Intelligent AI positioning method and system for multi-blank-type and multi-fixed-length steel billets of heating furnace

Abstract

The invention discloses an intelligent AI positioning method and system for a multi-blank-type multi-fixed-length steel billet of a heating furnace, and relates to the field of steel billet positioning, and the method comprises the steps: obtaining image data and point cloud data corresponding to the multi-blank-type multi-fixed-length steel billet, and generating a steel billet state three-dimensional model in combination with a registration and alignment technology, steel billet positioning parameters including posture and position variables are obtained; establishing a time sequence data set based on billet positioning parameters, analyzing a billet positioning deviation state in combination with a deep learning model, adjusting an electric roller way according to a deviation state result, and calibrating the position state of the billet; and the calibrated billet position information is sent to a pusher controller to generate an action sequence instruction, the action sequence instruction is combined with a closed-loop adjusting mechanism to push billets into a heating furnace, and automatic charging operation of the billets is achieved. According to the method, the three-dimensional model combination is constructed by using the image data and the point cloud data, accurate modeling can be carried out on the steel billets of any specification and any arrangement state, and intelligent positioning of the steel billets can be conveniently realized in the later period.

Description

Technical Field

[0001] The present invention relates to the field of steel billet positioning, and in particular to an intelligent AI positioning method and system for multi-type and multi-size steel billets in a heating furnace. Background Art

[0002] In recent years, the rapid development of industrial automation and intelligent manufacturing technologies has promoted the important role of heating furnaces in the metallurgical industry. In particular, efficient and accurate positioning has become key to improving production efficiency and energy utilization in the processing of multi-shape and multi-size billets. Traditional methods rely on manual operation or simple sensor detection, but as process complexity and quality requirements increase, their limitations gradually become apparent. Therefore, intelligent positioning technologies based on computer vision, point cloud processing, and artificial intelligence (AI) have emerged. Algorithms such as support vector machines (SVM) and deep learning are used for posture recognition and position prediction, significantly improving positioning accuracy. At the same time, the combination of three-dimensional modeling and point cloud data processing can comprehensively describe the spatial characteristics of the billet, providing a reliable basis for furnace loading.

[0003] However, the existing technology still has shortcomings: first, when faced with steel billets with complex postures, a single feature extraction method is difficult to accurately capture detailed features such as end position and edge contour, affecting positioning accuracy; second, it lacks dynamic deviation correction capabilities. In actual production, steel billets may be displaced or their posture changes due to factors such as vibration. The existing technology cannot adjust parameters in real time, resulting in poor loading effects, reduced production efficiency, and increased energy consumption and quality risks.

[0004] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0005] In response to the problems in the related art, the present invention proposes an intelligent AI positioning method and system for multi-type and multi-size steel billets in a heating furnace to overcome the above-mentioned technical problems existing in the existing related art.

[0006] To this end, the specific technical solutions adopted in the present invention are as follows:

[0007] In a first aspect, the present invention proposes an intelligent AI positioning method for multi-type and multi-size billets in a heating furnace, comprising:

[0008] Acquire image data and point cloud data corresponding to multiple types and sizes of billets, and generate a 3D model of the billet state using registration and alignment technology to obtain billet positioning parameters including posture and position variables;

[0009] A time series dataset is established based on the billet positioning parameters, and the billet positioning deviation status is analyzed using a deep learning model. The electric roller conveyor is adjusted based on the deviation status results to calibrate the billet position status.

[0010] The calibrated billet position information is sent to the pusher controller to generate action sequence instructions, and combined with the closed-loop adjustment mechanism, the billet is pushed into the heating furnace to realize the automated billet loading operation.

[0011] Preferably, the image data and point cloud data corresponding to the multi-blank type and multi-size steel billets are obtained, and the three-dimensional model of the steel billet state is generated by combining the registration and alignment technology to obtain the steel billet positioning parameters including posture and position variables, including:

[0012] Obtain the original 2D image and 3D point cloud data of multiple-type and multiple-size steel billets, and grayscale, remove noise, and resize the original 2D image to obtain the 2D image of the steel billet;

[0013] The dark channel prior theory and image enhancement algorithm are used to perform feature enhancement processing on the two-dimensional image of the steel billet, and the geometric feature points of the steel billet are obtained by aligning it with the three-dimensional point cloud data to generate regularized point cloud data.

[0014] The regularized point cloud data is initially aligned using a global registration strategy, and feature extraction is performed in combination with descriptors to generate an initial transformation matrix. A three-dimensional model of the billet state is constructed based on the initial transformation matrix.

[0015] Based on the feature histogram algorithm, the geometric features including the spatial posture and position coordinates of the billet are extracted from the three-dimensional model of the billet state, and the geometric features are used to train the hybrid intelligent model to output the billet positioning parameters.

[0016] Preferably, the dark channel prior theory and image enhancement algorithm are used to perform feature enhancement processing on the two-dimensional image of the steel billet, and the geometric feature points of the steel billet are obtained by aligning with the three-dimensional point cloud data to generate regularized point cloud data, which includes:

[0017] Based on the dark channel prior theory, the global atmospheric light value and transmittance map of the two-dimensional image of the steel billet are estimated. The dynamic blur estimation model and regularization technology are combined to extract blur features and eliminate the dynamic blur effect.

[0018] The defogging result is used to obtain a two-dimensional image of the steel billet. The reflection vector and illumination component of the two-dimensional image of the steel billet are decomposed based on the image enhancement algorithm to enhance the contrast between the steel billet and the background.

[0019] An edge-preserving filter is used to remove the halo effect of the 2D image of the steel billet to obtain an enhanced 2D image of the steel billet. At the same time, denoising and downsampling are performed on the 3D point cloud data.

[0020] Project the 3D point cloud data onto the 2D image of the steel billet, establish the correspondence between the point cloud and the image using geometric constraints, and dynamically estimate the relative posture of the point cloud and the image in combination with an adjustment strategy;

[0021] According to the estimation results, geometric feature points are extracted from the three-dimensional point cloud data and the two-dimensional image of the steel billet, and feature descriptors are generated to represent the local geometric features, which are combined with the fusion strategy to generate regularized point cloud data.

[0022] Preferably, extracting geometric features including the spatial posture and position coordinates of the billet from the three-dimensional model of the billet state based on a feature histogram algorithm, and using the geometric features to train a hybrid intelligent model to output billet positioning parameters includes:

[0023] The geometric features of the spatial posture and position coordinates of the billet in the 3D model of the billet state are extracted based on the fast point feature histogram, and high-dimensional feature information is generated by combining the adaptive weight distribution strategy;

[0024] Integrate high-dimensional feature information with corresponding labels to form a structured dataset, and use a hierarchical partitioning strategy to divide the dataset into training set, validation set, and test set in proportion;

[0025] Use training sets, validation sets, and test sets to train and adjust the hybrid intelligent model that includes deep learning and support vector machines, and gradually optimize the model parameters based on forward propagation and backpropagation;

[0026] The geometric features are input into the trained hybrid intelligent model to extract high-level features using deep learning technology, and combined with the positioning parameters of the support vector machine containing the spatial posture and position coordinates of the steel billet.

[0027] Preferably, the calculation formula of the positioning parameter is:

[0028]

[0029] Where P represents the positioning parameter, X represents the spatial posture and position coordinate vector to be optimized, φ represents the nonlinear mapping function of the deep neural network, F represents the geometric feature, θ represents the parameter geometry of the hybrid intelligent model, φ(F;θ) represents the predicted value of the geometric feature F, T represents the target reference value, represents the Euclidean norm squared, T i represents the i-th target value, K(F,T i ) represents the capture of geometric features F and target reference value T i The nonlinear relationship between represents the adaptive weight of the i-th target reference value, λ represents the adjustment coefficient, N represents the number of target values ​​in the set, and i represents the index of the target reference value.

[0030] Preferably, a time series data set is established based on the billet positioning parameters, and the billet positioning deviation state is analyzed in combination with a deep learning model. The electric roller is adjusted according to the deviation state result, and the position state of the billet is calibrated, which includes:

[0031] The billet positioning parameters are recorded at set time intervals to obtain a positioning time series dataset, and a deep learning model is trained to analyze the differences between the positioning time series dataset and the standard positioning parameters.

[0032] The positioning deviation of the billet is quantified based on the difference results, and the relative posture deviation of the billet is determined based on the positioning deviation. At the same time, the relative posture of the electric roller is analyzed using a pre-integration algorithm.

[0033] The relative posture deviation of the billet and the relative posture of the electric roller are cross-correlated and calculated, and the time offset between the billet and the electric roller is estimated based on the calculation results and the rotation constraint technology.

[0034] The speed and angle of the electric roller are adjusted based on the time offset estimation results. According to the adjustment results, the billet is adjusted from the deviation state to the standard positioning state to complete the position state of the billet, and the hybrid intelligent model is used to output the calibration positioning parameters of the billet.

[0035] Preferably, performing cross-correlation calculation on the relative posture deviation of the billet and the relative posture of the electric roller, and obtaining a time offset estimation result between the billet and the electric roller according to the calculation result and the rotation constraint technology includes:

[0036] Based on the recursive least squares classifier, the kernel correlation filters of the relative posture deviation of the billet and the relative posture of the electric roller table are judged respectively, and the maximum kernel correlation filter response value corresponding to the two is obtained;

[0037] Analyze the difference between the maximum kernel correlation filter response values ​​of the two, and find the corresponding attitude position data of the billet during the positioning calibration process within the billet positioning parameters based on the difference result;

[0038] An attitude error compensation model is constructed based on the attitude position data. After extracting the motion parameters of the electric roller conveyor, an auxiliary calibration model is constructed in combination with the kernel correlation filter to output the time offset estimation result.

[0039] Preferably, a posture error compensation model is constructed based on the posture position data, and an auxiliary calibration model is constructed by combining a kernel correlation filter after extracting the motion parameters of the electric roller, and the output time offset estimation result includes:

[0040] Analyze the identification function of the billet moving point and the center of gravity of the static point in the posture position data, perform affine transformation on the posture position data to determine the affine area, and combine the identification function and the center of gravity of the static point to generate a posture error compensation model;

[0041] Extract motion parameters based on the working principle of the electric roller, analyze the motion posture sample items of the electric roller according to the motion parameters, and integrate them with the posture error compensation model to obtain the motion posture error of the electric roller;

[0042] According to the motion principle of the electric roller table, the likelihood function of the electric roller table and the local gradient energy term of the motion are obtained to generate the reconstructed feature of the electric roller table motion. The motion parameter distribution of the electric roller table is analyzed in combination with the motion posture error.

[0043] Based on the distribution of electric roller table motion parameters and kernel correlation filter, an electric roller table posture auxiliary calibration model is established, and the electric roller table motion auxiliary correction result is output as the time offset estimation result.

[0044] Preferably, the calibrated billet position information is sent to the pusher controller to generate an action sequence instruction, and the billet is pushed into the heating furnace in combination with the closed-loop regulation mechanism to realize the automatic billet loading operation, which includes:

[0045] The calibrated billet position information is mapped to the pusher operation coordinate system. At the same time, the pusher controller extracts the billet position information and determines the pushing path between the billet and the heating furnace.

[0046] According to the pushing path, the mechanical constraints of the steel pusher during the pushing process are judged, and the optimal driving path of the steel pusher is generated in combination with the pushing parameters of the steel pusher, and the optimal driving path is converted into action sequence instructions;

[0047] The steel billets are pushed into the heating furnace using action sequence instructions to realize the automatic charging operation of the steel billets, and the action parameters of the pusher are dynamically adjusted during the pushing process in combination with model predictive control.

[0048] In a second aspect, the present invention further provides an intelligent AI positioning system for multiple-type and multiple-size steel billets in a heating furnace, the positioning system comprising:

[0049] The billet positioning parameter acquisition module is used to obtain image data and point cloud data corresponding to billets of multiple types and sizes, and generate a three-dimensional model of the billet state by combining registration and alignment technology to obtain billet positioning parameters including posture and position variables;

[0050] The billet position state calibration module is used to establish a time series data set based on the billet positioning parameters, analyze the billet positioning deviation state in combination with the deep learning model, adjust the electric roller conveyor according to the deviation state results, and calibrate the billet position state;

[0051] The billet control automatic loading module is used to send the calibrated billet position information to the pusher controller to generate action sequence instructions, and combine with the closed-loop adjustment mechanism to push the billet into the heating furnace to realize the automatic loading operation of the billet.

[0052] The beneficial effects of the present invention are:

[0053] 1. The present invention uses image data and point cloud data to build a three-dimensional model, which can accurately model steel billets of any specifications and any arrangement status, facilitating the intelligent positioning of the steel billets in the later stage. At the same time, the time series generated by the steel billet positioning results is used to construct the steel billet motion trajectory. Combined with the deep learning model, it is determined whether the steel billet is offset, and the position and posture of the steel billet are automatically adjusted according to the offset status, realizing real-time and accurate calibration, and ensuring the success rate of the steel billet entering the furnace.

[0054] 2. Through image enhancement, the present invention realizes the generation of high-quality images in complex industrial environments, thereby providing reliable basic data for subsequent three-dimensional point cloud alignment and modeling, ultimately improving the overall accuracy and robustness of the system. Through hybrid intelligent model positioning, efficient modeling and accurate identification of complex geometric features are achieved, thereby completing automated furnace loading control and significantly improving production efficiency and positioning accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 This is a flow chart of an intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to an embodiment of the present invention;

[0057] Figure 2 The present invention is a block diagram of the principle of an intelligent AI positioning system for multiple-type and multiple-size steel billets in a heating furnace according to an embodiment of the present invention.

[0058] In the picture:

[0059] 1. Billet positioning parameter acquisition module; 2. Billet position status calibration module; 3. Billet control automatic furnace loading module. DETAILED DESCRIPTION

[0060] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and the advantages of the present invention.

[0061] According to an embodiment of the present invention, an intelligent AI positioning method and system for multi-type and multi-size steel billets in a heating furnace are provided.

[0062] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1As shown, the intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to an embodiment of the present invention includes:

[0063] Step S1, obtaining image data and point cloud data corresponding to multi-blank and multi-size steel billets, and generating a three-dimensional model of the steel billet state by combining registration and alignment technology to obtain steel billet positioning parameters including posture and position variables.

[0064] In one embodiment, image data and point cloud data corresponding to multiple-type and multiple-size steel billets are obtained, and a three-dimensional model of the steel billet state is generated by combining registration and alignment technology. The steel billet positioning parameters including posture and position variables are obtained, including:

[0065] Obtain the original 2D image and 3D point cloud data of multiple-type and multiple-size steel billets, and grayscale, remove noise, and resize the original 2D image to obtain the 2D image of the steel billet;

[0066] The dark channel prior theory and image enhancement algorithm are used to perform feature enhancement processing on the two-dimensional image of the steel billet, and the geometric feature points of the steel billet are obtained by aligning it with the three-dimensional point cloud data to generate regularized point cloud data.

[0067] The regularized point cloud data is initially aligned using a global registration strategy, and feature extraction is performed in combination with descriptors to generate an initial transformation matrix. A three-dimensional model of the billet state is constructed based on the initial transformation matrix.

[0068] Based on the feature histogram algorithm, the geometric features including the spatial posture and position coordinates of the billet are extracted from the three-dimensional model of the billet state, and the geometric features are used to train the hybrid intelligent model to output the billet positioning parameters.

[0069] Among them, the dark channel prior theory and image enhancement algorithm are used to perform feature enhancement processing on the two-dimensional image of the steel billet, and the geometric feature points of the steel billet are obtained by alignment with the three-dimensional point cloud data to generate regularized point cloud data including:

[0070] Based on the dark channel prior theory, the global atmospheric light value and transmittance map of the two-dimensional image of the steel billet are estimated. The dynamic blur estimation model and regularization technology are combined to extract blur features and eliminate the dynamic blur effect.

[0071] The defogging result is used to obtain a two-dimensional image of the steel billet. The reflection vector and illumination component of the two-dimensional image of the steel billet are decomposed based on the image enhancement algorithm to enhance the contrast between the steel billet and the background.

[0072] An edge-preserving filter is used to remove the halo effect of the 2D image of the steel billet to obtain an enhanced 2D image of the steel billet. At the same time, denoising and downsampling are performed on the 3D point cloud data.

[0073] Project the 3D point cloud data onto the 2D image of the steel billet, establish the correspondence between the point cloud and the image using geometric constraints, and dynamically estimate the relative posture of the point cloud and the image in combination with an adjustment strategy;

[0074] According to the estimation results, geometric feature points are extracted from the three-dimensional point cloud data and the two-dimensional image of the steel billet, and feature descriptors are generated to represent the local geometric features, which are combined with the fusion strategy to generate regularized point cloud data.

[0075] Specifically, the geometric features including the spatial posture and position coordinates of the billet are extracted from the three-dimensional model of the billet state based on the feature histogram algorithm, and the hybrid intelligent model is trained using the geometric features to output the billet positioning parameters, including:

[0076] The geometric features of the spatial posture and position coordinates of the billet in the 3D model of the billet state are extracted based on the fast point feature histogram, and high-dimensional feature information is generated by combining the adaptive weight distribution strategy;

[0077] Integrate high-dimensional feature information with corresponding labels to form a structured dataset, and use a hierarchical partitioning strategy to divide the dataset into training set, validation set, and test set in proportion;

[0078] Use training sets, validation sets, and test sets to train and adjust the hybrid intelligent model that includes deep learning and support vector machines, and gradually optimize the model parameters based on forward propagation and backpropagation;

[0079] The geometric features are input into the trained hybrid intelligent model to extract high-level features using deep learning technology, and combined with the positioning parameters of the support vector machine containing the spatial posture and position coordinates of the steel billet.

[0080] It should be explained that in the process of obtaining the positioning parameters, a high-definition camera is used to collect the original two-dimensional image data of the steel billet in real time, and the original two-dimensional image data is preliminarily processed. The image is enhanced by combining the dark channel prior algorithm, convolutional neural network and Retinex algorithm. The original two-dimensional image data of the steel billet includes the position, contour, size, posture, surface texture information and end features of the steel billet.

[0081] At the same time, high-definition cameras are installed at the openings on the furnace roof at both ends of the heating furnace entrance side to ensure that the camera's field of view can completely cover the entire furnace area and collect the original two-dimensional image data inside the heating furnace in real time.

[0082] Preliminary processing includes grayscale processing, denoising processing and resizing. First, grayscale processing is performed to convert the color image into a grayscale image to reduce data redundancy and enhance contrast. Denoising processing is used to remove noise interference in the image and improve image quality. Finally, resizing is performed to unify the image to a standard size to facilitate subsequent algorithm processing and feature extraction.

[0083] Based on the preprocessed two-dimensional image data, the dark channel prior algorithm is used in combination with the dynamic blur estimation model to accurately remove the complex blur effects caused by industrial environmental factors and generate a defogged image. The specific process is to use the dark channel prior theory to estimate the global atmospheric light value and transmittance map, restore the basic structure of the fog-free image, and at the same time use the dynamic blur estimation model to extract blur features and generate blur kernel parameters, and eliminate the dynamic blur effect through inverse convolution or regularization methods. The whole process realizes the synergistic effect of defogging and blur removal through iterative optimization, and is supplemented by post-processing steps to improve image quality, and finally generates a clear and detail-rich defogging image suitable for complex industrial application scenarios.

[0084] Based on the dehazed image, an end-to-end optimization framework is constructed through a deep convolutional neural network (DCNN), which simultaneously realizes high-frequency noise suppression and steel billet key feature extraction, and generates high-fidelity images. A multi-branch network structure is designed, in which the high-frequency noise suppression branch extracts low-frequency information through the attention mechanism and residual learning module, while the steel billet key feature extraction branch uses deep convolutional layers and void convolution to capture surface texture and edge features. The results of the two branches are integrated through a fusion layer and optimized under the guidance of a multi-task loss function, including noise suppression loss, feature extraction loss and high-fidelity image reconstruction loss. Through training and optimization, high-quality and high-fidelity images are finally generated to meet the strict requirements for steel billet detection in industrial scenarios.

[0085] At the same time, an adaptive Retinex algorithm is used to adjust the brightness distribution. By decomposing the reflection component and the illumination component and dynamically adjusting the parameters, the contrast between the steel billet and the background is enhanced. At the same time, an edge-preserving filter (such as bilateral filtering or guided filtering) is combined to remove the halo effect, avoid blurring or over-enhancement of local image details, and further improve image quality. Finally, through the comprehensive processing of the brightness adjustment and halo removal results, a high-contrast fine structure image is generated, highlighting the key features of the steel billet, meeting the strict requirements for image clarity and contrast in industrial inspection.

[0086] At the same time, high-precision laser scanners at both ends of the heating furnace collect the distance, shape, and surface details of the steel billet, generating high-density 3D point cloud data. The distance of the steel billet is expressed as:

[0087]

[0088] Where d represents the distance from the target object to the sensor, c represents the speed of light in a vacuum, and t represents the total time delay from the laser signal being emitted to being reflected back by the target.

[0089] The specific process is to transmit lasers through the scanner and receive reflected signals, calculate the spatial coordinates of each point based on the time-of-flight method, and record additional information such as the intensity of reflected light to generate high-density three-dimensional point cloud data containing a large number of geometric and material features. By performing optimization processing such as denoising, interpolation and downsampling on the point cloud data, the data quality is further improved, providing accurate basic data support for subsequent three-dimensional modeling, dimensional detection and defect analysis.

[0090] A projection transformation algorithm and an adaptive preliminary adjustment strategy are used to perform geometrically constrained alignment of high-contrast fine structure images and three-dimensional point cloud data. It should be noted that the adaptive preliminary adjustment strategy is a dynamic optimization method. By analyzing the characteristics of the input data, it automatically adjusts the key parameters of the algorithm or the state of the model, thereby providing a better starting point for subsequent processing. Its core lies in "dynamic" and "feedback", that is, dynamically setting parameters according to data characteristics and continuously optimizing through error evaluation. This strategy is widely used in point cloud alignment, denoising, model training, online learning and other fields. It can significantly reduce initial errors, improve convergence speed, and enhance the flexibility and robustness of the system.

[0091] Specifically, through the projection transformation algorithm, the three-dimensional point cloud data is projected onto the two-dimensional image plane, and the correspondence between the point cloud and the image is established using the geometric constraint relationship. Combined with the adaptive preliminary adjustment strategy, the relative posture of the point cloud and the image is dynamically estimated, and outliers are eliminated to reduce errors. On this basis, the alignment accuracy is further improved through iterative optimization methods, and finally the geometric constraint alignment of high-contrast fine structure images and three-dimensional point cloud data is achieved, providing a reliable basic support for subsequent data fusion and analysis.

[0092] After geometric constraint alignment, the DeepFusion algorithm is used to extract geometric feature points from the high-contrast fine-structure image and 3D point cloud data. Feature point registration is then combined with the RANSAC algorithm to eliminate outliers and achieve precise alignment. Specifically, based on geometric constraint alignment, the DeepFusion algorithm is used to extract geometric feature points from the high-contrast fine-structure image and 3D point cloud data, and corresponding feature descriptors are generated to characterize local geometric characteristics. After establishing a preliminary correspondence through feature point matching, the RANSAC algorithm is used to eliminate outliers and screen out reliable matching point pairs. Finally, based on the screening results, further optimization is performed, using the ICP algorithm or other nonlinear optimization methods to minimize the overall error, thereby achieving precise alignment of the high-contrast fine-structure image and 3D point cloud data, providing high-quality fused data support for subsequent 3D modeling and analysis.

[0093] Statistical filtering is used to remove noise points, a grid-based method is used to remove duplicate points, and bilateral filtering is used to smooth the point cloud surface and retain geometric details. The specific process is to use statistical filtering, grid-based method and bilateral filtering techniques on point cloud data in sequence to achieve noise point removal, duplicate point elimination and point cloud surface smoothing. Specifically, statistical filtering removes isolated noise points by calculating local statistical characteristics; the grid-based method reduces redundant information by dividing the three-dimensional space and merging duplicate points; bilateral filtering combines spatial distance and normal vector consistency to smooth the point cloud surface while retaining important geometric details. After the above steps, regularized and high-quality point cloud data is generated, providing a reliable basic support for subsequent three-dimensional modeling and analysis.

[0094] Based on image alignment and data preprocessing, regularized three-dimensional point cloud data is generated through a multi-scale fusion strategy. It should be noted that the multi-scale fusion strategy extracts information from different scales (such as spatial resolution, feature hierarchy, etc.) and effectively integrates it to achieve a comprehensive description of the data; specifically, the strategy includes multi-scale feature extraction, feature alignment and standardization, fusion method selection, optimization and verification, etc. In practical applications, the multi-scale fusion strategy is widely used in three-dimensional point cloud processing, image recognition, hybrid intelligent models and other fields, which can significantly improve the robustness and accuracy of the system while taking into account global and local characteristics.

[0095] Specifically, a multi-scale pyramid of point cloud data is first constructed, and global and local geometric features are extracted at each scale. Then, features at different scales are fused through a dynamic weight distribution mechanism to ensure a balance between global shape and local details. Finally, the fused point cloud data is resampled and topologically optimized to generate regularized point cloud data with uniform distribution and high quality, providing a solid foundation for subsequent 3D modeling and feature analysis.

[0096] In the process of generating a three-dimensional model, the global registration strategy FGR is used to perform initial alignment on the regularized three-dimensional point cloud data. By constructing a global distance matrix and robust optimization matching, the rough transformation relationship between point clouds is quickly estimated. On this basis, the multi-scale geometric feature descriptor FPFH is used to extract the local and overall geometric characteristics of the point cloud, and the correspondence between point clouds is established through feature matching. Finally, the RANSAC algorithm is combined to eliminate outliers, and the screened matching point pairs are used to estimate the initial transformation matrix, thereby providing a reliable basic support for subsequent precise registration.

[0097] Based on the initial transformation matrix, the Truncated Least Squares (ICP) algorithm is used to align the three-dimensional point cloud data to the same coordinate system. The specific process is to first use the initial transformation matrix to roughly align the point cloud, then establish a matching relationship through iterative nearest point search, and introduce a truncated error function to screen valid point pairs to improve robustness. In each iteration, the rotation and translation parameters are optimized by minimizing the overall error of the matching point pairs until the convergence conditions are met. Finally, an accurate transformation matrix is ​​generated, and all point cloud data are aligned to the same coordinate system, providing high-quality data support for subsequent fusion and modeling.

[0098] Based on the registered 3D point cloud data, geometric defects are repaired by fusing overlapping areas, removing duplicate points and smoothing, and combined with topology optimization algorithms to generate high-quality 3D models. Specifically, redundant information and noise interference are eliminated by fusing overlapping areas, removing duplicate points and smoothing. At the same time, topology optimization algorithms are used to detect and repair geometric defects (such as holes, fractures, etc.) to ensure the integrity and geometric continuity of the model. Finally, through the refined processing and reconstruction of point cloud data, high-quality 3D models are generated to provide accurate data support for subsequent applications (such as detection, analysis or manufacturing).

[0099] In the process of identifying the spatial pose and position coordinates of high-quality 3D models, the Fast Point Feature Histogram (FPFH) algorithm is used to extract the geometric features of the spatial pose and position coordinates of the steel billet in the 3D model. Combined with an adaptive weight allocation strategy, high-dimensional feature information is generated. It should be noted that the adaptive weight allocation strategy is a dynamic optimization method that automatically adjusts the weights of different features, parameters, or modules based on data characteristics or task requirements, thereby enhancing system performance and robustness. The specific process includes feature or module importance assessment, weight initialization, dynamic weight adjustment, and verification and optimization. In practical applications, this strategy is widely used in fields such as multi-scale feature fusion, hybrid intelligent models, and online learning. It can significantly improve the system's flexibility and adaptability while ensuring efficient resource utilization.

[0100] Among them, through local neighborhood analysis and multi-scale feature extraction, a high-dimensional feature vector containing information such as normal vector and curvature is generated. At the same time, combined with an adaptive weight allocation strategy, the importance of different features is dynamically evaluated and weights are assigned to achieve weighted fusion of features, thereby generating more accurate and robust high-dimensional feature information, providing high-quality data support for the intelligent positioning and classification of steel billets.

[0101] Based on high-dimensional feature information, a data set is constructed and divided into training set, validation set and test set. The specific process is to first integrate the geometric features extracted by the FPFH algorithm with the corresponding labels into a structured data set, and perform preprocessing operations such as normalization, data cleaning and enhancement on the feature vector. Subsequently, random partitioning or hierarchical partitioning methods are used to divide the data set into training set, validation set and test set in proportion, which are used for model training, hyperparameter adjustment and final performance evaluation respectively. Throughout the process, the consistency of data distribution is ensured to avoid category imbalance or data leakage problems, thereby providing high-quality data support for the training and verification of hybrid intelligent models.

[0102] Use the training set to train the hybrid intelligent model, the validation set to optimize the hyperparameters, and the test set to evaluate the performance of the hybrid intelligent model. Combined with the online learning mechanism, the hybrid intelligent model can dynamically adjust the parameters according to real-time data; use the training set to train the hybrid intelligent model (deep learning combined with SVM), and gradually optimize the model parameters through forward propagation and backpropagation; use the validation set to evaluate the performance of different hyperparameter combinations, and select the optimal hyperparameters to avoid overfitting; finally, comprehensively evaluate the generalization ability and performance indicators of the model on the test set. At the same time, combined with the online learning mechanism, by collecting new data in real time and dynamically adjusting the model parameters, the hybrid intelligent model can adapt to the ever-changing industrial environment, continuously improve positioning accuracy and robustness, and meet actual production needs.

[0103] After training is completed, the extracted geometric features are input into the hybrid intelligent model to generate positioning parameters, which are expressed as:

[0104]

[0105] Where P represents the positioning parameter, X represents the spatial posture and position coordinate vector to be optimized, φ represents the nonlinear mapping function of the deep neural network, F represents the geometric feature, θ represents the parameter geometry of the hybrid intelligent model, φ(F;θ) represents the predicted value of the geometric feature F, T represents the target reference value, represents the Euclidean norm squared, T i represents the i-th target value, K(F,T i ) represents the capture of geometric features F and target reference value T i The nonlinear relationship between represents the adaptive weight of the i-th target reference value, λ represents the adjustment coefficient, N represents the number of target values ​​in the set, and i represents the index of the target reference value.

[0106] The specific process is that after the training is completed, the high-dimensional geometric feature vector extracted and standardized by the FPFH algorithm is input into the hybrid intelligent model. The model extracts high-level features through the deep learning part, and the support vector machine SVM part generates the positioning parameters of the spatial posture and position coordinates of the steel billet. The output positioning parameters are post-processed and calibrated to ensure that their accuracy and stability meet the requirements of industrial applications. This process realizes the efficient mapping from geometric features to positioning parameters, and provides accurate data support for automated furnace loading control.

[0107] Step S2: Establish a time series data set based on the billet positioning parameters, and analyze the billet positioning deviation state in combination with the deep learning model. Adjust the electric roller according to the deviation state results to calibrate the position state of the billet.

[0108] In one embodiment, a time series dataset is established based on the billet positioning parameters, and the billet positioning deviation state is analyzed in combination with a deep learning model. The electric roller conveyor is adjusted according to the deviation state result. The billet position state calibration includes:

[0109] The billet positioning parameters are recorded at set time intervals to obtain a positioning time series dataset, and a deep learning model is trained to analyze the differences between the positioning time series dataset and the standard positioning parameters.

[0110] The positioning deviation of the billet is quantified based on the difference results, and the relative posture deviation of the billet is determined based on the positioning deviation. At the same time, the relative posture of the electric roller is analyzed using a pre-integration algorithm.

[0111] The relative posture deviation of the billet and the relative posture of the electric roller are cross-correlated and calculated, and the time offset between the billet and the electric roller is estimated based on the calculation results and the rotation constraint technology.

[0112] The speed and angle of the electric roller are adjusted based on the time offset estimation results. According to the adjustment results, the billet is adjusted from the deviation state to the standard positioning state to complete the position state of the billet, and the hybrid intelligent model is used to output the calibration positioning parameters of the billet.

[0113] Specifically, the relative posture deviation of the billet and the relative posture of the electric roller are cross-correlated and calculated, and the time offset estimation result between the billet and the electric roller is obtained based on the calculation result and the rotation constraint technology, including:

[0114] Based on the recursive least squares classifier, the kernel correlation filters of the relative posture deviation of the billet and the relative posture of the electric roller table are judged respectively, and the maximum kernel correlation filter response value corresponding to the two is obtained;

[0115] Analyze the difference between the maximum kernel correlation filter response values ​​of the two, and find the corresponding attitude position data of the billet during the positioning calibration process within the billet positioning parameters based on the difference result;

[0116] An attitude error compensation model is constructed based on the attitude position data. After extracting the motion parameters of the electric roller conveyor, an auxiliary calibration model is constructed in combination with the kernel correlation filter to output the time offset estimation result.

[0117] It should be explained that in the process of obtaining posture position data, kernel correlation filters (KCFs) are constructed for the steel billet and the electric roller respectively to efficiently track their relative posture changes: KCF efficiently performs convolution operations in the frequency domain and can quickly match the peak response in the image or state change, and the response value at each moment reflects the similarity between the current posture and the historical model.

[0118] Based on the kernel correlation filter response, the RLS classifier is used to dynamically model the response sequences of the billet and the electric roller table. The output result is the maximum kernel correlation response value sequence of the two: the maximum response of the billet and the maximum response of the electric roller table.

[0119] The delayed difference sequence between the response values ​​of the two is calculated. The difference reflects the time misalignment in the attitude response of the billet and the electric roller. The local minimum (representing the moment of minimum difference) is found in the response difference sequence, which is the candidate point for the attitude alignment moment. This moment corresponds to the state in which the billet may have a highly coordinated attitude change with the roller.

[0120] Then, by combining the kernel correlation filter with the RLS classifier, the kernel correlation response difference can be mapped into the attitude error, which improves the accuracy of the compensation model. At the same time, the roller motion constraints are integrated to construct a physical rationality optimizer for time bias estimation, which facilitates the acquisition of attitude position data.

[0121] Specifically, an attitude error compensation model is constructed based on the attitude position data, and an auxiliary calibration model is constructed by combining the kernel correlation filter after extracting the motion parameters of the electric roller. The output time offset estimation results include:

[0122] Analyze the identification function of the billet moving point and the center of gravity of the static point in the posture position data, perform affine transformation on the posture position data to determine the affine area, and combine the identification function and the center of gravity of the static point to generate a posture error compensation model;

[0123] Extract motion parameters based on the working principle of the electric roller, analyze the motion posture sample items of the electric roller according to the motion parameters, and integrate them with the posture error compensation model to obtain the motion posture error of the electric roller;

[0124] According to the motion principle of the electric roller table, the likelihood function of the electric roller table and the local gradient energy term of the motion are obtained to generate the reconstructed feature of the electric roller table motion. The motion parameter distribution of the electric roller table is analyzed in combination with the motion posture error.

[0125] Based on the distribution of electric roller table motion parameters and kernel correlation filter, an electric roller table posture auxiliary calibration model is established, and the electric roller table motion auxiliary correction result is output as the time offset estimation result.

[0126] It should be explained that in the process of obtaining the time bias estimation results, the moving point (the key moving point of the billet) refers to the part with a higher dynamic response during the movement of the billet. According to the trajectory and velocity information of the billet, the key moving points on the billet can be identified through dynamic curve fitting and identification functions based on physical models. The form of the identification function is usually a nonlinear model based on the billet position change rate and acceleration data, which is used to quantify the dynamic response of various parts of the billet.

[0127] The static point refers to the point where the billet is relatively fixed and does not move with the movement of other parts. It is usually the center of the billet or a specific calibration point. The center of gravity position of the static point can be calculated by calculating the mass distribution and static position of each component during the movement of the billet.

[0128] Through the affine transformation algorithm, the posture position data of the billet is converted from the original coordinate system to the standard coordinate system for further analysis of the posture error. The affine transformation includes translation, rotation, scaling and other transformations, which are used to correct the relative motion error of the billet and ensure the consistency of the data in the standard coordinate system. By analyzing the transformed posture data, the motion range of the billet is determined, and the affine area is identified, that is, all the motion and posture changes of the billet within this range. The affine area will be used as the input of the subsequent error compensation model to ensure that the model is effective within the dynamic area of ​​the billet. Based on the analysis results of the moving point identification function and the static point center of gravity, combined with the affine area, a posture error compensation model is constructed, which can correct the posture error of the billet in real time according to the motion changes of the billet in actual production.

[0129] The electric roller conveyor drives and controls the movement of the steel billet through an electric motor. Its working process involves trajectory control, speed regulation and angle adjustment. The motion parameters of the roller conveyor are obtained in real time through sensors installed on the roller conveyor (such as fiber optic sensors, accelerometers, rotary encoders, etc.), including the movement speed of the steel billet on the roller conveyor, the small rotation or tilt changes generated by the roller conveyor during operation, and the acceleration information on the roller conveyor. The motion posture sample of the electric roller conveyor can be analyzed based on the motion data obtained from the sensor. The sample can describe the relationship between the steel billet and the roller conveyor through a kinematic model, such as the sliding speed or rotation angle of the steel billet on the roller conveyor.

[0130] The motion likelihood function is used to describe the motion state of the roller within a specific time range, that is, the combined effect of the motion driving intensity and external disturbances on the steel billet on the electric roller. By optimizing the likelihood function of the motion model, the reliability of the roller motion is evaluated. Assuming that the motion of the steel billet is a Gaussian process, the probability distribution of the motion process is calculated by maximum likelihood estimation, thereby deriving the expected path of the motion.

[0131] The local gradient energy term measures the small-scale errors that occur during the movement of the steel billet on the roller, such as the slight offset and slight bending of the roller in actual operation. By analyzing the local gradient changes and energy distribution, the source of the roller error can be inferred, thereby optimizing the motion posture calibration.

[0132] Based on the motion likelihood function and the local gradient energy term, the roller motion reconstruction feature is generated to describe the dynamic motion characteristics of the roller, including the long-term and short-term motion patterns of the steel billet on the roller, and the errors caused by external interference or system imbalance on the roller. The feature is used as the input of the auxiliary calibration model to calculate the motion posture error of the electric roller.

[0133] Then, the motion reconstruction feature of the electric roller and the posture error of the electric roller are combined with the kernel correlation filter to construct an auxiliary calibration model for the electric roller posture. The calibration model will optimize the posture adjustment of the electric roller according to the real-time error feedback through the filter weight and time synchronization mechanism, so as to achieve precise control of the matching between the roller and the steel billet, and output the auxiliary correction result of the electric roller, which represents the real motion posture error of the steel billet on the roller.

[0134] Furthermore, by analyzing the billet moving point identification function and the static point center of gravity, combined with affine transformation, a posture error compensation model can be generated. At the same time, the motion parameters of the electric roller, such as speed, angle, acceleration, etc., are extracted. The roller posture samples are analyzed to obtain the motion likelihood function and the local gradient energy term, and the motion reconstruction feature quantity is generated. The kernel correlation filter and the motion feature quantity are combined to construct a calibration model, output the electric roller posture auxiliary correction result, and estimate the time offset. The time offset estimation between the billet and the electric roller can be completed accurately, thereby facilitating the positioning and scheduling of the billet in the automated production process.

[0135] In step S3, the calibrated billet position information is sent to the pusher controller to generate an action sequence instruction, and combined with the closed-loop adjustment mechanism, the billet is pushed into the heating furnace to realize the automatic loading operation of the billet.

[0136] In one embodiment, the calibrated billet position information is sent to the pusher controller to generate an action sequence instruction, and the closed-loop regulation mechanism is combined to push the billet into the heating furnace to realize the automatic billet loading operation, which includes:

[0137] The calibrated billet position information is mapped to the pusher operation coordinate system. At the same time, the pusher controller extracts the billet position information and determines the pushing path between the billet and the heating furnace.

[0138] According to the pushing path, the mechanical constraints of the steel pusher during the pushing process are judged, and the optimal driving path of the steel pusher is generated in combination with the pushing parameters of the steel pusher, and the optimal driving path is converted into action sequence instructions;

[0139] The steel billets are pushed into the heating furnace using action sequence instructions to realize the automatic charging operation of the steel billets, and the action parameters of the pusher are dynamically adjusted during the pushing process in combination with model predictive control.

[0140] It needs to be explained that the standard positioning parameters and the corresponding dynamic deviation threshold range are defined based on historical production data, equipment operating characteristics and quality control requirements. It should be explained that the typical range of billet positioning parameters is extracted through statistical analysis and pattern recognition based on historical production data, and combined with the equipment operating characteristics and quality control requirements, the standard positioning parameters are further optimized and defined. At the same time, according to the dynamic changes in the production process (such as equipment status, environmental conditions, etc.), the dynamic deviation threshold range is set to ensure that the positioning parameters meet the equipment operation requirements and guarantee product quality.

[0141] When the deviation between positioning parameters and standard positioning parameters exceeds a dynamic threshold, an alarm is triggered and a closed-loop adjustment mechanism is initiated. This feedback loop instructs the hybrid intelligent model to recalibrate the positioning parameters. Specifically, when the deviation between positioning parameters and standard positioning parameters exceeds a dynamic threshold, the system triggers an alarm, alerts the operator through visual or auditory signals, and initiates the closed-loop adjustment mechanism. The closed-loop feedback loop transmits this deviation information to the hybrid intelligent model, instructing it to dynamically adjust parameters and recalibrate positioning results based on real-time data. This process, through multiple iterative optimizations, ensures that positioning parameters remain within a reasonable range, thereby improving the stability and automation level of the production system.

[0142] The calibrated positioning parameters are sent to the control unit of the pusher, and action sequence instructions are generated and executed. Specifically, the positioning parameters calibrated by the hybrid intelligent model are sent to the control unit of the pusher through the industrial communication protocol, parsed and converted into an action instruction format suitable for execution. Based on the mechanical characteristics and kinematic model of the pusher, the action sequence instructions are generated and optimized to ensure efficient and safe completion of the billet pushing task. Finally, the control unit drives the pusher to execute the action sequence, and through real-time feedback and closed-loop control mechanism, ensures positioning accuracy and production stability, and realizes automated furnace loading operation.

[0143] like Figure 2 As shown, according to another embodiment of the present invention, there is also provided an intelligent AI positioning system for multiple-type and multiple-size steel billets in a heating furnace, the positioning system comprising:

[0144] The billet positioning parameter acquisition module is used to obtain image data and point cloud data corresponding to billets of multiple types and sizes, and generate a three-dimensional model of the billet state by combining registration and alignment technology to obtain billet positioning parameters including posture and position variables;

[0145] The billet position state calibration module is used to establish a time series data set based on the billet positioning parameters, analyze the billet positioning deviation state in combination with the deep learning model, adjust the electric roller conveyor according to the deviation state results, and calibrate the billet position state;

[0146] The billet control automatic loading module is used to send the calibrated billet position information to the pusher controller to generate action sequence instructions, and combine with the closed-loop adjustment mechanism to push the billet into the heating furnace to realize the automatic loading operation of the billet.

[0147] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent AI positioning method for multi-type and multi-size billets in a heating furnace, characterized in that: The positioning method includes: Acquire image data and point cloud data corresponding to multiple types and sizes of billets, and generate a 3D model of the billet state using registration and alignment technology to obtain billet positioning parameters including posture and position variables; A time series dataset is established based on the billet positioning parameters, and the billet positioning deviation status is analyzed using a deep learning model. The electric roller conveyor is adjusted based on the deviation status results to calibrate the billet position status. The calibrated billet position information is sent to the pusher controller to generate action sequence instructions, and combined with the closed-loop adjustment mechanism, the billet is pushed into the heating furnace to realize the automated billet loading operation.

2. The intelligent AI positioning method for multiple-type and multiple-size billets in a heating furnace according to claim 1 is characterized in that: The method of acquiring image data and point cloud data corresponding to multi-type and multi-size steel billets, and generating a three-dimensional model of the steel billet state by combining registration and alignment technology, and obtaining steel billet positioning parameters including posture and position variables includes: Obtain the original 2D image and 3D point cloud data of multiple-type and multiple-size steel billets, and grayscale, remove noise, and resize the original 2D image to obtain the 2D image of the steel billet; The dark channel prior theory and image enhancement algorithm are used to perform feature enhancement processing on the two-dimensional image of the steel billet, and the geometric feature points of the steel billet are obtained by aligning it with the three-dimensional point cloud data to generate regularized point cloud data. The regularized point cloud data is initially aligned using a global registration strategy, and feature extraction is performed in combination with descriptors to generate an initial transformation matrix. A three-dimensional model of the billet state is constructed based on the initial transformation matrix. Based on the feature histogram algorithm, the geometric features including the spatial posture and position coordinates of the billet are extracted from the three-dimensional model of the billet state, and the geometric features are used to train the hybrid intelligent model to output the billet positioning parameters.

3. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 2 is characterized in that: The method of performing feature enhancement processing on the two-dimensional image of the steel billet using the dark channel prior theory and the image enhancement algorithm, and aligning the two-dimensional image with the three-dimensional point cloud data to obtain geometric feature points of the steel billet and generate regularized point cloud data includes: Based on the dark channel prior theory, the global atmospheric light value and transmittance map of the two-dimensional image of the steel billet are estimated. The dynamic blur estimation model and regularization technology are combined to extract blur features and eliminate the dynamic blur effect. The defogging result is used to obtain a two-dimensional image of the steel billet. The reflection vector and illumination component of the two-dimensional image of the steel billet are decomposed based on the image enhancement algorithm to enhance the contrast between the steel billet and the background. An edge-preserving filter is used to remove the halo effect of the 2D image of the steel billet to obtain an enhanced 2D image of the steel billet. At the same time, denoising and downsampling are performed on the 3D point cloud data. Project the 3D point cloud data onto the 2D image of the steel billet, establish the correspondence between the point cloud and the image using geometric constraints, and dynamically estimate the relative posture of the point cloud and the image in combination with an adjustment strategy; According to the estimation results, geometric feature points are extracted from the three-dimensional point cloud data and the two-dimensional image of the steel billet, and feature descriptors are generated to represent the local geometric features, which are combined with the fusion strategy to generate regularized point cloud data.

4. The intelligent AI positioning method for multiple-type and multiple-size billets in a heating furnace according to claim 3 is characterized in that: The method of extracting geometric features including the spatial posture and position coordinates of the billet from the three-dimensional model of the billet state based on the feature histogram algorithm and using the geometric features to train the hybrid intelligent model to output the billet positioning parameters includes: The geometric features of the spatial posture and position coordinates of the billet in the 3D model of the billet state are extracted based on the fast point feature histogram, and high-dimensional feature information is generated by combining the adaptive weight distribution strategy; Integrate high-dimensional feature information with corresponding labels to form a structured dataset, and use a hierarchical partitioning strategy to divide the dataset into training set, validation set, and test set in proportion; Use training sets, validation sets, and test sets to train and adjust the hybrid intelligent model that includes deep learning and support vector machines, and gradually optimize the model parameters based on forward propagation and backpropagation; The geometric features are input into the trained hybrid intelligent model to extract high-level features using deep learning technology, and combined with the positioning parameters of the support vector machine containing the spatial posture and position coordinates of the steel billet.

5. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 4 is characterized in that: The calculation formula of the positioning parameter is: Where P represents the positioning parameter, X represents the spatial posture and position coordinate vector to be optimized, φ represents the nonlinear mapping function of the deep neural network, F represents the geometric feature, θ represents the parameter geometry of the hybrid intelligent model, φ(F;θ) represents the predicted value of the geometric feature F, T represents the target reference value, represents the Euclidean norm squared, T i represents the i-th target value, K(F,T i ) represents the capture of geometric features F and target reference value T i The nonlinear relationship between represents the adaptive weight of the i-th target reference value, λ represents the adjustment coefficient, N represents the number of target values ​​in the set, and i represents the index of the target reference value.

6. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 1 is characterized in that: The method of establishing a time series data set based on the billet positioning parameters, analyzing the billet positioning deviation state in combination with a deep learning model, adjusting the electric roller conveyor according to the deviation state result, and calibrating the billet position state includes: The billet positioning parameters are recorded at set time intervals to obtain a positioning time series dataset, and a deep learning model is trained to analyze the differences between the positioning time series dataset and the standard positioning parameters. The positioning deviation of the billet is quantified based on the difference results, and the relative posture deviation of the billet is determined based on the positioning deviation. At the same time, the relative posture of the electric roller is analyzed using a pre-integration algorithm. The relative posture deviation of the billet and the relative posture of the electric roller are cross-correlated and calculated, and the time offset between the billet and the electric roller is estimated based on the calculation results and the rotation constraint technology. The speed and angle of the electric roller are adjusted based on the time offset estimation results. According to the adjustment results, the billet is adjusted from the deviation state to the standard positioning state to complete the position state of the billet, and the hybrid intelligent model is used to output the calibration positioning parameters of the billet.

7. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 6 is characterized in that: The cross-correlation calculation of the relative posture deviation of the steel billet and the relative posture of the electric roller table, and obtaining the time offset estimation result between the steel billet and the electric roller table based on the calculation result and the rotation constraint technology includes: Based on the recursive least squares classifier, the kernel correlation filters of the relative posture deviation of the billet and the relative posture of the electric roller table are judged respectively, and the maximum kernel correlation filter response value corresponding to the two is obtained; Analyze the difference between the maximum kernel correlation filter response values ​​of the two, and find the corresponding attitude position data of the billet during the positioning calibration process within the billet positioning parameters based on the difference result; An attitude error compensation model is constructed based on the attitude position data. After extracting the motion parameters of the electric roller conveyor, an auxiliary calibration model is constructed in combination with the kernel correlation filter to output the time offset estimation result.

8. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 7 is characterized in that: The posture error compensation model is constructed based on the posture position data, and after extracting the motion parameters of the electric roller, an auxiliary calibration model is constructed in combination with a kernel correlation filter. The output time offset estimation result includes: Analyze the identification function of the billet moving point and the center of gravity of the static point in the posture position data, perform affine transformation on the posture position data to determine the affine area, and combine the identification function and the center of gravity of the static point to generate a posture error compensation model; Extract motion parameters based on the working principle of the electric roller, analyze the motion posture sample items of the electric roller according to the motion parameters, and integrate them with the posture error compensation model to obtain the motion posture error of the electric roller; According to the motion principle of the electric roller table, the likelihood function of the electric roller table and the local gradient energy term of the motion are obtained to generate the reconstructed feature quantity of the electric roller table motion. The motion parameter distribution of the electric roller table is analyzed in combination with the motion posture error. Based on the distribution of electric roller table motion parameters and kernel correlation filter, an electric roller table posture auxiliary calibration model is established, and the electric roller table motion auxiliary correction result is output as the time offset estimation result.

9. The intelligent AI positioning method for multiple-type and multiple-size steel billets in a heating furnace according to claim 1 is characterized in that: The calibrated billet position information is sent to the pusher controller to generate an action sequence instruction, and the billet is pushed into the heating furnace in combination with the closed-loop regulation mechanism to realize the automatic billet loading operation, which includes: The calibrated billet position information is mapped to the pusher operation coordinate system. At the same time, the pusher controller extracts the billet position information and determines the pushing path between the billet and the heating furnace. According to the pushing path, the mechanical constraints of the steel pusher during the pushing process are judged, and the optimal driving path of the steel pusher is generated in combination with the pushing parameters of the steel pusher, and the optimal driving path is converted into action sequence instructions; The billets are pushed into the heating furnace using action sequence instructions to realize the automatic loading operation of the billets, and the action parameters of the pusher are dynamically adjusted during the pushing process in combination with model predictive control.

10. An intelligent AI positioning system for multiple-blank and multiple-size steel billets in a heating furnace, used to implement the intelligent AI positioning method for multiple-blank and multiple-size steel billets in a heating furnace according to any one of claims 1 to 9, characterized in that: The positioning system includes: The billet positioning parameter acquisition module is used to obtain image data and point cloud data corresponding to billets of multiple types and sizes, and generate a three-dimensional model of the billet state by combining registration and alignment technology to obtain billet positioning parameters including posture and position variables; The billet position state calibration module is used to establish a time series data set based on the billet positioning parameters, analyze the billet positioning deviation state in combination with the deep learning model, adjust the electric roller conveyor according to the deviation state results, and calibrate the billet position state; The billet control automatic loading module is used to send the calibrated billet position information to the pusher controller to generate action sequence instructions, and combine with the closed-loop adjustment mechanism to push the billet into the heating furnace to realize the automatic loading operation of the billet.

Images