A metal component quality deviation prediction method based on edge computing

Abstract

This invention relates to the field of metal component technology, and more particularly to a method for predicting the quality deviation of metal components based on edge computing. This method constructs an edge-end sensing array comprising a high-speed infrared thermal imager and an emissivity detection unit to synchronously acquire spatiotemporal data of the cooling process to characterize the temperature field evolution and radiation characteristic gradient. It dynamically compensates the data by combining process and operating parameters to eliminate water film and environmental interference. A thermo-mechanical coupled field tomography reconstruction algorithm is applied to invert the internal phase transition structure-texture distribution map. Through spatiotemporal adaptive threshold segmentation and microstructure dynamics rule filtering, it accurately identifies regions of true phase transition inhomogeneity and texture anomalies. Finally, a hyperbolic tangent-Avramie exponent joint model is used to calculate key microstructure parameters and quantitatively determine the quality deviation risk. This invention achieves high-fidelity real-time inversion from the macroscopic thermal field to the microstructure, effectively improving the accuracy of predicting the microstructure uniformity and performance consistency of metal components.

Description

Technical Field

[0001] This invention relates to the field of metal component technology, and in particular to a method for predicting the quality deviation of metal components based on edge computing. Background Technology

[0002] In modern steel and high-end metal material manufacturing, laminar cooling is a key process for controlling the microstructure and properties of metals after hot rolling. It directly affects the phase transformation behavior, grain size, texture evolution, and final anisotropy of mechanical properties of components. Especially for materials such as high-strength steel, pipeline steel, and silicon steel, which have strict requirements for microstructure uniformity and texture control, the non-uniformity of the temperature field during cooling can easily lead to spatial differences between the phase transformation initiation time and the termination temperature. This can cause microstructural defects such as mixed grains, banded structures, and abnormal texture strength, resulting in quality deviations such as yield plateau fluctuations, ear formation, or fracture tendency during tensile, stamping, or fatigue service. Traditional cooling control relies heavily on empirical models that set cooling rates and flow distributions, lacking the ability to perceive and provide feedback on the evolution of the surface and internal microstructure of metal components during actual cooling. This makes it difficult to accurately predict and control microscopic performance indicators such as texture anisotropy.

[0003] In recent years, with the development of infrared thermometry, edge computing, and digital twin technologies, online monitoring methods based on multi-source sensing have been gradually applied to the hot rolling cooling process. However, existing systems are mostly limited to the acquisition of a single parameter of the surface temperature field, failing to fully consider the impact of factors such as oxide film changes, water film coverage, and environmental radiation on temperature measurement accuracy, and lack a high-fidelity inversion mechanism from the macroscopic thermal field to the microstructure (such as grain orientation, phase transformation volume fraction, and texture intensity). Furthermore, due to the strong non-steady-state and high dynamic response characteristics of the laminar cooling process, the central control system often struggles to achieve millisecond-level data processing and closed-loop feedback, resulting in lag in microstructure control and failing to meet the requirements of precise control of the cooling path for high-end metal components.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method for predicting the quality deviation of metal components based on edge computing. This method aims to solve the technical problem that existing laminar cooling technology lacks real-time perception and high-fidelity inversion capabilities for microstructure evolution and is limited by the high latency of centralized processing, making it difficult to accurately predict and control the anisotropic quality deviation of phase transformation texture in metal components caused by non-uniform cooling.

[0006] To achieve the above objectives, this invention provides a method for predicting the quality deviation of metal components based on edge computing, the method comprising: S1. Construct a multi-dimensional sensing array facing the edge of the metal component to be cooled. The multi-dimensional sensing array includes a high-speed infrared thermal imager and a surface emissivity detection unit. Use the multi-dimensional sensing array to synchronously acquire the spatiotemporal dynamic data sequence of the component cooling process. S2. Obtain the rolling mill production process parameters and the real-time flow rate and operating conditions of the cooling water. Based on the production process parameters and real-time operating conditions, perform dynamic compensation processing on the spatiotemporal dynamic data sequence to eliminate the interference of water film residue and the temperature measurement deviation effect caused by changes in environmental heat radiation, and generate a compensated thermodynamic state feature set. S3. Based on the compensated thermodynamic state feature set, the phase change structure-texture distribution map inside the metal component is constructed by applying the thermo-mechanical coupled field tomography reconstruction algorithm. S4. Perform spatiotemporal adaptive threshold segmentation processing on the phase change structure-texture distribution map to obtain a candidate set of mixed crystal defects and texture abnormal regions. Then, according to the preset microstructure dynamics feature screening rules, the candidate set of defects is judged to determine the true phase change inhomogeneity region and the true texture anisotropy region. S5. For the real phase transformation non-uniform region and the real texture anisotropy region, the hyperbolic tangent-Avramie index joint model is used to calculate the phase transformation volume fraction parameter, texture intensity coefficient parameter and grain size distribution parameter corresponding to each region. When the phase transformation volume fraction parameter of any region exceeds the preset process target range or the texture intensity coefficient parameter exceeds the preset anisotropy critical threshold, it is determined that the metal component has a risk of microstructure and performance deviation that affects subsequent forming or service performance.

[0007] Optionally, step S2 specifically includes: S21. The cooling water impact pressure and water flow fluctuation data on the surface of the metal component are collected in real time by the fluid field monitoring network deployed below the cooling manifold. The pre-constructed water flow parameter-surface heat transfer coefficient mapping relationship table is queried according to the cooling water impact pressure and water flow fluctuation data to obtain the thermal interference compensation coefficient corresponding to the convective heat transfer capacity and film boiling probability at each component segment. S22. For the spatiotemporal data collected by each sensor channel, the thermal interference compensation coefficient is applied to the temperature correction processing of the infrared thermal image signal, and the compensation amount of the effective thermal radiation path is calculated in combination with the oxide film thickness parameter on the surface of the metal component. S23. The corrected spatiotemporal data is fused with the effective thermal radiation path compensation amount at the data level to generate a compensated thermodynamic state feature set including the latent heat release rate of phase change, cooling rate gradient and surface radiation temperature correction value.

[0008] Optionally, step S3 specifically includes: S31. The compensated thermodynamic state feature set is used as an input variable and transmitted to the pre-constructed thermo-mechanical coupling iterative tomography reconstruction model. The thermo-mechanical coupling iterative tomography reconstruction model has built-in prior physical constraints based on finite element phase transition simulation analysis. S32. By constructing an objective function to minimize the weighted mean square error between the actual measured temperature and the model simulation temperature, the spatial distribution matrix of the phase transition field and texture field inside the metal component is solved by iterative inversion, thereby generating a high-resolution phase transition structure-texture distribution map. S33. Perform spatiotemporal joint filtering on the phase transition structure-texture distribution map to effectively suppress artifacts caused by oxide scale shedding or water vapor interference during the tomographic reconstruction process while preserving the sharpness of grain boundary and phase interface features.

[0009] Optionally, step S4 specifically includes: S41. Based on the statistical distribution characteristics of the normal matrix region of the metal component in the phase change structure-texture distribution diagram, a dynamic segmentation threshold is determined. The dynamic segmentation threshold is spatiotemporally adaptively adjusted according to the local temperature variance and the target value of the final cooling temperature. S42. Based on the dynamic segmentation threshold, perform a multidimensional connected region labeling algorithm on the phase change organization-texture distribution map to obtain an initial set of candidate regions for phase change inhomogeneity and an initial set of candidate regions for texture anomalies. S43. For each candidate region in the initial set of candidate regions for phase transformation inhomogeneity, extract its phase transformation dynamics characteristic parameters and thermal hysteresis loop characteristic parameters; for each candidate region in the initial set of candidate regions for texture anomalies, extract its elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters; and match and verify the extracted characteristic parameters with the pre-constructed microstructure dynamics characteristic database. When the phase transformation dynamics characteristic parameters are in the unsteady-state range and the thermal hysteresis loop characteristic parameters are lower than a preset threshold, the candidate region is confirmed as a true phase transformation inhomogeneity region; when the elastic modulus directionality characteristic parameters show significant differences and the grain orientation concentration characteristic parameters exceed the preset texture critical value, the candidate region is confirmed as a true texture anisotropy region.

[0010] Optionally, the process of constructing the microstructure dynamics feature database includes: B1. Collect metal component samples with standard materials, prepare microstructure samples with known phase transformation type and texture type, and accurately measure the phase transformation kinetic characteristic parameters, thermal hysteresis loop characteristic parameters, elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters of each microstructure sample under simulated laminar flow cooling environment. B2. Based on the phase transformation volume fraction parameter and texture intensity parameter, establish a hierarchical mapping table of microstructure kinetic characteristic parameters, wherein ferrite-pearlite mixed crystals correspond to the first kinetic characteristic range, bainite-martensite mixed crystals correspond to the second kinetic characteristic range, weak texture regions correspond to the first orientation concentration range, and strong texture regions correspond to the second orientation concentration range. B3. Establish a correlation between the microstructure dynamics characteristic parameters in the hierarchical mapping table and the critical influence threshold of the anisotropy index of the mechanical properties of metal components to form a microstructure dynamics characteristic database for distinguishing the authenticity of tissue abnormalities and assessing their severity.

[0011] Optionally, the following steps, performed after step S5, may also be included: S6. Based on the spatial location and distribution density of the real phase transition inhomogeneity region and the real texture anisotropy region, generate a component microstructure performance heat map, and evaluate the performance deviation coefficient of the metal component based on the component microstructure performance heat map combined with the Hall-Page relationship model. S7. When the performance deviation coefficient exceeds the preset quality control threshold, an automatic process optimization analysis report is generated, which includes the spatial distribution characteristic parameters of tissue abnormalities, the severity level, and corresponding cooling process parameter adjustment suggestions.

[0012] Optionally, step S1 specifically includes: S11. Obtain the chemical composition parameters and phase transition critical point parameters of the metal component to be cooled. Based on the chemical composition parameters and phase transition critical point parameters, query the pre-built material-sensor configuration mapping table to determine the infrared thermal imager sampling frame rate and surface emissivity detection wavelength range suitable for this type of metal component. S12. Under the sampling frame rate and wavelength range, control the multi-dimensional sensing array to synchronously scan the component to be cooled, and capture the transient temperature response characteristics caused by the release of latent heat of phase change and the dynamic change sequence of surface radiation caused by texture change in real time through the edge computing node.

[0013] Optionally, the spatiotemporal dynamic data sequence is used to characterize the time-frequency evolution of the temperature field of a metal component under laminar cooling with a high heat transfer rate caused by the release of latent heat from phase change, as well as the gradient change of surface radiation characteristics caused by preferred grain orientation.

[0014] Optionally, the phase transformation structure-texture distribution map reflects the microscopic response characteristics of metallic materials under the coupling effect of temperature gradient and phase transformation stress.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an edge computing-based metal component quality deviation prediction program, wherein the edge computing-based metal component quality deviation prediction program, when executed by a processor, implements the steps of the edge computing-based metal component quality deviation prediction method as described above.

[0016] This invention provides a method for predicting quality deviations in metal components based on edge computing. The method constructs a multi-dimensional sensing array at the edge and integrates production process and cooling condition parameters to achieve high-precision, high-spatial-resolution synchronous sensing of the temperature field and surface radiation characteristics of metal components during laminar cooling, effectively overcoming temperature measurement deviations caused by water film interference and environmental radiation. Combining a thermo-mechanical coupled field tomography reconstruction algorithm with a microstructure dynamics feature discrimination mechanism, it can faithfully invert the internal phase transition structure and texture distribution from macroscopic thermal field data, accurately identifying regions of true phase transition inhomogeneity and texture anisotropy. Furthermore, a hyperbolic tangent-Avramy index joint model is introduced to quantitatively evaluate key microstructure parameters, enabling early prediction and quantitative assessment of quality deviation risks such as mixed crystals and texture anomalies. The overall edge computing architecture significantly improves the real-time performance of data processing and system response speed, providing a reliable basis for dynamic optimization of the cooling process and effectively improving the microstructure uniformity, performance consistency, and subsequent forming and service reliability of metal components. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an embodiment of the metal component quality deviation prediction method based on edge computing of the present invention. Figure 2 This is a flowchart illustrating another embodiment of the metal component quality deviation prediction method based on edge computing of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the edge computing-based method for predicting the quality deviation of metal components according to the present invention.

[0021] In one embodiment, the edge computing-based method for predicting the quality deviation of metal components includes: Step S1: Construct a multi-dimensional sensing array facing the edge of the metal component to be cooled. The multi-dimensional sensing array includes a high-speed infrared thermal imager and a surface emissivity detection unit. The multi-dimensional sensing array is used to synchronously acquire spatiotemporal dynamic data sequences of the component's cooling process. These spatiotemporal dynamic data sequences characterize the time-frequency evolution of the temperature field caused by the release of latent heat of phase change under laminar cooling at high heat transfer rates, as well as the gradient changes in surface radiation properties caused by preferred grain orientation.

[0022] The edge-end multi-dimensional sensing array can be a hardware combination deployed at the edge computing node of the laminar flow cooling site and integrating multiple physical quantity sensing units to simultaneously acquire the surface thermal state and radiation characteristics of metal components. It can be used to achieve in-situ high-precision acquisition of multi-dimensional physical information of the cooling process in environments with strong water vapor interference. In this embodiment, the edge-end multi-dimensional sensing array can collaborate with a high-speed infrared thermal imager and a surface emissivity detection unit, and ensure pixel-level matching of temperature field and radiation characteristic data through a spatiotemporal alignment mechanism. For example, the edge-end multi-dimensional sensing array can include, but is not limited to, thermal radiation sensing arrays, spectral reflectivity detection arrays, and polarized light imaging arrays. The high-speed infrared thermal imager can be an infrared radiation imaging device with high frame rate response capability, which can be used to record the time-frequency evolution characteristics of the temperature field caused by the release of latent heat from phase transitions. Furthermore, the high-speed infrared thermal imager can use an infrared detector to receive the spontaneous radiation energy of an object and generate a continuous thermal image sequence through photoelectric conversion and signal amplification. In an exemplary embodiment, the high-speed infrared thermal imager can be a cooled photon detector thermal imager, an uncooled microbolometer thermal imager, or a quantum well infrared detector thermal imager.

[0023] The surface emissivity detection unit can be a sensing module specifically designed for real-time measurement of the surface radiation characteristics of metal components, providing a radiation characteristic reference for correcting infrared thermometry. In this embodiment, the surface emissivity detection unit can calculate the dynamic emissivity value of the surface under current operating conditions through active light source illumination or passive multi-band radiation comparison. For example, the surface emissivity detection unit can employ a laser-induced fluorescence detection unit, a multispectral colorimetric detection unit, or a polarization difference detection unit. The spatiotemporal dynamic data sequence can be a multi-channel synchronous observation dataset containing timestamps and spatial coordinate information output by a multi-dimensional sensing array, which can be used to comprehensively describe the joint evolution of the temperature field and radiation characteristics of the metal component along the cooling path. In a specific embodiment, the spatiotemporal dynamic data sequence can include, but is not limited to, time-series thermogram sequences, radiation characteristic spectrum sequences, and multi-source fusion tensor sequences. The time-frequency evolution characteristics of the temperature field can be composite feature quantities extracted from the spatiotemporal dynamic data sequence, characterizing the temperature change with time and frequency, which can be used to reveal the intrinsic correlation between the latent heat release rate of phase transition and the nonlinearity of the cooling path. Furthermore, the time-frequency evolution characteristics of the temperature field can be characterized using short-time Fourier transform features, wavelet packet energy spectrum features, or empirical mode decomposition features. The gradient variation characteristics of surface radiation properties can be a quantitative indicator reflecting the degree of difference in the spatial distribution of the surface emissivity of a metallic component, and can be used to indicate the anisotropic radiation behavior caused by preferred grain orientation. In this embodiment, the gradient variation characteristics of surface radiation properties can be characterized using first-order spatial derivative features, second-order Laplace features, or structural tensor features.

[0024] Constructing a multi-dimensional sensing array at the edge of the metal component to be cooled can be achieved by arranging a high-speed infrared thermal imager and a surface emissivity detection unit along the laminar flow cooling zone from the inlet to the outlet, and completing optical axis calibration and clock synchronization configuration. Furthermore, the multi-dimensional sensing array at the edge of the metal component to be cooled can be constructed by using a rigid bracket for fixed installation and periodic calibration, using an adjustable gimbal for automatic focusing calibration, or using an embedded window installation with a purge device for fog prevention. This establishes a multi-dimensional sensing capability covering the entire cooling path and ensures the spatial continuity and temporal synchronization of data acquisition. Synchronously acquiring the spatiotemporal dynamic data sequence of the component cooling process can be achieved by triggering the multi-dimensional sensing array to continuously record thermal images and radiation characteristic data with a unified clock reference and storing them in timestamp alignment. In an exemplary embodiment, the synchronous acquisition of the spatiotemporal dynamic data sequence of the component cooling process can be achieved through microsecond-level synchronous acquisition based on hardware trigger signals, millisecond-level alignment based on software timestamp interpolation, or spatial synchronous acquisition based on encoder position signals. This allows for the acquisition of original multi-dimensional information reflecting the latent heat of phase change and grain orientation evolution, while avoiding data mismatch caused by asynchronous sampling.

[0025] Step S2: Obtain the production process parameters of the rolling line and the real-time flow rate parameters of the cooling water. Based on the production process parameters and real-time operating parameters, perform dynamic compensation processing on the spatiotemporal dynamic data sequence to eliminate the interference of water film residue and the temperature measurement deviation effect caused by changes in environmental thermal radiation, and generate a compensated thermodynamic state feature set.

[0026] The rolling mill production process parameters can be a set of set values ​​and measured values ​​describing the operating state of the hot rolling production line, used to provide the initial thermodynamic state and deformation history of the metal component before cooling. In this embodiment, the rolling mill production process parameters may include, but are not limited to, roughing mill parameters, finishing mill parameters, coiling mill parameters, etc. The real-time flow rate parameters of the cooling water can be online monitoring data of the valve opening, pump pressure, and nozzle flow rate in the laminar flow cooling system, which can be used to characterize the actual heat transfer intensity distribution. For example, the real-time flow rate parameters of the cooling water can be the total pipeline flow rate parameters, the branch flow rate parameters, or the single nozzle flow rate parameters. The compensated thermodynamic state feature set can be a data set that accurately reflects the true thermodynamic state of the metal component after dynamic compensation processing to eliminate environmental and water film interference, which can be used to ensure the physical consistency of the internal structure inversion results. In a specific embodiment, the compensated thermodynamic state feature set may include, but is not limited to, corrected temperature field characteristics, equivalent heat transfer coefficient characteristics, net radiation flux characteristics, etc.

[0027] Dynamic compensation processing of spatiotemporal dynamic data sequences can be achieved by using rolling mill production process parameters and real-time cooling water flow rate parameters to correct water film and environmental radiation interference in the original data through a compensation model. Furthermore, dynamic compensation processing of spatiotemporal dynamic data sequences can be implemented by calculating compensation values ​​pixel-by-pixel based on a physical mechanism model, outputting correction results end-to-end based on a machine learning model, or rapidly compensating using a lookup table method combined with real-time operating condition indexing. This can eliminate the influence of external interference factors on temperature measurement accuracy and generate a high-fidelity thermodynamic state feature set.

[0028] Step S3: Based on the compensated thermodynamic state feature set, the thermo-mechanical coupled field tomography reconstruction algorithm is applied to construct the phase transformation structure-texture distribution map inside the metal component. The phase transformation structure-texture distribution map reflects the microscopic response characteristics of the metal material under the coupling effect of temperature gradient and phase transformation stress.

[0029] The thermo-mechanical coupled field tomography reconstruction algorithm can be a computational model that inverts the surface thermodynamic state into the internal microstructure distribution based on the coupled equations of heat transfer and solid mechanics. It can be used to establish the mapping relationship between the macroscopic measurable thermal field and the microscopic phase transition structure and texture that cannot be directly observed. In this embodiment, the thermo-mechanical coupled field tomography reconstruction algorithm can solve a system of nonlinear partial differential equations considering latent heat of phase transition, thermal stress, and plastic strain, and combine regularization constraints to achieve a stable solution to the ill-conditioned inverse problem. Exemplarily, the thermo-mechanical coupled field tomography reconstruction algorithm can employ iterative optimization tomography algorithms, neural network surrogate tomography algorithms, or Bayesian inference tomography algorithms. The phase transition structure-texture distribution map can be a spatial map representing the type, volume fraction, and grain orientation distribution of phase transition products inside a metal component in two-dimensional or three-dimensional form. It can be used to visually present the microscopic response characteristics under the coupling effect of temperature gradient and phase transition stress. In an exemplary embodiment, the phase transition structure-texture distribution map can include, but is not limited to, phase composition distribution maps, orientation density distribution maps, and grain boundary feature distribution maps.

[0030] The application of a thermo-mechanical coupled-field tomography reconstruction algorithm to construct the phase transition structure-texture distribution map inside a metal component can be achieved by substituting the compensated thermodynamic state feature set into a coupled-field inversion model and solving for the spatial distribution of internal phase transitions and textures. Furthermore, the application of this algorithm to construct the phase transition structure-texture distribution map inside a metal component can be achieved through finite element discretization to solve partial differential equations, direct mapping of surface features to internal distribution using deep neural networks, or sparse representation combined with dictionary learning. This enables high-fidelity inversion from macroscopically measurable thermal fields to microscopically invisible structures, overcoming the limitations of traditional single temperature monitoring.

[0031] Step S4: Perform spatiotemporal adaptive threshold segmentation on the phase transformation structure-texture distribution map to obtain a candidate set of mixed crystal defects and texture abnormal regions. Then, based on the preset microstructure dynamics feature screening rules, the candidate defect set is judged to determine the true phase transformation inhomogeneity region and the true texture anisotropy region.

[0032] The candidate set of mixed crystal defects and texture anomalies can be a set of regions suspected of having structural inhomogeneity or texture deviation, initially screened through threshold segmentation, which can be used to narrow down the range of true and false defects. The microstructure dynamics feature screening rules can be discrimination criteria constructed based on phase transformation dynamics and texture evolution theory to verify whether candidate regions conform to physical laws, and can be used to eliminate false defect signals caused by noise or artifacts. For example, the microstructure dynamics feature screening rules can adopt phase transformation rate consistency rules, texture continuity rules, or thermal history compatibility rules. The true phase transformation inhomogeneity region can be a local area of ​​a metal component where the phase transformation process significantly deviates from the expected path, confirmed by the true / false defect discrimination, and can be used to identify risk locations that may lead to mixed crystals or banded structures. In a specific embodiment, the true phase transformation inhomogeneity region may include, but is not limited to, a slightly hysteretic phase transformation region, a moderately deviated phase transformation region, and a severely stagnant phase transformation region. A true texture anisotropy region can be a localized area in a metal component where the preferred grain orientation intensity exceeds the allowable range, as confirmed by a true / false identification process. It can be used to identify texture anomaly locations that may cause fluctuations in the stamping lug or yield plateau. Furthermore, the true texture anisotropy region can be a γ-fiber anomaly region, an α-fiber anomaly region, or a rotational cubic texture anomaly region.

[0033] Performing spatiotemporally adaptive threshold segmentation on the phase change structure-texture distribution map can involve dynamically adjusting the segmentation threshold and extracting suspected anomalous regions based on the local statistical characteristics of the distribution map. Furthermore, this spatiotemporally adaptive threshold segmentation can be achieved by adaptively determining the global threshold based on the Otsu algorithm, determining the regional threshold based on maximizing local entropy, or determining the boundary threshold based on gradient manifold analysis. This adapts to the non-stationarity of tissue evolution during cooling and improves the accuracy of candidate region extraction. Determining the authenticity of the defect candidate set based on preset microstructure dynamics feature screening rules can involve verifying whether the evolution trajectory of the candidate regions conforms to phase change dynamics and texture development laws, retaining compliant regions as genuine anomalies. In an exemplary embodiment, determining the authenticity of the defect candidate set based on preset microstructure dynamics feature screening rules can be achieved by verifying the rationality of the cooling path based on the phase change CCT curve, verifying the continuity of orientation changes based on the texture evolution ODE model, or verifying the consistency of the phase change sequence based on thermal history backtracking. This effectively filters out false defect signals and improves the credibility and engineering usability of the final identification results.

[0034] Step S5: For the real phase transformation inhomogeneity region and the real texture anisotropy region, the hyperbolic tangent-Avramie index joint model is used to calculate the phase transformation volume fraction parameter, texture intensity coefficient parameter and grain size distribution parameter corresponding to each region. When the phase transformation volume fraction parameter of any region exceeds the preset process target range or the texture intensity coefficient parameter exceeds the preset anisotropy critical threshold, it is determined that the metal component has a risk of microstructure and property deviation that affects subsequent forming or service performance.

[0035] The hyperbolic tangent-Avrammy index joint model can be a composite mathematical model integrating the saturation characteristics of the hyperbolic tangent function and the phase transition kinetics of the Avrammy equation. It can be used to quantitatively characterize the evolution of phase transition volume fraction, texture intensity, and grain size under non-isothermal cooling conditions. In this embodiment, the hyperbolic tangent-Avrammy index joint model can describe the asymptotic behavior of phase transition completion using the hyperbolic tangent function and characterize the nucleation and growth mechanisms using the Avrammy index, and obtain model parameters by jointly fitting experimental data. The phase transition volume fraction parameter can be a dimensionless quantity calculated by the joint model, representing the proportion of a specific phase transition product in a local region, and can be used to quantify the degree of phase transition completion. The texture intensity coefficient parameter can be a quantitative index derived from the joint model reflecting the degree of grain orientation concentration, and can be used to evaluate the anisotropy level of material mechanical properties. Furthermore, the texture intensity coefficient parameter can be the γ-fiber strength coefficient, α-fiber strength coefficient, or Goss texture intensity coefficient. Grain size distribution parameters can be a set of parameters estimated by a joint model to describe the statistical characteristics of grain size in a local region, and can be used to characterize the degree of microstructure refinement. In this embodiment, grain size distribution parameters may include, but are not limited to, average grain size, grain size standard deviation, and grain size skewness. The preset process target range can be a qualified range of phase transformation volume fraction parameters pre-set according to product specifications and performance requirements, and can be used as a benchmark threshold for determining whether the phase transformation has been completed normally. For example, the preset process target range can be a target range for high-strength steel, pipeline steel, or silicon steel. The preset anisotropy critical threshold can be the maximum allowable value of the texture strength coefficient parameter determined according to subsequent forming or service requirements, and can be used to define the boundary between texture abnormalities and non-abnormalities.

[0036] The hyperbolic tangent-Avramie index joint model is used to calculate the phase transformation volume fraction, texture intensity coefficient, and grain size distribution parameters for each region. This can be achieved by inputting the thermal history and stress state of the actual anomalous region into the joint model and solving for the three key microstructure parameters. Furthermore, the calculation of these parameters can be achieved through analytically solving the closed-form expression of the joint model, numerically iterating to solve the implicit equations, or rapidly inferring parameter values ​​using a pre-trained surrogate model. This allows for quantitative characterization of mixed crystals and texture anomalies, supporting refined quality assessment. Determining whether a metal component has a risk of microstructure deviation affecting subsequent forming or service performance can be achieved by comparing the calculated parameters with preset thresholds and marking any deviation risk when any threshold is exceeded. In one specific embodiment, determining that the metal component has a risk of microstructure deviation that may affect subsequent forming or service performance can be achieved through single-parameter independent over-limit judgment, multi-parameter weighted comprehensive scoring judgment, or decision tree-based hierarchical risk judgment, so that the microstructure state can be transformed into an operable quality early warning signal and guide the dynamic adjustment of the cooling process.

[0037] Taking the laminar flow cooling quality control of high-end pipeline steel as an example, the edge computing-based metal component quality deviation prediction method in this embodiment can be implemented by deploying a multi-dimensional sensor array at the edge end in the laminar flow cooling section of an X80 pipeline steel production line in a steel plant and simultaneously collecting thermal images and emissivity data of the strip surface. Combined with real-time parameters of the finishing mill outlet temperature and cooling water flow rate, the temperature measurement deviation caused by the water film is dynamically compensated. The thermo-mechanical coupled field tomography reconstruction algorithm is used to inversely determine the bainite phase transformation distribution and γ-fiber texture intensity in the thickness direction of the strip. Through adaptive threshold segmentation and dynamic rule screening, the true phase transformation lag zone in the middle region and the true texture abnormal zone at the edge are accurately located. The hyperbolic tangent-Avramie index joint model is used to calculate that the bainite volume fraction in this region is lower than the lower limit of the process target and the γ-fiber strength coefficient exceeds the deep drawing critical threshold. The system immediately determines that there is a risk of mixed crystals and stamping ears and feeds the results back to the cooling control system to adjust the water distribution in the middle section, thereby avoiding batch quality accidents.

[0038] In one embodiment, step S2 specifically includes: S21. Real-time data on cooling water impact pressure and water flow fluctuation on the surface of metal components are collected through a fluid field monitoring network deployed below the cooling manifold. Based on the cooling water impact pressure and water flow fluctuation data, a pre-built water flow parameter-surface heat transfer coefficient mapping table is queried to obtain the thermal interference compensation coefficient corresponding to the convective heat transfer capacity and film boiling probability at each component segment. The fluid field monitoring network can be an online monitoring array deployed below the laminar cooling manifold and composed of multiple distributed fluid sensing nodes. The fluid field monitoring network can be used to sense the fluid dynamics of the interaction between cooling water and the surface of the metal component in real time. In this embodiment, the fluid field monitoring network can directly acquire raw physical quantities reflecting the water film coverage and heat transfer state, such as cooling water impact pressure and water flow fluctuations. The fluid field monitoring network can be coordinated with a water flow parameter-surface heat transfer coefficient mapping table to convert the real-time collected fluid data into compensation coefficients characterizing heat transfer characteristics. Exemplary examples include, but are not limited to, piezoresistive fluid monitoring networks, ultrasonic flow monitoring networks, and capacitive liquid level monitoring networks. Cooling water impact pressure can be the instantaneous dynamic pressure value generated when the cooling water jet impacts the surface of the metal component. Cooling water impact pressure can be used to characterize the intensity of forced convection heat transfer. In an exemplary embodiment, cooling water impact pressure can include, but is not limited to, steady-state impact pressure, pulsating impact pressure, and peak impact pressure. Water flow fluctuation data can be a set of time-series signals describing the amplitude, frequency, and spatial distribution non-uniformity of cooling water during its flow on the surface of the metal component. Water flow fluctuation data can be used to indicate dynamic behaviors such as water film rupture, splashing, or localized dry spots. Furthermore, water flow fluctuation data can include, but is not limited to, standing wave type fluctuation data, traveling wave type fluctuation data, and turbulent fluctuation data. The water flow parameter-surface heat transfer coefficient mapping table can be a lookup table established based on experimental calibration or numerical simulation.

[0039] A water flow parameter-surface heat transfer coefficient mapping table can be used to transform the complex water film heat transfer effect, which is difficult to measure directly, into a quantitative compensation factor that can be quickly queried. In a specific embodiment, the water flow parameter-surface heat transfer coefficient mapping table can be obtained by simultaneously measuring temperature and fluid under a large number of operating conditions, fitting the nonlinear response relationship between water flow parameters and measured heat transfer coefficients, and discretizing and storing it. For example, the water flow parameter-surface heat transfer coefficient mapping table may include, but is not limited to, experimental calibration mapping tables, CFD simulation mapping tables, hybrid drive mapping tables, etc. The thermal interference compensation coefficient can be a dimensionless correction factor obtained from the mapping table. The thermal interference compensation coefficient can be used to quantify the attenuation and scattering effect of the water film on the infrared radiation transmission path. Further, the thermal interference compensation coefficient may include, but is not limited to, radiation attenuation compensation coefficient, emissivity shift compensation coefficient, background reflection compensation coefficient, etc. Convective heat transfer capacity can be a comprehensive index characterizing the intensity of forced convective heat transfer between cooling water and the surface of metal components. Convective heat transfer capacity can be used to reflect the actual efficiency of the water film in removing heat under the current operating conditions. In one exemplary embodiment, convective heat transfer capacity may include, but is not limited to, single-phase convective heat transfer capacity, nucleation boiling heat transfer capacity, and transition boiling heat transfer capacity. Film boiling probability can be a measure of the likelihood that a certain area on the surface of a metal component is in a stable vapor film coverage state. Film boiling probability can be used to predict the risk of localized cooling failure caused by vapor film insulation.

[0040] Real-time acquisition of cooling water impact pressure and flow fluctuation data on the surface of metal components via a fluid field monitoring network deployed below the cooling manifold can be achieved by continuously sensing the dynamic mechanical signals of the interaction between cooling water and the component surface using distributed sensors within the fluid field monitoring network and uploading them to edge computing nodes at a fixed sampling rate. Furthermore, this real-time acquisition of cooling water impact pressure and flow fluctuation data on the surface of metal components via a fluid field monitoring network deployed below the cooling manifold can be achieved by using a high-frequency pressure sensor array to achieve millisecond-level impact pressure acquisition, thereby avoiding model mismatch caused by relying on empirical assumptions. Querying a pre-built water flow parameter-surface heat transfer coefficient mapping table based on the cooling water impact pressure and flow fluctuation data can be achieved by using the real-time acquired impact pressure and fluctuation data as index keys to retrieve the corresponding entries in the pre-stored mapping table. In a specific embodiment, querying the pre-built water flow parameter-surface heat transfer coefficient mapping table based on the cooling water impact pressure and flow fluctuation data can be achieved by using bilinear interpolation to obtain continuous compensation values ​​from a discrete table, simplifying the complex multiphase flow heat transfer problem into an efficient table lookup operation. Obtaining the thermal interference compensation coefficients corresponding to the convective heat transfer capacity and film boiling probability at each component segment can be achieved by extracting compensation coefficients that match the current segment position from the mapping results. For example, obtaining the thermal interference compensation coefficients corresponding to the convective heat transfer capacity and film boiling probability at each component segment can be achieved by independently taking values ​​for each segment along the component length direction, thereby enabling spatially resolved quantitative characterization of thermal interference.

[0041] S22. For the spatiotemporal data collected by each sensor channel, the thermal interference compensation coefficient is applied to the temperature correction processing of the infrared thermal image signal, and the compensation amount of the effective thermal radiation path is calculated in combination with the oxide film thickness parameter on the surface of the metal component. The sensor channel can be the signal path corresponding to each independent data acquisition unit in a multi-dimensional sensor array. The sensor channel can be used to ensure the independence and synchronization of the acquisition of each physical quantity. In this embodiment, the sensor channel can include, but is not limited to, infrared radiation channels, visible light reflection channels, and polarized light channels. The infrared thermal image signal can be an uncorrected raw voltage or digital count value sequence output by a high-speed infrared thermal imager. The infrared thermal image signal can be used as a direct input for temperature correction processing. Furthermore, the infrared thermal image signal can include, but is not limited to, RAW format thermal image signals, 14-bit digital thermal image signals, and analog video thermal image signals. The oxide film thickness parameter on the surface of the metal component can be a process-related parameter characterizing the geometric thickness of the oxide layer formed on the surface of the metal component during cooling. The oxide film thickness parameter on the surface of the metal component can be used to reflect the degree of influence of the oxide film on infrared radiation transmission and absorption. In an exemplary embodiment, the oxide film thickness parameter on the surface of the metal component can include, but is not limited to, model-predicted oxide film thickness, offline detected oxide film thickness, and spectral inversion oxide film thickness. The compensation amount for the effective thermal radiation path can be a correction value that comprehensively considers the oxide film thickness and water film disturbance. The compensation amount for the effective thermal radiation path can be used to eliminate errors in radiation temperature measurement systems caused by changes in the optical properties of the dielectric layer. For example, the compensation amount for the effective thermal radiation path may include, but is not limited to, path length compensation, transmittance compensation, and interface reflection compensation.

[0042] Applying the thermal interference compensation coefficient to the temperature correction process of infrared thermal imaging signals can be achieved by embedding the compensation coefficient as a gain or offset factor into the radiation-temperature conversion formula of the infrared thermal imaging signal. Furthermore, applying the thermal interference compensation coefficient to the temperature correction process of infrared thermal imaging signals can be achieved by introducing the compensation coefficient into the radiometric calibration curve to correct the slope and intercept, thereby eliminating the negative impact of dynamic changes in the water film on temperature measurement accuracy from the source. Calculating the compensation amount for the effective thermal radiation path based on the oxide film thickness parameter of the metal component surface can be based on the oxide film thickness and its optical constants, and using a radiative transfer model to calculate the effective penetration depth and attenuation coefficient of the infrared band within the oxide film. In a specific embodiment, calculating the compensation amount for the effective thermal radiation path based on the oxide film thickness parameter of the metal component surface can be achieved by analytically calculating the compensation amount of a single-layer film using the Beer-Lambert law, thus correcting the filtering effect of the oxide film on infrared radiation.

[0043] S23. The corrected spatiotemporal data is fused with the effective thermal radiation path compensation at the data level to generate a compensated thermodynamic state feature set that includes the latent heat release rate of phase change, cooling rate gradient and surface radiation temperature correction value.

[0044] The latent heat release rate of phase transformation can be the rate at which heat is released or absorbed per unit volume of metal due to phase transformation per unit time. The latent heat release rate can be used to directly characterize the phase transformation kinetics process. In this embodiment, the latent heat release rate may include, but is not limited to, the latent heat release rate of ferrite transformation, bainite transformation, and martensite transformation. The cooling rate gradient can be the rate of change of the cooling rate of the compensated temperature field in the spatial direction. The cooling rate gradient can be used to reveal the degree of difference in microstructure evolution in the thickness or width direction. Further, the cooling rate gradient may include, but is not limited to, the cooling rate gradient in the thickness direction, the cooling rate gradient in the width direction, and the longitudinal cooling rate gradient. The surface radiation temperature correction value can be the radiation temperature estimate obtained after dual compensation of fluid disturbance and oxide film. The surface radiation temperature correction value can be the output of the compensated thermodynamic state feature set. For example, the surface radiation temperature correction value may include, but is not limited to, the primary fluid compensation temperature correction value, the secondary oxide film compensation temperature correction value, and the fusion optimization temperature correction value. Data-level fusion of the corrected spatiotemporal data with the effective thermal radiation path compensation can be achieved by performing algebraic operations or weighted synthesis on the fluid-compensated temperature signal and the oxide film path compensation within the same spatiotemporal grid. Furthermore, data-level fusion of the corrected spatiotemporal data with the effective thermal radiation path compensation can be achieved by using Kalman filtering to fuse multi-source compensation results, thereby integrating correction information from both the fluid and medium dimensions.

[0045] Taking the optimization of laminar flow cooling temperature control for high-strength automotive steel sheets as an example, the edge computing-based method for predicting the quality deviation of metal components in this embodiment can be implemented by installing a piezoresistive fluid field monitoring network below the laminar flow cooling manifold in the production of DP980 duplex steel in a steel plant, and collecting real-time data on the impact pressure and flow fluctuation of the cooling water on the strip surface. When a sudden drop in impact pressure and an increase in the fluctuation amplitude are detected in the middle region, the system automatically queries the water flow parameter-surface heat transfer coefficient mapping table and obtains the thermal interference compensation coefficient that causes a 30% decrease in convective heat transfer capacity and an increase in the film boiling probability to 0.7 at that segment. The coefficient is substituted into the infrared thermal imaging signal correction model and combined with the effective thermal radiation path compensation calculated based on the current oxide film thickness of 15μm to perform double correction on the original temperature data. The latent heat release rate curve of the phase transformation in the compensated thermodynamic state feature generated after fusion clearly shows the ferrite transformation peak and bainite transformation shoulder, and the cooling rate gradient shows a reasonable distribution in the thickness direction. The deviation between the surface radiation temperature correction value and the thermocouple verification point is less than 2℃, thereby accurately identifying the real phase transformation lag zone caused by the instability of the water film at the edge and avoiding unnecessary shutdowns caused by misjudging it as equipment failure.

[0046] In one embodiment, step S3 specifically includes: S31. The compensated thermodynamic state feature set is used as an input variable and transmitted to the pre-constructed thermo-mechanical coupling iterative tomographic reconstruction model. The thermo-mechanical coupling iterative tomographic reconstruction model has built-in prior physical constraints based on finite element phase transition simulation analysis. The thermo-mechanical coupled iterative tomographic reconstruction model can be a computational framework that integrates heat transfer and solid mechanics equations and uses iterative optimization algorithms to invert the surface thermodynamic state into the internal microstructure distribution. This model can be used to stably solve the spatial distribution of phase transitions and textures within metallic components under limited surface observation data conditions, overcoming the physical inconsistencies of purely data-driven methods. In this embodiment, the thermo-mechanical coupled iterative tomographic reconstruction model can be coordinated with prior physical constraints based on finite element phase transition simulation analysis to limit the solution space range through physical laws and improve the physical reliability of the inversion results. For example, the thermo-mechanical coupled iterative tomographic reconstruction model may include, but is not limited to, heat conduction-constrained tomographic models, phase transition kinetic-constrained tomographic models, and elastoplastic stress-constrained tomographic models.

[0047] The prior physical constraints based on finite element phase transition simulation analysis can be mathematical expressions or parameter sets obtained through offline finite element simulation that describe the phase transition evolution and thermodynamic response of metal components under typical cooling paths. These constraints can be embedded into the model as prior knowledge for the inversion process and provide physically reasonable solution space boundaries for iterative inversion to prevent non-physical solutions caused by measurement noise or model simplification. Furthermore, the prior physical constraints based on finite element phase transition simulation analysis can be obtained by performing thermo-mechanically coupled phase transition simulations on a large number of standard operating conditions and extracting the constitutive relations between the phase transition rate and stress / strain and the temperature field, which can then be transformed into differentiable constraint terms or regularized functionals. In an exemplary embodiment, the prior physical constraints based on finite element phase transition simulation analysis can take the form of differential equation constraints, inequality constraints, energy functional constraints, etc. The compensated thermodynamic state feature set is transmitted as input variables to the pre-constructed thermo-mechanically coupled iterative tomographic reconstruction model. This can be achieved by formatting the feature set, including the latent heat release rate of phase transition, the cooling rate gradient, and the surface radiation temperature correction value, according to the model interface specification, loading it into memory, and initializing the inversion iteration process. Furthermore, this transmission process can be achieved by transferring data through zero-copy shared memory, asynchronously transmitting feature sets through message queues, or loading across processes through tensor serialization protocols, thereby enabling high-fidelity thermodynamic information to enter the inversion model without loss and laying the data foundation for tissue reconstruction under physical constraints.

[0048] S32. By constructing an objective function to minimize the weighted mean square error between the actual measured temperature and the model simulation temperature, the spatial distribution matrix of the phase transition field and texture field inside the metal component is solved by iterative inversion, thereby generating a high-resolution phase transition structure-texture distribution map. The objective function can be a scalar function used to quantify the degree of matching between the inverted solution and the measured data. This function can be used as an optimization criterion to guide the inversion direction and balance the data fitting accuracy and the satisfaction of physical constraints. In this embodiment, the objective function may include, but is not limited to, data fidelity terms, regularization terms, and physical constraint penalty terms. The weighted mean square error can be an error measure that averages the squared difference between the actual measured temperature and the model simulation temperature after applying spatial or temporal weights. This error measure can be used to highlight the fitting importance of key areas or high-confidence data points and suppress the interference of low-quality data on the overall inversion. For example, the weighted mean square error can be weighted using adaptive signal-to-noise ratio, priority weighting for phase transition sensitive areas, or spatiotemporal smoothness weighting. The actual measured temperature can be a sequence of radiation temperature correction values ​​representing the true thermal state of the metal component surface, obtained after dynamic compensation processing. This temperature sequence can be used as a benchmark input for the data fitting term in the objective function and to provide the external observation constraints required for the inversion.

[0049] The model simulation temperature can be the predicted surface temperature calculated by the thermo-coupled model in the current inversion iteration step based on the assumed internal microstructure distribution. This predicted value can be compared with the actual measured temperature to form an error signal and drive the update of the internal microstructure distribution parameters. Furthermore, the model simulation temperature can include stage data such as the initial guessed simulation temperature, intermediate iteration simulation temperature, and convergence final state simulation temperature. The spatial distribution matrix of the phase transition field and texture field inside the metal component can be a multidimensional array representing the phase transition volume fraction and grain orientation intensity at various locations inside the metal component in the form of a discrete grid. This matrix can be used to fully characterize the evolution state of the microstructure in three-dimensional space and support subsequent defect identification and performance evaluation. The high-resolution phase transition microstructure-texture distribution map can be an image of the microstructure with fine spatial details obtained after visually rendering the spatial distribution matrix. This distribution map can be used to intuitively present the local features of internal phase transitions and textures and provide a high-quality visual carrier for manual interpretation and automatic analysis.

[0050] By constructing an objective function to minimize the weighted mean square error between the actual measured temperature and the model simulated temperature, the weighted sum of squares of the difference between the two can be calculated in each iteration and used as the optimization objective to drive parameter updates. Furthermore, this construction process can be achieved by using the Levenberg-Marquardt algorithm to solve nonlinear least squares, or by using the conjugate gradient method combined with line search optimization, or by using the ADMM splitting operator for alternating minimization. This ensures that the inversion results closely match the measured data, and the fitting priority of key regions is guaranteed through weight design. Iterative inversion to solve for the spatial distribution matrix of the phase transition field and texture field inside the metal component can be achieved by repeatedly adjusting the internal microstructure parameters under the combined action of physical constraints and the objective function until the convergence criterion is met and the final distribution matrix is ​​output. In a specific embodiment, this solution process can be achieved by using the Newton-Raphson method for fast convergence, using Bayesian inference to obtain the posterior distribution, or using deep learning agents to accelerate forward simulation. This allows for a physically consistent and data-aligned quantitative characterization of the internal microstructure and achieves a high-fidelity mapping from macroscopic to microscopic levels. Generating high-resolution phase transition tissue-texture distribution maps can be achieved by interpolating and enhancing the converged spatial distribution matrix, performing color mapping, and generating a visualization image. Furthermore, this generation process can be achieved by using cubic spline interpolation to improve spatial resolution, employing super-resolution reconstruction networks to enhance details, or using multi-channel fusion rendering to highlight tissue contrast. This provides a finely detailed tissue map that is both human-readable and machine-analyzable, supporting subsequent accurate defect localization.

[0051] S33. Perform spatiotemporal joint filtering on the phase transformation structure-texture distribution map to effectively suppress artifacts caused by oxide scale shedding or water vapor interference during the tomographic reconstruction process while preserving the sharpness of grain boundary and phase interface features.

[0052] The spatiotemporal joint filtering process can be an operation that selectively smooths the phase transition structure-texture distribution map simultaneously in the spatial and temporal domains. This process can be used to improve the signal-to-noise ratio and structural continuity of the reconstructed image and avoid artifacts misleading subsequent defect identification. In this embodiment, the spatiotemporal joint filtering process can include, but is not limited to, bilateral spatiotemporal filtering, nonlocal mean spatiotemporal filtering, and guided filtering. The sharpness of grain boundary and phase interface features can be a quantitative indicator characterizing the clarity of the boundaries of different tissue regions in the phase transition structure-texture distribution map. This indicator can be used to ensure that the filtering process does not blur the boundaries of key microstructures and maintains tissue resolution. For example, the sharpness of grain boundary and phase interface features can be measured by gradient magnitude sharpness, edge response width, contrast retention rate, etc. Artifacts can be false structural signals unrelated to the real tissue introduced during tomographic reconstruction due to oxide scale shedding, water vapor disturbance, or sensor anomalies. These signals, as interference objects that need to be suppressed, can reduce the accuracy of defect identification. In an exemplary embodiment, artifacts can include oxide scale shedding artifacts, water vapor scattering artifacts, motion blur artifacts, etc. Performing spatiotemporal joint filtering on phase transition structure-texture distribution maps can involve applying filters that balance spatial edge preservation and temporal continuity to denoise and enhance the structure of the distribution map sequence. Furthermore, this processing can be achieved by using guided filtering to preserve interfaces based on the original heatmap, using 3D block-matched filtering to utilize spatiotemporal redundancy for denoising, or using anisotropic diffusion filtering to smooth along the structure direction. This approach can suppress artifacts while maintaining the clarity of grain boundaries and phase interfaces, thus improving the usability and reliability of the reconstruction results.

[0053] For example, in the scenario of controlling the cooling texture of hot-rolled silicon steel, the edge computing-based metal component quality deviation prediction method of this embodiment can be used in the production of high-grade non-oriented silicon steel in a steel plant. The compensated thermodynamic state feature set is input into a thermo-mechanical coupling iterative tomographic reconstruction model with built-in phase transformation kinetic constraints. By minimizing the weighted mean square error between the actual measured temperature and the model simulation temperature, the spatial distribution matrix of γ-fiber texture intensity and recrystallization volume fraction in the thickness direction of the strip steel is iteratively inverted. The generated phase transformation structure-texture distribution map, after spatiotemporal joint filtering, clearly shows the sharp interface between the central layer texture weakening zone and the surface strong γ-fiber band, and effectively eliminates the strip artifacts caused by cooling water splashing. This high-quality distribution map is used for subsequent microstructure kinetic feature screening and accurately identifies the anisotropic region of the real texture, avoiding misjudging water vapor interference as texture abnormality and incorrectly adjusting the cooling strategy, thus ensuring the stability of the magnetic properties of silicon steel.

[0054] In one embodiment, step S4 specifically includes: S41. Based on the statistical distribution characteristics of the normal matrix region of the metal component in the phase transformation structure-texture distribution diagram, determine the dynamic segmentation threshold. The dynamic segmentation threshold is adjusted spatiotemporally and adaptively according to the local temperature variance and the target value of the final cooling temperature. The normal matrix region of the metal component can be a set of reference regions representing a uniform microstructure state that meets process expectations and has no significant defects in the phase transition structure-texture distribution map. This can be used to provide a statistical benchmark to distinguish between abnormal and normal states and support the adaptive generation of dynamic thresholds. In this embodiment, the normal matrix region of the metal component may include, but is not limited to, the global mean matrix region, the local sliding window matrix region, and the historical batch matrix region. The statistical distribution characteristics can be the probability density function and its moment characteristics obtained by statistically analyzing parameters such as phase transition volume fraction or texture intensity within the normal matrix region. This can be used to quantify the fluctuation range of normal microstructure and provide a data-driven boundary basis for the dynamic segmentation threshold. For example, the statistical distribution characteristics can adopt Gaussian distribution characteristics, Weibull distribution characteristics, kernel density estimation distribution characteristics, etc. Determining the dynamic segmentation threshold based on the statistical distribution characteristics of the normal matrix region of the metal component in the phase transition structure-texture distribution map can be achieved by calculating the mean and standard deviation of the parameters of the normal matrix region and combining them with the confidence level to generate an initial segmentation threshold. Furthermore, the dynamic segmentation threshold can be determined based on the statistical distribution characteristics of the normal matrix region of the metal component in the phase transformation structure-texture distribution map. This can be achieved by setting the threshold boundary using the 3σ principle, determining the robust threshold using the percentile method, and separating the normal and abnormal distributions using a Gaussian mixture model. This can achieve the technical effect of establishing a data-driven adaptive segmentation benchmark and avoiding the subjectivity and lag of manually setting the threshold.

[0055] The dynamic segmentation threshold can be an adaptively generated discrimination boundary value for separating abnormal regions based on the statistical characteristics of the normal matrix and real-time operating conditions. It can overcome the poor adaptability of fixed thresholds in unsteady cooling processes and improve the spatiotemporal consistency of candidate region extraction. In an exemplary embodiment, the dynamic segmentation threshold can be coordinated with the local region temperature variance and the target final cooling temperature, and the threshold can be corrected in real time through a feedback mechanism to adapt to operating condition drift. The local region temperature variance can be a statistical quantity characterizing the degree of temperature field dispersion within a specific spatial neighborhood of the metal component and reflecting the cooling uniformity state. It can serve as a perturbation factor for dynamic threshold adjustment; the larger the temperature variance, the looser the threshold to avoid over-segmentation. The target final cooling temperature can be the ideal temperature point that the metal component should reach at the end of cooling, as set by the process specifications. It can serve as a benchmark anchor point for dynamic threshold adjustment and ensure that the segmentation results are aligned with the process target. For example, the target final cooling temperature can be the target final cooling temperature for high-strength steel, pipeline steel, silicon steel, etc.

[0056] The dynamic segmentation threshold can be spatiotemporally adaptively adjusted based on the local temperature variance and the target final cooling temperature. This can be achieved by using the local temperature variance as a perturbation factor and the target final cooling temperature as a baseline offset to correct the initial threshold in real time. Furthermore, the spatiotemporal adaptive adjustment of the dynamic segmentation threshold based on the local temperature variance and the target final cooling temperature can be achieved through methods such as linear weighted fusion adjustment, dynamic adjustment using fuzzy logic reasoning, and optimization of adjustment coefficients using reinforcement learning strategies. This allows the segmentation threshold to dynamically change with the cooling conditions and adapt to the strong non-steady-state characteristics of the laminar cooling process.

[0057] S42. Based on the dynamic segmentation threshold, a multidimensional connected region labeling algorithm is performed on the phase change organization-texture distribution map to obtain the initial set of candidate regions for phase change inhomogeneity and the initial set of candidate regions for texture anomalies. The multidimensional connected component labeling algorithm is an image processing algorithm that identifies and labels topologically connected anomalous regions in the multidimensional feature space of a phase transition organization-texture distribution map. It can be used to aggregate discrete anomalous pixels into candidate regions with complete spatial structure and avoid fragmented false detections. Applying a multidimensional connected component labeling algorithm to a phase transition organization-texture distribution map based on a dynamic segmentation threshold involves applying connectivity criteria in the multidimensional feature space and aggregating adjacent voxels exceeding the dynamic threshold into independent candidate regions. Furthermore, applying a multidimensional connected component labeling algorithm to a phase transition organization-texture distribution map based on a dynamic segmentation threshold can be implemented using a two-pass scanning labeling algorithm, a disjoint-set data structure (DFS) efficient labeling algorithm, or a GPU-parallel connected component labeling algorithm. This allows for the acquisition of spatially topologically complete candidate regions and provides structured input for subsequent feature extraction.

[0058] The initial set of candidate regions for phase transition inhomogeneity can be a set of regions suspected of deviating from the phase transition process, obtained after dynamic threshold segmentation and connectivity labeling. This set can serve as the first-stage input for authenticity determination and narrow down the scope of subsequent dynamic feature verification. The initial set of candidate regions for texture anomalies can be a set of regions suspected of having concentrated grain orientation anomalies, obtained after dynamic threshold segmentation and connectivity labeling. This set can be used to identify potential risk locations requiring further verification of texture authenticity. Obtaining the initial set of candidate regions for phase transition inhomogeneity and the initial set of candidate regions for texture anomalies can be achieved by classifying phase transition-related and texture-related connected regions separately based on the labeling results, forming two independent candidate sets. Furthermore, obtaining the initial set of candidate regions for phase transition inhomogeneity and the initial set of candidate regions for texture anomalies can be achieved through automatic classification based on principal component contribution, differentiation based on the angle between feature vectors, and classification based on preset label template matching. This achieves preliminary decoupling of phase transition and texture anomalies and supports differentiated authenticity determination strategies.

[0059] S43. For each candidate region in the initial set of candidate regions for phase transformation inhomogeneity, extract its phase transformation dynamics characteristic parameters and thermal hysteresis loop characteristic parameters; for each candidate region in the initial set of candidate regions for texture anomalies, extract its elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters; and match and verify the extracted characteristic parameters with the pre-constructed microstructure dynamics characteristic database. When the phase transformation dynamics characteristic parameters are in the unsteady-state range and the thermal hysteresis loop characteristic parameters are lower than the preset threshold, the candidate region is confirmed as a true phase transformation inhomogeneity region; when the elastic modulus directionality characteristic parameters show significant differences and the grain orientation concentration characteristic parameters exceed the preset texture critical value, the candidate region is confirmed as a true texture anisotropy region.

[0060] Among them, phase transition kinetic characteristic parameters can be quantitative indicators describing the kinetic behavior of the candidate region, such as phase transition rate, nucleation density, and growth mechanism. They can be used to characterize whether the phase transition process conforms to physical laws and to distinguish between real phase transition inhomogeneity and noise interference. Hysteresis loop characteristic parameters can be parameters characterizing the temperature response hysteresis characteristics of the candidate region in the heating-cooling cycle and reflecting the asymmetry of latent heat release and absorption during phase transition. They can be used to verify the thermodynamic consistency of phase transition behavior, as spurious defects usually lack reasonable hysteresis loop characteristics. Extracting the phase transition kinetic characteristic parameters and hysteresis loop characteristic parameters of each candidate region from the initial set of candidate regions for phase transition inhomogeneity can be achieved by performing kinetic modeling and hysteresis analysis on the time series data of each candidate region and outputting the corresponding feature vector. Furthermore, the extraction of phase transition kinetic characteristic parameters and hysteresis loop characteristic parameters of each candidate region from the initial set of candidate regions for phase transition inhomogeneity can be achieved by using JMAK equation fitting to extract kinetic parameters, using Hilbert transform to analyze hysteresis features, and using wavelet packet decomposition to extract transient features. This allows for the acquisition of multi-dimensional features characterizing the physical essence of phase transition and supports mechanism-based verification of authenticity.

[0061] The elastic modulus directionality characteristic parameter can be an anisotropy measure characterizing the degree of change of elastic modulus with crystallographic orientation within a candidate region. It can be used to reflect the intensity of the influence of texture on mechanical properties and to determine whether texture anomalies affect serviceability. The grain orientation concentration characteristic parameter can be a statistical measure quantifying the degree to which the preferred orientation of grains deviates from a random distribution within a candidate region. It can directly characterize the texture intensity, and exceeding a critical value indicates a significant anisotropy risk. For example, the grain orientation concentration characteristic parameter can be the γ-fiber concentration parameter, the Cube texture concentration parameter, or the Goss texture concentration parameter. Extracting the elastic modulus directionality characteristic parameter and the grain orientation concentration characteristic parameter of each candidate region in the initial set of candidate regions for texture anomalies can be done by calculating the elastic anisotropy index and orientation density peak based on texture distribution data and generating a texture feature vector. Furthermore, the elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters of each candidate region in the initial texture anomaly candidate region set can be extracted by using the weighted average of single crystal elastic constants to calculate the directionality, using ODF integral to calculate the concentration, and using spherical harmonic function expansion to extract features. This can quantify the degree of influence of texture on performance and provide a physical basis for confirming texture anomalies.

[0062] A microstructure dynamics feature database can be a knowledge base storing the actual microstructure evolution characteristic parameters and their valid ranges of various metallic materials under different processing conditions. It can serve as an authoritative reference system for authenticity determination and support feature matching verification based on physical mechanisms. In a specific embodiment, the microstructure dynamics feature database can be constructed through laboratory calibration, production big data mining, and simulation verification, and includes the legal range and association rules of phase transformation and texture features. Furthermore, the microstructure dynamics feature database can include, but is not limited to, experimental calibration databases, production performance databases, and multi-scale simulation databases. Matching and verifying the extracted feature parameters with the pre-constructed microstructure dynamics feature database can involve querying the database for the legal feature ranges of the corresponding materials and working conditions and determining whether the candidate region features fall within the valid range. Furthermore, matching and verifying the extracted feature parameters with the pre-constructed microstructure dynamics feature database can be achieved using nearest neighbor distance matching, Maslow's Leyenoff stability criterion verification, and Bayesian probability evaluation of the matching degree. This allows for the use of prior knowledge to filter out false defects that do not conform to physical laws and improves the reliability of the determination.

[0063] The unsteady-state region can be a range of parameter values ​​defined in the microstructure dynamics database, representing a phase transition process in a non-equilibrium or anomalous evolution state. It can be used to identify the degree to which phase transition dynamics deviates from the steady-state path and serves as one of the necessary conditions for confirming genuine phase transition heterogeneity. The preset threshold can be a minimum allowable value set for the thermal hysteresis loop characteristic parameter to distinguish between genuine phase transitions and spurious signals. It can be used to ensure that candidate regions have sufficient thermodynamic response intensity and exclude false candidates caused by pure noise. When the phase transition dynamics characteristic parameter is in the unsteady-state region and the thermal hysteresis loop characteristic parameter is below the preset threshold, the candidate region is confirmed as a genuine phase transition heterogeneity region. This can be determined as a genuine phase transition anomaly when both the dynamics are unsteady and the thermal hysteresis loop response is weak. Furthermore, confirming that the candidate region is a genuine phase transition heterogeneity region when the phase transition dynamics characteristic parameter is in the unsteady-state region and the thermal hysteresis loop characteristic parameter is below the preset threshold can be achieved through methods such as strict judgment using logical AND gates, comprehensive judgment using weighted scoring, and hierarchical judgment using decision trees. This allows for confirmation of genuine phase transition heterogeneity through dual physical constraints and avoids misjudgment based on a single indicator.

[0064] The preset texture threshold can be an upper limit threshold set for the grain orientation concentration characteristic parameter to determine whether a texture anomaly constitutes a quality risk. It can be used to define the boundary between acceptable texture intensity and harmful texture intensity and support the assessment of forming performance risks. In an exemplary embodiment, the preset texture threshold may include, but is not limited to, the texture threshold for deep drawing, the texture threshold for fatigue service, and the texture threshold for electromagnetic performance. When the elastic modulus directional characteristic parameter shows a significant difference and the grain orientation concentration characteristic parameter exceeds the preset texture threshold, the candidate region is confirmed as a true texture anisotropy region. This can be determined as a true texture anomaly when both the elastic anisotropy and the orientation concentration are too high. Furthermore, confirming that the candidate region is a true texture anisotropy region when the elastic modulus directional characteristic parameter shows a significant difference and the grain orientation concentration characteristic parameter exceeds the preset texture threshold can be achieved by using threshold cross-validation, principal component projection separation verification, and machine learning classifier comprehensive judgment, thereby accurately identifying true texture anomalies that are harmful to forming or service and supporting quality risk warnings.

[0065] For example, in the scenario of controlling the stamping ears of IF steel used in automotive outer panels, the edge computing-based metal component quality deviation prediction method of this embodiment can determine the dynamic segmentation threshold based on the statistical distribution of γ fiber strength in the normal matrix region during IF steel production in a steel plant, and automatically relax the threshold according to the increase of temperature variance at the edge of the strip. After performing multidimensional connectivity labeling on the phase transformation structure-texture distribution map, 12 initial texture anomaly candidate regions are obtained. For each candidate region, the elastic modulus directionality feature parameter and grain orientation concentration feature parameter are extracted and matched with the IF steel microstructure dynamics feature database. Among them, 3 candidate regions, although having high orientation concentration, are judged as measurement artifacts and are eliminated because their elastic anisotropy is not significant. The remaining 9 regions simultaneously satisfy the significant difference in elastic modulus directionality and the γ fiber concentration exceeding the texture critical value for deep drawing, and are confirmed as true texture anisotropy regions. Based on this, the system warns of stamping ear risks and feeds back to adjust the fine rolling reduction distribution to effectively avoid batch stamping cracking accidents.

[0066] In one embodiment, the process of constructing the microstructure dynamics feature database includes: B1. Collect metal component samples with standard materials, prepare microstructure samples with known phase transformation type and texture type, and accurately measure the phase transformation kinetic characteristic parameters, thermal hysteresis loop characteristic parameters, elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters of each microstructure sample under simulated laminar flow cooling environment. The standard material metal component samples can be metal material samples whose chemical composition, purity, and initial microstructure all meet national standards or enterprise internal control standards. Standard material metal component samples can be used to eliminate the interference of material fluctuations on characteristic parameter measurements and ensure that the constructed database reflects a pure microstructure-property mapping relationship. In this embodiment, standard material metal component samples may include, but are not limited to, low-carbon aluminum killed steel samples, microalloyed high-strength steel samples, and non-oriented silicon steel samples. Microstructure samples with known phase transformation types and texture types can be calibrated samples that have undergone precise control through laboratory heat treatment and deformation processes, resulting in a definite phase composition ratio and a clear texture component. Microstructure samples with known phase transformation types and texture types can provide truth labels between characteristic parameters and microstructure types to support the supervised construction of the hierarchical mapping table. The simulated laminar flow cooling environment can be a controllable thermal simulation device or test platform that reproduces industrial laminar flow cooling heat transfer conditions in the laboratory. The simulated laminar flow cooling environment ensures that the measured kinetic characteristic parameters are consistent with the physical response under actual production conditions and guarantees the engineering applicability of the database. In one exemplary embodiment, the simulated laminar cooling environment can be a Gleeble thermal simulator environment, a water spray cooling test bench environment, a numerical simulation virtual environment, etc.

[0067] Collecting metal component samples with standard materials and preparing microstructure samples with known phase transformation and texture types can be achieved by selecting standard samples with qualified composition and preparing calibration samples with clear microstructure labels through controlled heat treatment and deformation processes. Furthermore, collecting metal component samples with standard materials and preparing microstructure samples with known phase transformation and texture types can be achieved by using salt bath quenching and tempering to prepare two-phase samples, rolling and annealing to prepare texture samples, or directional solidification to prepare single samples. This allows for obtaining physically authentic and accurately labeled benchmark datasets and lays the material foundation for database construction. Accurately measuring the phase transformation kinetics, thermal hysteresis loop, elastic modulus directionality, and grain orientation concentration characteristics of each microstructure sample under simulated laminar flow cooling can be achieved by placing the calibration samples in a simulated cooling device and simultaneously acquiring thermal radiation signals and extracting four-dimensional characteristic parameters. Furthermore, the phase transformation dynamics characteristic parameters, thermal hysteresis loop characteristic parameters, elastic modulus directionality characteristic parameters, and grain orientation concentration characteristic parameters of each tissue sample can be accurately measured in situ using a Gleeble thermal simulator, monitored online using synchrotron X-ray diffraction, and tested for elastic modulus using laser ultrasonic non-destructive testing. This allows for the acquisition of high-precision characteristic data consistent with industrial operating conditions and ensures the engineering representativeness of the database.

[0068] B2. Based on the phase transformation volume fraction parameter and texture intensity parameter, establish a hierarchical mapping table of microstructure kinetic characteristic parameters, where ferrite-pearlite mixed crystals correspond to the first kinetic characteristic range, bainite-martensite mixed crystals correspond to the second kinetic characteristic range, weak texture regions correspond to the first orientation concentration range, and strong texture regions correspond to the second orientation concentration range. The phase transition volume fraction parameter can be a quantitative indicator characterizing the volume proportion of a specific phase transition product in a microstructure sample. It serves as a quantitative benchmark for defining the type and degree of mixed crystals and supports the division of kinetic characteristic ranges. In this embodiment, the phase transition volume fraction parameter can be metallographic volume fraction, XRD volume fraction, EBSD volume fraction, etc. The texture intensity parameter can be a quantitative indicator describing the degree of preferred orientation concentration of grains in a microstructure sample. It can distinguish between weak and strong texture regions and support the calibration of the orientation concentration range. For example, the texture intensity parameter can include ODF peak density, inverse pole figure multiple, texture component volume fraction, etc. The hierarchical mapping table of microstructure kinetic characteristic parameters can be a structured lookup table indexed by phase transition volume fraction and texture intensity, classifying multidimensional kinetic characteristic parameters into corresponding intervals for different microstructure types. This hierarchical mapping table can be obtained by establishing a nonlinear mapping relationship between characteristic parameters and microstructure types based on measured data from a large number of calibrated samples through cluster analysis or expert rules, and then discretizing and storing the data. A hierarchical mapping table of microstructure dynamics characteristic parameters enables rapid retrieval from a continuous feature space to discrete microstructure categories and supports online authenticity judgment and type identification. Ferrite-pearlite mixed crystals can be a two-phase microstructure composed of proeutectoid ferrite and pearlite. Ferrite-pearlite mixed crystals can be used to identify the risk of banded microstructures in conventional hot-rolled steel.

[0069] The first kinetic characteristic range can be the legal value range of characteristic parameters specific to ferrite-pearlite mixed crystal structures in the microstructure kinetic characteristic parameter classification mapping table. The first kinetic characteristic range can provide the physical boundary for distinguishing the authenticity of this type of mixed crystal. For example, the first kinetic characteristic range may include the phase transformation rate range, the thermal hysteresis area range, and the nucleation activation energy range. Bainite-martensite mixed crystals can be low-temperature transformation mixed structures composed of bainite and martensite. Bainite-martensite mixed crystals can be used to identify the risk of uneven hardness in high-strength steel. The second kinetic characteristic range can be the legal value range of characteristic parameters specific to bainite-martensite mixed crystal structures in the microstructure kinetic characteristic parameter classification mapping table. The second kinetic characteristic range can define the physical reasonable domain of low-temperature transformation mixed crystals and support the accurate identification of phase transformation anomalies in high-strength steel. Further, the second kinetic characteristic range may include the shear driving force range, the Ms point offset range, and the self-tempering degree range. Weak texture regions can be microstructure regions where the grain orientation is nearly randomly distributed and the anisotropy is below the engineering concern threshold. Weak texture regions can be used to eliminate harmless background texture fluctuations. In this embodiment, weak texture regions may include, but are not limited to, randomly oriented weak texture regions, γ-fiber weakened regions, and cube texture weakened regions.

[0070] The first orientation concentration range can be the upper limit of the grain orientation concentration parameter characterizing a weak texture state as defined in the hierarchical mapping table. The first orientation concentration range can set the initial criterion for texture anomalies. For example, the first orientation concentration range can include the MRD value range, J-index range, entropy value range, etc. A strong texture region can be a region of microstructure with highly preferred grain orientation that may lead to significant anisotropy in mechanical properties. A strong texture region can trigger subsequent severity assessments. In a specific embodiment, a strong texture region may include, but is not limited to, a strong texture region in deep-drawn lugs, a strong texture region with yield plateau fluctuations, and a strong texture region with decreased magnetic flux density. The second orientation concentration range can be the lower limit and above the grain orientation concentration parameter characterizing a strong texture state as defined in the hierarchical mapping table. The second orientation concentration range can define the texture anomaly domains that require key attention and support hierarchical early warning of forming or service risks. The hierarchical mapping table establishing microstructure dynamic characteristic parameters based on phase change volume fraction parameters and texture intensity parameters can be a structured mapping table generated by clustering or regularly dividing the measured characteristic parameters using phase change volume fraction and texture intensity as classification axes. Furthermore, a hierarchical mapping table of microstructure dynamic characteristic parameters based on phase change volume fraction parameters and texture intensity parameters can be established by using K-means clustering for automatic binning, manually defining boundaries using expert experience, and learning the optimal split point using decision trees. This enables refined classification and characterization of complex tissue types and supports rapid online retrieval and discrimination.

[0071] B3. Establish a correlation between the microstructure dynamics characteristic parameters in the hierarchical mapping table and the critical influence threshold of the anisotropy index of the mechanical properties of metal components, and form a microstructure dynamics characteristic database for distinguishing the authenticity of tissue abnormalities and assessing their severity.

[0072] The critical impact threshold of the anisotropy index of mechanical properties can be the minimum unacceptable directional difference in the mechanical properties of materials, determined according to product service requirements. The critical impact threshold of the anisotropy index of mechanical properties can directly correlate microstructural characteristics with macroscopic service failure and enable the database to assess engineering risks. In this embodiment, the critical impact threshold of the anisotropy index of mechanical properties may include, but is not limited to, the critical threshold of plastic strain ratio, the critical threshold of yield strength anisotropy, and the critical threshold of fatigue life directionality. The correlation can be a mathematical model or logical expression describing the quantitative or qualitative correspondence rules between the dynamic characteristic parameters and the critical thresholds of anisotropy in the hierarchical mapping table. The correlation can be obtained by establishing a mapping function between the combination of characteristic parameters and the anisotropy index through regression analysis, machine learning, or mechanism derivation. The correlation can achieve end-to-end inference from structural characteristics to performance risk and support the quantitative assessment of quality deviations. For example, the correlation may include, but is not limited to, linear regression correlation, neural network correlation, and fuzzy rule correlation. Establishing a correlation between the microstructure dynamics characteristic parameters in the hierarchical mapping table and the critical influence thresholds of anisotropic indices of the mechanical properties of metal components can be achieved by binding each characteristic interval with the anisotropic critical threshold through statistical analysis or mechanistic modeling to form risk mapping rules. Furthermore, establishing this correlation can be achieved by using multiple regression to fit performance prediction models, using SVM classifiers to establish risk boundaries, and using physical constitutive equations to derive critical conditions. This endows the database with engineering risk assessment capabilities and allows the results of microstructure anomaly identification to directly guide process adjustments.

[0073] For example, in the scenario of preventing edge hardening defects in DP980 duplex steel, the edge computing-based metal component quality deviation prediction method of this embodiment can be as follows: Standard composition samples containing different proportions of bainite-martensite are collected during DP980 production at a steel plant. The phase transformation kinetics and thermal hysteresis characteristics are measured under Gleeble simulated laminar flow cooling. A graded mapping table is established based on the phase transformation volume fraction, and the range of 30% to 50% bainite content corresponds to the second kinetic characteristic range. Simultaneously, the lower limit of the thermal hysteresis area within this range is correlated with the critical influence threshold of the edge yield strength ratio of 1.15. During online monitoring, when the characteristic parameters of the candidate region at the edge of the strip fall into the second kinetic characteristic range and the thermal hysteresis area is lower than the correlated threshold, the system determines that there is a risk of edge hardening caused by bainite-martensite mixed crystals and automatically reduces the edge cooling water volume, successfully controlling the edge hardness fluctuation within ±15 HV.

[0074] In one embodiment, the following steps are also included after step S5: S6. Based on the spatial location and distribution density of the real phase transition inhomogeneity region and the real texture anisotropy region, generate a component microstructure performance heat map, and evaluate the performance deviation coefficient of the metal component based on the component microstructure performance heat map combined with the Hall-Page relationship model.

[0075] The component microstructure and performance thermogram can be a visualized map that maps the distribution density and location information of the actual phase transformation inhomogeneity region and the actual texture anisotropy region to a continuous field based on spatial coordinates. It can be used to transform discrete microscopic defect detection results into a spatially structured characterization reflecting the overall quality state of the component and serve as an input carrier for performance deviation assessment. In this embodiment, the component microstructure and performance thermogram can be used in conjunction with the Hall-Page relation model to provide spatially resolved microstructural feature input and support the quantitative calculation of local performance deviations. The Hall-Page relation model can be a physical constitutive model describing the inverse square root relationship between the yield strength and grain size of metallic materials. It can be used to achieve a quantitative mapping from microscopic grain size distribution to macroscopic mechanical properties and provide a physical basis for performance deviation coefficient assessment. In an exemplary embodiment, the Hall-Page relation model can establish a mathematical expression of the grain boundary strengthening mechanism based on dislocation pile-up theory and obtain material-specific Hall-Page coefficients and frictional stress parameters through experimental calibration.

[0076] The performance deviation coefficient, a dimensionless index representing the degree to which actual mechanical properties deviate from the target value, can be calculated based on the component microstructure performance heatmap and the Hall-Page relationship model. It can be used to quantify the overall performance fluctuation level of metal components caused by microstructure inhomogeneity and serve as a criterion for triggering process optimization. Generating the component microstructure performance heatmap based on the spatial location and distribution density of the actual phase transition inhomogeneity region and the actual texture anisotropy region can be achieved by interpolating or aggregating the identified actual anomaly regions according to their spatial coordinates and distribution density into a continuous field map. Furthermore, generating the component microstructure performance heatmap based on the spatial location and distribution density of the actual phase transition inhomogeneity region and the actual texture anisotropy region can be achieved by using kernel density estimation to generate a smooth heatmap, using grid statistics to generate a resolution-controllable heatmap, or using Gaussian process regression to generate a heatmap with uncertainty estimation. This transforms discrete defect information into a structured spatial representation and supports subsequent performance evaluation based on a physical model.

[0077] The performance deviation coefficient of metal components can be evaluated by combining the component microstructure and property thermogram with the Hall-Page relationship model. This can be achieved by substituting the grain size distribution in the thermogram into the Hall-Page relationship model, calculating the yield strength of each region, and comparing it with the target value to obtain the deviation coefficient. Furthermore, the evaluation of the performance deviation coefficient of metal components based on the component microstructure and property thermogram combined with the Hall-Page relationship model can be achieved by calculating the deviation coefficient pixel-by-pixel and then taking a weighted average, by performing partitioned statistical analysis and then summarizing by volume weight, or by solving the overall deviation based on the integral of the probability distribution function. This allows for a quantitative mapping from microstructure morphology to macroscopic mechanical properties and overcomes the accuracy limitations of inferring performance indirectly based solely on the temperature field.

[0078] S7. When the performance deviation coefficient exceeds the preset quality control threshold, an automatic process optimization analysis report is generated, which includes the spatial distribution characteristics of the tissue abnormality, the severity level, and corresponding cooling process parameter adjustment suggestions.

[0079] The preset quality control threshold can be the maximum allowable limit of the performance deviation coefficient pre-set according to product technical standards and downstream application requirements. It can be used to define the dividing point between quality qualification and non-qualification, and to initiate the process optimization analysis report generation process when the value is exceeded. The spatial distribution characteristic parameters of organizational anomalies can be a set of quantitative indicators describing the geometric shape and aggregation degree of the real phase change non-uniform region and the real texture anisotropy region in the component. It can be used to accurately characterize the spatial topological features of defects and provide targeted positioning basis for process adjustment. The severity level can be a quality risk classification identifier comprehensively defined based on the performance deviation coefficient and the spatial distribution characteristic parameters of organizational anomalies. It can be used to prioritize quality deviations and guide the formulation of differentiated response strategies. The cooling process parameter adjustment suggestions can be a set of executable control commands for the laminar flow cooling system generated for the identified organizational anomaly types and spatial locations. It can be used to transform quality diagnosis results into specific process intervention measures and support closed-loop feedback control. In a specific embodiment, the cooling process parameter adjustment suggestions may include, but are not limited to, water distribution adjustment suggestions, cooling rate adjustment suggestions, and manifold start-up and shutdown sequence adjustment suggestions. Process optimization analysis reports can be automatically generated structured documents containing spatial distribution characteristics of tissue abnormalities, severity levels, and suggestions for adjusting cooling process parameters. They can be used as information carriers to form a closed-loop feedback chain of perception-evaluation-decision and support rapid response by operators or automatic execution by the system.

[0080] When the performance deviation coefficient exceeds a preset quality control threshold, a process optimization analysis report is automatically generated. This can involve automatically extracting spatial distribution characteristic parameters of tissue abnormalities and assessing their severity when the performance deviation coefficient exceeds the limit, then matching cooling process parameter adjustment suggestions and assembling them into a structured report. Furthermore, the automatic generation of the process optimization analysis report when the performance deviation coefficient exceeds the preset quality control threshold can be achieved through matching predefined adjustment strategies based on a rule engine, retrieving historical similar operating condition handling plans based on case reasoning, or dynamically generating optimal adjustment suggestions based on a reinforcement learning model. This forms a closed-loop feedback chain of perception-evaluation-decision, improving the intelligent control level and response timeliness of the laminar flow cooling process.

[0081] Taking the closed-loop control of laminar flow cooling quality of high-strength automotive steel sheets as an example, the edge computing-based metal component quality deviation prediction method in this embodiment can identify a real phase transformation inhomogeneity region in the middle of the strip and a real texture anisotropy region at the edge during the production of DP980 dual-phase steel in a steel plant. Based on this, a component microstructure performance thermogram is generated. Combined with the Hall-Page relationship model, the yield strength deviation coefficient is calculated to be 1.18, which exceeds the preset quality control threshold of 1.10. The system automatically generates a process optimization analysis report. The microstructure abnormality spatial distribution characteristic parameters show that the area of ​​the abnormal zone in the middle accounts for 12% and the centroid is located 30mm to the right of the width centerline. The severity level is rated as moderate deviation. The cooling process parameter adjustment suggestion is to increase the water volume of the third group of manifolds in the middle section by 15% and delay the opening time of the edge shielding valve by 200ms. The report is immediately pushed to the cooling control system and the system automatically executes the adjustment command. The performance deviation coefficient of the next coil of strip steel is reduced to 1.05, effectively ensuring the consistency of stamping forming performance.

[0082] In one embodiment, step S1 specifically includes: S11. Obtain the chemical composition parameters and phase transition critical point parameters of the metal component to be cooled. Based on the chemical composition parameters and phase transition critical point parameters, query the pre-built material-sensor configuration mapping table to determine the sampling frame rate and surface emissivity detection wavelength range of the infrared thermal imager suitable for this type of metal component.

[0083] The chemical composition parameters can be a set of physicochemical indicators characterizing the alloy element composition and mass fraction of the metal component to be cooled. These parameters can serve as an intrinsic basis for determining the phase transformation kinetics and radiation characteristics of the material and can be matched with the sensor acquisition configuration. In this embodiment, the chemical composition parameters may include, but are not limited to, carbon equivalent parameters, microalloying element content parameters, and residual element content parameters. The phase transformation critical point parameters can be thermodynamic characteristic values ​​describing the onset and end temperatures of solid-state phase transformation in a metallic material during heating or cooling. These parameters can be used to define the time window and temperature range for the release of latent heat of phase transformation and guide the setting of the infrared sampling frequency to avoid signal aliasing. For example, the phase transformation critical point parameters may include, but are not limited to, austenitization critical point parameters, ferrite transformation critical point parameters, and bainite transformation critical point parameters.

[0084] The material-sensor configuration mapping table can be a pre-built structured database storing the correspondence between different material property combinations and optimal sensing acquisition parameters. It can be used to adaptively configure the sensor array's operating parameters and ensure high-quality raw data that meets inversion requirements for different steel grades. In an exemplary embodiment, the material-sensor configuration mapping table can be generated based on historical experimental data and theoretical model calibration, and associated with recommended frame rates and wavelength ranges using chemical composition and phase transition critical points as index keys. Furthermore, the material-sensor configuration mapping table can include, but is not limited to, rule lookup tables, multidimensional interpolation mapping tables, and machine learning regression mapping tables. The infrared thermal imager sampling frame rate can be the number of thermal images acquired per unit time by the high-speed infrared thermal imager and characterizes the temporal resolution level. It can be used to ensure distortion-free capture of transient temperature fluctuations caused by rapid phase transition processes and avoid losing key thermal response details due to undersampling. The surface emissivity detection wavelength range can be the spectral response interval selected when the surface emissivity detection unit is operating. It can be used to match the radiation peak bands under the coverage of specific material oxide films and water films, maximize the signal-to-noise ratio, and suppress environmental interference.

[0085] Obtaining the chemical composition and phase transition critical point parameters of the metal component to be cooled can be achieved by reading the material identification and thermodynamic parameters of the current component from the production management system or online testing equipment. Furthermore, this can be done automatically via the L2 system API interface, by scanning material labels with a barcode scanner and querying a local database, or by manually entering key parameters by the operator. This provides accurate prior material information for adaptive matching of sensor configurations. Querying a pre-built material-sensor configuration mapping table based on the chemical composition and phase transition critical point parameters can be done by using the obtained parameters as index keys and retrieving the corresponding recommended sensor configuration item from the mapping table. For example, querying the pre-built material-sensor configuration mapping table based on the chemical composition and phase transition critical point parameters can be achieved through precise key-value matching, fuzzy interval matching, or nearest neighbor retrieval based on similarity. This allows for quickly obtaining the optimal acquisition parameters for the current material characteristics and avoids manual trial-and-error parameter tuning. Determining the appropriate infrared thermal imager sampling frame rate and surface emissivity detection wavelength range for this type of metal component can be achieved by parsing the query results into specific frame rate values ​​and upper and lower wavelength limits, and then sending them to the sensor array controller. In an exemplary embodiment, determining the appropriate infrared thermal imager sampling frame rate and surface emissivity detection wavelength range for this type of metal component can be achieved by directly using the recommended values ​​in the table, fine-tuning the recommended values ​​based on real-time operating conditions, or combining multiple candidate values ​​through weighted fusion. This allows the sensor array's operating state to be precisely adapted to the material's phase transition and radiation characteristics.

[0086] S12. Under the sampling frame rate and wavelength range, control the multi-dimensional sensing array to scan the component to be cooled synchronously, and capture the transient temperature response characteristics caused by the release of latent heat of phase change and the dynamic change sequence of surface radiation caused by texture change in real time through edge computing nodes.

[0087] The transient temperature response features can be short-time temperature change patterns extracted from high-frequency thermal imaging data, reflecting the rate and intensity of latent heat release during phase transition. These patterns can be used to directly characterize the real-time process of phase transition dynamics and provide high-time-resolution heat source input for tomographic reconstruction. In this embodiment, the transient temperature response features may include, but are not limited to, temperature rise slope features, latent heat peak area features, and cooling curve inflection point features. The surface radiation dynamic change sequence can be a data stream continuously recorded within an adapted wavelength range, reflecting the time-varying laws of surface radiation characteristics caused by texture evolution. This data can be used to capture the optical response trajectory during grain orientation adjustment and support the dynamic identification of texture anisotropy. For example, the surface radiation dynamic change sequence may include, but is not limited to, single-band radiation intensity sequences, multi-band ratio sequences, and polarization degree time series sequences. The edge computing node can be an embedded computing unit deployed in the laminar cooling environment with local data processing capabilities. It can be used to convert high-bandwidth raw sensor data in-situ into low-dimensional key feature quantities, eliminate transmission delays, and ensure real-time closed-loop control requirements. In one specific embodiment, edge computing nodes can directly connect to a multi-dimensional sensor array via a dedicated interface and perform feature extraction and compression algorithms at the data source. Furthermore, edge computing nodes may include, but are not limited to, FPGA acceleration nodes, GPU inference nodes, and ASIC dedicated processing nodes.

[0088] Controlling the multi-dimensional sensor array to synchronously scan the component to be cooled within the sampling frame rate and wavelength range can be achieved by sending configuration commands to the infrared thermal imager and emissivity detection unit and initiating a synchronous acquisition sequence. Furthermore, controlling the multi-dimensional sensor array to synchronously scan the component to be cooled within the sampling frame rate and wavelength range can be achieved through hardware-triggered synchronous startup, software protocol handshake synchronization, or master-slave clock cascade synchronization, thereby ensuring strict temporal and spatial alignment of the multi-dimensional data and laying the foundation for subsequent fusion processing. Real-time capture of the transient temperature response characteristics caused by the release of latent heat from phase change and the dynamic change sequence of surface radiation caused by texture changes via edge computing nodes can be achieved by the edge nodes performing online feature extraction algorithms on the convective sensing data and outputting a compressed feature sequence. In an exemplary embodiment, real-time capture of the transient temperature response characteristics caused by the release of latent heat from phase change and the dynamic change sequence of surface radiation caused by texture changes via edge computing nodes can be achieved through real-time feature calculation based on sliding windows, event-triggered feature capture, or pipelined parallel processing feature extraction, thereby enabling the extraction of high-value information at the data source, reducing transmission load, and improving the real-time response of the system.

[0089] Taking the cooling monitoring of a flexible rolling production line with multiple steel grades as an example, the edge computing-based metal component quality deviation prediction method in this embodiment can automatically acquire the chemical composition parameters and phase transformation critical point parameters of the current strip steel and query the material-sensor configuration mapping table when DP980 duplex steel and 3% Si silicon steel are alternately produced on the same hot rolling production line. For DP980 steel, due to its fast phase transformation rate and concentrated latent heat release, the infrared thermal imager sampling frame rate is automatically set to 500Hz and the emissivity detection wavelength range is 3-5μm. When switching to silicon steel, due to its slow texture evolution and sensitive radiation characteristics, the frame rate is automatically adjusted to 200Hz and the wavelength range is 8-12μm. The edge computing node then synchronously captures the transient temperature response characteristics and surface radiation dynamic change sequence under the corresponding parameters, ensuring that both different materials can obtain high-quality sensing data that matches their physical properties without the need for machine shutdown or manual intervention.

[0090] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing an edge computing-based metal component quality deviation prediction program, wherein the edge computing-based metal component quality deviation prediction program, when executed by a processor, implements the steps of the edge computing-based metal component quality deviation prediction method as described above.

[0091] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting quality deviation of a metal member based on edge computing, characterized by, The method includes: S1. Construct a multi-dimensional sensing array facing the edge of the metal component to be cooled. The multi-dimensional sensing array includes a high-speed infrared thermal imager and a surface emissivity detection unit. Use the multi-dimensional sensing array to synchronously acquire the spatiotemporal dynamic data sequence of the component cooling process. S2. Obtain the rolling mill production process parameters and the real-time flow rate and operating conditions of the cooling water. Based on the production process parameters and real-time operating conditions, perform dynamic compensation processing on the spatiotemporal dynamic data sequence to eliminate the interference of water film residue and the temperature measurement deviation effect caused by changes in environmental heat radiation, and generate a compensated thermodynamic state feature set. S3. Based on the compensated thermodynamic state feature set, the phase change structure-texture distribution map inside the metal component is constructed by applying the thermo-mechanical coupled field tomography reconstruction algorithm. S4. Perform spatiotemporal adaptive threshold segmentation processing on the phase change structure-texture distribution map to obtain a candidate set of mixed crystal defects and texture abnormal regions. Then, according to the preset microstructure dynamics feature screening rules, the candidate set of defects is judged to determine the true phase change inhomogeneity region and the true texture anisotropy region. S5. For the real phase transformation non-uniform region and the real texture anisotropy region, the hyperbolic tangent-Avramie index joint model is used to calculate the phase transformation volume fraction parameter, texture intensity coefficient parameter and grain size distribution parameter corresponding to each region. When the phase transformation volume fraction parameter of any region exceeds the preset process target range or the texture intensity coefficient parameter exceeds the preset anisotropy critical threshold, it is determined that the metal component has a risk of microstructure and performance deviation that affects subsequent forming or service performance.

2. The method for predicting the quality deviation of metal components based on edge computing as described in claim 1, characterized in that, Step S2 specifically includes: S21. The cooling water impact pressure and water flow fluctuation data on the surface of the metal component are collected in real time by the fluid field monitoring network deployed below the cooling manifold. The pre-constructed water flow parameter-surface heat transfer coefficient mapping relationship table is queried according to the cooling water impact pressure and water flow fluctuation data to obtain the thermal interference compensation coefficient corresponding to the convective heat transfer capacity and film boiling probability at each component segment. S22. For the spatiotemporal data collected by each sensor channel, the thermal interference compensation coefficient is applied to the temperature correction processing of the infrared thermal image signal, and the compensation amount of the effective thermal radiation path is calculated in combination with the oxide film thickness parameter on the surface of the metal component. S23. The corrected spatiotemporal data is fused with the effective thermal radiation path compensation amount at the data level to generate a compensated thermodynamic state feature set including the latent heat release rate of phase change, cooling rate gradient and surface radiation temperature correction value.

3. The method for predicting the quality deviation of metal components based on edge computing as described in claim 2, characterized in that, Step S3 specifically includes: S31. The compensated thermodynamic state feature set is used as an input variable and transmitted to the pre-constructed thermo-mechanical coupling iterative tomography reconstruction model. The thermo-mechanical coupling iterative tomography reconstruction model has built-in prior physical constraints based on finite element phase transition simulation analysis. S32. By constructing an objective function to minimize the weighted mean square error between the actual measured temperature and the model simulation temperature, the spatial distribution matrix of the phase transition field and texture field inside the metal component is solved by iterative inversion, thereby generating a high-resolution phase transition structure-texture distribution map. S33. Perform spatiotemporal joint filtering on the phase transition structure-texture distribution map to effectively suppress artifacts caused by oxide scale shedding or water vapor interference during the tomographic reconstruction process while preserving the sharpness of grain boundary and phase interface features.

4. The method for predicting the quality deviation of metal components based on edge computing as described in claim 3, characterized in that, Step S4 specifically includes: S41. Based on the statistical distribution characteristics of the normal matrix region of the metal component in the phase change structure-texture distribution diagram, a dynamic segmentation threshold is determined. The dynamic segmentation threshold is spatiotemporally adaptively adjusted according to the local temperature variance and the target value of the final cooling temperature. S42. Based on the dynamic segmentation threshold, perform a multidimensional connected region labeling algorithm on the phase change organization-texture distribution map to obtain an initial set of candidate regions for phase change inhomogeneity and an initial set of candidate regions for texture anomalies. S43. For each candidate region in the initial set of candidate regions for phase transformation inhomogeneity, extract its phase transformation dynamics characteristic parameters and thermal hysteresis loop characteristic parameters; for each candidate region in the initial set of candidate regions for texture anomalies, extract its elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters; and match and verify the extracted characteristic parameters with the pre-constructed microstructure dynamics characteristic database. When the phase transformation dynamics characteristic parameters are in the unsteady-state range and the thermal hysteresis loop characteristic parameters are lower than a preset threshold, the candidate region is confirmed as a true phase transformation inhomogeneity region; when the elastic modulus directionality characteristic parameters show significant differences and the grain orientation concentration characteristic parameters exceed the preset texture critical value, the candidate region is confirmed as a true texture anisotropy region.

5. The method for predicting the quality deviation of metal components based on edge computing as described in claim 4, characterized in that, The process of constructing the microscopic tissue dynamics feature database includes: B1. Collect metal component samples with standard materials, prepare microstructure samples with known phase transformation type and texture type, and accurately measure the phase transformation kinetic characteristic parameters, thermal hysteresis loop characteristic parameters, elastic modulus directionality characteristic parameters and grain orientation concentration characteristic parameters of each microstructure sample under simulated laminar flow cooling environment. B2. Based on the phase transformation volume fraction parameter and texture intensity parameter, establish a hierarchical mapping table of microstructure kinetic characteristic parameters, wherein ferrite-pearlite mixed crystals correspond to the first kinetic characteristic range, bainite-martensite mixed crystals correspond to the second kinetic characteristic range, weak texture regions correspond to the first orientation concentration range, and strong texture regions correspond to the second orientation concentration range. B3. Establish a correlation between the microstructure dynamics characteristic parameters in the hierarchical mapping table and the critical influence threshold of the anisotropy index of the mechanical properties of metal components to form a microstructure dynamics characteristic database for distinguishing the authenticity of tissue abnormalities and assessing their severity.

6. The method for predicting the quality deviation of metal components based on edge computing as described in claim 1, characterized in that, It also includes the following steps performed after step S5: S6. Based on the spatial location and distribution density of the real phase transition inhomogeneity region and the real texture anisotropy region, generate a component microstructure performance heat map, and evaluate the performance deviation coefficient of the metal component based on the component microstructure performance heat map combined with the Hall-Page relationship model. S7. When the performance deviation coefficient exceeds the preset quality control threshold, an automatic process optimization analysis report is generated, which includes the spatial distribution characteristic parameters of tissue abnormalities, the severity level, and corresponding cooling process parameter adjustment suggestions.

7. The method for predicting the quality deviation of metal components based on edge computing as described in claim 1, characterized in that, Step S1 specifically includes: S11. Obtain the chemical composition parameters and phase transition critical point parameters of the metal component to be cooled. Based on the chemical composition parameters and phase transition critical point parameters, query the pre-built material-sensor configuration mapping table to determine the infrared thermal imager sampling frame rate and surface emissivity detection wavelength range suitable for this type of metal component. S12. Under the sampling frame rate and wavelength range, control the multi-dimensional sensing array to synchronously scan the component to be cooled, and capture the transient temperature response characteristics caused by the release of latent heat of phase change and the dynamic change sequence of surface radiation caused by texture change in real time through the edge computing node.

8. The method for predicting the quality deviation of metal components based on edge computing as described in claim 1, characterized in that, The spatiotemporal dynamic data sequence is used to characterize the time-frequency evolution of the temperature field of a metal component under laminar cooling with a high heat transfer rate caused by the release of latent heat from phase change, as well as the gradient change of surface radiation properties caused by preferred grain orientation.

9. The method for predicting the quality deviation of metal components based on edge computing as described in claim 1, characterized in that, The phase transformation structure-texture distribution map reflects the microscopic response characteristics of metallic materials under the coupling effect of temperature gradient and phase transformation stress.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a metal component quality deviation prediction program based on edge computing, which, when executed by a processor, implements the steps of the metal component quality deviation prediction method based on edge computing as described in any one of claims 1 to 9.

Images