction, automobiles, aerospace, etc. due to their excellent properties such as light weight, high strength and corrosion resistance. As the market's requirements for the quality of aluminum profiles continue to increase, building an efficient and accurate quality control system has become the key to improving the competitiveness of aluminum profile manufacturers. This article, based on actual production cases, deeply analyzes the complete implementation process of aluminum profile production quality control systems and methods, providing a reference for industry development.

[0037] Example: The technical solution in the embodiment of the present application is to solve the problem that the data between different factory areas are independent of each other, forming information islands, and cannot realize collaborative analysis and shared utilization of data, which limits the improvement of overall production quality. The overall idea is as follows:

[0038] In response to the problems existing in the prior art, the present invention provides an aluminum profile production quality control system, the specific contents of which are as follows:

[0039] (1) Quantum dot spectrum detection chip detection:

[0040] 1. Sample and chip preparation:

[0041] During the production of aluminum alloy profiles, the quality of raw materials directly affects the performance of the final product. To ensure the accuracy and reliability of testing, strict sample preparation is carried out on the aluminum raw materials. A high-precision CNC five-axis linkage machining center is used to accurately process the aluminum raw materials into cylindrical samples with a diameter of φ1.000±0.002mm and a thickness of φ1.000±0.002mm in accordance with the international standard ISO 286-2. After processing, the sample surface is tested using a surface roughness measuring instrument. The relationship between the test results and the production process requirements is used to determine whether the production raw materials are qualified, laying a good foundation for subsequent testing.

[0042] For the pretreatment of quantum dot spectral detection chips, every link is crucial. First, the chip is placed in deionized water with specific parameters for ultrasonic cleaning. The cleaning is maintained at a power of 40kHz, a temperature of 30°C, and a sound pressure level of 80dB for 15 minutes. The cavitation effect of ultrasound is used to effectively remove impurities, dust, and organic residues on the chip surface. Subsequently, the chip is transferred to a vacuum drying oven and dried for 2 hours at a vacuum degree of 10-3Pa, a temperature of 60°C, and a heating rate of 5°C / min to ensure that the chip surface is dry and clean, and to prevent moisture and other factors from interfering with subsequent coating and detection.

[0043] Then, the core-shell structure quantum dot material is coated on the chip surface by spin coating. During the coating process, the precise control of the film thickness has an important influence on the test results. According to the formula

[0044] Where d is the film thickness, V is the volume of the coating material, r is the chip radius, T is the current ambient temperature, T0 is the standard temperature, is the temperature change rate, and α, β, γ, and δ are the primary, secondary, time-temperature coupled, and time-squared thermal expansion coefficients of the material, respectively.

[0045] For example, assuming that the coating material volume V = 10 μL, the chip radius r = 5 mm, the current ambient temperature T = 25 ° C, the standard temperature T0 = 20 ° C, and the thermal expansion coefficient of the material α = 2.3 × 10 -5 / ℃, β=1.2×10 -7 / ℃ 2 , γ=5×10 -6 / (℃·s),δ=3×10 -8 / (℃ 2 ·s 2 )(Assumptions: Obtained in real time by environmental monitoring equipment), the film thickness was calculated to be approximately 0.127 μm.

[0046] In addition, to ensure coating uniformity, a centrifugal acceleration compensation algorithm is introduced, and the formula is as follows:

[0047]

[0048] Where k is the compensation coefficient, ω is the angular velocity of rotation, μ is the dynamic viscosity of the coating material, and ρ is the material density. is the radial velocity gradient, is the second-order gradient of radial velocity, is the radial gradient of density, ν L , η is the compensation adjustment coefficient, and This term takes into account the effect of the second-order derivative of radial velocity on the compensation velocity, ν LThe weight used to adjust the influence reflects the effect of the acceleration of the speed change during the spin coating process on the coating uniformity; The influence of the coupling relationship between material density change and speed change on the compensation speed is considered. η is used to adjust the weight of this coupling effect, reflecting the comprehensive influence of the interaction between material density change with position and speed change on coating uniformity. By reasonably adjusting ν L , η, can optimize the spin coating equipment parameters to ensure that the core-shell structure quantum dot material is evenly coated on the chip surface, laying the foundation for the accuracy of subsequent spectral detection.

[0049] For example, assuming that the compensation coefficient k = 0.1, the rotational angular velocity ω = 314 rad / s (corresponding to 3000 r / min), and the dynamic viscosity of the coating material μ = 5×10 -3 Pa·s, material density ρ=1200kg / m 3 , radial velocity gradient The compensation speed was calculated to be approximately 0.717 m / s, and the spin coating equipment parameters were precisely adjusted accordingly to ensure that the quantum dot material was evenly coated on the chip surface, providing a stable and reliable basis for subsequent spectral detection.

[0050] 2. Testing process:

[0051] After the sample and chip are prepared, the test phase begins. The prepared sample is accurately delivered to the chip's microfluidic channel at a stable flow rate of 0.200 μL / min through a high-precision microfluidic syringe pump (double-plunger pump structure). This microfluidic syringe pump uses advanced flow control technology to achieve precise delivery of small flows, ensuring stable flow of the sample in the channel. Subsequently, the chip is irradiated with ultraviolet light of specific parameters: a wavelength of 380.0 nm, a half-peak width of 5 nm, and an illumination intensity of 50.0 mW / cm 2 , spot diameter 1mm; ultraviolet light irradiates and excites the quantum dot material to generate a fluorescent signal, which is collected by an avalanche photodiode (APD) at a frequency of 10.00kHz and converted into an electrical signal.

[0052] In the signal acquisition process, in order to effectively remove noise and improve signal quality, a sliding average filtering method that integrates adaptive weights, wavelet transform and variational mode decomposition is used. The formula is as follows:

[0053]

[0054] Where, is the filtered signal, y i is the original signal, w i is the weight, t is the current time, t i is the i-th sampling point time, σ is the time constant, n is the window size, is the signal y i The wavelet transform coefficients at scale a and translation b are: is the signal y i The kth modal component after variational modal decomposition; this filtering method fully combines the advantages of various signal processing technologies, can adaptively adjust the weight according to the characteristics of the signal, effectively suppress noise interference, highlight useful signal characteristics, and provide high-quality signal data for subsequent data analysis.

[0055] 3. Data Analysis

[0056] The signal processing unit uses the least squares spectrum fitting algorithm based on regularization, Bayesian estimation, sparse representation and non-negative matrix decomposition to conduct in-depth analysis of the collected signals; assuming that the measured spectrum data is y i , the fitting function is Among them, y is the predicted value obtained by fitting function, a j is the fitting coefficient, φ j (x i ) is the basis function (pre-set function form, used to construct the basic structure of the fitting function. In spectral fitting, common basis functions such as Lorentz function and Gaussian function are used to describe the shape of the spectral peak. Taking Gaussian function as an example, its form is f(x) = a·exp(-((xb) 2 ) / (2c 2 )), where a, b, and c are parameters. a determines the peak height of the function, b is the center position of the function, and c controls the width of the function. In practical applications, the choice of basis function can be determined by preliminary fitting of the spectral data of known standard samples to observe which basis function form can better approximate the data; different basis function combinations can simulate complex spectral characteristics, x i is the independent variable, representing the measured data points of the sample), construct the error function:

[0057]

[0058] Where y i is the actual measurement data (e.g. in spectrum detection, y i is the spectral intensity actually measured at the i-th wavelength), λ is the regularization parameter, P(a j ) is the fitting coefficient a j Prior probability distribution, τ is the sparse constraint parameter, A is the sparse basis matrix, d=[a1,a2,…,a m ] T , is the constraint parameter of non-negative matrix decomposition, W and H are the basis matrix and coefficient matrix of non-negative matrix decomposition, respectively, Y is the measured spectrum data matrix, ||·|| Fis the Frobenius norm; the solution was obtained by iteratively using the alternating direction multiplier method (ADMM), Markov chain Monte Carlo (MCMC) method and multiplication update rule. The iterations were 10,000 times, and the results were saved every 10 times. Each element was measured 5 times and the average value was taken to determine the contents of 15 trace elements.

[0059] The value of the regularization parameter can be determined by cross-validation. The training data set is divided into multiple subsets, and different regularization parameter values ​​are used on different subsets to train and validate the model. By comparing the performance indicators of the model on the validation set (such as mean square error, mean absolute error, etc.), the regularization parameter value that optimizes the performance is selected. For example, a series of candidate values ​​can be selected in a reasonable range (such as 10-5 to 105) with logarithmic intervals, such as 10 -5 , 10 -4 , 10 -3 ,…,10 4 , 10 5 , perform cross-validation tests respectively and finally determine the optimal regularization parameters;

[0060] The determination of the sparsity constraint parameters can be combined with the Bayesian estimation principle. In the Bayesian framework, the sparsity constraint parameters can be regarded as a control of the prior distribution of the model parameters. Taking the Laplace prior distribution as an example, its probability density function is P(θ)∝exp(-λ de |θ de |), where λ de is the sparse constraint parameter, θ de is the model parameter, λ de The larger the value, the more the model parameters tend to take zero values, thus achieving a sparser representation. In actual operation, we can observe the effect of different λ by multiple experiments on the training data. de The model fits the data well under the value and the sparsity of the model. Select λ that can ensure the model fits the data well and make the model reach the appropriate sparsity. de value.

[0061] also, Is the core part of the error function, called the residual sum of squares, is the fitting function at x i The predicted value at , this formula calculates the sum of the squares of the difference between the actual value and the predicted value at each measurement data point, which is used to measure the degree of deviation between the fitting function and the actual data. The smaller the difference, the better the fitting effect.

[0062] For example, given the regularization parameter λ = 0.01, the sparse constraint parameter τ = 0.005, and the non-negative matrix factorization constraint parameter Constructing the error function

[0063] And satisfy the constraints a≥0, W≥0, H≥0.

[0064] The algorithm is solved iteratively through the alternating direction method of multipliers (ADMM), Markov chain Monte Carlo (MCMC) method and multiplication update rule. The number of iterations is set to 10,000, and the result is saved every 10 times. Taking the silicon element in this batch of aluminum raw materials as an example, in order to ensure the accuracy and reliability of the test results, it is repeatedly measured 5 times. The spectral fitting algorithm can fully consider the influence of various factors on the spectral data. Through regularization and sparse representation, the accuracy and stability of the fitting are effectively improved. At the same time, Bayesian estimation and non-negative matrix decomposition are used to better mine the information in the spectral data, providing strong support for the accurate analysis of the raw material composition.

[0065] (2) Atomic force microscopy detection:

[0066] 1. Equipment and sample settings:

[0067] As a high-precision surface morphology detection device, the atomic force microscope (AFM) plays an important role in the surface quality inspection of aluminum profile raw materials. Before use, the AFM is fully and strictly calibrated to ensure that the elastic coefficient of the probe cantilever beam is stable at 0.100N / m and the resonant frequency is calibrated to 300kHz. The calibration process uses professional calibration tools and standard samples. Through precise measurement and adjustment, it ensures that all parameters of the equipment are in the optimal state to meet the requirements of high-precision detection.

[0068] The prepared sample is precisely fixed to the center of the sample stage with a special conductive glue to ensure that the sample remains stable during the detection process to avoid affecting the detection results due to sample movement; the scanning range is set to 10.00μm×10.00μm, the tapping mode is used, the scanning speed is 1.00Hz, and the initial distance between the probe and the sample is set to 5.00nm; during the scanning process, due to the complexity of the sample surface morphology, the distance between the probe and the sample needs to be adjusted in real time to ensure the accuracy and safety of the detection; a PID feedback control system based on fuzzy adaptation and reinforcement learning is used, which can monitor the spacing error e and error change rate in real time. The Q-learning algorithm is used to automatically calculate the control parameter adjustment amount, and the control parameter K p , K i , K d Dynamic adjustment based on fuzzy reasoning of sample surface morphology:

[0069]

[0070] Where K p , K i , K dare the proportional coefficient, integral coefficient and differential coefficient respectively, are the mapping functions of the proportional parameter, the integral parameter, and the differential parameter, respectively. e is the spacing error. is the error change rate.

[0071] In the initial state, the spacing error e = 0.2nm, the error change rate The proportional coefficient K is obtained by calculation p Adjust to 0.12, integral coefficient K i Adjusted to 0.05, differential coefficient K d Adjust it to 0.03 and update the system parameters in real time to achieve precise control of the distance between the probe and the sample, ensuring that the AFM can stably and accurately collect the atomic force change signal on the sample surface.

[0072] 2. Data collection and processing:

[0073] During the scanning process, the AFM collects the atomic force variation signals between the probe and the sample surface at high resolution. The signal acquisition system has a resolution of 0.1nN, capable of capturing extremely small force variation information on the sample surface. After acquisition, an adaptive algorithm that integrates wavelet transform, median filtering, non-local mean denoising, and morphological filtering is first applied:

[0074]

[0075] in, is the denoised signal, WT is wavelet transform, MF is median filtering, and NLM is non-local means denoising;

[0076] Image enhancement uses a composite algorithm based on Retinex theory, multi-scale decomposition, and histogram matching to enhance image contrast and details. Using an improved SIFT algorithm, key point detection fuses Gaussian difference pyramid, Laplacian pyramid, and DoG-LoG hybrid features. Key point description combines rotation, scale, illumination invariance, and local structure tensor features to identify surface defects through a support vector machine (SVM) classification model (the kernel function is an improved radial basis function, and local adaptive bandwidth adjustment is introduced). Model parameters are optimized using a hybrid method based on particle swarm optimization and genetic algorithm (PSO-GA).

[0077] For example, first, the adaptive algorithm integrating wavelet transform, median filtering, non-local mean denoising and morphological filtering is used to preprocess the signal; for a certain section of the original signal y i , first perform wavelet transform WT(y i ), using the multi-resolution analysis characteristics of wavelet transform, the signal is decomposed into different frequency scales, thereby achieving a preliminary separation of noise and useful components in the signal; then median filtering MF(WT(yi )), by replacing the current data with the median of the data in the window, the impulse noise in the signal is effectively removed; then the non-local mean denoising NLM(MF(WT(y i This method uses the self-similarity of the signal to find similar local areas in the entire signal space for weighted averaging, further removing noise and retaining signal details; finally, morphological filtering (opening operation followed by closing operation) is performed to remove isolated noise points and fill tiny holes in the signal by corroding and dilating the signal to obtain the denoised signal.

[0078] Subsequently, a composite algorithm based on Retinex theory, multi-scale decomposition, histogram matching, and generative adversarial networks was used to enhance the image. Retinex theory, based on the characteristics of the human visual system, can effectively remove the effects of uneven illumination and enhance image contrast and detail. Multi-scale decomposition decomposes the image into different scales and processes each scale separately, thereby better preserving image detail information. Histogram matching adjusts the image histogram distribution to match the reference histogram, further enhancing the image contrast. The generative adversarial network learns the image feature distribution through adversarial training of the generator and discriminator to produce high-quality enhanced images. An improved SIFT algorithm was used for surface defect recognition. In the key point detection stage, Gaussian difference pyramid, Laplacian pyramid, and DoG-LoG hybrid features were integrated to improve the detection accuracy and stability of key points. In the key point description stage, rotation, scale, and illumination invariance, local structure tensor features, and depth features were combined for analysis using a support vector machine (SVM) classification model (the kernel function is an improved radial basis function, introducing local adaptive bandwidth adjustment and multi-kernel fusion). No obvious defects were found on the sample surface.

[0079] (3) Comprehensive judgment:

[0080] Comprehensive judgment of raw material testing is a key step in ensuring that only qualified raw materials enter the production process. It integrates the results of element content testing and surface image testing, and draws a final conclusion through scientific and rigorous judgment methods. A complex comprehensive judgment model is constructed based on the element content data of quantum dot spectral detection and the surface image data of atomic force microscopy. The element content judgment is based on uncertainty propagation, Monte Carlo simulation, Bayesian decision theory and interval analysis:

[0081]

[0082] Where x is the test value, x0 is the standard value, u is the measurement uncertainty, ρ xu is the correlation coefficient, δ0 is the threshold, P(x|D) is the posterior probability based on the measured data D, P this the decision threshold, [x L ,x U ] is a reasonable interval determined by interval analysis, and the posterior probability is calculated by Monte Carlo simulation combined with Bayesian theorem; surface image judgment fuses the structural similarity index (SSIM), local binary pattern (LBP) texture feature similarity, Fourier shape feature similarity, deep learning semantic feature similarity and graph neural network structure similarity to construct the weighted comprehensive similarity FSIM:

[0083] FSIM=ζ1×SSIM+ζ2×TSIM+ζ3×TFSIM+ζ4×DSSIM+ζ5×GNSIM;

[0084] Where SSIM is the structural similarity index, TSIM is the local binary pattern texture feature similarity, TFSIM is the Fourier shape feature similarity, DSSIM is the deep learning semantic feature similarity, GNSIM is the graph neural network structure similarity, ζ1, ζ2, ζ3, ζ4, and ζ5 are the weights of SSIM, TSIM, TFSIM, DSSIM, and GNSIM, respectively. When FSIM is lower than the threshold FSIM0, it is determined that there are defects on the surface. Only raw materials that meet both standards can enter the production process.

[0085] For the determination of element content, take a certain element in aluminum raw material as an example, assuming that its standard value x0 = 0.6%, the detection value x = 0.65%, the measurement uncertainty u = 0.02%, and the correlation coefficient ρ xu =0.1, threshold δ0 = 0.05; Monte Carlo simulation combined with Bayesian theorem to calculate the posterior probability P(x|D) = 0.8, the decision threshold P th =0.7; calculated And 0.8>0.7, which indicates that the deviation between the test value and the standard value is within an acceptable range, and based on the existing measurement data, it is concluded that the probability of the element content meeting the standard is high, so the element content is judged to be qualified.

[0086] For surface image judgment, the structural similarity index SSIM = 0.92, the local binary pattern texture feature similarity TSIM = 0.90, the Fourier shape feature similarity TFSIM = 0.91, the deep learning semantic feature similarity DSSIM = 0.93, and the graph neural network structure similarity GNSIM = 0.92 were obtained by calculation; the weights ζ1 = 0.2, ζ2 = 0.2, ζ3 = 0.2, ζ4 = 0.2, and ζ5 = 0.2 were set, and the weighted comprehensive similarity was calculated by the formula FSIM = ζ1×SSIM+ζ2×TSIM+ζ3×TFSIM+ζ4×DSSIM+ζ5×GNSIM, and the result was 0.916>FSIM0 (assuming FSIM0 = 0.9), indicating that the similarity between the sample surface image and the standard image was high, no obvious defects were found on the surface, and the surface image was judged to be qualified; based on the comprehensive element content and surface image detection results, the batch of raw materials was judged to be qualified as a whole and allowed to enter the production process, thus providing reliable quality raw material guarantee for subsequent production links.

[0087] 3. Implementation of production process monitoring:

[0088] (1) Construction of three-dimensional dynamic model:

[0089] In the production of aluminum profiles, extrusion and smelting are two key processes. The precise control of their process parameters has a decisive impact on product quality. By constructing a three-dimensional dynamic model, it is possible to simulate and predict the physical phenomena and changing trends in the production process in real time, providing strong support for production process optimization and quality control.

[0090] Based on the equipment design drawings, operation and process parameters, ANSYS software was used to build a three-dimensional dynamic model with over one million computing units. The extrusion process was based on the Navier-Stokes equation and the Hill yield criterion, taking into account the strain rate effect, temperature softening effect, grain growth effect, dynamic recrystallization effect and frictional heat effect. The plastic strain rate correction formula is:

[0091]

[0092] Where, is the plastic strain rate tensor (reflecting the change of plastic strain rate in the i and j directions), σ ij is the stress tensor, is the equivalent stress, γ is the material constant, n is the strain rate sensitivity index, T is the current temperature, T0 is the standard temperature, T m is the melting point of the material, m is the temperature softening index, is the grain growth rate, N0 is the initial number of grains, α is the grain growth influence coefficient, is the rate of change of dynamic recrystallization volume fraction, X0 is the initial dynamic recrystallization volume fraction, β is the dynamic recrystallization influence coefficient, μf is the friction coefficient, v s is the extrusion speed, k is the Boltzmann constant, θ is the friction heat influence coefficient, σ ij =C, C is the elastic tensor, D is the plastic constitutive function;

[0093] The smelting process combines the energy conservation equation, Stefan-Boltzmann law, discrete coordinate method, gray body radiation correction and turbulent heat transfer model, and the radiation heat transfer equation is:

[0094]

[0095] Where, is the negative value of the divergence of the radiation intensity I along the direction Ω, σ s is the scattering coefficient, ε is the material emissivity (characterizing the ability of the material surface to emit radiation, ranging from 0 to 1, the larger the value, the stronger the material's ability to emit radiation), σ is the Stefan-Boltzmann constant, which is 5.67×10~ 8 W(m 2 ·K 4 ),∫4π, For location The radiation intensity along the Ω′ direction at Φ(Ω′,Ω) is the scattering phase function, which describes the probability distribution of radiation scattered from the Ω′ direction to the Ω direction, h c is the convective heat transfer coefficient, T is the material temperature (unit: K), T ∞ is the ambient temperature; It is the radiation emittance of the material surface per unit area and unit solid angle, that is, the radiation energy emitted by the material itself into the surrounding space due to temperature. This formula is based on the Stefan-Boltzmann law and reflects the key influence of temperature on the emission of radiation energy. The higher the temperature, the stronger the radiation energy emitted by the material.

[0096] By comparing with actual production data, a hybrid optimization method based on multi-objective genetic algorithm-simulated annealing algorithm-particle swarm optimization (MOGA-SA-PSO) is used to minimize the model error E m , maximize the prediction accuracy P m , minimize the computation time T c and minimize energy consumption E 能耗 The objective function is:

[0097] minF=w1E m +w2(1-P m )+w3T c +w4E 能耗 Among them, w1, w2, w3, and w4 are weights, and the Pareto front solution is used to adjust the model parameters.

[0098] In the melting process, it is known that the material emissivity ε = 0.3 and the Stefan-Boltzmann constant σ = 5.67×10 -8 W / (m 2 ·K 4 ), scattering coefficient σ s =0.1, convection heat transfer coefficient h c =10W / (m 2 ·K), ambient temperature T ∞ =25℃=298K, the temperature at a certain location T=600℃=873K; according to the radiation heat transfer equation (assuming The radiation heat transfer is actually calculated by the discrete coordinate method (DOP), thereby accurately simulating the heat transfer and temperature distribution during the smelting process.

[0099] In the extrusion process, taking the production data at a certain moment as an example, a plastic strain rate calculation model is established by comprehensively considering various physical effects. The known stress tensor σ ij , equivalent stress Material constant γ = 0.01, strain rate sensitivity index n = 0.1, current temperature T = 450 ° C, standard temperature T0 = 400 ° C, material melting point T m =660℃, temperature softening index m=0.2, grain growth rate Initial number of grains N0 = 1 × 10 12 / m 3 , grain growth influence coefficient α=0.3, dynamic recrystallization volume fraction change rate Initial dynamic recrystallization volume fraction X0 = 0.1, dynamic recrystallization influence coefficient β = 0.4, friction coefficient μ f =0.1, extrusion speed v s =0.5m / s, Boltzmann constant k = 1.38×10 -23 J / K, temperature T = 450 + 273 = 723K, friction heat influence coefficient θ = 0.2; plastic strain rate correction formula (assuming The calculated plastic strain rate is about 1.368×10 -5 / s, providing data support for real-time understanding of material deformation status.

[0100] The plastic strain rate reflects the deformation rate of the aluminum profile material under the current complex working conditions, providing key data support for mold design, extrusion speed adjustment, etc. For example, if the calculated plastic strain rate is too high, it may mean that the mold wear is aggravated or the extrusion speed is too fast, and the process parameters need to be adjusted in time to avoid quality problems such as profile surface cracks and dimensional deviations.

[0101] In the melting process, it is known that the material emissivity ε = 0.3 and the Stefan-Boltzmann constant σ = 5.67×10 -8 W / (m2 ·K 4 ), scattering coefficient σ s =0.1, convection heat transfer coefficient h c =10W / (m 2 ·K), ambient temperature T ∞ =25℃=298K, the temperature at a certain location T=600℃=873K; according to the radiation heat transfer equation (assuming The radiation heat transfer is actually calculated by the discrete coordinate method (DOP), thereby accurately simulating the heat transfer and temperature distribution during the smelting process.

[0102] This equation describes the complex relationship between heat transfer and radiation during the smelting process. By solving and analyzing this equation, we can accurately understand the temperature distribution and heat changes of the molten aluminum during the smelting process. If the calculation results show abnormal heat transfer in a certain area, parameters such as heating power and stirring speed can be adjusted in a timely manner to ensure uniform composition of the molten aluminum and avoid quality problems such as segregation caused by uneven temperature.

[0103] (2) Data collection and processing:

[0104] The equipment is equipped with a variety of sensors, including fiber Bragg grating temperature sensors and piezoresistive pressure sensors, which collect key data from the production process in real time at a frequency of 20.00±0.10Hz. These sensors feature high precision, high sensitivity, and strong anti-interference capabilities, accurately capturing subtle changes in parameters such as temperature, pressure, and flow. The collected analog signals are converted to digital signals by a 24-bit high-precision ADC chip and transmitted to the data processing unit.

[0105] In the data processing unit, a fusion algorithm based on unscented Kalman filter (UKF), particle filter (PF), strong tracking filter (STF) and cubature Kalman filter (CKF) is used to process data; this fusion algorithm fully utilizes the advantages of each filter algorithm and is suitable for data processing requirements under different complex working conditions. The specific formula is:

[0106]

[0107] in, is the state estimation after fusion, They are the state estimation of UKF, PF, STF, and CKF algorithms respectively, and Ψ1, Ψ2, Ψ3, and Ψ4 are weights, which are determined by the adaptive weight adjustment algorithm based on fuzzy neural network-reinforcement learning.

[0108] Taking temperature data processing as an example, at a certain production stage, the UKF algorithm estimates the temperature data as 448°C, the PF algorithm estimates it as 452°C, the STF algorithm estimates it as 450°C, and the CKF algorithm estimates it as 449°C. Using the adaptive weight adjustment algorithm of the fuzzy neural network-reinforcement learning algorithm, the weights Ψ1 = 0.25, Ψ2 = 0.2, Ψ3 = 0.3, and Ψ4 = 0.25 are obtained. The fused temperature estimate is:

[0109] The processed data updates the three-dimensional dynamic model in real time, enabling the model to more accurately reflect the actual status of the production process and provide a reliable basis for quality monitoring and decision-making.

[0110] (3) Quality problem warning:

[0111] In the aluminum profile production process, a hybrid early warning model based on a fuzzy neural network-long short-term memory network-attention mechanism (FNN-LSTM-AM) has emerged to achieve accurate and timely early warning of quality issues. This model deeply integrates the advantages of three network architectures to effectively address the ambiguity, timing, and key feature differences of production parameters. The following details its construction process.

[0112] 1. Clarify the overall model architecture and warning trigger conditions:

[0113] The hybrid early warning model takes production parameters such as temperature and pressure as input, and outputs quality risk warning values ​​through collaborative processing by FNN, LSTM and AM. Strict early warning trigger conditions are set: when the temperature T>1.1T0 and the temperature change rate Or pressure |P-P0|>0.15P0 and pressure change acceleration When , the system will input the corresponding parameters into the model for calculation; among them, T0, P0, These are the standard reference values ​​for temperature, temperature change rate, pressure, and pressure change acceleration, which are determined based on the aluminum profile production process standards and historical stable production data. They accurately define the range of abnormal parameters and ensure that the model initiates early warning at key nodes.

[0114] 2. Constructing Fuzzy Neural Network (FNN):

[0115] 2.1. Select membership function:

[0116] Using the triangular membership function, continuous production parameters such as temperature and pressure are mapped to fuzzy sets such as "low", "medium" and "high". Taking temperature as an example, the triangular membership function formula is:

[0117]

[0118] Where μ A(T) is the membership of temperature T to the fuzzy set A (such as the "high temperature" set), and its value range is between 0 and 1. The closer the value is to 1, the higher the degree to which temperature T belongs to the fuzzy set; the closer the value is to 0, the lower the degree to which temperature T belongs to the set. In the model, it is used to quantify the fuzziness of the temperature parameter and convert the precise temperature value into a membership value that can be processed by fuzzy logic. T is the actual temperature value, and T low 、T mid 、T high They are the low temperature value, medium temperature value and high temperature value corresponding to the fuzzy set, which are determined by analyzing the parameter range corresponding to the qualified products in the historical production. For example, for the extrusion temperature of aluminum alloy profiles, after statistical analysis, T is set low =380℃、T mid =400℃、T high =420°C, achieving fuzzy processing of temperature parameters. Key parameters such as low, medium, and high temperature values ​​within the fuzzy neural network membership function can be determined through statistical analysis of extensive historical production data. For example, temperature data collected over a period of time during production can be subjected to statistical histogram analysis to identify the key distribution intervals. The boundary of the low-temperature interval, where temperature data is less frequently distributed, is designated as the low temperature value; the boundary of the middle interval, where data is more concentrated, is designated as the medium temperature value; and the boundary of the high-temperature interval, where data is less frequently distributed, is designated as the high temperature value. Appropriate adjustments can also be made based on the varying temperature requirements in actual production and expert experience.

[0119] 2.2. Determine the inference algorithm and rule base:

[0120] Using the Mamdani inference algorithm, expert experience and historical production data were combined to construct a fuzzy rule base. Rules such as "If the temperature is 'high' and the pressure is 'high', the quality risk level is 'high'" were formed. Through data mining of more than 500 groups of qualified and unqualified production samples, and revisions by industry experts, a rule base containing 27 core rules was finally formed. These rules comprehensively cover complex scenarios involving the coupling of multiple parameters such as temperature, pressure and their rate of change, and convert fuzzy inputs into preliminary quality risk assessment results.

[0121] 3. Build a long short-term memory network (LSTM):

[0122] The LSTM network efficiently processes time series data through forget gates, input gates, and output gates. Its core calculation formula is as follows:

[0123] Forget gate: f t =σ(W f ·[h t-1 ,x t ]+b f );

[0124] Input gate: i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0125] Output gate: o t =σ(W o ·[h t-1 ,x t ]+b o ), h t =o t tanh(Ct);

[0126] Where, f t is the output value of the forget gate at time t, ranging from 0 to 1; σ is the Sigmoid activation function; W f is the weight matrix of the forget gate; [h t-1 ,x t ] is the data input to the forget gate, b f is the bias term of the forget gate, which is used to adjust the baseline value of the forget gate output;

[0127] i t is the output value of the input gate at time t, ranging from 0 to 1; W i is the weight matrix of the input gate; b i It is the bias term of the input gate, which has a similar function to the bias term of the forget gate and is used to adjust the benchmark of the input gate output. is the candidate cell state generated at time t, which is obtained by processing the input data through the tanh activation function, W C is the weight matrix used to calculate the candidate cell state, b C is the bias term for calculating the candidate cell state;

[0128] o t is the output value of the output gate at time t, ranging from 0 to 1; W i is the weight matrix of the output gate; b o The bias term of the output gate has a similar function to the bias term of the forget gate and is used to adjust the output benchmark of the output gate. t is the output of the hidden layer at time t, which is the output o of the output gate t And the cell state C after tanh activation function processing t Multiply to get;

[0129] For long short-term memory networks, the weight matrix can be initialized using the Xavier initialization method. This method initializes the weights based on the number of input and output neurons, so that the outputs of each layer of the network have similar variances at the beginning, which helps to accelerate the convergence of the network. For example, for the weight matrix Wih from the input layer to the hidden layer, the initialization value of its element Wij can be uniformly distributed. Random sampling in, where n in is the number of input neurons, n hidden is the number of hidden layer neurons; the bias term can be initialized to a zero vector.

[0130] In the aluminum profile production scenario, the fuzzy feature vector output by FNN is used as x t Input LSTM; the forget gate selects long-term information that is valuable for quality analysis based on historical data and current input, and filters out redundant data in the stable production stage; the input gate updates the key parameter information at the current moment to the cell state; the output gate extracts and outputs time series features that are critical for quality warning, effectively capturing the dependency and dynamic change trends of production parameters in long time series, such as the cumulative impact of temperature fluctuations for several hours on the final microstructure performance of the profile.

[0131] 4. Introducing the Attention Mechanism (AM):

[0132] The scaled dot product attention mechanism is used, and its calculation formula is:

[0133]

[0134] Among them, Q is the query vector, K is the key vector, and V is the value vector, and they are all obtained by the feature map after FNN-LSTM processing. Tis the transposed matrix of the key vector K, used for matrix multiplication with the query vector Q, and dk is the dimension of the key vector K. In the aluminum profile quality warning scenario, Q focuses on the current quality risk prediction target, while K and V contain full parameter features such as temperature, pressure, and rate of change. Softmax is the Softmax activation function, which converts the input vector into a probability distribution so that the output value ranges from 0 to 1, and the sum of all elements is 1. By calculating attention weights, key features such as high temperature and pressure mutations are given higher weights (actual testing shows that the weight of abnormal temperature features can reach 0.6-0.8), allowing the model to focus on parameters that play a dominant role in quality issues and improve the pertinence and accuracy of warnings. Among them, during initialization, the attention mechanism can initialize the attention weight matrix to an all-one matrix and then update it through the backpropagation algorithm during training. During training, the hyperparameters of the attention mechanism (such as the number of attention heads and the type of attention mechanism) can be adjusted to observe the performance changes of the model on the validation set, and the optimal hyperparameter settings can be selected to improve the model's ability to capture key features of quality issues.

[0135] 5. Model training and optimization:

[0136] The model weights are trained using the Adaptive Moment Estimation-Stochastic Gradient Descent-Second-Order Optimization (Adam-SGD-SO) hybrid optimization algorithm. The training is divided into two stages:

[0137] Pre-training phase: The Adam algorithm is mainly used, with the learning rate α set to 0.001, and the momentum parameters β1 to 0.9 and β2 to 0.999. The Adam algorithm is used to adaptively adjust the learning rate and momentum, quickly searching in the parameter space and initializing the network weights to a better area, thus shortening the exploration time in the early stages of training. This is usually completed within the first 20 training cycles.

[0138] Fine-tuning phase: Switch to the SGD algorithm (learning rate α = 0.0005) and work in conjunction with the second-order optimization algorithm (such as the BFGS algorithm); the SGD algorithm ensures the efficiency of the training process, and the second-order optimization algorithm uses the Hessian matrix information to accurately optimize the complex loss function (which comprehensively considers multiple indicators such as false alarm rate and false negative rate). Within 20-100 training cycles, the model training accuracy and generalization ability are balanced to make the model converge to the optimal weight configuration.

[0139] Through the above steps, the construction of the FNN-LSTM-AM hybrid early warning model is completed. This model can effectively deal with quality risks in the aluminum profile production process and provide strong technical support for production quality control.

[0140] For example, suppose that during a production process, the monitored temperature is T = 460°C (T0 = 400°C), and the temperature change rate is Pressure P = 120 MPa (P0 = 100 MPa), pressure change acceleration When the warning conditions are met, these parameters are input into the model; after calculation by the model, the output value is 0.8 (the warning threshold is set to 0.7), and the system immediately issues a warning signal; after receiving the warning, the operator adjusts the heating power and extrusion speed in time to avoid quality problems such as profile size deviation and surface bubbles caused by excessive temperature and abnormal pressure.

[0141] 4. Quality analysis and early warning implementation:

[0142] (1) Analysis of factors affecting quality:

[0143] First, we conduct an in-depth analysis of the many factors that affect the quality of aluminum profiles; in addition to raw material composition and production process parameters, factors such as equipment status and environmental conditions also have a significant impact on product quality.

[0144] In terms of equipment status, the degree of screw wear of the extruder will affect the material conveying and plasticizing effect. By installing vibration sensors and torque sensors at key locations on the screw, data such as vibration frequency and torque are collected in real time. A multivariate regression model is established to analyze screw wear and vibration and torque parameters:

[0145] W=a1V+a2T+a3, where W is the screw wear, V is the vibration frequency, T is the torque, and a1, a2, and a3 are regression coefficients; it is determined by least squares fitting; when the calculated screw wear exceeds the set threshold (such as 0.5mm), the system prompts that the screw needs to be repaired or replaced to prevent problems such as cavities inside the profile and increased surface roughness due to screw wear.

[0146] In terms of environmental conditions, workshop temperature and humidity significantly affect paint drying and adhesion. Temperature and humidity sensors were evenly distributed throughout the workshop to monitor environmental parameters in real time. A prediction model based on a support vector machine (SVM) was established by collecting paint drying time and adhesion test data under different temperature and humidity conditions. The model was constructed using relative humidity (RH) and temperature (T) as input variables and paint drying time (t) and adhesion grade (A) as output variables.

[0147] t = f1(RH, T), A = f2(RH, T); for example, when the relative humidity in the workshop is predicted to rise to 80% and the temperature is 20°C, the model predicts that the paint drying time will be extended to 8 hours (normal drying time is 4 hours) and the adhesion level may drop to level 2 (normal is level 5); at this time, the system recommends taking dehumidification measures and adjusting the paint formula or construction process to ensure product surface quality.

[0148] Similarly, corresponding detection models can be established for other states of the equipment and other environmental conditions according to detection needs, and targeted detection and early warning can be carried out. This is only an example to illustrate and will not be elaborated on.

[0149] (2) Construction of quality early warning model:

[0150] Based on the analysis results of quality influencing factors, a multi-dimensional quality early warning model is constructed; this model integrates machine learning and statistical methods and can accurately predict potential quality problems.

[0151] An ensemble learning method is used to fuse algorithms such as support vector regression (SVR), random forest regression (RFR) and gradient boosting decision tree (GBDT). For each quality indicator (such as tensile strength, dimensional accuracy, etc.), different models are trained separately and their prediction results are combined by weighted average. Taking tensile strength prediction as an example, let y SVR 、y RFR 、y GBDT The predicted values ​​of SVR, RFR, and GBDT models are w1=0.4, w2=0.3, and w3=0.3 respectively, so the fused predicted value y=0.4y SVR +0.3y RFR +0.3y GBDT .

[0152] During the model training process, production data from the past three years (a total of 5,000 groups) were used and divided into training, validation, and test sets in a ratio of 7:2:1. The model parameters were adjusted through cross-validation to minimize the mean square error of the model's prediction on the validation set. Ultimately, the mean square error of the tensile strength prediction of the fusion model on the test set was reduced, significantly improving the prediction accuracy compared to a single model. When the predicted value is lower than 95% of the standard value, the system issues a warning signal and provides an analysis report on factors that may cause quality problems, helping operators quickly locate the root cause of the problem and take appropriate measures.

[0153] 5. Implementation of finished product testing:

[0154] In the finished product inspection module, when the terahertz time-domain spectrometer, microfocus X-ray imager and laser-induced breakdown spectrometer are combined, the signal interference problem can be solved from both hardware and software aspects. In terms of hardware, the power supply of different instruments is isolated, an independent voltage-regulated power supply is used, and a filter is installed on the power input line to reduce power supply interference. For example, a ferrite magnetic ring filter is used to suppress the propagation of high-frequency interference signals along the power line. At the same time, the signal transmission line of the instrument is shielded, a double-layer shielded line is used, and the shielding layer is well grounded to reduce the impact of external electromagnetic interference on signal transmission. In terms of software, a signal filtering algorithm is used to pre-process the collected data. For example, for the data of the terahertz time-domain spectrometer, a wavelet filtering algorithm can be used to select the appropriate wavelet basis function and decomposition layer according to the frequency characteristics of the terahertz signal to remove noise and interference signals. For the data of the laser-induced breakdown spectrometer, an adaptive filtering algorithm can be used to adjust the filter parameters in real time according to the statistical characteristics of the signal to improve the signal-to-noise ratio of the signal.

[0155] To ensure the overall stability and detection accuracy of the combined instrument, each instrument must undergo individual performance testing and calibration before system integration. For example, a terahertz time-domain spectrometer requires frequency calibration. This can be done using a terahertz source of known frequency to ensure the accuracy of its frequency measurement. A microfocus X-ray imager requires imaging resolution calibration. A standard resolution test card is used for imaging testing, and imaging parameters are adjusted to achieve optimal imaging resolution. A laser-induced breakdown spectrometer requires wavelength and intensity calibration. Standard spectral lines are used to calibrate the wavelength, and standard samples of known concentrations are used to calibrate the intensity. After system integration, the overall performance of the combined instrument should be regularly verified. Standard samples can be used for multiple tests to statistically analyze the repeatability and accuracy of the test results. When performance degradation is detected, each instrument should be recalibrated and maintained promptly.

[0156] The terahertz time-domain spectrometer emits terahertz waves to penetrate the interior of the finished aluminum profile, and detects pores and delamination defects inside the finished aluminum profile based on the time delay and intensity difference between wave reflection and transmission; the microfocus X-ray imager scans the finished aluminum profile and reconstructs the image through the filtered back projection algorithm to obtain the dimensions of the finished aluminum profile; the laser-induced breakdown spectrometer emits laser pulses to bombard the surface of the finished product, exciting the surface material to produce a plasma spectrum, and by analyzing the intensity and position of the characteristic spectral lines in the spectrum, the element distribution on the surface is obtained. Traditional methods can also be used for detection, as follows:

[0157] (1) Mechanical properties testing:

[0158] In the inspection of finished products, mechanical properties testing is one of the key links; a universal material testing machine is used to conduct tensile tests on aluminum profiles to test their tensile strength, yield strength, elongation and other indicators.

[0159] The finished profiles are processed into standard specimens according to the test method specified in the national standard GB / T228.1-2021; the specimen shape is dumbbell-shaped, the gauge length is 50mm, the parallel section width is 10mm, and the thickness is the actual wall thickness of the profile; during the test, the specimen is installed on the fixture of the universal material testing machine and stretched at a constant speed of 5mm / min until the specimen breaks.

[0160] During the test, the tension and displacement data are collected in real time by sensors and transmitted to the computer for processing; according to the formula σ b =S0F b Calculate the tensile strength, where F b is the maximum tensile force when the sample is broken, S0 is the original cross-sectional area of ​​the sample; through the formula σ s =S0F s Calculate the yield strength, F s is the tensile force at which the specimen yields.

[0161] For example, taking aluminum alloy profile as an example, the original cross-sectional area of ​​the sample S0 = 10mm 2 , maximum tensile force F when breaking b =2500N, tensile force F when yield occurs s =2000N, then the tensile strength σ b =2500 / 10=250MPa, yield strength σ s =2000 / 10=200MPa, which meets the standard requirements of this type of profile: the tensile strength is not less than 205MPa and the yield strength is not less than 170MPa; the elongation is calculated by measuring the gauge length L1 of the sample after fracture according to the formula δ=(L1-L0) / L0×100%, where L0=50mm is the original gauge length of the sample. If L1=60mm, the elongation δ=(60-50) / 50×100%=20%, which meets the standard requirements.

[0162] The tensile strength and yield strength are calculated through curve analysis. For example, the tensile strength test value of a batch of profiles is 210MPa and the yield strength is 180MPa, which meets the quality standard requirements of aluminum alloy profiles.

[0163] The hardness test is carried out using a Brinell hardness tester or a Vickers hardness tester. Multiple test points are selected on the surface of the profile for measurement. Taking the Brinell hardness tester as an example, a carbide ball with a diameter of 10mm is pressed into the profile surface with a test force of 3000kg. After holding for 10 seconds, the test force is removed and the indentation diameter is measured. The Brinell hardness calculation formula is used to calculate the indentation diameter. (Where F is the test force, D is the indenter diameter, and d is the indentation diameter) the hardness value is calculated; the hardness test results of different parts of this batch of profiles show that the hardness value is between 80-90HB, with good uniformity, meeting the product quality requirements.

[0164] (2) Dimensional accuracy detection:

[0165] After the production of finished aluminum alloy profiles is completed, dimensional accuracy is one of the important indicators to measure whether the product meets the design requirements. The finished products are inspected using a high-precision three-dimensional coordinate measuring machine equipped with an advanced laser interferometer and a high-precision grating scale, with a measurement accuracy of up to ±0.001mm. Taking a batch of aluminum profiles with a length of 6000mm and a rectangular cross-section (size of 80mm×40mm) as an example, the key dimensions of the profiles, such as length, width, height, and wall thickness, are measured in accordance with the requirements for dimensional deviation in the national standard GB / T6892-2015.

[0166] When measuring the length, place the profile on the workbench of the three-dimensional coordinate measuring machine and use its automatic positioning system to measure the profile at both ends and in the middle. A total of three measurement values ​​are obtained: 5999.98mm, 5999.99mm, and 6000.01mm. The average value of the length dimension is calculated according to the formula:

[0167] Compared with the standard length of 6000mm, the dimensional deviation is 6000.00-6000=0mm, which meets the standard requirement of length dimensional deviation of ±1.5mm.

[0168] For cross-sectional dimensions, select five measurement points at different locations on the profile and measure the width and height respectively. Taking width measurement as an example, the five measured values ​​are 79.99mm, 80.01mm, 80.00mm, 79.98mm, and 80.02mm respectively. Calculate the average value of the width dimension:

[0169] The dimensional deviation is 80.00-80=0mm, which meets the standard requirement of ±0.3mm in width. Through precise measurement and strict calculation of multiple key dimensions, the dimensional accuracy of the finished aluminum profile is comprehensively evaluated to ensure that the product meets the design and use requirements.

[0170] (3) Surface quality inspection:

[0171] Surface quality directly affects the appearance and subsequent performance of aluminum profiles. Finished product surfaces are inspected using a machine vision inspection system consisting of a high-resolution industrial camera, a ring light source, an image acquisition card, and image processing software. The industrial camera has a resolution of 5 megapixels and can clearly capture subtle defects on the profile surface. The ring light source uses uniform diffuse reflection technology to eliminate shadows and ensure uniform surface illumination.

[0172] During the inspection process, the aluminum profile passes through the inspection area at a speed of 1m / s, and the industrial camera captures images at a frequency of 20 frames / second. Using image processing software, the captured images are first grayscaled to convert the color images into grayscale images to facilitate subsequent feature extraction and analysis. Then, a convolutional neural network (CNN) algorithm based on deep learning is used to identify defects in the images. This algorithm has been trained with a large number of image samples containing different types of surface defects (such as scratches, bubbles, pits, etc.) and can accurately identify various surface defects.

[0173] Taking one of the profiles as an example, a slight scratch about 5mm in length and 0.2mm in width was found on the surface. According to the surface quality standards formulated by the company, slight scratches less than 10mm in length and less than 0.3mm in width can be repaired by polishing or other methods without affecting the performance of the product and then judged as qualified products. If the defect exceeds the standard range, it will be judged as an unqualified product. Through the efficient detection and accurate judgment of the machine vision inspection system, surface quality problems can be discovered and dealt with in a timely manner, thereby improving the appearance quality and market competitiveness of the product.

[0174] 6. Data Management Implementation

[0175] (1) Data storage and backup

[0176] During the operation of the aluminum profile production quality control system, a large amount of data is generated, including raw material inspection data, production process monitoring data, finished product inspection data, etc. To ensure the security and integrity of the data, a distributed storage system was established. The Ceph distributed storage architecture is used to store data in a dispersed manner across multiple storage nodes, and data is transmitted and synchronized between each node through a high-speed network. This architecture has the characteristics of high reliability, high scalability and high performance, and can meet the needs of large-scale data storage.

[0177] At the same time, a strict data backup strategy has been formulated; a full backup is performed at 2 a.m. every day, and all data generated that day is completely backed up to an off-site data center; incremental backups are performed every hour, and only data that has changed since the last backup is backed up; backup data is protected by encryption technology, and the encryption algorithm uses AES-256 to ensure the security of data during storage and transmission; through regular data backup and encryption protection measures, data loss and leakage are effectively prevented, providing a solid guarantee for the company's data security.

[0178] (2) Data Analysis and Mining

[0179] Utilize big data analysis technology to conduct in-depth analysis and mining of stored data to discover the underlying information and patterns in the data and support the company's production decisions. Build a data analysis platform using the Hadoop and Spark big data processing frameworks. Hadoop is used to store and manage large-scale data, while Spark is used for efficient data processing and analysis.

[0180] By analyzing raw material testing data, we can establish a correlation model between raw material quality and finished product quality, identify raw material factors that have a greater impact on finished product quality, and thus optimize raw material procurement strategies. For example, analysis has found that the fluctuation in the content of a certain trace element in a batch of raw materials is correlated with changes in the mechanical properties of the finished product. Based on this, companies can strictly control the content range of this element in subsequent procurements.

[0181] Mining production process monitoring data can identify the relationship between key process parameters in the production process and product quality, thereby optimizing the production process. By analyzing the impact of extrusion temperature, speed and other parameters on profile dimensional accuracy and surface quality, the optimal process parameter combination can be determined to improve production efficiency and product quality. In addition, machine learning algorithms such as support vector machines (SVM) and random forests are used to train historical data and establish quality prediction models to predict product quality trends in advance, take preventive measures in a timely manner, and reduce the defective rate.

[0182] (3) Data sharing and interaction

[0183] In order to realize data sharing and interaction between departments within the enterprise and between the enterprise and its partners, a data sharing platform based on Web services has been established; this platform adopts the RESTful API interface specification and has good compatibility and scalability; the production department, quality inspection department, R&D department, etc. within the enterprise can obtain the required data in real time through this platform, realizing timely information communication and collaborative work.

[0184] For example, after the quality inspection department uploads the finished product inspection data to the platform, the production department can immediately view the inspection results. If quality problems are found, the production process can be adjusted in a timely manner; the R&D department can use this data to develop and improve new products, and improve product quality and performance; at the same time, the company can also exchange data with suppliers, customers and other partners to achieve collaborative management of the supply chain; suppliers can understand the company's raw material needs and quality requirements through the platform and provide standard-compliant raw materials in a timely manner; customers can query the product's production process and quality inspection data to enhance their trust in the product.

[0185] Finally, it should be noted that the above embodiments are merely examples for the purpose of illustrating the present invention and are not intended to limit the embodiments. Those skilled in the art will readily appreciate that other variations or modifications based on the above description are possible. It is not necessary and impossible to provide an exhaustive list of all embodiments. However, obvious variations or modifications arising therefrom remain within the scope of protection of the present invention.

Claims

1. A quality control system for aluminum profile production, characterized in that: The control system includes: The raw material detection module is used to detect the chemical composition and surface microstructure of aluminum raw materials, and compare them with the preset quality standard threshold to determine whether the aluminum raw materials are qualified; The production process monitoring module is used to build a virtual model of the production line, collect operating parameters and internal status data of production equipment, update the virtual model and predict quality risks; The quality analysis and early warning module is used to analyze operating parameters and status data, build a quality prediction model based on historical data from multiple plant areas, and provide evidence and early warning for parameters that exceed the quality control range; Finished product inspection module is used to detect internal defects, dimensional accuracy and surface chemical composition of finished aluminum profiles using a combination of multiple inspection technologies, and compare them with the finished product quality standards to determine whether they are qualified; The data management module is used to store data from the entire production process and build a data association network for intelligent data retrieval and in-depth analysis.

2. The aluminum profile production quality control system according to claim 1, characterized in that: The raw material detection module uses a combination of spectral detection technology and microscopic observation equipment to detect the chemical composition and surface microstructure of aluminum raw materials. The specific contents are as follows: Among them, the spectral detection technology uses a microfluidic chip to integrate a quantum dot spectral sensor to detect the chemical composition of trace aluminum raw material samples; Microscopic observation equipment is used to obtain microscopic defect data on the surface of raw materials.

3. The aluminum profile production quality control system according to claim 1, characterized in that: The production process monitoring module builds a virtual model based on digital twin technology, collects physical parameters of production equipment including temperature, pressure, speed and vibration amplitude through sensors, and uses tomography equipment to detect defects in internal components of production equipment; The construction process of the virtual model is as follows: The design drawings and operating parameters of the production equipment serve as the basic information for building the virtual model; Install sensors on production equipment to collect real-time physical parameters during the operation of production equipment; Using modeling technology and based on the principles of finite element analysis, the entire aluminum profile production process was broken down and a three-dimensional dynamic model was constructed. When building the extrusion process model in the production process, boundary conditions were set based on the principles of metal flow and deformation to simulate the production process of the extrusion process. The collected data is processed to remove noise and interference, filter out valid data, and then input the processed valid data into the 3D dynamic model in real time to update the model status and synchronize the virtual model with the operating status of the actual production equipment. By comparing the operating status in the virtual model with the pre-set standard process parameter range, quality problems in the production process are predicted based on the comparison results, and the operating status is simulated according to the preset adjustment strategy. By comparing the simulation results with the quality standards, the optimal adjustment plan is selected using a multi-objective optimization algorithm.

4. The aluminum profile production quality control system according to claim 1, characterized in that: The quality analysis and early warning module performs multi-factory data security sharing and intelligent early warning through the following steps. The specific steps are as follows: Deploy edge computing nodes at the production site to perform preliminary cleaning and feature extraction on the raw data collected in real time by production equipment; After completing the training of the quality prediction model locally, each factory uploads the encrypted updated values ​​of the quality prediction model parameters. The system uses a secure aggregation protocol to perform weighted fusion of the parameters of each quality prediction model to generate a global quality prediction model. Homomorphic encryption technology is used during data transmission, and Gaussian noise is added for differential privacy protection; By presetting the fluctuation threshold of parameters, when the edge node detects that the production equipment parameters exceed the threshold, it immediately triggers the automatic recording instruction, records the time of the abnormality, equipment status, and process parameters, and links the recorded data to the distributed ledger through the hash algorithm; After the early warning is triggered, the digital twin model is called for virtual verification to evaluate the effects of different adjustment plans, select the optimal plan, generate process adjustment instructions, and synchronize them to the production equipment.

5. The aluminum profile production quality control system according to claim 1, characterized in that: The specific process of combining multiple detection technologies in the finished product detection module is as follows: Preprocessing of detection data obtained by different detection technologies; Using image synthesis algorithm, we assign weights to different detection data according to the characteristics of each detection technology; The processed multi-source data is fused to generate a comprehensive evaluation image; then the comprehensive evaluation image is compared with the preset finished product quality standard image to determine whether the finished aluminum profile is qualified.

6. The aluminum profile production quality control system according to claim 5, characterized in that: The detection technology in the finished product detection module includes a terahertz time-domain spectrometer, a microfocus X-ray imager, and a laser-induced breakdown spectrometer, and the three are combined together in an optical coaxial calibration manner; Among them, the terahertz time-domain spectrometer emits terahertz waves to penetrate the interior of the finished aluminum profile, and detects the pores and delamination defects inside the finished aluminum profile based on the time delay and intensity difference between wave reflection and transmission; The micro-focus X-ray imager scans the finished aluminum profile and reconstructs the image using the filtered back-projection algorithm to obtain the dimensions of the finished aluminum profile. The laser-induced breakdown spectrometer emits laser pulses to bombard the surface of the finished product, exciting the surface material to produce a plasma spectrum. By analyzing the intensity and position of the characteristic spectral lines in the spectrum, the element distribution on the surface is obtained.

7. The aluminum profile production quality control system according to claim 1, characterized in that: The data management module adopts a blockchain distributed storage architecture for data storage, and sets up a number of storage nodes that are connected to each other to form a ring network; When new quality data is generated, the hash value is calculated using the consistent hashing algorithm. The corresponding storage node is found based on the node distribution on the hash ring. Then, the parallel writing technology is used to store the data in multiple nodes at the same time. When storing data, a practical Byzantine fault-tolerant consensus algorithm is used, and the data is encrypted using the SHA-256 hash algorithm to form a chain data structure.

8. The aluminum profile production quality control system according to claim 7, characterized in that: In terms of data association analysis, the data management module uses the graph neural network algorithm to convert the data of the entire production process into a data association network composed of nodes and edges; Use named entity recognition technology in natural language processing to extract key information, and then use the graph convolutional neural network algorithm to continuously update the node connection weights to extract the correlation data between raw material characteristics, process parameters, equipment status and finished product quality; Based on the data association network, intelligently retrieve data and deeply analyze the relationship between production factors.

9. A method for controlling the production quality of aluminum profiles based on the system according to any one of claims 1 to 8, characterized in that: The control method includes the following steps: Before aluminum profile production, aluminum raw materials are tested. If the raw materials meet the quality standards, they will enter the subsequent production process; if not, the aluminum raw materials will be rejected and the test data will be uploaded; During the production process, the operating parameters and internal status data of production equipment are collected in real time, the virtual model of the production line is updated, and the real-time data is analyzed. The quality prediction model constructed by combining data from multiple plant areas performs in-depth predictions. If the parameters are within the preset range, production continues; if they are outside the range, the early warning information is recorded and the equipment maintenance or process adjustment process is triggered. At the same time, the effectiveness of the adjustment plan is pre-verified through the virtual model. After aluminum profile production is completed, the finished product is fully inspected. If the finished product is qualified, it is packaged and stored, and the finished product quality data is uploaded to the data management module. If it is unqualified, the unqualified products are classified and recorded, and the production process data is traced through the data association network. The reasons for the failure are analyzed, and improvement measures are taken. The improvement plan is used to optimize the quality prediction model. Regularly conduct joint analysis and processing of quality data from multiple factories, optimize the quality prediction model, and synchronize the updated model to each production node.

10. The aluminum profile production quality control method according to claim 9, characterized in that: During the process adjustment process, the adjustment plan is pre-verified through the virtual model of the production line to simulate the parameter changes of the production process after the adjustment; When tracing production process data through the data association network, the association analysis function is used to sort out the relationship between various factors and quality issues from a large amount of production data and conduct quality problem analysis; Throughout the entire production process, it is based on distributed storage and data association network technology.

Images