Rolling mill production line self-adaptive control method and computer equipment

Abstract

The present application relates to a rolling production line adaptive control method and computer equipment, by acquiring the multi-modal original production data of the rolling production line, extracting the multi-modal production physical characteristics for identifying the roll eccentricity, surface defects or temperature abnormal production risk, and performing semantic fusion on the multi-modal production physical characteristics to generate a joint semantic vector, and then acquiring the production risk data based on the joint semantic vector, and finally adaptively adjusting the rolling parameters according to the preset control strategy and the production risk data. Compared with the extensive control mode in the prior art which depends on the preset rolling procedure or a single sensor, the present application can realize real-time perception of various production risks such as roll eccentricity, surface defects and temperature abnormalities from multiple dimensions, accurately understand the internal relationship between different modal information through semantic fusion, and realize adaptive adjustment of rolling parameters based on the fused risk data, so as to realize accurate adaptive control of the rolling production line based on process information.

Description

Technical Field

[0001] This invention relates to the field of automation control in iron and steel metallurgy, and in particular to an adaptive control method and computer equipment for a steel rolling production line. Background Technology

[0002] In the iron and steel metallurgical industry, the steel rolling production line is a core and crucial link in realizing the deep processing of steel billets and the production of high-precision strip steel products. Its operating status directly determines the dimensional accuracy, surface quality, yield, and production efficiency of the strip steel products, and has a vital impact on the market competitiveness and economic benefits of steel enterprises. The steel rolling production process includes core processes such as hot rolling and cold rolling. The production conditions have significant characteristics such as strong nonlinearity, strong coupling, time-varying nature, and uncertainty. In actual production operation, it is highly susceptible to interference from various abnormal process factors. Among them, roll eccentricity, strip surface defects, and abnormal rolling temperature are the three most frequent and serious types of core faults on site, and are also the main bottlenecks restricting the high-quality and high-efficiency operation of the steel rolling production line.

[0003] The three types of process anomalies mentioned above not only directly disrupt the quality stability of strip steel products, leading to quality problems such as excessive strip thickness, surface scratches, and excessive iron oxide scale, resulting in a large number of defective products being scrapped, but also cause serious production accidents such as steel piling, strip breakage, and mill roll jamming. This, in turn, leads to unplanned mill shutdowns, disrupting normal production rhythms and causing significant capacity losses and equipment wear. According to industry statistics, the scrap rate and unplanned downtime caused by untimely compensation for roll eccentricity and ineffective control of rolling temperature fluctuations alone account for more than 30% of the total losses in steel rolling production. This economic loss is particularly pronounced in the production of high-end precision strip steel. Therefore, steel rolling production lines urgently need efficient, precise, and real-time process sensing capabilities, as well as adaptive and intelligent rolling parameter adjustment functions, to cope with complex and ever-changing rolling conditions and ensure the stability of the production process and the consistency of product quality.

[0004] However, most steel companies in China still use traditional process sensing and rolling parameter control methods for their rolling production lines. This method has obvious technical limitations and cannot meet the stringent requirements of modern high-end strip steel production. The specific drawbacks are mainly reflected in the following aspects: First, the process perception is limited in scope and suffers from significant lag. Traditional control methods primarily rely on preset rolling schedules or single-type sensors for condition monitoring. For example, they might collect rolling force data using pressure sensors or monitor the rolling zone temperature using single-point temperature sensors, using this as the basis for setting parameters such as roll gap and rolling speed. This single-dimensional monitoring model cannot comprehensively capture the multi-physical field coupling information during steel rolling production. It cannot identify latent changes in operating conditions such as roll wear, fluctuations in incoming material composition, and changes in material deformation resistance in real time, nor can it promptly detect early signs of faults such as roll eccentricity and surface defects. In actual production, a steel plant once relied solely on rolling force monitoring without configuring roll condition monitoring devices, resulting in the failure to detect roll eccentricity faults in a timely manner and the inability to provide targeted compensation for roll gaps. Ultimately, this led to the entire batch of high-end automotive strip steel exceeding thickness tolerances, resulting in direct economic losses exceeding one million yuan. Furthermore, traditional monitoring methods often employ offline sampling or lagging data processing, leading to significant time lags in fault identification and operating condition judgment, failing to provide timely and accurate basis for real-time parameter adjustment.

[0005] Secondly, modal information is isolated, resulting in poor generality and interpretability of feature extraction. Existing process control systems mainly provide overall process status information at the production line level. While they can reflect the macroscopic operation of the rolling process, they cannot achieve refined monitoring of individual stands or key rolling nodes, making it difficult to analyze the specific impact of local process parameter anomalies on product quality. In long-process steel rolling production lines, there are significant differences in process parameters such as rolling force, roll gap, temperature, and tension between different stands. Macroscopic monitoring at the production line level cannot provide accurate process data support for individual stands, leading to the inability of the rolling mill to take effective process compensation measures for local anomalies. In addition, traditional machine learning modeling schemes often blindly fit the original monitoring data without combining the core physical mechanisms of metal plastic deformation, heat transfer, and mass transfer in steel rolling. The extracted fault features have poor interpretability, and the information of each monitoring mode (force, temperature, vibration, image) is isolated from each other, making it impossible to achieve deep coupling of multi-source data. It is difficult to characterize the evolution law of multi-fault coupling, resulting in extremely low accuracy in identifying latent and coupled faults, and a persistently high rate of missed detection for high-risk sparse faults.

[0006] Third, the control strategies are rigid and inefficient, lacking adaptive adjustment capabilities. Traditional steel rolling control often employs open-loop or simple closed-loop adjustment methods with fixed parameters. Rolling parameters are mostly preset based on historical production experience, failing to dynamically adapt to real-time monitoring of abnormal operating conditions and risk levels. Faced with complex and variable operating conditions such as fluctuating incoming material temperature, roll wear, and changes in material deformation resistance, this rigid control strategy is prone to problems such as adjustment lag and large overshoot, leading to quality and safety accidents such as thickness deviations, surface scratches, and strip breakage. Furthermore, traditional control methods have low precision in adjusting process parameters, making it difficult to meet the stringent requirements of high-end precision strip steel for dimensional accuracy and surface quality, thus hindering the domestic substitution process for high-end strip steel products. According to relevant industry reports, under traditional control methods, the dimensional tolerance fluctuation of high-strength thin-gauge strip steel often exceeds ±1.2mm, and the batch scrap rate far exceeds the industry's advanced level, resulting in direct and hidden economic losses of up to billions of yuan annually due to product precision issues.

[0007] In summary, existing process sensing and rolling parameter control technologies for steel rolling production lines generally suffer from core technical defects such as sensing lag, modal isolation, and coarse control. They cannot achieve full-dimensional, high-precision, and real-time condition sensing of the rolling process, nor can they adaptively adjust rolling parameters based on precise process information. They are unable to adapt to the complex working conditions of modern steel rolling production, which are characterized by strong nonlinearity, strong coupling, and time-varying fluctuations, and they cannot meet the high-precision, high-stability, and intelligent production requirements of high-end strip steel.

[0008] As the steel industry transforms towards high-end, intelligent, and green development, market demand for high-precision strip steel products continues to rise, placing higher demands on the operational stability and product quality consistency of rolling mill production lines. Against this backdrop, breaking through the limitations of traditional control technologies and developing a control method for rolling mill production lines that enables deep coupling of multi-source data, real-time perception of all-dimensional operating conditions, and adaptive and precise adjustment of rolling parameters, has become a critical technical challenge urgently needing to be addressed in the steel metallurgy field. This method is of significant practical importance and application value in promoting the upgrading of rolling mill production processes, improving product quality, reducing production costs, and enhancing the core competitiveness of enterprises. Summary of the Invention

[0009] To solve at least one of the above-mentioned technical problems, the present invention provides an adaptive control method for a steel rolling production line, characterized in that the method includes: S100: Acquire multimodal raw production data from the steel rolling production line; S200: Extract the multimodal production physical features corresponding to the multimodal raw production data. The multimodal production physical features are used to identify production risks such as roll eccentricity, surface defects, or abnormal temperature. S300: Perform semantic fusion on the multimodal production physical features to generate a joint semantic vector; S400: Based on the joint semantic vector, obtain the production risk data of the steel rolling production line; S500: Adaptively adjust the rolling parameters of the steel rolling production line according to the preset control strategy and the production risk data.

[0010] Furthermore, the multimodal raw production data includes: rolling force data, roll gap profile data, infrared thermal imaging images, and visible light images; The extraction of multimodal production physical features corresponding to the multimodal raw production data includes: extracting the roll eccentric vibration intensity corresponding to the rolling force data; extracting the surface defect movement speed corresponding to the roll gap contour data; extracting the temperature gradient map corresponding to the infrared thermal imaging image; and extracting the texture feature map corresponding to the visible light image.

[0011] Further, the step of extracting the roll eccentric vibration intensity corresponding to the rolling force data includes: S211: Preprocess the original rolling force time series data, remove invalid working conditions such as strip threading, tail throwing, and acceleration / deceleration, and use wavelet threshold noise reduction to remove high-frequency electromagnetic interference and impact noise; S212: Calculate the fundamental frequency of the roll based on the real-time rotation speed of the roll; set a bandpass filter range based on the fundamental frequency to retain the 1st to 6th harmonic signals of the roll and filter out low-frequency material thickness fluctuation interference. S213: Perform a fast Fourier transform on the filtered rolling force time-series signal to obtain the spectral amplitude distribution and extract the spectral peak amplitude corresponding to the roll fundamental frequency; S214: The peak amplitude of the fundamental frequency spectrum is used as the instantaneous eccentric vibration intensity of the roll. At the same time, the proportion of harmonic energy per unit time is statistically analyzed to obtain the steady-state eccentric vibration intensity, thus completing the feature extraction of eccentric vibration intensity.

[0012] Further, the step of extracting the surface defect movement speed corresponding to the roll gap profile data includes: S221: Baseline calibration is performed on continuous multi-frame roll gap profile data to eliminate profile reference offset caused by mill roll-down drift and stand vibration, and a standardized profile sequence is obtained. S222: Perform a difference operation between the standardized contour of each frame and the standard flat roll gap contour, segment the defect area by a preset height threshold, and extract the defect pixel clusters and defect coordinates of the contour depression / protrusion. S223: The sparse optical flow matching algorithm is used to track and match the same defect region in consecutive frames to obtain the inter-frame displacement of the defect in the rolling advance direction and the frame acquisition time interval. S224: The instantaneous moving speed of the surface defect is calculated based on the inter-frame displacement and the frame acquisition time interval. The multi-frame speed data is smoothed and filtered to obtain a stable surface defect moving speed characteristic.

[0013] Further, the step of extracting the temperature gradient map corresponding to the infrared thermal imaging image includes: S231: Perform radiometric calibration, bad pixel repair and Gaussian smoothing correction on the original infrared thermal imaging image, and convert the grayscale image into a global temperature matrix. S232: Calculate the lateral and longitudinal partial derivatives of the temperature matrix to obtain the lateral and longitudinal temperature gradients. S233: Generate a global temperature gradient magnitude map based on the gradient magnitude, and simultaneously generate a gradient direction map; S234: Extract the effective temperature measurement ROI area of ​​the rolls and strip, remove background interference from the frame and water cooling equipment, and finally obtain a refined temperature gradient map for temperature anomaly identification.

[0014] Further, the step of extracting the texture feature map corresponding to the visible light image includes: S241: Perform grayscale conversion, illumination equalization, and noise reduction preprocessing on the original visible light image to eliminate uneven lighting and dust noise interference on site; S242: Use a 16×16 sliding window to traverse the entire image and construct a gray-level co-occurrence matrix window by window; S243: Based on the gray-level co-occurrence matrix, extract four types of texture parameters: contrast, energy, entropy, and correlation, and generate corresponding contrast texture maps, energy texture maps, entropy texture maps, and correlation texture maps pixel by pixel; S244: Fuse four types of single-channel texture maps to obtain a multi-dimensional fused texture feature map, which is used to characterize the surface roughness, defects, and oxide scale distribution characteristics of strip steel.

[0015] Furthermore, the semantic fusion of the multimodal production physical features to generate a joint semantic vector includes: S310: The multimodal production physical features are compressed into modal high-order features of a preset dimension through a modal-specific backbone network; the modal-specific backbone network adopts an architecture combining CNN and Transformer; for one-dimensional roll eccentric vibration intensity and surface defect movement speed, 1D-CNN combined with Transformer encoder is used to extract temporal high-order features; for two-dimensional temperature gradient map and texture feature map, 2D-CNN combined with Transformer encoder is used to extract spatial high-order features; S320: Concatenate the higher-order features of each modality to form a combined feature vector; S330: Input the combined feature vector into the cross-modal Transformer, perform semantic interaction and alignment between modalities through a multi-head self-attention mechanism, and output a joint semantic vector; the different dimensional segments of the joint semantic vector correspond to the roll eccentric vibration intensity scale, surface defect motion trend scale, temperature gradient scale and texture feature scale, respectively, so as to achieve unified representation and cross-modal comparison of multimodal production physical features in the same semantic space.

[0016] Furthermore, obtaining the production risk data of the steel rolling production line based on the joint semantic vector includes: S410: Combining the modal coupling feature distribution of the joint semantic vector, the system adaptively determines whether the current single fault condition or multi-fault coupling condition is met. Through a self-designed scenario-layered dynamic weight iterative allocation mechanism, the system assigns differentiated and nonlinear dynamic weights to each modal feature within the joint semantic vector. The layered dynamic weight iterative allocation mechanism calculates the modal weights in real time based on the three dimensions of fault physical hazard, feature signal-to-noise ratio, and fault coupling correlation strength, thus abandoning fixed weights and general attention mapping logic. S420: Input the joint semantic vector after dynamic weighted optimization into the pre-trained production risk level assessment model, and output the independent risk probability and coupled risk coefficient corresponding to three types of faults: roll eccentricity, temperature abnormality and surface defects, respectively, and quantify the four levels of production risk: low, medium, high and extremely dangerous.

[0017] Furthermore, the production risk level assessment model employs a composite loss function of modal hierarchical adaptive FocalLoss and fault classification cost-sensitive matrix loss; the method for constructing the composite loss function includes: S431: Modal layering FocalLoss construction: targeting roll eccentric vibration intensity, surface defect movement speed, temperature gradient map, and texture features. Figure 4 Differential modulation factors are set based on the signal-to-noise ratio of modal features and the differences in sample distribution. For latent fault samples corresponding to temperature gradient and texture features, large modulation factors are set to reduce the loss weight of regular samples. For explicit fault samples corresponding to rolling force vibration and defect velocity, small modulation factors are set to balance the feature learning weights. Based on the differential modulation factors, the independent FocalLoss of each modal branch is calculated and fused to obtain the modal hierarchical focusing loss. S432: Construction of Fault Classification Cost Sensitive Matrix: For three types of faults, namely roll eccentricity, abnormal temperature, and surface defects, independent four-level risk cost sensitive sub-matrices are constructed respectively; for low, medium, high, and extremely dangerous risk conditions in each sub-matrice, exponentially differentiated penalty coefficients are set for missed detection, false detection, and misjudgment, respectively, where the penalty coefficient for missed detection of extremely dangerous faults is 15 to 30 times that of false alarm penalty coefficient for minor faults; S433: Composite Loss Fusion: The modal hierarchical FocalLoss and the fault classification cost-sensitive loss are weighted and fused together. Dynamic fusion weights are set. In the early stage of model training, the weight of the cost-sensitive loss is increased to quickly converge the risk classification error. In the later stage of training, the weight of the modal hierarchical FocalLoss is increased to refine the mining of hidden fault features.

[0018] On the other hand, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of any of the above methods.

[0019] This invention addresses a series of common technical bottlenecks in existing steel rolling control systems, including single perception dimensions, lack of process mechanism support for feature extraction, isolated and non-integrated multi-source heterogeneous data, low fault identification accuracy, easy omission of high-risk sparse faults, rigid and extensive control strategies, and poor adaptability to operating conditions. It innovatively proposes an adaptive intelligent control system for steel rolling production lines based on the coupling and fusion of multimodal mechanism features. The system boasts strong overall technical innovation, original core algorithms, significant technological advancements, and extremely high engineering value. Firstly, based on the unique mechanisms of three types of high-incidence, high-hazard, and strongly coupled core faults in steel rolling, this invention constructs a four-dimensional heterogeneous complementary perception system encompassing mechanics, geometry, global temperature, and microstructure. This achieves precise binding of "fault mechanism - data source - feature type," with each feature extraction step specifically adapted to the evolutionary characteristics of the corresponding fault. This completely solves the shortcomings of traditional technologies, such as data redundancy, invalid features, and missing dimensions, achieving a comprehensive, refined, and mechanism-based upgrade in operating condition perception. Secondly, this invention innovatively designs a high-order cross-modal fusion architecture of "modal-specific purification + cross-variance regularization + process prior embedding," breaking through the technical limitations of traditional simple splicing and fusion. It can accurately identify complex, implicit operating conditions involving the coupled evolution of multiple faults, solving the industry problem that existing technologies cannot characterize fault coupling patterns. At the risk identification level, this invention forms two core original innovations: first, it upgrades the hierarchical dynamic cross-modal attention mechanism, adapting to differentiated weight adjustments for single faults, coupled faults, and different risk levels, accurately focusing on core fault characteristics; second, it uniquely adapts to the hierarchical composite loss function of the multimodal heterogeneous features of this application. Through modal-differentiated focused training + fault-level cost-sensitive constraints, it solves the world-class industrial modeling problem from the root of the algorithm, addressing issues such as long-tail distribution of industrial samples, missed detection of high-risk faults, and imbalanced training of heterogeneous features. The model's identification accuracy and operating condition adaptability far exceed existing general-purpose models. At the control level, this invention constructs a fully closed-loop control system with four-level risk-level precise regulation and hardware safety safeguards, achieving millisecond-level and micrometer-level adaptive optimization of process parameters, balancing production accuracy, quality, and safety. Compared to existing technologies, this invention significantly reduces quality defects such as strip thickness deviation, surface imperfections, and material inhomogeneity, greatly improving product yield and dimensional accuracy. It effectively avoids production accidents such as steel piling, strip breakage, equipment overload, and abnormal roller wear, significantly reducing unplanned downtime and improving production line efficiency and stability. The overall solution has low hardware modification costs, wide adaptability, and strong portability, making it widely applicable to various hot and cold continuous rolling production lines, effectively reducing production energy consumption, equipment maintenance costs, and labor costs. This invention breaks through the technical ceiling of traditional rolling mill control, achieving a leapfrog upgrade from "passive threshold alarm and coarse fixed control" to "active mechanism prediction and intelligent adaptive fine control." The core algorithm innovations are prominent and not obvious, possessing extremely high engineering application value, economic value, and market promotion prospects. Attached Figure Description

[0020] Figure 1 This is an application environment diagram of the adaptive control method for a steel rolling production line in one embodiment; Figure 2 This is a flowchart illustrating an adaptive control method for a steel rolling production line in one embodiment. Figure 3 This is a flowchart illustrating the adaptive control method for a steel rolling production line in another embodiment; Figure 4 This is a structural block diagram of an adaptive control device for a steel rolling production line in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the execution order of the method. Those skilled in the art will understand that anything that does not violate the inventive concept should be included within the scope of protection of the present invention.

[0023] The adaptive control method for steel rolling production lines provided in this application can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on the cloud or other network servers. Terminal 102 sends an adaptive control request for the rolling mill production line to server 104. Server 104 responds to the request by receiving multimodal raw production data of the rolling mill production line uploaded by various data acquisition devices (such as rolling force sensors, laser rangefinders, infrared cameras, etc.); extracting multimodal production physical features corresponding to the multimodal raw production data, which are used to identify production risks such as roll eccentricity, surface defects, or temperature anomalies; performing semantic fusion on the multimodal production physical features to generate a joint semantic vector; obtaining production risk data of the rolling mill production line based on the joint semantic vector; and adaptively adjusting the rolling parameters of the rolling mill production line according to a preset control strategy and the production risk data. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0024] In one embodiment, such as Figure 2 As shown, an adaptive control method for a steel rolling production line is provided, which is then applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps: S100: Acquire multimodal raw production data from the steel rolling production line.

[0025] Specifically, raw production data can be collected by deploying various sensors around the rolling mill, providing a comprehensive data foundation for subsequent production status perception. In particular, multimodal raw production data refers to operating condition information from different sensing methods and types. Information acquired by a single sensor has limitations; for example, a pressure sensor alone cannot detect the presence of roll eccentricity vibration or surface defects. Therefore, this step uses multimodal data acquisition to obtain raw data that reflects the production status from multiple dimensions. Specifically, multimodal raw production data may include, but is not limited to: sensor data reflecting process elements such as roll eccentricity and rolling force fluctuations obtained through rolling force sensors; roll gap profile data obtained through laser rangefinders, which reflects the location, size, and movement of surface defects in the strip; infrared thermal imaging images obtained through infrared cameras, which reflect the temperature distribution of the strip and rolls during rolling; and visible light images obtained through visible light cameras, which reflect visual information such as strip surface texture and color. By acquiring the aforementioned multimodal raw production data, a data foundation is provided for the subsequent extraction of physical features related to production risks such as roll eccentricity, surface defects, and temperature anomalies.

[0026] In one embodiment, obtaining multimodal raw production data from a steel rolling production line includes: S110: Collects rolling force data of the steel rolling production line through a rolling force sensor.

[0027] Rolling force data is collected by rolling force sensors deployed around the rolling mill. These sensors detect changes in pressure between the rolls and the strip during rolling, enabling precise sensing of rolling force fluctuations. Analysis of the sensor signals reveals process parameters such as roll eccentricity vibration intensity, force disturbances caused by surface defects, and changes in deformation resistance due to temperature anomalies. These parameters reflect the existence and severity of production risks such as roll eccentricity, surface defects, and temperature anomalies, providing crucial input for subsequent risk identification.

[0028] In one specific embodiment, the rolling force sensor is a piezoelectric rolling force sensor. Piezoelectric rolling force sensors have the advantages of high sensitivity and fast response, enabling real-time monitoring of process parameters such as rolling force fluctuations, roll eccentric vibration, and changes in deformation resistance caused by temperature anomalies. For example, this sensor can detect production anomalies such as roll eccentric vibration intensity exceeding 0.5 mm / s, surface defect size exceeding 5 mm, and temperature anomalies exceeding 15°C, providing the system with comprehensive rolling force information to promptly identify potential production risks. It is understood that the specific type and parameters of the rolling force sensor can be adjusted according to the actual application scenario; this embodiment is only a preferred example and not a limitation.

[0029] S120: Collects roll gap profile data of the steel rolling production line through a laser rangefinder sensor.

[0030] Laser rangefinders collect roll gap profile data of the environment surrounding the rolling mill. By emitting a laser beam and receiving the reflected signal, the laser rangefinder generates profile data containing a large number of three-dimensional spatial points. This roll gap profile data can precisely describe the geometry of the roll gap around the mill, including its position, shape, direction of change, and rate of change. This data is crucial for identifying surface defects, as surface defects form clusters with specific characteristics in the profile data during strip movement. Analyzing the changing characteristics of these profile data allows for the assessment of the severity of surface defects. Furthermore, the profile data can be used to identify changes in roll gap thermal expansion caused by temperature anomalies, providing a basis for roll gap compensation control.

[0031] In one specific embodiment, the laser rangefinder sensor employs a high-precision laser displacement sensor, which combines a time-of-flight method with triangulation. This measurement system can generate high-precision roll gap profile data with a measurement accuracy of ±10μm or higher, accurately identifying minute changes in the roll gap. The high-precision roll gap profile data can provide more detailed defect information, including the location, depth, and shape of defects, helping the system accurately determine the safety status of the rolling process. It is understood that the specific measurement system and accuracy parameters of the laser rangefinder sensor can be selected according to actual application requirements; this embodiment is merely a preferred example and not a limitation.

[0032] S130: Acquires infrared thermal imaging and visible light images of the steel rolling production line.

[0033] Infrared thermal imaging and visible light images are simultaneously acquired by optical sensors deployed around the rolling mill. Infrared thermal imaging reflects the temperature distribution of the strip and rolls during rolling. Analysis of temperature gradient maps can detect abnormal temperature changes, such as localized frictional temperature rise caused by roll eccentricity or rapid temperature changes due to surface defects passing through the mill. This temperature information is closely related to production risks. Visible light images reflect the appearance characteristics of the strip. Analysis of texture feature maps can identify changes in surface roughness and the degree of oxide scale adhesion. This visual information helps determine the actual impact of the rolling environment on strip quality. The combination of infrared thermal imaging and visible light images allows for the simultaneous acquisition of thermal and visual information about the rolling process, forming a dual optical perception of strip surface quality.

[0034] In one specific embodiment, the visible light-infrared binocular camera is configured as follows: the infrared sensor has a resolution of 1280×1024 and a frame rate of 30Hz, providing high-definition infrared thermal imaging images to accurately detect the temperature distribution of the strip; the visible light camera supports HDR (High Dynamic Range) imaging, with a dynamic range exceeding 120dB, enabling clear imaging in both strong and low light environments. The combined use of the binocular cameras fully leverages the advantages of both visible light and infrared thermal imaging, providing richer information about the rolling process, including the strip's temperature distribution, texture details, and edge contours. It is understood that the specific resolution, frame rate, dynamic range, and other parameters of the camera can be adjusted according to the actual application scenario; this embodiment is merely a preferred example and not a limitation.

[0035] S200: Extracts multimodal production physical features corresponding to multimodal raw production data. These multimodal production physical features are used to identify production risks such as roll eccentricity, surface defects, or abnormal temperatures.

[0036] Specifically, multimodal production physical features refer to physical parameters extracted from raw data of different modes that reflect production conditions. Specifically, for roll eccentricity risk, roll eccentricity vibration intensity can be extracted from rolling force sensor data; this feature reflects the periodic rolling force fluctuation amplitude caused by roll eccentricity. For surface defect risk, surface defect movement speed can be extracted from laser ranging sensor data; this feature reflects the movement state of strip surface defects as they pass through the mill and their impact on product quality. For temperature anomaly risk, temperature field distribution parameters can be extracted from laser ranging sensor data, and temperature gradient maps can be extracted from infrared thermal imaging images; this feature reflects abnormal changes in temperature distribution during rolling and is related to roll gap changes caused by roll thermal expansion. Furthermore, texture feature maps are extracted from visible light images; this feature reflects changes in strip surface roughness and iron oxide scale distribution, and is related to process anomalies such as roll wear and poor lubrication. By extracting these multimodal production physical features, the raw sensor data is transformed into feature parameters with clear physical meaning that can be directly used for production risk identification, providing input for subsequent semantic fusion.

[0037] In one embodiment, the multimodal raw production data includes: rolling force data, roll gap profile data, infrared thermal imaging images, and visible light images; The rolling force data is acquired in real time by piezoelectric rolling force sensors deployed on the rolling mill stand, including the steady-state average rolling force, rolling force fluctuation peak value, periodic fluctuation time sequence signal, and impact disturbance signal; the roll gap profile data is acquired by scanning with a laser rangefinder sensor, including the roll gap longitudinal height profile, strip width direction profile curve, and continuous frame profile change data; the infrared thermal imaging image is acquired synchronously by an infrared thermal imager at the rolling mill entrance and exit, containing pixel data of the full-area temperature of the strip and roll surfaces; the visible light image is acquired by an industrial camera, containing full-area texture imaging data of the strip surface. The extraction of multimodal production physical features corresponding to the multimodal raw production data includes: extracting the roll eccentric vibration intensity corresponding to the rolling force data; extracting the surface defect movement speed corresponding to the roll gap contour data; extracting the temperature gradient map corresponding to the infrared thermal imaging image; and extracting the texture feature map corresponding to the visible light image.

[0038] In this embodiment, a preferred implementation of step S200 is provided, employing a distributed collaborative synchronous acquisition scheme using four types of heterogeneous sensing devices. Dedicated high-precision sensors are matched according to the physical characteristics of different rolling parameters to achieve full-coverage, high-fidelity, and highly synchronized data acquisition. Specifically, piezoelectric rolling force sensors are precisely deployed at key bearing points on the upper and lower stands of the rolling mill. Relying on the dynamic response characteristics of piezoelectric crystals, they acquire mechanical signals throughout the entire rolling process at millisecond levels. This not only captures the average pressure value during steady-state rolling but also fully preserves instantaneous fluctuation peaks, periodic vibration timing signals, and strip-threading impact disturbance signals, comprehensively reconstructing the mechanical stress and vibration state of the rolling mill. High-precision laser ranging sensors are deployed in the detection areas on both sides of the roll gap, continuously acquiring the longitudinal height profile of the roll gap, the undulation curve in the strip width direction, and the inter-frame profile dynamics using a high-frequency scanning method. The system uses dynamic change data to accurately capture geometric morphological changes such as micro-deformation of the roll gap, uneven strip thickness, and local bulges and depressions. Infrared thermal imagers are symmetrically installed at the inlet and outlet of the rolling mill to collect infrared radiation grayscale images of the strip and roll surfaces in a full-domain pixel temperature measurement mode. After calibration, these images are converted into full-domain temperature pixel data, completely covering the temperature distribution in the rolling contact area, cooling area, and friction temperature rise area. A high-definition industrial camera is positioned directly in front of the roll gap forming area to collect full-domain visible light texture images of the strip surface around the clock, completely preserving micro-morphological information such as surface roughness, scratches, roll marks, oxide scale, and pitting. At the feature extraction level, this invention strictly adheres to the physical mechanism of rolling to achieve precise binding of data, features, and faults: Based on rolling force temporal vibration data, the eccentric vibration intensity of the rolls is quantified by combining the roll rotation characteristics, accurately corresponding to mechanical faults such as roll wear, assembly eccentricity, and excessive bearing clearance; based on dynamic change data of the roll gap profile, surface defect location and dynamic speed measurement are achieved through inter-frame comparison and differential calculation, accurately characterizing the generation, migration, and diffusion patterns of surface defects in strip steel; based on infrared full-domain temperature pixel data, a refined temperature gradient map is constructed to accurately identify temperature faults such as localized overheating during rolling, uneven cooling, and abnormal friction; based on multi-dimensional feature fusion of visible light texture images, the microscopic quality state of the strip steel surface is accurately quantified, achieving comprehensive capture of minute defects. This solution possesses at least the following technical effects: Based on the specific physical causes and evolution characteristics of the three core faults in steel rolling, the data collection type and feature extraction dimension are matched in a targeted manner, solving the core problems of mismatch between traditional technical data and fault mechanism and feature redundancy and ineffectiveness. 1. For roll eccentricity faults: This fault is caused by the periodic mechanical vibration of the roll, which is only reflected in the time-series fluctuation characteristics of the rolling force. Temperature and image data cannot characterize the mechanical eccentricity characteristics. Therefore, this application specifically extracts the eccentric vibration intensity corresponding to the rolling force to accurately match the mechanical vibration mechanism of the eccentricity fault and avoid dimensional mismatch. 2. For strip surface macroscopic defect evolution faults: The movement, diffusion, and deformation of defects can only be quantified by the dynamic changes of the roll gap contour in continuous frames. Static images cannot capture the dynamic evolution law. Therefore, the defect movement speed corresponding to the roll gap contour data is specifically extracted to achieve dynamic tracking of the fault. 3. For rolling temperature abnormality faults: The core hazards of temperature faults are thermal deformation and material degradation caused by excessive local temperature difference and sudden temperature changes. Single-point temperature values ​​cannot characterize the potential risks of gradient changes. Therefore, an infrared temperature gradient map is specifically constructed to match the hidden evolution mechanism of temperature abnormalities. 4. For surface microscopic defect faults: Scratches, pits, oxide scale, and other minute defects do not have obvious contour deformation and can only be identified by surface texture differences. Therefore, a visible light texture feature map is specifically extracted to fill the blind spot of microscopic detection. This application completely abandons the traditional extensive mode of blindly collecting and extracting features without distinction by precisely binding "fault mechanism - data source - feature type". Each type of feature corresponds to the core judgment criteria of a specific fault, which improves the accuracy and interpretability of fault identification from the source and builds a four-dimensional fault perception system with complete mechanism, complementary dimensions, no redundancy and no blind spots.

[0039] In one embodiment, S210: Extracting the roll eccentricity vibration intensity corresponding to the rolling force data is a key physical quantity reflecting the degree of roll eccentricity, and its value directly characterizes the severity of the roll eccentricity fault. Roll eccentricity has a modulating effect on the rolling force signal, and the intensity of the sensing signal received by the rolling force sensor is related to factors such as the amplitude and frequency distribution of the roll eccentricity. By extracting the roll eccentricity vibration intensity, a quantitative basis can be provided for subsequent identification of roll eccentricity risk. Preferably, the following options are available: S211: Preprocess the original rolling force time series data, remove invalid working conditions such as strip threading, tail throwing, and acceleration / deceleration, and use wavelet threshold noise reduction to remove high-frequency electromagnetic interference and impact noise; S212: Calculate the fundamental frequency of the roll based on the real-time rotation speed of the roll; set a bandpass filter range based on the fundamental frequency to retain the 1st to 6th harmonic signals of the roll and filter out low-frequency material thickness fluctuation interference. S213: Perform a fast Fourier transform on the filtered rolling force time-series signal to obtain the spectral amplitude distribution and extract the spectral peak amplitude corresponding to the roll fundamental frequency; S214: The peak amplitude of the fundamental frequency spectrum is used as the instantaneous eccentric vibration intensity of the roll. At the same time, the proportion of harmonic energy per unit time is statistically analyzed to obtain the steady-state eccentric vibration intensity, thus completing the feature extraction of eccentric vibration intensity.

[0040] Specifically, addressing the industry challenges of concealed roll eccentricity faults, susceptibility to on-site interference, and difficulty in identifying minute eccentricities, this invention designs a complete refined feature extraction process encompassing condition screening, noise reduction and purification, adaptive filtering, and spectrum analysis. First, the original rolling force time-series data undergoes precise condition screening. Based on mill operating speed, tension signals, and start / stop status labels, non-steady-state invalid condition data such as strip threading, tail throwing, acceleration / deceleration, and start / stop transitions are automatically eliminated, retaining only stable rolling valid data to avoid invalid condition data interfering with feature extraction accuracy. Subsequently, a wavelet threshold noise reduction algorithm is used to perform layered noise reduction on the valid time-series data. Through multi-layer wavelet decomposition, low-frequency valid signals are separated from high-frequency noise signals, accurately filtering out high-frequency noise points caused by on-site frequency converter electromagnetic interference, instantaneous impacts on the stand, and circuit noise, preserving the true vibration characteristics of the rolling force to the greatest extent possible. Based on this, the real-time fundamental frequency of the roll rotation (n) is calculated using the formula f0=n / 60, based on the real-time feedback of the roll speed (n) from the mill encoder. This adapts to the dynamic changes in rolling speed and avoids the limitation of fixed filtering parameters being unable to adapt to variable-speed rolling. An adaptive bandpass filter window is constructed around the real-time fundamental frequency to accurately retain the 1st to 6th times the roll rotation harmonics. This frequency band is the characteristic frequency band of roll eccentricity faults. At the same time, low-frequency interference signals such as uneven material thickness and rolling load fluctuations are thoroughly filtered out, achieving accurate purification of eccentricity fault characteristics. Finally, a fast Fourier transform is performed on the purified pure characteristic time-series signal to accurately convert the time-domain vibration signal into a frequency-domain spectral distribution. The peak amplitude of the spectrum corresponding to the fundamental frequency is extracted as the instantaneous eccentricity vibration intensity, accurately characterizing the real-time eccentricity imbalance of the roll. Simultaneously, the total energy ratio of the eccentricity harmonics per unit time is statistically analyzed to quantify the average eccentricity under steady-state conditions, achieving dual quantitative characterization of instantaneous dynamic eccentricity and steady-state continuous eccentricity, fully covering the dynamic evolution state of roll eccentricity faults.

[0041] In this embodiment, a preferred embodiment of step S210 is provided. A dedicated adaptive frequency domain purification and extraction algorithm is designed to address the unique characteristics of roll eccentricity faults, including variable speed, strong interference, concealed micro-faults, and periodic evolution. This algorithm overcomes the technical shortcomings of traditional fixed filtering and time-domain threshold detection, which cannot adapt to the characteristics of eccentricity faults. Roll eccentricity is a periodic fault in rotating machinery, with fault characteristics concentrated only in the fundamental and harmonic frequency ranges of the roll, and dynamically changing in real time with the rolling speed. Traditional fixed-parameter filtering is prone to feature rejection or residual interference; simultaneously, a large amount of electromagnetic interference and incoming material fluctuation noise can completely drown out the minute eccentricity signal. This application uses a dedicated process—including working condition screening to remove invalid data, wavelet denoising to remove complex noise, speed-adaptive dynamic frequency domain locking, and time-frequency dual-dimensional quantization—to accurately match the periodicity, variable working conditions, and low signal-to-noise ratio characteristics of eccentricity faults, retaining only the specific effective features of eccentricity and completely isolating irrelevant interference. Compared to traditional solutions, this feature extraction method is specifically adapted to the rolling force time-series vibration data collected in this application. It addresses the pain points of difficulty in identifying and misjudging eccentric micro-faults in this scenario, greatly improving the purity and recognizability of eccentric features. This provides exclusive, accurate, and interference-free core feature support for subsequent multimodal fusion and eccentricity risk assessment.

[0042] In one embodiment, S220: Extract the surface defect movement speed corresponding to the roll gap profile data; the surface defect movement speed refers to the movement speed of surface defects on the strip as the strip passes through the rolling mill, and is a key physical quantity for identifying surface defects and their impact on product quality. By analyzing the movement speed of defect features in the profile data, defect groups with abnormal movement characteristics can be identified, and their speed can be calculated, reflecting the development trend and severity of defects, and providing a basis for early warning of production anomalies such as surface defects. Preferably, the following options are available: S221: Baseline calibration is performed on continuous multi-frame roll gap profile data to eliminate profile reference offset caused by mill roll-down drift and stand vibration, and a standardized profile sequence is obtained. S222: Perform a difference operation between the standardized contour of each frame and the standard flat roll gap contour, segment the defect area by a preset height threshold, and extract the defect pixel clusters and defect coordinates of the contour depression / protrusion. S223: The sparse optical flow matching algorithm is used to track and match the same defect region in consecutive frames to obtain the inter-frame displacement ΔL of the defect in the rolling advance direction and the frame acquisition time interval Δt. S224: The instantaneous moving speed of the surface defect is calculated according to the formula v=ΔL / Δt. The multi-frame speed data is smoothed and filtered to obtain a stable surface defect moving speed characteristic.

[0043] Specifically, addressing the issues of distorted roll gap profile detection data and inaccurate defect detection caused by mechanical drift of the rolling mechanism, stand vibration, and installation reference offset during long-term operation of rolling mills, this invention designs a full-process dynamic defect speed measurement algorithm encompassing baseline calibration, differential segmentation, optical flow tracking, and velocity filtering. First, frame-by-frame baseline calibration is performed on multiple frames of raw roll gap profile data continuously acquired by the laser sensor. The current frame's profile offset is corrected in real-time using historical steady-state profile reference data, offsetting reference deviations caused by mill mechanical drift, high-frequency stand vibration, and temperature deformation. This generates a standardized profile sequence with a unified reference and time-series comparability, eliminating detection deviations caused by equipment mechanical errors at their source. Subsequently, each frame of standardized profile data is compared point-by-point with a pre-stored standard flat roll gap theoretical profile using differential calculation. The profile height difference is calculated, and combined with a preset adaptive height threshold based on strip steel production process parameters, precisely segmenting defect areas such as local depressions, protrusions, and warping on the strip steel surface. Key information such as the defect pixel cluster range, defect center coordinates, and defect area are accurately extracted, achieving precise defect localization. To capture the dynamic evolution of defects, this invention introduces a sparse optical flow matching algorithm. This algorithm tracks and matches feature points within the same defect region in consecutive image frames. By using neighborhood pixel association constraints, it achieves precise cross-frame pairing of defects, effectively avoiding tracking failures caused by interference from adjacent defects. This accurately obtains the inter-frame displacement and sampling time interval of the defect in the rolling direction. Based on displacement and time parameters, the instantaneous velocity of the defect is calculated using a physical velocity measurement formula. Finally, a moving average filtering algorithm is used to smooth the instantaneous velocity across multiple frames, filtering out detection jitter noise and outputting a stable, accurate, and representative feature of the defect's true migration state. This fully reconstructs the dynamic process of defect generation, migration, and diffusion.

[0044] This embodiment provides a preferred embodiment of step S220. Addressing the unique characteristics of scenarios involving dynamic migration of strip surface defects, superposition of equipment interference, easy adhesion and diffusion of defects, and static detection failure, a dedicated dynamic feature extraction logic is designed to precisely adapt to the attributes of the continuous scanning data of the roll gap profile in this application. Traditional defect detection only extracts static defect morphology features, failing to capture the migration, acceleration, and diffusion patterns of defects as the rolling process progresses. One of the core innovations of this application is the prediction of defect evolution risks. Therefore, a dedicated design includes baseline calibration to eliminate equipment drift errors, profile differential to accurately segment defects, and optical flow temporal tracking to quantify migration speed. This feature extraction method perfectly aligns with the fault mechanism of surface defects in this scenario—"dynamic evolution with strip movement, and progressive risk amplification"—upgrading traditional static morphology features to dynamic evolution features. It accurately characterizes the degree of danger and development trend of defects, providing unique dynamic fault dimension information for multimodal fusion in this application. This approach differs from the static feature extraction mode of existing technologies, demonstrating significant creativity and scenario adaptability.

[0045] In one embodiment, S230: Extract the temperature gradient map corresponding to the infrared thermal imaging image; the temperature gradient map is a feature map extracted from the infrared thermal imaging image that reflects the spatial rate of change of rolling temperature. It can characterize the degree of drastic temperature change in space and is of great significance for production risk identification. For example, roll eccentricity is often accompanied by local frictional temperature rise, and surface defects can cause rapid local temperature changes when passing through the rolling mill. These abnormal temperature changes will appear as significant temperature gradients in the infrared thermal imaging image. By extracting the temperature gradient map, the temperature information in the original infrared thermal imaging image can be transformed into a feature representation sensitive to production risks. Preferably, the following options are available: S231: Perform radiometric calibration, bad pixel repair and Gaussian smoothing correction on the original infrared thermal imaging image, and convert the grayscale image into a global temperature matrix T(x,y), where x is the coordinate of the strip width direction and y is the coordinate of the rolling forward direction. S232: Calculate the transverse and longitudinal partial derivatives of the temperature matrix respectively to obtain the transverse temperature gradient G. x =∂T / ∂x, longitudinal temperature gradient Gᵧ=∂T / ∂y; S233: Calculation formula for gradient magnitude Generate a global temperature gradient magnitude map and a gradient direction map simultaneously; S234: Extract the effective temperature measurement ROI area of ​​the rolls and strip, remove background interference from the frame and water cooling equipment, and finally obtain a refined temperature gradient map for temperature anomaly identification.

[0046] Specifically, addressing the shortcomings of traditional single-point and local temperature measurements, which cannot characterize the full-domain distribution of the rolling temperature field and have a high rate of missed detection of hidden temperature anomalies, this invention designs a temperature feature extraction scheme that includes refined infrared image correction, global matrix transformation, bidirectional gradient solving, and effective area cropping. First, the original grayscale images acquired by the infrared thermal imager undergo professional preprocessing, including equipment radiation accuracy calibration, image defect pixel repair, and Gaussian smoothing denoising correction. This eliminates acquisition defects such as infrared equipment temperature measurement deviation, image noise, and pixel failure, ensuring the authenticity and accuracy of the image data. Based on the calibrated infrared grayscale images, each pixel grayscale value is accurately converted into the corresponding spatial coordinate's true temperature value according to the infrared temperature measurement calibration formula. This constructs a two-dimensional global temperature matrix T(x,y) covering the strip width and rolling length directions, achieving a dimensional upgrade from "single-point temperature value" to a "global temperature field matrix." Based on this, the partial derivatives of the temperature matrix are solved in both the transverse and longitudinal directions to calculate the transverse temperature gradient in the strip width direction and the longitudinal temperature gradient in the rolling advance direction, accurately characterizing the rate of temperature change and the degree of abrupt change in both dimensions. The global temperature gradient amplitude is calculated using the gradient amplitude synthesis formula, simultaneously generating a gradient direction distribution map to fully reconstruct the detailed information such as the temperature distribution, abrupt change locations, and diffusion directions in the rolling area. Finally, the effective temperature measurement area of ​​the ROI is preset according to the rolling process range, and the core rolling area of ​​the strip and rolls is precisely trimmed, eliminating invalid interference areas such as the stand, water cooling pipes, and environmental background. This results in a highly targeted, highly accurate, and directly usable fine-grained temperature gradient feature map that can accurately identify various hidden temperature faults such as localized overheating, uneven cooling, abnormal frictional temperature rise, and abrupt temperature changes at the beginning and end of the rolling mill.

[0047] This embodiment provides a preferred embodiment of step S230. Addressing the unique characteristics of rolling temperature anomalies, including localized abrupt changes, gradient imbalances, latent propagation, and global correlation, and leveraging the advantages of the infrared global imaging data presented in this application, a gradient feature extraction scheme is designed to solve the problem of traditional single-point temperature measurement failing to capture latent temperature risks. The core hazard of rolling temperature faults is not absolute temperature exceeding limits, but rather thermal deformation, stress concentration, and material degradation caused by excessive local temperature differences, sudden temperature changes, and uneven temperature field distribution. Such risks cannot be identified through single-point temperature values. This application constructs a two-dimensional temperature matrix based on global infrared images and solves the temperature gradient through bidirectional partial derivatives, accurately capturing the rate of temperature change and the location of abrupt changes, perfectly adapting to the latent evolution mechanism of temperature faults in this scenario. This feature extraction method upgrades static temperature values ​​to dynamic gradient change features, accurately matching the global perception advantage of the temperature modal data presented in this application. It provides microscopic temperature anomaly features that traditional technologies cannot obtain for multimodal fusion, enabling advance prediction of temperature risks. It demonstrates strong scenario adaptability and technological innovation.

[0048] In one embodiment, S240: Extract the texture feature map corresponding to the visible light image; the texture feature map is a feature map extracted from the visible light image that reflects the variation law of the strip surface texture, characterizing the visual attributes of the strip surface such as roughness, regularity, and directionality, and plays an important role in identifying process anomalies such as iron oxide scale adhesion and roll surface wear. For example, surface defects can cause the strip surface texture to change from smooth to rough; changes in iron oxide scale caused by temperature anomalies can also change the strip surface texture features. By extracting the texture feature map, the visual information in the original visible light image can be transformed into a feature representation sensitive to production risks. Preferably, the following options are available: S241: Perform grayscale conversion, illumination equalization, and noise reduction preprocessing on the original visible light image to eliminate uneven lighting and dust noise interference on site; S242: Use a 16×16 sliding window to traverse the entire image and construct the gray-level co-occurrence matrix (GLCM) window by window; S243: Based on the gray-level co-occurrence matrix, extract four types of texture parameters: contrast, energy, entropy, and correlation, and generate corresponding contrast texture maps, energy texture maps, entropy texture maps, and correlation texture maps pixel by pixel; S244: Fuse four types of single-channel texture maps to obtain a multi-dimensional fused texture feature map, which is used to characterize the surface roughness, defects, and oxide scale distribution characteristics of strip steel.

[0049] Specifically, addressing the industrial challenges of high dust levels, uneven lighting, strong water mist interference, and difficulty in detecting minute surface defects in steel rolling workshops, this invention designs an anti-interference multi-dimensional texture feature fusion algorithm to achieve refined detection of the microscopic quality of strip steel surfaces. First, the color visible light images captured by industrial cameras undergo standardized preprocessing. Grayscale processing simplifies image computation dimensions, and an adaptive illumination equalization algorithm counteracts the interference from strong light, shadows, and uneven lighting in the workshop. An adaptive filtering algorithm removes image noise caused by dust, water mist, and suspended particles, maximizing the restoration of the true texture details of the strip steel surface and ensuring the authenticity of image feature extraction. Subsequently, a 16×16 pixel fixed sliding window is used to perform a full-coverage, non-overlapping traversal of the entire image, constructing a grayscale co-occurrence matrix (GLCM) window by window. This accurately mines the spatial distribution patterns and correlation features of pixel grayscale within the image, adapting to the characteristics of the fine and uniform texture of the strip steel surface. Based on the gray-level co-occurrence matrix, four types of core process-related texture parameters are extracted: contrast parameter characterizes the surface texture unevenness and defect undulation; energy parameter characterizes the uniformity and smoothness of the strip surface texture; entropy parameter characterizes the surface texture complexity and defect disorder; and correlation parameter characterizes the pixel spatial correlation law and texture distribution consistency. Four independent single-channel texture feature maps are generated through pixel-by-pixel mapping, each corresponding to different dimensions of surface quality information. Finally, the multi-channel texture maps are fused using feature weighting, retaining the core effective information of each dimension to generate a multi-dimensional fused texture feature map. This map can comprehensively and accurately characterize various micro-defects on the strip surface, such as roughness, fine scratches, roll marks, pitting, oxide scale distribution, and local texture disorder, achieving refined quality inspection that macroscopic contour detection cannot cover.

[0050] In this embodiment, a preferred embodiment of step S250 is provided. Considering the characteristics of micro-defects on the strip steel surface—small size, no obvious contour deformation, susceptibility to environmental interference, and extremely high concealment—a multi-dimensional GLCM texture fusion extraction scheme is designed to adapt to the microscopic perception capabilities of visible light images in this application. Traditional contour detection can only identify macroscopic unevenness defects, completely failing to capture microscopic defects such as micron-level scratches, oxide scale, and texture disorder, which are core causes of scrapping high-end strip steel. This application quantifies microscopic features based on gray-level co-occurrence matrices from four dimensions: texture smoothness, complexity, correlation, and unevenness, accurately matching the texture information advantages of visible light images and compensating for the microscopic blind spots of roll gap contour detection. This feature extraction method is specifically designed for the harsh industrial environment and micro-defect detection needs of this scenario, complementing the other three types of macroscopic, mechanical, and temperature features in this application to construct a comprehensive defect feature system. This provides a unique microscopic quality dimension for multimodal fusion, significantly improving the comprehensiveness and precision of overall fault identification.

[0051] S300: Semantic fusion of multimodal production physical features to generate a joint semantic vector.

[0052] Specifically, the multimodal production physical features extracted in step S200 are semantically fused to generate a joint semantic vector capable of cross-modal comparison. Semantic fusion refers to mapping features from different sensors, with different physical meanings and data formats, to a unified semantic space for processing. Since features such as roll eccentric vibration intensity, surface defect movement speed, temperature gradient map, and texture feature map differ significantly in physical meaning, dimensions, and data dimensions, direct concatenation or simple weighting cannot reflect the inherent relationships between these features. For example, there may be a correlation between roll eccentric vibration intensity and temperature gradient map (roll thermal expansion exacerbates the eccentricity effect), but traditional fusion methods cannot capture this cross-modal semantic relationship. Therefore, this step employs semantic fusion technology. First, each modal feature is compressed into higher-order modal features of a preset dimension through an independent feature extraction network. These higher-order features retain the unique information of each modality while providing a unified feature dimension. Then, the higher-order modal features are concatenated to form a combined feature vector. Finally, the combined feature vector is input into a cross-modal Transformer, where a self-attention mechanism is used for semantic interaction and alignment between modalities, outputting a fixed-dimensional joint semantic vector. This joint semantic vector is interpretable; in practical applications, its different dimensional segments correspond to semantic scales for different modal features: some segments constitute a scale for the intensity of roll eccentricity vibration, reflecting the continuous change from weak to strong roll eccentricity; some segments constitute a scale for the movement trend of surface defects, reflecting the change from slight to severe defects; some segments constitute a scale for the temperature gradient, reflecting the distribution change from normal to abnormal temperature; and some segments constitute a scale for texture features, reflecting the texture change from smooth to rough strip surface. Through this design, a unified representation and cross-modal comparison of multimodal production physical features are achieved within the same semantic space, enabling the system to understand the multimodal feature combination patterns corresponding to different production risks.

[0053] In one embodiment, the step of semantically fusing the multimodal production physical features to generate a joint semantic vector includes: S310: The multimodal production physical features are compressed into modal high-order features of a preset dimension through a modal-specific backbone network; the modal-specific backbone network adopts an architecture combining CNN and Transformer; for one-dimensional roll eccentric vibration intensity and surface defect movement speed, 1D-CNN combined with Transformer encoder is used to extract temporal high-order features; for two-dimensional temperature gradient map and texture feature map, 2D-CNN combined with Transformer encoder is used to extract spatial high-order features; all modal high-order features are uniformly compressed to 128 dimensions; Specifically, a modality-specific backbone network refers to a feature extraction network independently designed for each modality's features. Because the physical characteristics of different production modalities vary significantly in data format, dynamic range, and physical meaning—for example, the eccentric vibration intensity of a roll is a one-dimensional signal, while a temperature gradient map is a two-dimensional image—using a uniform network structure for feature extraction might lead to some modal information being "overwhelmed" by others. Therefore, setting up a dedicated backbone network for each modality allows for the targeted extraction of core information for each modality, preserving its unique feature representation. The backbone network employs an architecture combining CNN and Transformer. Convolutional Neural Networks (CNNs) excel at extracting local spatial features, effectively capturing spatial information such as texture, edges, and temperature gradients in images; Transformers excel at capturing global dependencies, establishing long-range associations between features. Combining the two allows CNNs to extract local detailed features while Transformers establish global contextual relationships, thereby generating more expressive high-order modality features.

[0054] In one specific embodiment, the modality-specific backbone network adopts a structure combining CNN and Transformer to process the eccentric vibration intensity of the roll, the surface defect velocity, the temperature gradient map, and the texture feature map, respectively. The eccentric vibration intensity and surface defect velocity of the roll are one-dimensional signals, while the temperature gradient map and texture feature map are two-dimensional images. For the one-dimensional signals, a one-dimensional CNN is used to extract local temporal features, and then a Transformer encoder is used to establish global dependencies. For the two-dimensional images, a two-dimensional CNN is used to extract spatial features, and then a Transformer encoder is used to capture global spatial relationships. Each backbone network compresses the input two-dimensional or one-dimensional signals into 128-dimensional modal high-order features, preserving modal-specific information while providing a unified spatiotemporal resolution for subsequent fusion. It is understood that the specific dimensions of the modal high-order features can be adjusted according to actual application requirements; for example, 256 dimensions or 64 dimensions can be used, as long as they can effectively represent modal information, and there is no limitation on this.

[0055] S320: Concatenate the higher-order features of each modality to form a 4×128-dimensional combined feature vector; Specifically, the higher-order features of each mode generated in step S310 are concatenated to form a combined feature vector containing information from all modes. Specifically, the higher-order features corresponding to the roll eccentric vibration intensity, the surface defect velocity, the temperature gradient map, and the texture feature map are concatenated along the feature dimension. This concatenation method preserves the independence of each modal feature while integrating multimodal information into a unified vector, providing input for subsequent cross-modal interactions.

[0056] In one specific embodiment, four modality-specific backbone networks each output 128-dimensional modality higher-order features. These four 128-dimensional features are concatenated to form a 4×128-dimensional combined feature vector, i.e., a 512-dimensional combined feature vector. It is understood that the dimension of the combined feature vector depends on the dimension of each modality's higher-order features. If the higher-order features of each modality use other dimensions (such as 64 dimensions), the dimension of the concatenated vector will change accordingly.

[0057] S330: Input the combined feature vector into the cross-modal Transformer, perform semantic interaction and alignment between modalities through a multi-head self-attention mechanism, and output a 512-dimensional fixed-dimensional joint semantic vector; the different dimensional segments of the joint semantic vector correspond to the roll eccentric vibration intensity scale, the surface defect motion trend scale, the temperature gradient scale, and the texture feature scale, respectively, so as to achieve a unified representation and cross-modal comparison of multimodal production physical features in the same semantic space.

[0058] Specifically, the cross-modal Transformer is a neural network structure based on a self-attention mechanism, capable of performing global interactive computation on input features. Unlike traditional feature concatenation or weighted fusion, the cross-modal Transformer, through its self-attention mechanism, enables features from different modalities to interact at the semantic level, capturing the correlations between modalities. For example, the system can learn the correlation between "increased intensity of roll eccentric vibration" and "abnormal change in temperature gradient," or the correlation between "increased surface defect movement speed" and "change in texture feature map," thereby achieving cross-modal semantic alignment.

[0059] After cross-modal Transformer processing, a fixed-dimensional joint semantic vector is output. This joint semantic vector is interpretable, with different dimensional segments corresponding to semantic scales of different modal features. Specifically, some dimensional segments constitute a scale for the intensity of roll eccentricity vibration, reflecting the continuous change of roll eccentricity from weak to strong; some dimensional segments constitute a scale for the movement trend of surface defects, reflecting the change of defects from slight to severe; some dimensional segments constitute a scale for temperature gradient, reflecting the distribution change of temperature from normal to abnormal; and some dimensional segments constitute a scale for texture features, reflecting the texture change of strip surface from smooth to rough. Through this design, a unified representation and cross-modal comparison of multimodal production physical features are achieved within the same semantic space, enabling the system to understand the multimodal feature combination patterns corresponding to different production risks.

[0060] In one specific embodiment, the cross-modal Transformer employs a multi-head self-attention mechanism and a feedforward neural network (FFN) structure. The input is a 4×128-dimensional combined feature vector (i.e., a 512-dimensional token). The multi-head self-attention mechanism calculates the attention weights between features to achieve semantic interaction between modalities. Then, a nonlinear transformation is performed through the feedforward neural network to output a 512-dimensional joint semantic vector. Different dimensions of this 512-dimensional joint semantic vector correspond to different physical meaning scales.

[0061] More preferably, based on the above, this invention innovatively introduces a high-order fusion strategy of modal cross-covariance regularization constraint + fault coupling prior semantic embedding, which significantly improves the coupling representation capability and scenario adaptability of the joint semantic vector. Specifically, step S300 further includes: S340: After generating the joint semantic vector (the initial 512-dimensional joint semantic vector obtained in S330), calculate the cross-modal cross-covariance matrix between the higher-order features of the four modalities, quantify the pairwise fault coupling correlation degree of eccentricity-temperature, eccentricity-defect, temperature-defect, and defect-texture; combine the pre-set fault coupling prior weight matrix based on the rolling process mechanism, and regularize the cross-covariance matrix to suppress irrelevant noise coupling between modes and enhance the real fault correlation coupling features; S350: The causal evolution prior knowledge of the three types of faults is embedded into the joint semantic vector as a semantic bias. Adaptive gain correction is performed on the modal feature dimension with strong causal coupling. It can be selected, but not limited to, the classic process coupling link of "intensified roll eccentric vibration → roll gap offset → local temperature gradient anomaly → surface texture defect diffusion". The semantic weight of the link-related features is automatically enhanced, and pseudo-coupling features without process association are weakened. S360: The semantic vector dimension is calibrated through layer normalization operation to generate the final coupled joint semantic vector.

[0062] This embodiment presents a preferred embodiment of step S300, which overcomes the inherent defects of traditional multimodal fusion technology from the underlying architecture. The basic solution, through a time-series / spatial dual-dedicated network architecture, completely solves the problems of poor adaptability of traditional general-purpose networks to heterogeneous features and the easy subtraction of subtle features, achieving accurate purification and unified representation of heterogeneous features. A further optimized high-order fusion scheme, distinct from the unconstrained and mechanistic blind fusion mode of existing technologies, achieves the following: 1. Through cross-modal cross-covariance regularization, it can accurately distinguish between real process fault coupling and pseudo-coupling of detection noise, solving the core problem of traditional fusion easily generating false feature associations and distorted coupling representations; 2. Through the embedding of causal prior semantics of rolling process faults, it integrates industry mechanism knowledge into the fusion process, achieving a dual-driven fusion of "data-driven + process mechanism-driven," giving the joint semantic vector clear physical interpretability and accurately restoring the real working conditions of the gradual evolution and mutual coupling of multiple faults in steel rolling; 3. It specifically strengthens the characteristics of classic fault coupling links, significantly improving the representation ability of composite, progressive, and latent faults, completely solving the shortcoming of existing technologies in being unable to identify coupled faults. The refined joint semantic vectors generated by this solution significantly improve fault coupling identification, latent feature retention, and process adaptability compared to traditional fusion results. This provides highly innovative and high-quality feature support for subsequent high-precision risk assessment, which cannot be easily replaced by existing technologies.

[0063] S400: Based on joint semantic vectors, obtain production risk data of steel rolling production lines.

[0064] Specifically, the joint semantic vector generated in step S300 is input into the risk analysis module, which outputs quantified production risk data. This data is used to guide the adaptive adjustment of rolling parameters. First, the current production anomaly type is obtained. Since the joint semantic vector itself is a fused feature representation and does not directly contain a production anomaly type label, this step obtains the current production anomaly type information through external means. For example, real-time process information of the steel rolling production line can be obtained by calling the process status message of the process control system, including the type and severity level of production anomaly events such as roll eccentricity, temperature anomaly, and surface defects. Second, a cross-modal dynamic attention algorithm is used to assign dynamic attention weights to different modal features of the joint semantic vector. The core idea of ​​this algorithm is that under different production anomaly conditions, the importance of different modal features for risk judgment is different. For example, under roll eccentricity anomaly, the weight of the roll eccentricity vibration intensity feature should be increased accordingly; under surface defect anomaly, the weight of the surface defect movement speed feature should be increased accordingly; under temperature anomaly, the weights of the temperature gradient map feature and texture feature map should be increased accordingly. Through a dynamic attention mechanism, the system can automatically adjust the contribution of each modal feature in subsequent risk assessment based on the current production anomaly type, making risk assessment more focused on features strongly correlated with the current production risk. Finally, the weighted joint semantic vector is input into a pre-trained production risk level assessment model. This model is a classifier pre-trained with a large amount of sample data and can output the corresponding risk level based on the input joint semantic vector. Since samples of extreme production anomalies such as roll eccentricity and surface defects are relatively sparse in practical applications, the model is optimized using Focal Loss and a cost-sensitive matrix during training, setting higher cost weights for misjudgments of key risks to ensure that the model maintains high recognition accuracy even with extremely sparse samples. The model outputs real-time levels for three risk categories: roll eccentricity, temperature anomalies, and surface defects. Each risk category can be divided into multiple levels, such as low, medium, high, and critical, to quantitatively represent the severity of the current production risk.

[0065] In one embodiment, obtaining production risk data for a steel rolling production line based on joint semantic vectors includes: S410: Based on the production anomaly type label input from the external input, combined with the modal coupling feature distribution of the joint semantic vector, it adaptively determines the current single fault condition or multi-fault coupling condition. Through a self-designed scenario-layered dynamic weight iterative allocation mechanism, it assigns differentiated and nonlinear dynamic weights to each modal feature within the joint semantic vector. The layered dynamic weight iterative allocation mechanism calculates the modal weights in real time based on the three dimensions of fault physical hazard, feature signal-to-noise ratio, and fault coupling correlation strength, abandoning fixed weights and general attention mapping logic. Specifically, the system first acquires externally input production anomaly type labels to determine the main production anomaly types (such as roll eccentricity, temperature anomalies, surface defects, etc.) in the current steel rolling production line area. These production anomaly type labels originate from data sources independent of this system, such as production early warning service interfaces and process status messages from the steel plant's process control system. By using external input, prior information on the current production anomaly type can be quickly obtained, thereby guiding subsequent modal feature weighting.

[0066] After obtaining the current production anomaly type, a cross-modal dynamic attention algorithm is used to process the joint semantic vector generated in step S300. The core idea of ​​the cross-modal dynamic attention algorithm is that the importance of different modal features in the joint semantic vector varies under different production risk scenarios. For example, under a roll eccentricity fault, the dimension segment related to the roll eccentricity vibration intensity should receive higher weight; under a surface defect anomaly, the dimension segment related to the surface defect movement trend should receive higher weight; under a temperature anomaly, the dimension segments related to the temperature gradient scale and texture feature scale should receive higher weight. Through the dynamic attention mechanism, the contribution of each modal feature in subsequent risk assessment can be adaptively adjusted according to the current production anomaly type, making risk assessment more focused on features strongly correlated with the current production risk. Specifically, the cross-modal dynamic attention algorithm uses the externally input production anomaly type label as the query condition and the joint semantic vector as the query key-value pair to calculate the attention weight distribution. This weight distribution reflects the importance of each dimension segment in the joint semantic vector under the current production anomaly type. The calculated attention weights are weighted with the joint semantic vector to obtain the weighted joint semantic vector.

[0067] In one specific embodiment, the cross-modal dynamic attention algorithm uses production anomaly labels as query vectors and the joint semantic vector as both key and value vectors, employing scaled dot-product attention for computation. During the training phase, production anomaly labels are derived from process status messages in the steel plant's process control system, which provide standardized production observation data. During the inference phase, production status data from the rolling mill production line is retrieved in real-time via a lightweight production data API, updated every 30 minutes to ensure the timeliness of production anomaly types. This dynamic attention mechanism automatically adjusts the weights of each modal feature based on the current production anomaly type, and the attention weights are interpretable—for example, they can intuitively tell maintenance personnel that "in this roll gap adjustment decision, 62% was due to roll eccentricity and 28% was due to temperature anomalies," meeting the needs of auditing and fault tracing. It is understood that the source of the production anomaly type label can be any reliable external data source, and the attention algorithm can also adopt other forms of attention mechanism (such as multi-head attention, additive attention, etc.), as long as it can dynamically adjust the weight distribution of the joint semantic vector according to the current production anomaly type. This invention does not limit this.

[0068] S420: Input the joint semantic vector after dynamic weighted optimization into the pre-trained production risk level assessment model, and output the independent risk probability and coupled risk coefficient corresponding to three types of faults: roll eccentricity, temperature abnormality and surface defects, respectively, and quantify the four levels of production risk: low, medium, high and extremely dangerous.

[0069] Specifically, the weighted joint semantic vector output by S410 is input into a pre-trained production risk level assessment model, which outputs the real-time levels for three types of production risks: roll eccentricity, temperature anomaly, and surface defects.

[0070] More specifically, the production risk level assessment model is a pre-trained classification model. Its input is a weighted joint semantic vector, and its output is the level of each risk type. Risk levels can be multi-level, such as low, medium, and high, to quantitatively represent the severity of the current production risk. The specific structure of this model is not specifically limited in this invention; a 3-layer multilayer perceptron (MLP) structure can be selected. Considering the limited computing resources of edge devices, the model is quantized to INT8 precision during deployment, enabling a single-frame inference time of less than 6 milliseconds on edge computing platforms such as the RK3588 NPU, meeting real-time requirements.

[0071] However, due to the scarcity of extreme production anomaly samples during the model training phase, this application further improves the training method. Focal Loss combined with a cost-sensitive matrix is ​​preferably used for optimization, where misjudgments of key risks (such as misclassifying extremely dangerous risks as low-risk) can be weighted with a cost of up to 100 times. In this way, even with few extreme samples in the training data, the model can still learn to effectively identify extreme risks. It is understood that the production risk level assessment model iteratively trains and converges through a self-developed hierarchical differentiated composite loss constraint mechanism that adapts to multimodal heterogeneous features. The entire training constraint logic is customized based on the distribution characteristics, long-tail characteristics of samples, and operating condition cost characteristics of the four types of specific physical features in this application.

[0072] In one embodiment, the production risk level assessment model employs a composite loss function of modal hierarchical adaptive Focal Loss and fault classification cost-sensitive matrix loss; the method for constructing the composite loss function includes: S431: Modal layering Focal Loss construction: targeting roll eccentric vibration intensity, surface defect movement speed, temperature gradient map, and texture features. Figure 4 Differential modulation factors are set based on the signal-to-noise ratio and sample distribution differences of modal features. For latent fault samples corresponding to temperature gradients and texture features, large modulation factors are set to reduce the loss weight of regular samples. For explicit fault samples corresponding to rolling force vibration and defect velocity, small modulation factors are set to balance the feature learning weights. Independent Focal Loss of each modal branch is calculated based on the differential modulation factors, and the modal hierarchical focusing loss is obtained by fusion. This enables targeted training and optimization of the multimodal heterogeneous features of this application.

[0073] S432: Fault Classification Cost Sensitive Matrix Construction: For three types of faults—roll eccentricity, temperature abnormality, and surface defects—independent four-level risk cost sensitive sub-matrices are constructed. For low, medium, high, and extremely dangerous risk conditions in each sub-matrice, exponentially differentiated penalty coefficients are set for missed detection, false detection, and misjudgment. The penalty coefficient for missed detection of extremely dangerous faults is 15 to 30 times that of false alarms of minor faults. This can accurately match the differences in actual losses in steel rolling production.

[0074] S433: Composite Loss Fusion: The modal hierarchical Focal Loss and the fault classification cost-sensitive loss are weighted and fused together. Dynamic fusion weights are set. In the early stage of model training, the weight of the cost-sensitive loss is increased to quickly converge the risk classification error. In the later stage of training, the weight of the modal hierarchical Focal Loss is increased to refine the mining of hidden fault features. This can complete the training of a dedicated model adapted to the multimodal heterogeneous features of this scenario.

[0075] Specifically, this invention addresses the unique characteristics of the four types of specific physical features extracted in this application, namely strong heterogeneity, uneven distribution of fault samples, and significant differences in risk costs among different faults. It designs a fully customized composite loss function training scheme, distinct from existing general loss function schemes. First, a modal-layered focusing loss design is implemented: in this application, the faults corresponding to temperature gradient and surface texture features are mostly latent, minute, and sparse faults, with extremely few samples and extremely low signal-to-noise ratios, making them the most difficult feature dimensions for the model to learn and the easiest to miss. In contrast, the fault signal features corresponding to rolling force vibration and defect velocity features are obvious, with abundant samples and high signal-to-noise ratios. Therefore, this application abandons the general unified modulation factor, setting a large modulation factor of γ=2.0~2.5 for the two types of latent feature modes, significantly suppressing the gradient contribution of a large number of normal samples, forcing the model to focus on learning rare latent fault samples; and setting a small modulation factor of γ=0.5~1.0 for the two types of explicit feature modes, ensuring stable learning of explicit fault features and avoiding overfitting, thus achieving layered and accurate training of the four types of heterogeneous features. Secondly, a fault classification cost-sensitive matrix was constructed: based on the calibration of massive steel rolling fault cases, three types of core faults were modeled independently. Extremely critical faults, if missed, would cause equipment damage, complete line shutdown, and batch scrapping, therefore an extremely high penalty coefficient was set; high-risk faults were misclassified with medium-to-high penalties; and low-to-medium risk false alarms, which only cause minor process disturbances, were penalized with extremely low penalties, forming a three-dimensional differentiated penalty system of "fault type + risk level + error type". Finally, a dynamic weight fusion strategy was adopted. In the early stages of training, cost-sensitive loss was used primarily to quickly correct the model's risk classification bias and ensure basic classification accuracy; in the later stages of training, modal hierarchical Focal Loss was used primarily to finely mine latent heterogeneous features, achieving ultimate optimization of model accuracy and perfectly adapting to the risk modeling requirements after multimodal feature fusion in this application.

[0076] This embodiment presents a preferred implementation of the training process, which significantly improves upon the fundamental shortcomings of existing general loss functions in adapting to the multimodal heterogeneous features, long-tailed samples, and differentiated risk costs in steel rolling. The benefits are as follows: 1. Modal hierarchical Focal Loss achieves differentiated training based on the characteristics of the four types of proprietary physical features described in this application, specifically addressing the problems of insufficient learning of latent micro-features and overfitting of explicit features, maximizing the refined detection advantages of the multimodal feature system; 2. A three-dimensional fault classification cost-sensitive matrix breaks away from the traditional two-dimensional fixed penalty mode, accurately binding to the real production losses in steel rolling processes, making model training no longer pure data fitting, but mechanism-based training that aligns with the safety and quality costs of industrial production; 3. A dynamic fusion weight strategy achieves adaptive optimization of the training process, balancing model convergence speed and refined recognition accuracy. This proprietary composite loss function is entirely customized for the multimodal fault recognition scenario of this application and cannot be universally reused. It significantly improves the model's recognition accuracy for minor, latent, high-risk, and coupled faults from the algorithmic root, completely solving the long-standing problem of missed detection of high-risk faults in the industry. It demonstrates high technical barriers and outstanding innovation.

[0077] S500: Adaptively adjusts rolling parameters of the steel rolling production line based on preset control strategies and production risk data.

[0078] Specifically, based on the production risk data output in step S400, adaptive adjustment control of the roll gap is executed, forming a complete closed loop from sensing to execution. Optional features include: S510: Based on the production risk data, query the preset roll gap control strategy library to obtain the roll gap target value, response time, and additional action instructions; the strategy library stores the refined control parameters and protection logic corresponding to different risk levels; Specifically, the system uses production risk data (i.e., risk levels for roll eccentricity, temperature anomalies, and surface defects) as query criteria to match corresponding control parameters from a pre-defined roll gap control strategy library. This library pre-stores the mapping relationship between different risk levels and roll gap control actions. This mapping relationship can be a predefined rule table or a set of strategies generated through machine learning. Specifically, for each risk level (e.g., low, medium, high, extremely high) of each type of production risk (roll eccentricity, temperature anomalies, surface defects), the strategy library defines corresponding roll gap target values, response times, and additional action instructions. The roll gap target value determines the degree of roll gap adjustment required. For example, at low risk, only appropriate adjustment of the roll gap value is needed to maintain normal rolling, while at extremely high risk, adjustment to the target roll gap and activation of additional protective measures are required. Therefore, the strategy library pre-defines corresponding roll gap target values, response times, and additional action instructions for each risk level. The target value of the roll gap determines the set size of the roll gap, the response time determines the delay from risk identification to action execution, and the additional action instructions include auxiliary protection measures such as starting the laminar flow cooling system, shutting down unnecessary cooling branches, starting alarm prompts, and cutting off power to non-critical equipment.

[0079] In one specific embodiment, the preset roll gap control strategy library is implemented using a JSON-formatted rule engine, allowing for online editing and customization of strategy rules according to actual application needs. For example, for a hot strip mill production line, a user can adjust the roll gap value corresponding to the "medium" risk of roll eccentricity from 0.45mm to 0.30mm to provide stricter control. Because the strategy rules are decoupled from the underlying risk identification model, even if the model is upgraded or replaced later, the edited strategy rules do not need to be retested and verified, greatly improving the maintainability and flexibility of the system. The typical mapping relationships preset in the strategy library are as follows: When the risk level is "low", the target roll gap value is 0.80mm, the response time is 5 seconds, and the additional action is to start the laminar flow cooling system; when the risk level is "medium", the target roll gap value is 0.45mm, the response time is 2 seconds, and the additional action is to shut down unnecessary cooling branches; when the risk level is "high", the target roll gap value is 0.15mm, the response time is 1 second, and the additional action is to start a defect alarm; when the risk level is "extremely critical", the target roll gap value is the minimum safe roll gap value, the response time is 0.5 seconds, and the additional action is to cut off the power supply to non-critical equipment. It is understood that the specific values ​​and additional actions in the above mapping relationships are only examples, and in actual applications, they can be adjusted according to the physical characteristics of the rolling mill, rolling process conditions, and user needs. This invention does not limit these adjustments.

[0080] S520: Adjust the roll gap adjustment state of the steel rolling production line according to the roll gap target value, response time and additional action command, and realize micron-level stepless roll gap adjustment through the hydraulic AGC system; Specifically, the hydraulic AGC system performs precise roll gap adjustment based on the target roll gap value, bringing the roll gap to the set position. Since the requirements for roll gap accuracy vary under different abnormal production conditions, this hydraulic AGC system can optionally support micron-level stepless adjustment (±10μm), enabling precise positioning of any roll gap value within the set range. This avoids the coarse control problem caused by traditional control methods that only have limited ranges.

[0081] More specifically, based on the roll gap target value, response time, and additional action commands acquired by S510, the hydraulic AGC system is driven to complete the corresponding actions. Specifically, the hydraulic AGC system receives the roll gap target value command and adjusts the roll gap to the designated position within the specified response time. Ensuring the response time depends on the system's real-time processing capabilities and the rapid response characteristics of the actuator. The additional action commands synchronously trigger corresponding auxiliary equipment (such as laminar flow cooling systems, cooling branch control, defect alarm devices, and power cut-off devices for non-critical equipment).

[0082] In one specific embodiment, the hydraulic AGC system employs a hydraulic servo valve in conjunction with a roll gap fine-tuning device, supporting micron-level stepless adjustment (±10μm) and a repeatability accuracy ≤±10μm. Unlike traditional limited-gap control (such as only maximum roll gap, half roll gap, and minimum roll gap), micron-level stepless adjustment can continuously adjust the roll gap value according to the risk level, achieving refined control of rolling parameters. For example, when the temperature changes abnormally by 18°C, the system can adjust the roll gap value from 0.60mm to 0.38mm, reducing the rolling force fluctuation coefficient from 0.32 to 0.19. The repeatability accuracy ≤±10μm ensures that the roll gap can accurately reach the designated position during multiple adjustments, improving the system's reliability and stability. The achieved response time relies on the rapid inference of the lightweight edge model and the rapid actuation of the actuator; for example, under extremely critical conditions, the system can complete the entire process from risk identification to the start of roll gap adjustment within 0.5 seconds. Understandably, the specific mechanical structure (such as hydraulic servo valves, electric pressing, etc.) and roller gap adjustment method of the hydraulic AGC system can be selected according to the actual application scenario, as long as it can achieve micron-level stepless adjustment (±10μm) and high repeatability positioning accuracy.

[0083] S530: Real-time monitoring of the rolling force of the steel rolling production line. If the rolling force exceeds the preset rolling force threshold, it will automatically trigger a protective shutdown process to forcibly lock the roll gap and achieve hardware-level safety protection.

[0084] Specifically, during the roll gap adjustment process, the rolling force is monitored in real time. Rolling force is a physical quantity that measures the magnitude of the rolling load on the rolling mill. When the rolling force is too large, the rolls may be damaged due to excessive load, or abnormal fluctuations in rolling force may affect the quality of the strip steel product. Therefore, this step integrates a rolling force detection sensor to monitor the rolling force value on the rolling mill in real time. When the monitored rolling force exceeds a preset threshold (range 0-50000kN), the system automatically triggers the mill's protective shutdown procedure. Regardless of the current risk level, the mill is forcibly stopped and the roll gap is adjusted to ensure the safety of the rolls and strip steel products under extreme rolling force conditions. Through the above steps, this application achieves a complete closed loop from multimodal sensor data acquisition, production physical feature extraction, semantic fusion, risk identification to adaptive adjustment of rolling parameters, significantly improving the production stability and product quality assurance capabilities of the steel rolling production line under complex working conditions such as roll eccentricity, surface defects, and abnormal temperatures.

[0085] More specifically, the rolling force is monitored in real time during roll gap adjustment (including normal adjustment and static holding), and the monitored value is compared with a preset rolling force threshold. When the detected rolling force exceeds the threshold, regardless of the current risk level, the system automatically triggers a protective shutdown procedure, forcibly adjusting the roll gap to the minimum safe position. This protection mechanism, as a safety redundancy channel independent of the main control logic, can provide additional safety assurance under extreme production conditions (such as sudden strong temperature anomalies, equipment failures, etc.), preventing the mill from being damaged by excessive rolling force, or preventing abnormal operating conditions from seriously affecting the quality of strip steel products.

[0086] In one specific embodiment, rolling force detection is achieved through a six-dimensional force sensor integrated into the hydraulic AGC system. This sensor has a rolling force range of 0-50000kN and can monitor the forces and torques acting on the mill in different directions in real time, accurately sensing the magnitude and direction of the rolling force. The preset rolling force threshold can be set based on the mill's structural strength and historical rolling process data, for example, set to 80% of the mill's maximum rolling force. To ensure extremely high reliability of the protective shutdown, this embodiment employs a dual-channel protection design: when the digital channel outputs an "extremely dangerous" level from the risk level assessment model, it issues a minimum safe roll gap adjustment command through the hydraulic controller; the analog channel directly compares the analog signal from the rolling force sensor with the preset threshold using an independent comparator. When the rolling force exceeds the threshold, the comparator output signal directly pulls down the hardware enable, cutting off the hydraulic system power supply within 1 millisecond, causing the roll gap to automatically close to the minimum safe position under the action of the safety locking force. The analog channel is completely independent of the software system and processor; even if the software crashes or the processor freezes, the protective shutdown function can still be reliably triggered, thus achieving fault tolerance for software failures. It is understandable that other types of force sensors (such as strain gauges, piezoelectric sensors, etc.) can be used for rolling force detection, and other hardware logic circuits can be used for triggering protective shutdown, as long as rolling force over-threshold protection can be achieved independently of software.

[0087] More specifically, this invention establishes an integrated hierarchical adaptive control system encompassing "risk level - process parameters - safety protection," enabling refined, differentiated, and millisecond-level dynamic control for different operating conditions and risks. Based on extensive on-site rolling tests, fault cases, and process parameter calibration data, a dedicated roll gap control strategy library covering four risk levels—low, medium, high, and extremely dangerous—is built in advance. This strategy library independently configures dedicated roll gap adjustment amplitude, adjustment response speed, dynamic compensation coefficient, steady-state holding parameters, and emergency protection commands for each risk level and each type of fault, achieving a refined matching of one strategy per risk and one parameter per operating condition. After the system completes a quantitative assessment of the production risk level, it retrieves the corresponding standard control strategy in real time, accurately matching the roll gap target adjustment value, response time, and additional process control commands, abandoning the traditional coarse control mode of fixed parameters and uniform adjustment. Subsequently, the control commands are sent to the mill's hydraulic AGC automatic thickness control system. Relying on the high-precision response characteristics of the hydraulic servo mechanism, micron-level stepless continuous roll gap fine-tuning is achieved, dynamically compensating for roll gap offset and thickness deviation caused by eccentric vibration, temperature deformation, and surface defects, and adapting to the process requirements of the current fault condition in real time. While the process is adaptively optimized and controlled, the system collects rolling force operation data in real time at the millisecond level throughout the process, compares it with the equipment's rated safe rolling force threshold, and establishes a hardware-level safety fallback protection mechanism. Once dangerous conditions such as rolling force overload or sudden exceeding of the standard are detected, the conventional control logic is immediately interrupted, and the emergency lock-up and shutdown protection process is automatically triggered to forcibly fix the roll gap position, preventing major production safety accidents such as steel piling, strip breakage, and equipment overload damage, thus achieving an organic combination of flexible intelligent control and rigid safety protection.

[0088] This embodiment presents a preferred embodiment of step S500, completely overcoming the shortcomings of traditional rolling control systems with fixed parameters, lagging regulation, and lack of differentiation. It constructs a hierarchical, refined, millisecond-level response, and micron-level precision adaptive intelligent control system. Through precise mapping of risks and strategies, differentiated process control for different faults and risk levels is achieved, significantly improving the adaptability of rolling parameters to real-time operating conditions. This effectively corrects thickness deviations, strip shape fluctuations, and surface quality defects caused by dynamic operating conditions, significantly improving the precision and quality consistency of finished strip steel. Simultaneously, an innovative dual protection mechanism of "intelligent flexible parameter adjustment + hardware safety backup" is constructed. While achieving dynamic optimization of process parameters and improving product quality, it also prevents production accidents caused by overload and abnormal operating conditions from an equipment safety perspective, balancing production precision, efficiency, and safety. The overall control response speed is far faster than traditional manual and fixed logic control, perfectly adapting to the high-speed rolling production rhythm, significantly reducing the probability of unplanned downtime, and significantly improving the overall operating efficiency of the production line.

[0089] In summary, the aforementioned adaptive control method for a steel rolling production line acquires multimodal raw production data, extracts multimodal physical features for identifying production risks such as roll eccentricity, surface defects, or temperature anomalies, and performs semantic fusion on these multimodal physical features to generate a joint semantic vector. Based on this joint semantic vector, production risk data is obtained, and finally, rolling parameters are adaptively adjusted according to a preset control strategy and the production risk data. Compared to the coarse control methods in existing technologies that rely on preset rolling procedures or single sensors, this application perceives multiple production risks such as roll eccentricity, surface defects, and temperature anomalies in real time from multiple dimensions. Through semantic fusion, the system accurately understands the intrinsic relationships between different modal information, and adaptively adjusts rolling parameters based on the fused risk data. This enables precise adaptive control of the steel rolling production line based on process information.

[0090] Specifically, this invention adopts a fully intelligent control architecture encompassing "multimodal physical sensing, mechanism feature analysis, cross-modal deep fusion, differentiated risk assessment, and hierarchical adaptive closed-loop control," adapting to the complex industrial conditions of hot and cold continuous rolling production lines characterized by high speed, variable loads, multiple interferences, and coupled faults. During actual rolling production, the system is deployed around the clock with multiple types of sensing devices in the mill stand, roll gap area, and mill entrance / exit, synchronously and in parallel collecting raw production data across four dimensions: mechanical force, geometric contour, infrared temperature, and visible light texture. This eliminates the traditional single-sensor time-sharing acquisition mode with missing data dimensions, achieving comprehensive, highly synchronous, and high-density sampling of all elements of the production process. After data acquisition, the system does not directly perform model fitting and feature extraction on the raw data. Instead, based on the elastic-plastic deformation mechanism of steel rolling, the eccentric vibration mechanism of the roll body, the frictional heat generation mechanism of rolling, and the evolution mechanism of surface defects in strip steel, it performs mechanism-constrained precise feature extraction on four types of heterogeneous raw data. Redundant information unrelated to rolling faults is filtered out, and only highly correlated physical features that can directly characterize roll eccentricity imbalance, strip surface defect migration, and abnormal rolling temperature field are retained, ensuring the interpretability and process adaptability of the features. For the extracted one-dimensional temporal and two-dimensional spatial heterogeneous multimodal physical features, this invention abandons the traditional shallow fusion method of simple splicing and weighted summation. Instead, it uses a modality-specific deep learning backbone network to complete high-order feature purification. Through a cross-modal semantic interaction alignment mechanism, discrete features with different dimensions, different dimensions, and different physical meanings are mapped to a unified semantic space, generating a joint semantic vector with unified dimensions, coupled features, and quantifiable comparison, which accurately characterizes the complex working conditions of multi-fault coupling evolution on site. Based on the fused joint semantic vectors, combined with actual production line operating condition labels and prior industrial risk knowledge, a multi-dimensional and graded quantitative assessment of production risks is completed. This accurately outputs the probability of occurrence, risk level, and evolution trend of various faults, overcoming the limitation of traditional threshold judgments in identifying latent and early-stage faults. Finally, the system matches a preset refined graded control strategy according to the four-level risk levels, linking with the hydraulic AGC thickness control system to achieve dynamic adaptive fine-tuning of roll gap parameters. Simultaneously, it incorporates equipment overload safety protection logic, realizing a fully closed-loop, unmanned, and intelligent adaptive control system from operating condition perception and risk assessment to process control and safety protection. The entire process requires no manual intervention, and the response speed perfectly matches the high-speed rolling rhythm of steelmaking.

[0091] This solution precisely addresses the core shortcomings of existing technologies in the background field, namely "single-dimensional monitoring, isolated fault analysis, and inability to adapt to coupled faults." It innovatively selects three core faults—roll eccentricity, surface defects, and abnormal rolling temperature—that have the highest incidence, greatest impact, and strongest coupling in steel rolling operations as identification targets, constructing a dedicated multimodal perception-fusion-evaluation closed-loop system. From a production mechanism perspective, these three types of faults do not exist independently: roll eccentricity vibration causes periodic roll gap shifts, inducing uneven strip thickness and abnormal local frictional temperature rise; imbalance in the rolling temperature field leads to roll thermal deformation and fluctuations in strip plasticity, further exacerbating surface defects and equivalent eccentric disturbances; the dynamic diffusion of surface defects, in turn, alters the rolling contact load, causing continuous deterioration of operating conditions. Single fault monitoring cannot fully cover the actual coupled operating conditions in production. This application specifically extracts the dedicated physical features corresponding to the three types of faults and performs cross-modal semantic fusion to generate a joint semantic vector that can fully characterize the coupled evolution of faults, fundamentally solving the shortcomings of traditional technologies that can only monitor single operating conditions and cannot identify complex latent faults. At the same time, relying on unified semantic vectors to achieve global quantitative analysis of operating conditions, replacing traditional fragmented threshold judgment, the risk assessment is made to fit the real physical evolution mechanism of steel rolling, realizing early prediction of hidden faults, accurate location of coupled faults, and quantitative classification of risk levels. Finally, it matches adaptive control strategies, realizing a leapfrog upgrade of steel rolling control from "single-point passive alarm" to "global active prediction and control" at the architectural level, and significantly improving production line stability and product quality.

[0092] Based on the same inventive concept, this application also provides an adaptive control device for a rolling mill production line to implement the adaptive control method for the rolling mill production line described above. The solution provided by this device is based on the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the adaptive control device for a rolling mill production line provided below can be found in the limitations of the adaptive control method for the rolling mill production line described above, and will not be repeated here.

[0093] In one embodiment, such as Figure 4 As shown, an adaptive control device for a steel rolling production line is provided, comprising: The environmental data acquisition module 100 is used to acquire multimodal raw production data of the steel rolling production line; The feature extraction module 200 is used to extract the multimodal production physical features corresponding to the multimodal raw production data. The multimodal production physical features are used to identify production risks such as roll eccentricity, surface defects or abnormal temperature. The semantic fusion module 300 is used to semantically fuse multimodal production physical features to generate a joint semantic vector; Risk analysis module 400 is used to obtain production risk data of the steel rolling production line based on joint semantic vectors; The control module 500 is used to adaptively adjust the rolling parameters of the steel rolling production line according to the preset control strategy and production risk data.

[0094] In one embodiment, the environmental data acquisition module 100 is further configured to acquire rolling force data of the steel rolling production line via a rolling force sensor; acquire roll gap profile data of the steel rolling production line via a laser rangefinder; and acquire infrared thermal imaging images and visible light images of the steel rolling production line.

[0095] In one embodiment, the feature extraction module 200 is further used to extract the roll eccentric vibration intensity corresponding to the rolling force data; extract the surface defect movement speed corresponding to the roll gap contour data; extract the temperature gradient map corresponding to the infrared thermal imaging image; and extract the texture feature map corresponding to the visible light image.

[0096] In one embodiment, the semantic fusion module 300 is further used to compress multimodal production physical features into modal high-order features of a preset dimension through a modality-specific backbone network; the modality-specific backbone network adopts an architecture combining CNN and Transformer; the modal high-order features are concatenated to form a combined feature vector; the combined feature vector is input into a cross-modal Transformer, and semantic interaction and alignment between modalities are performed through a self-attention mechanism, outputting a joint semantic vector of a fixed dimension; different dimension segments of the joint semantic vector correspond to the roll eccentric vibration intensity scale, surface defect motion trend scale, temperature gradient scale and texture feature scale, respectively, so as to achieve unified representation and cross-modal comparison of multimodal production physical features in the same semantic space.

[0097] In one embodiment, the risk analysis module 400 is also used to identify the current production anomaly type based on the externally input production anomaly type label, and to assign dynamic attention weights to different modal features of the joint semantic vector using a cross-modal dynamic attention algorithm; the weighted joint semantic vector is input into a pre-trained production risk level assessment model to obtain the risk level for roll eccentricity, temperature anomaly and surface defects; the pre-trained production risk level assessment model is trained using Focal Loss and a cost-sensitive matrix.

[0098] In one embodiment, the control module 500 is further configured to query a preset roll gap control strategy library based on production risk data to obtain the roll gap target value, response time, and additional action instructions; adjust the roll gap adjustment state of the steel rolling production line according to the roll gap target value, response time, and additional action instructions; and detect the rolling force of the steel rolling production line in real time. If the rolling force exceeds the preset rolling force threshold, a protective shutdown process is automatically triggered.

[0099] Each module in the aforementioned adaptive control device for the steel rolling production line can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0100] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores preset data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements an adaptive control method for a steel rolling production line.

[0101] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0102] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described adaptive control method for a steel rolling production line.

[0103] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described adaptive control method for a steel rolling production line.

[0104] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An adaptive control method for a steel rolling production line, characterized in that, The method includes: S100: Acquire multimodal raw production data from the steel rolling production line; S200: Extract the multimodal production physical features corresponding to the multimodal raw production data. The multimodal production physical features are used to identify production risks such as roll eccentricity, surface defects, or abnormal temperature. S300: Perform semantic fusion on the multimodal production physical features to generate a joint semantic vector; S400: Based on the joint semantic vector, obtain the production risk data of the steel rolling production line; S500: Adaptively adjust the rolling parameters of the steel rolling production line according to the preset control strategy and the production risk data.

2. The adaptive control method for a steel rolling production line according to claim 1, characterized in that, The multimodal raw production data includes: rolling force data, roll gap profile data, infrared thermal imaging images, and visible light images; The extraction of multimodal production physical features corresponding to the multimodal raw production data includes: extracting the roll eccentric vibration intensity corresponding to the rolling force data; extracting the surface defect movement speed corresponding to the roll gap contour data; extracting the temperature gradient map corresponding to the infrared thermal imaging image; and extracting the texture feature map corresponding to the visible light image.

3. The adaptive control method for a steel rolling production line according to claim 2, characterized in that, The step of extracting the roll eccentric vibration intensity corresponding to the rolling force data includes: S211: Preprocess the original rolling force time series data, remove invalid working conditions such as strip threading, tail throwing, and acceleration / deceleration, and use wavelet threshold noise reduction to remove high-frequency electromagnetic interference and impact noise; S212: Calculate the fundamental frequency of the roll based on the real-time rotation speed of the roll; set a bandpass filter range based on the fundamental frequency to retain the 1st to 6th harmonic signals of the roll and filter out low-frequency material thickness fluctuation interference. S213: Perform a fast Fourier transform on the filtered rolling force time-series signal to obtain the spectral amplitude distribution and extract the spectral peak amplitude corresponding to the roll fundamental frequency; S214: The peak amplitude of the fundamental frequency spectrum is used as the instantaneous eccentric vibration intensity of the roll. At the same time, the proportion of harmonic energy per unit time is statistically analyzed to obtain the steady-state eccentric vibration intensity, thus completing the feature extraction of eccentric vibration intensity.

4. The adaptive control method for a steel rolling production line according to claim 2, characterized in that, The step of extracting the surface defect movement speed corresponding to the roll gap profile data includes: S221: Baseline calibration is performed on continuous multi-frame roll gap profile data to eliminate profile reference offset caused by mill roll-down drift and stand vibration, and a standardized profile sequence is obtained. S222: Perform a difference operation between the standardized contour of each frame and the standard flat roll gap contour, segment the defect area by a preset height threshold, and extract the defect pixel clusters and defect coordinates of the contour depression / protrusion. S223: The sparse optical flow matching algorithm is used to track and match the same defect region in consecutive frames to obtain the inter-frame displacement of the defect in the rolling advance direction and the frame acquisition time interval. S224: The instantaneous moving speed of the surface defect is calculated based on the inter-frame displacement and the frame acquisition time interval. The multi-frame speed data is smoothed and filtered to obtain a stable surface defect moving speed characteristic.

5. The adaptive control method for a steel rolling production line according to claim 2, characterized in that, The step of extracting the temperature gradient map corresponding to the infrared thermal imaging image includes: S231: Perform radiometric calibration, bad pixel repair and Gaussian smoothing correction on the original infrared thermal imaging image, and convert the grayscale image into a global temperature matrix. S232: Calculate the lateral and longitudinal partial derivatives of the temperature matrix to obtain the lateral and longitudinal temperature gradients. S233: Generate a global temperature gradient magnitude map based on the gradient magnitude, and simultaneously generate a gradient direction map; S234: Extract the effective temperature measurement ROI area of ​​the rolls and strip, remove background interference from the frame and water cooling equipment, and finally obtain a refined temperature gradient map for temperature anomaly identification.

6. The adaptive control method for a steel rolling production line according to claim 2, characterized in that, The step of extracting the texture feature map corresponding to the visible light image includes: S241: Perform grayscale conversion, illumination equalization, and noise reduction preprocessing on the original visible light image to eliminate uneven lighting and dust noise interference on site; S242: Use a 16×16 sliding window to traverse the entire image and construct a gray-level co-occurrence matrix window by window; S243: Based on the gray-level co-occurrence matrix, extract four types of texture parameters: contrast, energy, entropy, and correlation, and generate corresponding contrast texture maps, energy texture maps, entropy texture maps, and correlation texture maps pixel by pixel; S244: Fuse four types of single-channel texture maps to obtain a multi-dimensional fused texture feature map, which is used to characterize the surface roughness, defects, and oxide scale distribution characteristics of strip steel.

7. The adaptive control method for a steel rolling production line according to claim 1, characterized in that, The step of semantically fusing the multimodal production physical features to generate a joint semantic vector includes: S310: The multimodal production physical features are compressed into modal high-order features of a preset dimension through a modal-specific backbone network; the modal-specific backbone network adopts an architecture combining CNN and Transformer; for one-dimensional roll eccentric vibration intensity and surface defect movement speed, 1D-CNN combined with Transformer encoder is used to extract temporal high-order features; for two-dimensional temperature gradient map and texture feature map, 2D-CNN combined with Transformer encoder is used to extract spatial high-order features; S320: Concatenate the higher-order features of each modality to form a combined feature vector; S330: Input the combined feature vector into the cross-modal Transformer, perform semantic interaction and alignment between modalities through a multi-head self-attention mechanism, and output a joint semantic vector; the different dimensional segments of the joint semantic vector correspond to the roll eccentric vibration intensity scale, surface defect motion trend scale, temperature gradient scale and texture feature scale, respectively, so as to achieve unified representation and cross-modal comparison of multimodal production physical features in the same semantic space.

8. The adaptive control method for a steel rolling production line according to claim 1, characterized in that, The process of obtaining production risk data for the steel rolling production line based on the joint semantic vector includes: S410: Combining the modal coupling feature distribution of the joint semantic vector, the system adaptively determines whether the current single fault condition or multi-fault coupling condition is met. Through a self-designed scenario-layered dynamic weight iterative allocation mechanism, the system assigns differentiated and nonlinear dynamic weights to each modal feature within the joint semantic vector. The layered dynamic weight iterative allocation mechanism calculates the modal weights in real time based on the three dimensions of fault physical hazard, feature signal-to-noise ratio, and fault coupling correlation strength, thus abandoning fixed weights and general attention mapping logic. S420: Input the joint semantic vector after dynamic weighted optimization into the pre-trained production risk level assessment model, and output the independent risk probability and coupled risk coefficient corresponding to three types of faults: roll eccentricity, temperature abnormality and surface defects, respectively, and quantify the four levels of production risk: low, medium, high and extremely dangerous.

9. The adaptive control method for a steel rolling production line according to claim 8, characterized in that, The production risk level assessment model employs a composite loss function combining modal hierarchical adaptive FocalLoss and fault classification cost-sensitive matrix loss; the method for constructing the composite loss function includes: S431: Modal Layered FocalLoss Construction: Differential modulation factors are set based on the signal-to-noise ratio and sample distribution differences of four modal features: roll eccentric vibration intensity, surface defect movement speed, temperature gradient map, and texture feature map. For latent fault samples corresponding to temperature gradient and texture features, a large modulation factor is set to reduce the loss weight of conventional samples. For explicit fault samples corresponding to rolling force vibration and defect speed, a small modulation factor is set to balance the feature learning weights. Independent FocalLoss for each modal branch is calculated based on the differential modulation factors, and then fused to obtain the modal layered focusing loss. S432: Construction of Fault Classification Cost Sensitive Matrix: For three types of faults, namely roll eccentricity, abnormal temperature, and surface defects, independent four-level risk cost sensitive sub-matrices are constructed respectively; for low, medium, high, and extremely dangerous risk conditions in each sub-matrice, exponentially differentiated penalty coefficients are set for missed detection, false detection, and misjudgment, respectively, where the penalty coefficient for missed detection of extremely dangerous faults is 15 to 30 times that of false alarm penalty coefficient for minor faults; S433: Composite Loss Fusion: The modal hierarchical FocalLoss and the fault classification cost-sensitive loss are weighted and fused together. Dynamic fusion weights are set. In the early stage of model training, the weight of the cost-sensitive loss is increased to quickly converge the risk classification error. In the later stage of training, the weight of the modal hierarchical FocalLoss is increased to refine the mining of hidden fault features.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 9.

Images