Non-ferrous metal calendering quality management system and method

Abstract

The present application relates to the technical field of metal processing quality control, in particular to a non-ferrous metal calendering quality management system and method; the system comprises: a calendering working condition data acquisition unit for acquiring calendering process data and microstructure data to generate a calendering structure data tensor; a cavity risk intelligent identification unit for constructing a cavity nucleation potential map and a disturbance response map, and inputting the cavity nucleation potential map and the disturbance response map into a double-branch attention time sequence neural network model to output a cavity risk heat map and a cavity risk score matrix; an interference stripe closed-loop regulation unit for constructing a pseudo-membrane layer evolution map, generating a spray array response coefficient matrix and outputting to a spray cooling valve group as a control quantity; and a calendering quality grading evaluation unit for constructing a multi-grade calendering quality label set and generating a calendering quality evaluation table. The present application realizes intelligent identification of cavity risk and closed-loop quality control of cooling film thickness during metal calendering.

Description

Technical Field

[0001] This invention relates to the field of metal processing quality control technology, specifically to a non-ferrous metal rolling quality management system and method. Background Technology

[0002] In the field of non-ferrous metal processing, metal strip rolling is a crucial foundation for achieving high-precision thin plate manufacturing, and it is widely used in the manufacturing of foils and medium-thick plates of materials such as copper, aluminum, and their alloys. As high-end fields such as aerospace, electronic packaging, and new energy batteries place increasingly higher demands on the uniformity of metal strip microstructure, the integrity of surface film, and the stability of the rolling process, traditional rolling quality control methods that rely on manual experience and periodic sampling are gradually becoming insufficient to meet the real-time, refined, and closed-loop control requirements of modern intelligent manufacturing. Therefore, it is necessary to build an intelligent rolling quality management system.

[0003] Existing non-ferrous metal rolling quality control technologies still have three key problems: First, the acquisition of quality data during the rolling process is incomplete, especially at the microstructure level where there is a lack of structured mapping methods with rolling process parameters, making it impossible to identify defects such as voids, inclusions, and film abnormalities in real time during the rolling process; Second, current spray cooling control is mostly based on empirical parameter adjustments or offline film thickness measurements, making it difficult to dynamically correct control strategies based on local film deviations; Third, rolling quality assessment often relies on manual sampling or result-based testing, lacking quantitative level judgment based on process data. Summary of the Invention

[0004] The purpose of this invention is to provide a quality management system and method for non-ferrous metal rolling to solve the three key problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention aims to provide a non-ferrous metal rolling quality management system, comprising:

[0006] The rolling process data acquisition unit is used to collect rolling process data and microstructure data during hot rolling and cold rolling, and generate rolling structure data tensors based on the time series of data acquisition and the spatial position of the metal strip in the rolling direction and transverse direction.

[0007] The intelligent identification unit for void risk is used to extract the grain structure sub-tensor and tension temperature perturbation channel in the tensor of the rolled structure data to construct a void nucleation potential map and a perturbation response map. The void nucleation potential map and the perturbation response map are then input into a dual-branch attention temporal neural network model, and the output is a void risk heat map and a void risk scoring matrix.

[0008] The interference fringe closed-loop control unit is used to acquire the interference image sequence of the cold press outlet, extract the phase transfer features of the interference fringes and reconstruct the film thickness response map, and construct a pseudo film evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, the spray array response coefficient matrix is ​​generated and output to the spray cooling valve group as a control quantity.

[0009] The calendering quality grading assessment unit is based on the void risk scoring matrix and the spray array response coefficient matrix. It uses quality grade mapping rules to construct a multi-level calendering quality label set and combines it with the calendering structure data tensor to generate a calendering quality assessment table.

[0010] Preferably, in the rolling process data acquisition unit, the rolling process data includes the rolling force, roll speed, roll temperature, roll gap, and metal strip tension; the microstructure data includes grain size distribution, twin density, and phase boundary information obtained through integrated metallographic image analysis.

[0011] The calendering structure data tensor is obtained by mapping and fusing calendering process data and microstructure data, with time, longitudinal position and lateral position as three-dimensional index dimensions.

[0012] Preferably, the cavity risk intelligent identification unit includes a structural disturbance channel extraction module;

[0013] The structural perturbation channel extraction module is used to extract multi-channel data tensors of rolled structure data to obtain grain structure sub-tensors and tension temperature perturbation channels.

[0014] Among them, the grain structure subtensor is a part of the rolling structure data tensor, including grain size distribution, twin density and phase boundary number; the tension-temperature perturbation channel is a perturbation sequence matrix formed by tension data and temperature data during the rolling process, and is constructed based on the time series gradient change, used to characterize the response of tension and temperature to the metal microstructure.

[0015] Preferably, the cavity risk intelligent identification unit further includes a cavity potential mapping construction module;

[0016] The void potential mapping construction module is used to construct a void nucleation potential map based on the grain structure subtensor and to construct a perturbation response map based on the tension temperature perturbation channel. The specific method steps are as follows:

[0017] S22.1 Based on the grain size distribution, twin density and phase boundary number in the grain structure subtensor, calculate the local difference rate of grain size, twin density variability and phase boundary distribution gradient, and construct the void nucleation potential map by weighted linear combination.

[0018] S22.2 Perform first-order and second-order time difference operations on the tension and temperature data in the tension-temperature disturbance channel, respectively, and introduce a weighted response function based on the transverse spatial coordinates of rolling to generate a disturbance response diagram.

[0019] Preferably, the cavity risk intelligent identification unit further includes a risk assessment neural network module;

[0020] The risk assessment neural network module is based on a dual-branch attention temporal neural network model, and uses the cavity nucleus potential map and the disturbance response map as inputs to obtain a cavity risk heat map and a cavity risk scoring matrix.

[0021] The dual-branch attention temporal neural network model includes a structural feature branch and a perturbation response branch. The structural feature branch receives the cavity nucleation potential map and extracts the spatial evolution features of the microstructure through graph embedding and spatial attention mechanisms. The perturbation response branch receives the perturbation response map and extracts the perturbation intensity variation pattern through temporal convolution and channel attention mechanisms. The structural feature branch and the perturbation response branch are concatenated by a fusion layer, and a cavity risk heatmap and a cavity risk scoring matrix are generated by a fully connected layer.

[0022] The cavity risk heat map is a spatial distribution map used to indicate the location of cavity formation; the cavity risk scoring matrix is ​​a two-dimensional risk numerical matrix of the metal strip along the length and width directions, used to quantitatively calculate the cavity risk.

[0023] Preferably, the interference fringe closed-loop control unit includes an interference feature extraction module;

[0024] The interference feature extraction module is used to reconstruct the film thickness response map. The specific steps are as follows:

[0025] S31.1 Acquire the interference image sequence of rolled metal strip at the exit of the cold press, and register and correct the distortion of the interference image sequence according to the time axis;

[0026] S31.2 For all images in the interference image sequence, the phase shift interferometry is used to extract the phase shift features of the interference fringes and generate a pixel-level phase distribution map;

[0027] S31.3 Calculate the phase gradient and phase transition region based on the phase distribution map, and generate the film thickness response map;

[0028] The film thickness response map is a two-dimensional spatial distribution map obtained by converting the fringe phase transfer features in the interference image sequence through the phase thickness mapping relationship. It is used to characterize the relative thickness distribution and local gradient change of the film layer on the surface of the metal strip at the cold pressing exit.

[0029] The phase shift interferometry method includes calculating the interference phase distribution based on the phase shift interferometry principle and constructing a stable fringe evolution model through multi-frame comparison, which is used to characterize the spatial non-uniformity of the film thickness on the surface of the metal strip before and after cooling control.

[0030] Preferably, the interference fringe closed-loop control unit further includes a film evolution modeling module;

[0031] The membrane evolution modeling module is used to construct a pseudo membrane evolution map based on the cavity risk heatmap and membrane thickness response map. The specific steps are as follows:

[0032] S32.1. Use the cavity risk heat map as a regional weight map and overlay it on the film thickness response map to construct a pseudo film layer evolution map;

[0033] S32.2 Calculate the phase error field between the film thickness response diagram and the pseudo-film evolution diagram, and extract the film thickness phase residual function;

[0034] S32.3 Adjust the local fluctuation range of the film thickness phase residual function by using curvature normalization transformation and weighting the lateral position of the metal strip;

[0035] Among them, the pseudo-film evolution map is a two-dimensional mapping image constructed based on the cavity risk heat map, which is used to evolve the spatiotemporal variation trend of potential film thickness anomalies in specific regions during cold pressing.

[0036] The film thickness phase residual function is a function calculated from the phase difference between the film thickness response map and the pseudo-film evolution map. It is used to quantify the degree of deviation between the film thickness response and the potential anomalous region.

[0037] Preferably, the interference fringe closed-loop control unit further includes a spray response control module;

[0038] The spray response control module is used to calculate the spray array response coefficient matrix based on the film thickness phase residual function, and convert the spray array response coefficient matrix into a driving signal to dynamically regulate the cooling flow rate and spray distribution of the spray cooling valve group. The specific method steps are as follows:

[0039] S33.1 Calculate the heat flux adjustment value of each nozzle region in the transverse direction of the metal strip based on the film thickness phase residual function;

[0040] S33.2. Introduce a weighting function based on nozzle spacing and coverage area to construct the spray array response coefficient matrix;

[0041] S33.3 Output the spray array response coefficient matrix to the spray cooling valve group in real time to control the cooling flow rate and spray distribution.

[0042] Preferably, the calendering quality grading assessment unit calculates the calendering quality grade scoring vector based on the cavity risk scoring matrix and the spray array response coefficient matrix, and uses the quality grade mapping rule to convert the calendering quality grade scoring vector into a calendering quality label set, and combines the calendering structure data tensor to generate a calendering quality assessment table;

[0043] The rolling quality grade scoring vector is a quantitative representation of the quality assessment results of each rolling section based on defect risk and control response; the rolling quality label set is used to map the continuous scoring vector into discrete labels, which are divided into four levels: excellent, good, acceptable, and abnormal; the rolling quality assessment table is used to bind the label set with the corresponding spatiotemporal coordinates and output the rolling quality assessment table.

[0044] The rolling quality assessment form is used to display the quality of non-ferrous metal strips at each stage of rolling.

[0045] On the other hand, the present invention provides a method for quality management of non-ferrous metal rolling, used in the non-ferrous metal rolling quality management system described above, comprising the following steps:

[0046] S10.1 Collect rolling process data and microstructure data during hot rolling and cold rolling, and generate rolling structure data tensor based on the time series of data collection and the spatial position of metal strip in the rolling direction and transverse direction.

[0047] S10.2 Extract the grain structure sub-tensor and tension temperature perturbation channel from the tensor of the rolled structure data to construct the void nucleation potential map and perturbation response map. Input the void nucleation potential map and perturbation response map into the dual-branch attention temporal neural network model to output the void risk heat map and void risk score matrix.

[0048] S10.3 Acquire the interference image sequence of the cold pressing outlet, extract the phase transfer features of the interference fringes and reconstruct the film thickness response map, and construct the pseudo film evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, generate the spray array response coefficient matrix and output it to the spray cooling valve group as a control quantity.

[0049] S10.4 Based on the cavity risk scoring matrix and the spray array response coefficient matrix, a multi-level calendering quality label set is constructed using quality level mapping rules, and a calendering quality evaluation table is generated by combining the calendering structure data tensor.

[0050] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0051] 1. In this invention, by constructing a rolling structure data tensor and fusing the grain structure sub-tensor and tension temperature perturbation channel, it is possible to achieve synergistic perception of the microstructure of metal strip and rolling process perturbation, thereby accurately identifying the void formation potential and perturbation response characteristics in the rolling process, and improving the prediction accuracy and feedforward control capability of quality defects.

[0052] 2. In this invention, by constructing a film thickness phase residual function based on the film thickness response map and the void risk thermal map, and calculating the spray array response coefficient matrix, dynamic modeling and control feedback of the spatial distribution of film thickness can be realized during the rolling process, thereby improving the accuracy of cooling control and the ability to control the uniformity of the film layer of metal strip, and ultimately accurately managing and controlling the quality of metal strip. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of one embodiment of the present invention;

[0054] Figure labels: 1. Rolling condition data acquisition unit; 2. Cavity risk intelligent identification unit; 21. Structural disturbance channel extraction module; 22. Cavity potential mapping construction module; 23. Risk assessment neural network module; 3. Interference fringe closed-loop control unit; 31. Interference feature extraction module; 32. Film evolution modeling module; 33. Spray response control module; 4. Rolling quality grading assessment unit. Detailed Implementation

[0055] Example 1, as Figure 1 As shown, a non-ferrous metal rolling quality management system is provided, including:

[0056] The rolling process data acquisition unit 1 is used to acquire rolling process data and microstructure data during hot rolling and cold rolling, and generate rolling structure data tensor based on the time series of data acquisition and the spatial position of metal strip in the rolling direction and transverse direction.

[0057] In the rolling process data acquisition unit 1 of this embodiment, the rolling process data includes the rolling force, roll speed, roll temperature, roll gap and metal strip tension; the microstructure data includes the grain size distribution, twin density and phase boundary information obtained by integrated metallographic image analysis.

[0058] The calendering structure data tensor is obtained by mapping and fusing calendering process data and microstructure data, with time, longitudinal position and lateral position as three-dimensional index dimensions.

[0059] In this embodiment, the rolling process data acquisition unit 1 uses a multi-channel data acquisition method and an embedded metallographic imaging module to acquire rolling process data and microstructure data during hot rolling and cold rolling processes, respectively, and records them as a time series at a constant sampling frequency. Based on the time stamp, rolling direction and transverse direction of the same rolling region, the above two types of data are indexed and fused to form a data tensor with a three-dimensional spatial index as the structural skeleton, which is defined as the rolling structure data tensor. Each index point in the rolling structure data tensor contains the coupling result of thermodynamic parameters and microstructure information.

[0060] The intelligent identification unit 2 for void risk is used to extract the grain structure sub-tensor and tension temperature perturbation channel in the tensor of the rolled structure data to construct a void nucleation potential map and a perturbation response map. The void nucleation potential map and the perturbation response map are then input into a dual-branch attention temporal neural network model, and the output is a void risk heat map and a void risk scoring matrix.

[0061] In this embodiment, the cavity risk intelligent identification unit 2 includes a structural disturbance channel extraction module 21;

[0062] The structural disturbance channel extraction module 21 is used to extract multi-channel data tensors of rolled structure data to obtain grain structure sub-tensors and tension temperature disturbance channels.

[0063] Among them, the grain structure subtensor is a part of the rolling structure data tensor, including grain size distribution, twin density and phase boundary number; the tension-temperature perturbation channel is a perturbation sequence matrix formed by tension data and temperature data during the rolling process, and is constructed based on the time series gradient change, used to characterize the response of tension and temperature to the metal microstructure.

[0064] In this embodiment, multi-channel extraction of the rolling structure data tensor is based on the tensor channel decomposition strategy. The rolling structure data tensor is divided into channels in the physical parameter dimension, and local extraction is performed in combination with the spatial position and time series dimensions, so that the grain structure sub-tensor and the tension temperature perturbation channel maintain index correspondence and physical consistency in the tensor structure.

[0065] The perturbation sequence matrix is ​​based on continuous sensor sampling data of tension and temperature during the rolling process. The local gradient change rate is calculated using first-order time difference and then normalized and denoised to form a tension-temperature perturbation channel. This channel is used to quantify the thermal perturbation change trend of the metal strip at multiple locations.

[0066] In this embodiment, the tension-temperature perturbation channel captures the stress-strain state changes during the rolling process and, combined with the physical mechanisms of microstructure evolution, such as thermally induced recrystallization and dynamic recovery, indirectly reflects the potential trend of grain structure evolution with perturbation, which is to characterize the response of tension and temperature to the metal microstructure.

[0067] In this embodiment, the cavity risk intelligent identification unit 2 further includes a cavity potential mapping construction module 22;

[0068] The void potential mapping construction module 22 is used to construct a void nucleation potential map based on the grain structure sub-tensor and to construct a perturbation response map based on the tension temperature perturbation channel. The specific method steps are as follows:

[0069] S22.1 Based on the grain size distribution, twin density and phase boundary number in the grain structure subtensor, calculate the local difference rate of grain size, twin density variability and phase boundary distribution gradient, and construct the void nucleation potential map by weighted linear combination.

[0070] S22.2 Perform first-order and second-order time difference operations on the tension and temperature data in the tension-temperature disturbance channel, respectively, and introduce a weighted response function based on the transverse spatial coordinates of rolling to generate a disturbance response diagram.

[0071] In this embodiment, the local difference rate of grain size is the normalized difference value of the average diameter of grains at different sampling points within a unit interval of the rolling process. The sliding window method is used to calculate the degree of grain size variation between multiple adjacent points. The twin density variability is obtained by calculating the standard deviation of the twin density at all locations in the grain structure subtensor. The phase boundary distribution gradient is based on the phase boundary number density per unit area in different locations. The first-order spatial difference method is used to calculate its rate of change in the rolling process.

[0072] In this embodiment, the weight coefficients of each item in the weighted linear combination are set based on historical experience data and expert experience rules, and are used to construct a cavity nucleation potential map, representing the spatial distribution of the possibility of cavity formation caused by microstructure.

[0073] In this embodiment, the first-order difference is used to calculate the instantaneous rate of change of tension and temperature during the rolling process, and the second-order difference is used to reflect the intensity of the fluctuation. The weighted response function assigns different importance weights to different transverse coordinate points, emphasizing the difference in sensitivity of the central metal strip and the edge position to the void response, and finally generates a disturbance response map.

[0074] In this embodiment, the cavity risk intelligent identification unit 2 further includes a risk assessment neural network module 23;

[0075] The risk assessment neural network module 23 is based on a dual-branch attention temporal neural network model, and combines the cavity nucleus potential map and the disturbance response map as inputs to obtain a cavity risk heat map and a cavity risk scoring matrix.

[0076] The dual-branch attention temporal neural network model includes a structural feature branch and a perturbation response branch. The structural feature branch receives the cavity nucleation potential map and extracts the spatial evolution features of the microstructure through graph embedding and spatial attention mechanisms. The perturbation response branch receives the perturbation response map and extracts the perturbation intensity variation pattern through temporal convolution and channel attention mechanisms. The structural feature branch and the perturbation response branch are concatenated by a fusion layer, and a cavity risk heatmap and a cavity risk scoring matrix are generated by a fully connected layer.

[0077] The cavity risk heat map is a spatial distribution map used to indicate the location of cavity formation; the cavity risk scoring matrix is ​​a two-dimensional risk numerical matrix of the metal strip along the length and width directions, used to quantitatively calculate the cavity risk.

[0078] In this embodiment, the input void nucleation potential map is represented as a node feature matrix and an adjacency tensor using a graph embedding method. High-order spatial features of each node in the graph structure under local topological and microstructural indices are extracted through graph convolution operations and input into the spatial attention mechanism in tensor form. The spatial attention mechanism is used to construct a spatial weight map based on the graph embedding feature matrix, assigning higher weights to the spatial locations of regions with potential void nucleation potential. It consists of multi-layer dot-multiplication attention calculation, normalization function and residual connection, and is used to emphasize structural heterogeneous regions with void risk.

[0079] In this embodiment, the temporal convolution method adopts a one-dimensional convolutional network structure. The input is a stepwise sequence tensor of the perturbation response map in the time dimension. It extracts the local variation law of tension and temperature perturbation over time and uses it to predict the sudden change in perturbation intensity and potential causes. The channel attention mechanism uses the Squeeze-and-Excitation mechanism to model the importance of each channel of the perturbation response map. After compressing the spatial dimension, it generates channel weighting factors and then weights and enhances the original perturbation channels.

[0080] In this embodiment, the fusion layer concatenates the tensors output by the structural feature branch and the disturbance response branch along the channel dimension to form a unified joint feature representation, which is then input into the fully connected layer to generate a void risk heatmap and a void risk scoring matrix. The void risk heatmap is a two-dimensional spatial matrix whose size is consistent with the horizontal and vertical coordinates of the rolling structure data tensor, and is used for spatial location-level void risk visualization. The void risk scoring matrix is ​​a two-dimensional matrix obtained by discretizing the horizontal and vertical spatial components of the metal strip, where each matrix unit represents the risk score at the corresponding location, supporting hot zone clustering and defect trend assessment.

[0081] Interference fringe closed-loop control unit 3 is used to acquire the interference image sequence of the cold press outlet, extract the phase transfer features of the interference fringes and reconstruct the film thickness response map, and construct a pseudo film layer evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, the spray array response coefficient matrix is ​​generated and output to the spray cooling valve group as a control quantity.

[0082] In this embodiment, the interference fringe closed-loop control unit 3 includes an interference feature extraction module 31;

[0083] The interference feature extraction module 31 is used to reconstruct the film thickness response map. The specific method steps are as follows:

[0084] S31.1 Acquire the interference image sequence of rolled metal strip at the exit of the cold press, and register and correct the distortion of the interference image sequence according to the time axis;

[0085] S31.2 For all images in the interference image sequence, the phase shift interferometry is used to extract the phase shift features of the interference fringes and generate a pixel-level phase distribution map;

[0086] S31.3 Calculate the phase gradient and phase transition region based on the phase distribution map, and generate the film thickness response map;

[0087] The film thickness response map is a two-dimensional spatial distribution map obtained by converting the fringe phase transfer features in the interference image sequence through the phase thickness mapping relationship. It is used to characterize the relative thickness distribution and local gradient change of the film layer on the surface of the metal strip at the cold pressing exit.

[0088] The phase shift interferometry method includes calculating the interference phase distribution based on the phase shift interferometry principle and constructing a stable fringe evolution model through multi-frame comparison, which is used to characterize the spatial non-uniformity of the film thickness on the surface of the metal strip before and after cooling control.

[0089] In this embodiment, the interference image sequence is registered and distortion corrected according to the time axis. Specifically, the spatial position of the interference images of different time frames of the same metal strip region is aligned by using a sub-pixel-level time frame matching algorithm to eliminate displacement deviation caused by conveyor belt vibration or optical disturbance. Distortion correction uses a radial distortion model to calculate the optical axis deviation and nonlinear lens distortion, and performs geometric reconstruction of the image to ensure the consistency of fringe morphology and the accuracy of phase analysis.

[0090] In this embodiment, the phase shift interferometry method is based on an image sequence with a known phase difference introduced into a coherent light interferometric image. The phase information is calculated by the change of gray values ​​between multiple interferometric images. A five-step phase shift algorithm is used to reverse the gray level of the stripes, which can extract a high-precision phase distribution map and suppress the interference of illumination fluctuations.

[0091] In this embodiment, the phase thickness mapping relationship is established based on the principle of film optical path difference, and the mapping relationship is as follows:

[0092] ;

[0093] in, The thickness of the film layer; The wavelength of the interference light; The refractive index of the film; for ; For the spatial coordinates of the rolled metal strip in the transverse direction; The coordinates of the rolled metal strip are in the longitudinal direction.

[0094] Phase thickness mapping converts the phase distribution map into a film thickness response map, thereby reconstructing the spatial distribution of film thickness.

[0095] In this embodiment, the film thickness response map is not only used to reflect the average thickness of the film layer, but also to extract the local gradient changes of the film layer through the Laplacian gradient operator, and to identify regions of film discontinuity or abrupt thickness change caused by uneven cooling.

[0096] In this embodiment, the phase shift interferometry method adopts a strategy that combines multi-frame temporal averaging filtering with fringe stability modeling. The steps are as follows: first, pixel-level difference is performed on the fringe structure of different frames to construct a fringe evolution trajectory map; then, high-confidence frames are selected by regional stability judgment for phase superposition; finally, phase noise is suppressed by variational minimization model.

[0097] In this embodiment, the interference fringe closed-loop control unit 3 further includes a film evolution modeling module 32;

[0098] The membrane evolution modeling module 32 is used to construct a pseudo membrane evolution map based on the cavity risk heatmap and membrane thickness response map. The specific method steps are as follows:

[0099] S32.1. Use the cavity risk heat map as a regional weight map and overlay it on the film thickness response map to construct a pseudo film layer evolution map;

[0100] S32.2 Calculate the phase error field between the film thickness response diagram and the pseudo-film evolution diagram, and extract the film thickness phase residual function;

[0101] S32.3 Adjust the local fluctuation range of the film thickness phase residual function by using curvature normalization transformation and weighting the lateral position of the metal strip;

[0102] Among them, the pseudo-film evolution map is a two-dimensional mapping image constructed based on the cavity risk heat map, which is used to evolve the spatiotemporal variation trend of potential film thickness anomalies in specific regions during cold pressing.

[0103] The film thickness phase residual function is a function calculated from the phase difference between the film thickness response map and the pseudo-film evolution map. It is used to quantify the degree of deviation between the film thickness response and the potential anomalous region.

[0104] In this embodiment, the pseudo-film evolution map is a second-order mapping map that integrates surface film thickness data and microstructure risk indicators. Its construction logic is as follows: the risk intensity of each pixel in the void risk heat map is used as the local enhancement weight of the film thickness response map, and the film thickness response map is regionally weighted and superimposed to make the film changes in high-risk areas more prominent, thereby reflecting the dynamic abnormal evolution trend of film thickness at both the visual and computational levels.

[0105] In this embodiment, the film thickness phase residual function refers to the difference function of the corresponding phase values ​​between the film thickness response map and the pseudo film evolution map. It is used to quantify the ideal film state, that is, the cooling film thickness that the spray control should achieve. The film thickness phase residual function is used to reveal the location of local overcooling or undercooling of the film and to automatically adjust the cooling flow rate and spray distribution.

[0106] In this embodiment, curvature normalization transformation is used to eliminate evaluation bias caused by differences in film curvature in different spatial regions, and to prevent local thickness abrupt changes from being misjudged as abnormal. The lateral position weighting mechanism is based on the physical sensitivity weight of the lateral spatial position of the metal strip. Considering that film fluctuations at the edge of the metal strip have a more significant impact on the final product quality, a higher weight needs to be assigned in the film thickness phase residual function calculation.

[0107] In this embodiment, the interference fringe closed-loop control unit 3 further includes a spray response control module 33;

[0108] The spray response control module 33 is used to calculate the spray array response coefficient matrix based on the film thickness phase residual function, and convert the spray array response coefficient matrix into a driving signal to dynamically regulate the cooling flow rate and spray distribution of the spray cooling valve group. The specific method steps are as follows:

[0109] S33.1 Calculate the heat flux adjustment value of each nozzle region in the transverse direction of the metal strip based on the film thickness phase residual function;

[0110] S33.2. Introduce a weighting function based on nozzle spacing and coverage area to construct the spray array response coefficient matrix;

[0111] S33.3 Output the spray array response coefficient matrix to the spray cooling valve group in real time to control the cooling flow rate and spray distribution.

[0112] In this embodiment, the spray array response coefficient matrix is ​​a two-dimensional spatial mapping matrix, representing the spray intensity ratio that each nozzle control unit in the transverse direction of the metal strip should adjust within the current feedback control cycle; each element of the spray array response coefficient matrix corresponds to the heat flux compensation requirement of a nozzle in its coverage area, with the unit being the normalized response coefficient, and the value range being [0,1]; the spray array response coefficient matrix is ​​used to establish a direct response mapping relationship between the physical film thickness residual and the spatial spray execution action.

[0113] In this embodiment, the heat flux adjustment value is the local energy regulation requirement value derived from the film thickness phase residual function. The specific calculation method can be achieved by setting the sensitivity coefficient of film thickness change to heat flux, i.e., the film thermal conductivity inverse deduction coefficient, and establishing a mapping function.

[0114] In this embodiment, the nozzle spacing and coverage area weighting function is used to solve the regulation interference problem caused by the overlap of multiple nozzle control areas in spray control. The nozzle spacing and coverage area weighting function is a Gaussian function, which is used to weight and integrate the response coefficients according to the coverage efficiency of each nozzle in the lateral space, so as to avoid the generation of overcooled areas by overlapping nozzles at the boundary.

[0115] The calendering quality grading assessment unit 4 is based on the void risk scoring matrix and the spray array response coefficient matrix. It uses the quality grade mapping rule to construct a multi-level calendering quality label set and combines the calendering structure data tensor to generate a calendering quality assessment table.

[0116] In this embodiment, the calendering quality grading evaluation unit 4 calculates the calendering quality grade scoring vector based on the cavity risk scoring matrix and the spray array response coefficient matrix, and uses the quality grade mapping rule to convert the calendering quality grade scoring vector into a calendering quality label set, and combines the calendering structure data tensor to generate a calendering quality evaluation table;

[0117] The rolling quality grade scoring vector is a quantitative representation of the quality assessment results of each rolling section based on defect risk and control response; the rolling quality label set is used to map the continuous scoring vector into discrete labels, which are divided into four levels: excellent, good, acceptable, and abnormal; the rolling quality assessment table is used to bind the label set with the corresponding spatiotemporal coordinates and output the rolling quality assessment table.

[0118] The rolling quality assessment form is used to display the quality of non-ferrous metal strips at each stage of rolling.

[0119] In this embodiment, the quality grade mapping rule is set based on the joint influence range between the cavity risk score and the spray control response. A linear or nonlinear boundary function is used to divide the rolling quality grade score vector into intervals. The boundary of each grade is dynamically fine-tuned through regression analysis of historical rolling assessment data to achieve adaptive optimization of the quality grade determination. The foundation for constructing the quality grade mapping rule should include: the threshold interval division logic of the cavity risk score: a score of 0.0–0.25 is excellent, 0.25–0.5 is good, 0.5–0.75 is acceptable, and 0.75–1 is abnormal. The quality grade mapping rule is dynamically adjustable and is obtained through statistical training using historical quality assessment data.

[0120] In this embodiment, the rolling quality assessment table is a structured table containing the following fields: time, horizontal position, vertical position, quality label, score value, and residual trend, etc. It is used to perform visual analysis and data retention of the overall quality of the current rolling batch, and can serve as the basis for decision-making for subsequent quality traceability, parameter correction, and defect warning.

[0121] In this embodiment, the calendering section is divided into equally spaced sub-regions based on the horizontal and vertical indices of the calendering structure data tensor, and each calendering section corresponds to a small sub-tensor region in the tensor.

[0122] Example 2: This invention proposes a non-ferrous metal rolling quality management method, used in the non-ferrous metal rolling quality management system described in Example 1 above, comprising the following steps:

[0123] S10.1 Collect rolling process data and microstructure data during hot rolling and cold rolling, and generate rolling structure data tensor based on the time series of data collection and the spatial position of metal strip in the rolling direction and transverse direction.

[0124] S10.2 Extract the grain structure sub-tensor and tension temperature perturbation channel from the tensor of the rolled structure data to construct the void nucleation potential map and perturbation response map. Input the void nucleation potential map and perturbation response map into the dual-branch attention temporal neural network model to output the void risk heat map and void risk score matrix.

[0125] S10.3 Acquire the interference image sequence of the cold pressing outlet, extract the phase transfer features of the interference fringes and reconstruct the film thickness response map, and construct the pseudo film evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, generate the spray array response coefficient matrix and output it to the spray cooling valve group as a control quantity.

[0126] S10.4 Based on the cavity risk scoring matrix and the spray array response coefficient matrix, a multi-level calendering quality label set is constructed using quality level mapping rules, and a calendering quality evaluation table is generated by combining the calendering structure data tensor.

[0127] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A quality management system for non-ferrous metal rolling, characterized in that, include: The rolling process data acquisition unit (1) is used to acquire rolling process data and microstructure data during hot rolling and cold rolling, and to generate rolling structure data tensor based on the time sequence of data acquisition and the spatial position of metal strip in the rolling direction and transverse direction. The intelligent identification unit for void risk (2) is used to extract the grain structure sub-tensor and tension temperature perturbation channel in the tensor of the rolled structure data to construct the void nucleation potential map and perturbation response map. The void nucleation potential map and perturbation response map are input into the dual-branch attention temporal neural network model, and the void risk heat map and void risk score matrix are output. Interference fringe closed-loop control unit (3) is used to collect the interference image sequence of the cold press outlet, extract the phase transfer features of the interference fringe and reconstruct the film thickness response map, and construct the pseudo film layer evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, the spray array response coefficient matrix is ​​generated and output to the spray cooling valve group as the control quantity. The calendering quality grading assessment unit (4) is based on the void risk scoring matrix and the spray array response coefficient matrix. It uses the quality grade mapping rule to construct a multi-level calendering quality label set and combines the calendering structure data tensor to generate a calendering quality assessment table.

2. The non-ferrous metal rolling quality management system according to claim 1, characterized in that, In the rolling process data acquisition unit (1), the rolling process data includes the rolling force, roll speed, roll temperature, roll gap and metal strip tension; the microstructure data includes the grain size distribution, twin density and phase boundary information obtained by integrated metallographic image analysis. The calendering structure data tensor is obtained by mapping and fusing calendering process data and microstructure data, with time, longitudinal position and lateral position as three-dimensional index dimensions.

3. The non-ferrous metal rolling quality management system according to claim 2, characterized in that, The cavity risk intelligent identification unit (2) includes a structural disturbance channel extraction module (21). The structure disturbance channel extraction module (21) is used to extract the tensor of the rolled structure data in multiple channels to obtain the grain structure sub-tensor and the tension temperature disturbance channel; Among them, the grain structure subtensor is a part of the rolling structure data tensor, including grain size distribution, twin density and phase boundary number; the tension-temperature perturbation channel is a perturbation sequence matrix formed by tension data and temperature data during the rolling process, and is constructed based on the time series gradient change, used to characterize the response of tension and temperature to the metal microstructure.

4. The non-ferrous metal rolling quality management system according to claim 3, characterized in that, The cavity risk intelligent identification unit (2) also includes a cavity potential mapping construction module (22). The void potential mapping construction module (22) is used to construct a void nucleation potential map based on the grain structure sub-tensor and to construct a perturbation response map based on the tension temperature perturbation channel. The specific method steps are as follows: S22.1 Based on the grain size distribution, twin density and phase boundary number in the grain structure subtensor, calculate the local difference rate of grain size, twin density variability and phase boundary distribution gradient, and construct the void nucleation potential map by weighted linear combination. S22.2 Perform first-order and second-order time difference operations on the tension and temperature data in the tension-temperature disturbance channel, respectively, and introduce a weighted response function based on the transverse spatial coordinates of rolling to generate a disturbance response diagram.

5. The non-ferrous metal rolling quality management system according to claim 4, characterized in that, The cavity risk intelligent identification unit (2) also includes a risk assessment neural network module (23). The risk assessment neural network module (23) is based on a dual-branch attention temporal neural network model, and combines the cavity nucleus potential map and the disturbance response map as inputs to obtain a cavity risk heat map and a cavity risk scoring matrix. The dual-branch attention temporal neural network model includes a structural feature branch and a perturbation response branch. The structural feature branch receives the cavity nucleation potential map and extracts the spatial evolution features of the microstructure through graph embedding and spatial attention mechanisms. The perturbation response branch receives the perturbation response map and extracts the perturbation intensity variation pattern through temporal convolution and channel attention mechanisms. The structural feature branch and the perturbation response branch are concatenated by a fusion layer, and a cavity risk heatmap and a cavity risk scoring matrix are generated by a fully connected layer. The cavity risk heat map is a spatial distribution map used to indicate the location of cavity formation; the cavity risk scoring matrix is ​​a two-dimensional risk numerical matrix of the metal strip along the length and width directions, used to quantitatively calculate the cavity risk.

6. The non-ferrous metal rolling quality management system according to claim 5, characterized in that, The interference fringe closed-loop control unit (3) includes an interference feature extraction module (31); The interference feature extraction module (31) is used to reconstruct the film thickness response map. The specific steps are as follows: S31.1 Acquire the interference image sequence of rolled metal strip at the exit of the cold press, and register and correct the distortion of the interference image sequence according to the time axis; S31.2 For all images in the interference image sequence, the phase shift interferometry is used to extract the phase shift features of the interference fringes and generate a pixel-level phase distribution map; S31.3 Calculate the phase gradient and phase transition region based on the phase distribution map, and generate the film thickness response map; The film thickness response map is a two-dimensional spatial distribution map obtained by converting the fringe phase transfer features in the interference image sequence through the phase thickness mapping relationship. It is used to characterize the relative thickness distribution and local gradient change of the film layer on the surface of the metal strip at the cold pressing exit. The phase shift interferometry method includes calculating the interference phase distribution based on the phase shift interferometry principle and constructing a stable fringe evolution model through multi-frame comparison, which is used to characterize the spatial non-uniformity of the film thickness on the surface of the metal strip before and after cooling control.

7. The non-ferrous metal rolling quality management system according to claim 6, characterized in that, The interference fringe closed-loop control unit (3) also includes a film evolution modeling module (32). The membrane evolution modeling module (32) is used to construct a pseudo membrane evolution map based on the cavity risk heat map and the membrane thickness response map. The specific method steps are as follows: S32.

1. Use the cavity risk heat map as a regional weight map and overlay it on the film thickness response map to construct a pseudo film layer evolution map; S32.2 Calculate the phase error field between the film thickness response diagram and the pseudo-film evolution diagram, and extract the film thickness phase residual function; S32.3 Adjust the local fluctuation range of the film thickness phase residual function by using curvature normalization transformation and weighting the lateral position of the metal strip; Among them, the pseudo-film evolution map is a two-dimensional mapping image constructed based on the cavity risk heat map, which is used to evolve the spatiotemporal variation trend of potential film thickness anomalies in specific regions during cold pressing. The film thickness phase residual function is a function calculated from the phase difference between the film thickness response map and the pseudo-film evolution map. It is used to quantify the degree of deviation between the film thickness response and the potential anomalous region.

8. The non-ferrous metal rolling quality management system according to claim 7, characterized in that, The interference stripe closed-loop control unit (3) also includes a spray response control module (33). The spray response control module (33) is used to calculate the spray array response coefficient matrix based on the film thickness phase residual function, and convert the spray array response coefficient matrix into a driving signal to dynamically regulate the cooling flow rate and spray distribution of the spray cooling valve group. The specific method steps are as follows: S33.1 Calculate the heat flux adjustment value of each nozzle region in the transverse direction of the metal strip based on the film thickness phase residual function; S33.

2. Introduce a weighting function based on nozzle spacing and coverage area to construct the spray array response coefficient matrix; S33.3 Output the spray array response coefficient matrix to the spray cooling valve group in real time to control the cooling flow rate and spray distribution.

9. The non-ferrous metal rolling quality management system according to claim 8, characterized in that, The calendering quality grading assessment unit (4) calculates the calendering quality grade scoring vector based on the cavity risk scoring matrix and the spray array response coefficient matrix, and uses the quality grade mapping rule to convert the calendering quality grade scoring vector into a calendering quality label set, and combines the calendering structure data tensor to generate a calendering quality assessment table; The rolling quality grade scoring vector is a quantitative representation of the quality assessment results of each rolling section based on defect risk and control response; the rolling quality label set is used to map the continuous scoring vector into discrete labels, which are divided into four levels: excellent, good, acceptable, and abnormal; the rolling quality assessment table is used to bind the label set with the corresponding spatiotemporal coordinates and output the rolling quality assessment table. The rolling quality assessment form is used to display the quality of non-ferrous metal strips at each stage of rolling.

10. A method for quality management of non-ferrous metal rolling, used in the non-ferrous metal rolling quality management system as described in any one of claims 1-9, characterized in that: Includes the following steps: S10.1 Collect rolling process data and microstructure data during hot rolling and cold rolling, and generate rolling structure data tensor based on the time series of data collection and the spatial position of metal strip in the rolling direction and transverse direction. S10.2 Extract the grain structure sub-tensor and tension temperature perturbation channel from the tensor of the rolled structure data to construct the void nucleation potential map and perturbation response map. Input the void nucleation potential map and perturbation response map into the dual-branch attention temporal neural network model to output the void risk heat map and void risk score matrix. S10.3 Acquire the interference image sequence of the cold pressing outlet, extract the phase transfer features of the interference fringes and reconstruct the film thickness response map, and construct the pseudo film evolution map based on the cavity risk heat map. By calculating the film thickness phase residual function, generate the spray array response coefficient matrix and output it to the spray cooling valve group as a control quantity. S10.4 Based on the cavity risk scoring matrix and the spray array response coefficient matrix, a multi-level calendering quality label set is constructed using quality level mapping rules, and a calendering quality evaluation table is generated by combining the calendering structure data tensor.

Images