Intelligent factory modeling optimization collaboration method for adaptive adjustment of metallurgical process parameters

Abstract

The invention relates to the technical field of metallurgical factory material tracking and monitoring, and particularly discloses an intelligent factory modeling optimization collaboration method for adaptive adjustment of metallurgical process parameters, which comprises the following steps of: scanning the contour of a material pile in real time through three-dimensional laser equipment, and uploading the contour to a digital material yard module of a scheduling system; a real-time material pile three-dimensional digital model is established in the digital material yard module; the digital stockyard module is used for collecting equipment parameters of stockpiling and material taking, matching operation priorities by adopting an optimal path algorithm, and carrying out material uniform mixing and automatic compensation; the infrared camera identifies image data and temperature data and uploads the image data and the temperature data to a scheduling system; the scheduling system establishes a data management architecture based on an object storage service OBS, automatically calculates the particle size distribution of coke or sintered ore, and judges whether the particle size distribution reaches an early warning threshold value or not; according to the invention, accurate control of uniformly mixed ore components is realized; the detection efficiency is improved, the manual intervention and maintenance cost is reduced, and stable operation and efficient production of the blast furnace are assisted.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metallurgical plant material tracking monitoring, and in particular to an intelligent plant modeling optimization coordination method with adaptive adjustment of metallurgical process parameters. BACKGROUND

[0002] At present, under the leadership of large steel enterprises, domestic leading steel enterprises have gradually introduced intelligent batching systems, combined mechanism models (such as blast furnace thermodynamic models, material balance models) with real-time production data to achieve batching optimization. Related research focuses on the application of industrial internet platforms, which dynamically corrects by collecting blast furnace sensor data (temperature, pressure, gas composition, etc.) in real time combined with mechanism models. For example, a certain steel uses edge computing technology to achieve millisecond-level data feedback, optimizing the response speed of batching. In view of the problems of low silicon stability rate and frequent manual intervention, a "multi-objective optimization algorithm" and a "closed-loop compensation mechanism" are developed. However, at present, steel plants lack high-precision and high-efficiency implementation methods for pre-iron batching and feeding. In the process of pre-iron feeding and batching, the optimal batching structure cannot be found, and accurate charging raw material information cannot be obtained in time, affecting the stable operation of the blast furnace. Therefore, a metallurgical intelligent plant full-process dynamic optimization simulation and data interaction control method needs to be designed to solve the problem of lack of high-precision and high-efficiency feeding affecting the stable operation of the blast furnace in the existing system. SUMMARY

[0003] In view of the problems in the prior art, the purpose of the present application is to provide an intelligent plant modeling optimization coordination method with adaptive adjustment of metallurgical process parameters, a full-process tool chain to improve batching efficiency and optimal batching structure and to enable the charging raw material quality to be timely provided to the blast furnace process personnel.

[0004] The technical scheme adopted by the present application to solve its technical problems is: an intelligent plant modeling optimization coordination method with adaptive adjustment of metallurgical process parameters, comprising the following steps:

[0005] S1, a three-dimensional laser device is used to scan the pile profile in real time, and the digital yard module of the scheduling system is uploaded, and a real-time three-dimensional digital model of the pile is established in the digital yard module;

[0006] S2, the digital yard module collects the equipment parameters of the stacking and taking of materials, automatically identifies the material level and quality scheduling data of the stacking and taking equipment and the taking and stacking equipment, matches the job priority using the best path algorithm, and performs material mixing and automatic compensation;

[0007] S3, the infrared camera recognition image data and temperature data are uploaded to the scheduling system, the infrared camera is integrated with an infrared imager as a non-contact temperature measurement device, which captures the temperature distribution information of the sintering machine tail section in real time, and the fusion model of the scheduling system uses a Transformer architecture to realize the set function.

[0008] S4, the scheduling system establishes a data management architecture based on object storage service OBS, automatically calculates the particle size distribution of coke or sinter, and judges whether the warning threshold is reached.

[0009] Specifically, the digital yard module in step S1 uses image processing technology and high-precision three-dimensional image reconstruction technology to vectorize the entire yard modeling, and the three-dimensional laser equipment includes a 3D laser imager, a laser range finder, and a machine body attitude detector. The 3D laser imager is arranged on the stacker-reclaimer, and the vehicle-mounted front-end computer system of the stacker-reclaimer starts the 3D laser imager to scan the raw material yard. The 3D radar measures objects at high speed and high density and outputs 3D point cloud. After the point cloud data is preprocessed in the vehicle-mounted front-end computer system, it is transmitted to the graphic workstation in the control room through the network, and then the point cloud data is processed again through filtering and denoising, geometric fitting and data simplification algorithms. Finally, the processed point cloud data is registered and spliced, and the point cloud data is updated on the yard map of the scheduling system after being processed by surface reconstruction. After processing by the graphic workstation, the scheduling system obtains the volume, length, and characteristic point coordinate data of the stockpile, which are used in the scheduling system to guide the automatic operation of the stacker-reclaimer. In the process of taking material, the scheduling system automatically identifies the boundary data of the stockpile and adjusts the rotation and inching amount in time. The height and length of each layer are calculated to provide a basis for the opening layer strategy and the layer changing strategy. The basis for determining the cutting point of each layer of taking material is provided.

[0010] Specifically, the scheduling system uses multi-sensor data fusion, attitude positioning, and automatic driving to simulate the stockpile and material flow for the automatic operation of the stacker-reclaimer, which includes material flow simulation, automatic stockpiling, automatic material taking, and anti-collision.

[0011] Specifically, the best path algorithm in step S2 matches the job priority by analyzing the image recognized by the three-dimensional laser equipment through computer vision algorithms to determine the travel path of the stacker-reclaimer. The best path algorithm includes the following steps:

[0012] S21, color space conversion, using HSV color space to process images with large changes in illumination;

[0013] S22, grayscale and binarization processing, grayscale is the process of converting color images to grayscale images, and binarization is the process of converting grayscale images to black and white images. These two processing methods simplify image data.

[0014] S23, morphological processing is usually used for shape extraction and separation in the image, using two operations of dilation and erosion, dilation is used to fill the small holes inside the object, and erosion removes small objects; Gaussian filter and median filter are used for image smoothing, which is used to remove noise in the image, Gaussian filter is based on Gaussian distribution to make weighted average of the image, median filter is to replace the center pixel value with the median value of the pixel neighborhood;

[0015] S24, the processed image is calculated by SPFA algorithm to calculate the best path, using array d to record the shortest path estimate value of each node, using adjacency list to store the graph G, using dynamic approximation method to set a first-in-first-out queue to save the nodes to be optimized, when optimizing, the first node u is taken out of the queue, and the current shortest path estimate value of u point is used to relax the nodes pointed by u, if the shortest path estimate value of v point is adjusted, and v point is not in the current queue, v point is put into the tail of the queue, the nodes are taken out of the queue to perform relaxation operation until the queue is empty;

[0016] Let Dist represent the current shortest distance from S to I point, Fa represent the number of one point before I point in the current shortest path from S to I; at the beginning, Dist is all +∞, only Dist[S]=0, Fa is all 0;

[0017] Maintain a queue to store all points that need to be iterated, initially the queue has only one point S, use a boolean array to record whether each point is in the queue or not;

[0018] Each iteration, take out the head of the queue v, enumerate the edges v->u from v in turn, set the length of the edge as len, judge whether Dist[v]+len is less than Dist[u], if less than, improve Dist[u], record Fa[u] as v, and since the shortest distance from S to u has become smaller, judge u to improve other points, so if u is not in the queue, put u into the tail of the queue, iterate until the queue is empty, that is, the shortest distance from S to all points is determined, end the algorithm; if a point is in the queue more than n times, set a negative weight ring.

[0019] Specifically, the infrared camera in step S3 is installed on the sintering machine to collect the change of the FeO content of the sintering machine and the thermal field evolution of the sintering key area, and the collection, storage and processing mechanism of the infrared camera image and temperature data realizes the FeO content prediction; an infrared thermal imager is deployed at the tail of the sintering machine, which has a thermal imaging resolution of 384*288 pixels, a frame rate of more than 30Hz, supports continuous shooting and outputs original infrared images and corresponding temperature matrix; the collection frequency is set to 2 frames per second, the point data generated by the PLC control system is synchronized with the infrared camera through time stamping, and the image data and process parameters are aligned in time sequence, that is, the sintering machine speed, material thickness and wind box pressure, which is convenient for subsequent fusion analysis.

[0020] The temperature matrix is stored in a structured format, each infrared image contains a two-dimensional temperature matrix of 384*288 pixels, which is converted into a structured numerical matrix and additional metadata, time stamp, temperature value; ApacheParquet format is used for compressed storage, MinIO server is deployed as a transfer station, object storage service based on MinIO is constructed, which is used to accept original temperature data packets from the infrared camera, real-time receive and cache infrared images and temperature data, provide permission control and log recording function, storage cycle is stored and saved according to every minute as a sub file.

[0021] Specifically, the infrared image is a structured numerical matrix, which is specifically converted into a conversion matrix by calibrating the feature points of the RGB image and the infrared image, and the matrix transformation formula is as follows:

[0022]

[0023] (x',y') is the coordinate of the infrared image, (x,y) is the coordinate of the RGB image, a 1-4 The conversion factor between the two images is obtained, t x , t y represents the translation variable between the two images, at least 3 pairs of matching points are needed to calculate 6 variables, and at least 10 pairs of matching points are set for the image.

[0024] Specifically, the Transformer architecture in step S3 includes the following steps:

[0025] S31, image blocking: ViT first divides the input image into multiple small blocks, each of which is a 16*16 pixel small area;

[0026] S32, flattening and vectorization: each small block is flattened into a vector for subsequent processing;

[0027] S33, position coding: in order to retain the spatial information in the original image, position coding is added in the vector.

[0028] S34, global self-attention calculation: the vector with position encoding information is input into a standard Transformer encoder, is processed through a global self-attention mechanism, so as to capture long-distance dependencies in the entire image;

[0029] S35, classification output: using the last vector for image classification or other tasks.

[0030] Specifically, the data management architecture of the object storage service OBS in the step S4 realizes the automatic management of data uploading and downloading by using the browser tool obs-browser-plus and the command line tool obsutil provided by the OBS, the obs-browser-plus supports concurrent uploading of more than 500 images, meets the batch return demand after local labeling, the obsutil supports the realization of breakpoint resume and incremental synchronization through script configuration, adapts to the data processing scene of long-period task, and the whole data uploading process is recorded to a log file and combined with an MD5 check mechanism.

[0031] The data management architecture automatically calculates the particle size distribution of the coke or sinter according to the segmentation result, judges whether the early warning threshold is reached, whether the proportion of coke particle size less than 25mm is lower than 15%, analyzes the result in a structured form, and pushes the result to a production scheduling system in real time.

[0032] The present application has the following beneficial effects:

[0033] The intelligent factory modeling optimization collaboration method with adaptive adjustment of metallurgical process parameters designed by the present application, aiming at the problems of low efficiency of iron supply process and deviation of supply quality under high-capacity mode, carries out digital batching and intelligent supply research, constructs a process control model based on neural network through the application of deep learning and genetic algorithm in process optimization, realizes multi-node collaborative optimization and rapid response to abnormal working conditions, analyzes the large composition fluctuation in mixing and batching process, which affects the subsequent production factors, establishes a mixing and batching control model, and realizes accurate control of the composition of mixed ore.

[0034] The intelligent factory modeling optimization collaboration method with adaptive adjustment of metallurgical process parameters designed by the present application, aiming at the actual needs of ironmaking plant for online monitoring and process optimization of sinter quality, focuses on the key problems such as large fluctuation of ferrous oxide (FeO) content in sintering process, difficulty in real-time and accurate detection, and difficulty in modeling of multi-source heterogeneous data fusion, researches a joint feature extraction method based on process operation parameters and sinter machine tail infrared thermal imaging section temperature data, relies on the Pan Gu big model to construct an FeO prediction framework, realizes dynamic prediction of the FeO content of sinter, and realizes online real-time prediction and process feedback control support of the model through end-edge-cloud collaborative deployment.

[0035] The intelligent factory modeling optimization collaborative method designed by the metallurgical process parameter adaptive adjustment, aiming at the problem that the particle size and distribution of blast furnace raw materials directly affect the stable operation of blast furnace, studies the image OBS storage management and incremental synchronization technology. Real-time images of coke and sinter on the belt are collected by using Hikvision industrial camera and fill light, and an optimized instance segmentation model is deployed on the Pangudaimo inference server for particle size detection, realizing 24-hour online monitoring and real-time image display, automatic identification of particle size distribution, setting of alarm threshold overrun early warning according to production demand and particle size historical trend tracking, improving detection efficiency, reducing manual intervention and maintenance cost, ensuring job safety, helping blast furnace stable operation and efficient production. BRIEF DESCRIPTION OF DRAWINGS

[0036] Fig. 1 is a schematic diagram of raw material particle size identification technology.

[0037] Fig. 2 is a schematic diagram of a fusion model using a Transformer architecture. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be further clearly, completely and specifically explained in combination with the drawings in the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0039] As shown in Figs. 1-2 A metallurgical process parameter adaptive adjustment intelligent factory modeling optimization collaborative method includes the following steps:

[0040] 1. The three-dimensional laser equipment scans the real-time pile profile uploaded to the digital yard module of the scheduling system, and a real-time three-dimensional digital model of the pile is established in the digital yard module.

[0041] The digital stockyard module uses image processing technology and high-precision three-dimensional image reconstruction technology to vectorize model the entire stockyard. The three-dimensional laser equipment includes a 3D laser imager, a laser range finder and a machine body attitude detector. The 3D laser imager is arranged on the stacker-reclaimer, and a vehicle-mounted front-end computer system of the stacker-reclaimer starts the 3D laser imager to scan the stockyard. The 3D radar performs high-speed and high-density measurement on objects and outputs 3D point clouds. After the point cloud data is pretreated by the vehicle-mounted front-end computer system, the point cloud data is transmitted to a graphic workstation in a control room through a network, and the point cloud data is processed again by filtering and denoising, geometric fitting and data simplification algorithms. Finally, the processed point cloud data is registered and spliced, and is processed by surface reconstruction, and then the point cloud data is updated on a stockyard map of a dispatching system. After the graphic workstation processing, the dispatching system obtains data of the volume, length and characteristic point coordinates of the stockpile, and the data is used to guide the stacker-reclaimer to perform automatic operation in the dispatching system. In the material taking process, the dispatching system automatically identifies the boundary data of the stockpile, and adjusts the rotation and inching amount in time. The height and length of each layer are calculated to provide a basis for the layer opening strategy and the layer changing strategy, and to provide a basis for determining the cut-in point of each layer of the material taking operation.

[0042] The dispatching system uses multi-sensor data fusion, attitude positioning and automatic driving to simulate the stockpile and the material flow, and is used for automatic operation of the stacker-reclaimer. The automatic operation of the stacker-reclaimer includes material flow simulation, automatic stockpiling, automatic material taking and anti-collision.

[0043] Through an intelligent flow decision model, the best path algorithm of graph theory is used for the stockyard transportation equipment according to the production demand of material transportation, the available flow between the start equipment and the end equipment is searched, and the energy-saving and convenient optimal dynamic flow is intelligently decided by comprehensively considering the properties of each equipment. The main research contents of the intelligent flow decision model include: operation priority matching function, flow dynamic search function, flow intelligent decision function, flow intelligent control function, flow automatic tracking function and transportation performance automatic matching function.

[0044] Task priority matching studies the allocation of task priorities under different processes and operational conditions. All task flows in a given work area are dynamically adjusted based on linear programming. An intelligent mixing and batching control model generates corresponding batching plans based on received mixing and batching plans, and adjusts the feeder's discharge rate based on real-time production data to ensure stable SiO2 and TFe content in the mixed material. The intelligent mixing and batching control model includes: material batching plan generation, data acquisition for feeding and batching processes, feeder discharge speed calculation, and automatic mixing compensation. The automatic mixing compensation function receives planned data, tracks actual batching performance to form a series of time-series tracking data, and calculates the deviation between actual mixing and batching performance and the planned value in real time. Through smoothing the time-series data, the batching setpoints for each material type are recalculated. The recalculated batching setpoints are then sent to the equipment layer for deviation correction, either periodically or triggered by the mixing and stacking to a specific location.

[0045] 2. The digital material yard module collects equipment parameters for stacking and reclaiming, automatically identifies the material level and quality scheduling data of stacking and reclaiming equipment and material feeding equipment, and uses the best path algorithm to match the operation priority and automatically compensate for material mixing.

[0046] The optimal path algorithm matches job priorities. Images identified by 3D laser equipment are analyzed using computer vision algorithms to determine the stacker-reclaimer's travel path. The optimal path algorithm includes the following steps:

[0047] 1) Color space conversion: Use the HSV color space to process images with large changes in lighting.

[0048] 2) Grayscale and binarization processing: Grayscale is the process of converting a color image into a grayscale image, while binarization is the process of converting a grayscale image into a black and white image. These two processing methods simplify image data.

[0049] 3) Morphological processing is usually used for the extraction and separation of shapes in images. Two operations are dilation and erosion. Dilation is used to fill small holes inside objects, while erosion removes small objects. Gaussian filtering and median filtering are used for image smoothing to remove noise in images. Gaussian filtering is based on the Gaussian distribution to perform a weighted average on the image, while median filtering replaces the center pixel value with the median value of the pixel neighborhood.

[0050] 4) The processed image adopts the SPFA algorithm to calculate the optimal path, uses an array d to record the shortest path estimate value of each node, uses an adjacency list to store the graph G, and uses a dynamic approximation method to set up a first-in first-out queue to save the nodes to be optimized. When optimizing, the first node u in the queue is taken out each time, and the current shortest path estimate value of the u node is used to relax the nodes pointed to by the u node. If the shortest path estimate value of the v node is adjusted and the v node is not in the current queue, the v node is put into the tail of the queue. The nodes in the queue are taken out continuously to perform the relaxation operation until the queue is empty.

[0051] Let Dist represent the current shortest distance from S to I, and Fa represent the number of the point before I in the current shortest path from S to I. At the beginning, all Dist are +∞, only Dist[S]=0, and all Fa are 0.

[0052] A queue is maintained to store all points that need to be iterated. Initially, there is only one point S in the queue. A Boolean array is used to record whether each point is in the queue.

[0053] In each iteration, the head node v in the queue is taken out, and the edges v->u starting from v are enumerated in turn. Let the length of the edge be len. It is judged whether Dist[v]+len is less than Dist[u]. If it is less, Dist[u] is improved, Fa[u] is recorded as v, and since the shortest distance from S to u has become smaller, it is judged whether u improves other points. Therefore, if u is not in the queue, u is put into the tail of the queue. The iteration is continued until the queue is empty, that is, the shortest distances from S to all points are determined, and the algorithm ends. If a point is put into the queue more than n times, a negative weight ring is set.

[0054] 3. The infrared camera identifies image data and temperature data uploaded to the scheduling system. The infrared camera integrates an infrared imager as a non-contact temperature measurement device, which captures the temperature distribution information of the sintering machine tail section in real time. The fusion model of the scheduling system uses the Transformer architecture to realize the set function.

[0055] The infrared camera is installed on the sintering machine to collect the change of FeO content of the sintering ore and the thermal field evolution of the sintering key area. The collection, storage and processing mechanism of the infrared camera image and temperature data realizes the prediction of FeO content. The infrared thermal imager is deployed at the tail of the sintering machine, which has a thermal imaging resolution of 384*288 pixels and a frame rate of more than 30Hz, supports continuous shooting and outputs the original infrared image and the corresponding temperature matrix. The collection frequency is set to 2 frames per second. The point data generated by the PLC control system is synchronized with the infrared camera through time stamping. The image data and the process parameters, such as the sintering machine speed, the material thickness and the wind box pressure, are aligned in time sequence, which facilitates subsequent fusion analysis.

[0056] The temperature matrix structured storage format is designed, each frame of infrared image contains a two-dimensional temperature matrix of 384*288 pixels, which is converted into a structured numerical matrix and attached with metadata, timestamp and temperature value; Apache Parquet format is used for compressed storage, MinIO server is deployed as a transfer station, object storage service based on MinIO is constructed, which is used to accept raw temperature data packets from the infrared camera, receives and caches infrared images and temperature data in real time, provides permission control and log recording functions, and stores a sub-file every minute.

[0057] The infrared image is a structured numerical matrix, which is specifically obtained by converting the conversion relationship matrix through the feature points of the calibrated RGB image and infrared, and the matrix transformation formula is as follows:

[0058]

[0059] (x',y') is the coordinate of the infrared image, (x,y) is the coordinate of the RGB image, a 1-4 The conversion factor t x between the two images is obtained. y The translation variable between the two images is represented, and at least 3 pairs of matching points are needed to calculate 6 variables, and at least 10 pairs of matching points are set for the image.

[0060] The Transformer architecture includes the following steps:

[0061] 1) Image patching: ViT first divides the input image into multiple small patches, each of which is a 16x16 pixel small region.

[0062] 2) Flattening and vectorization: Each small patch is flattened into a vector for subsequent processing.

[0063] 3) Position encoding: In order to preserve the spatial information in the original image, position encoding is added to the vector.

[0064] 4) Global self-attention calculation: The vector with position encoding information is input into the standard Transformer encoder, and is processed through the global self-attention mechanism, so as to capture the long-distance dependency relationship in the whole image.

[0065] 5) Classification output: The last vector is used for image classification or other tasks.

[0066] The fusion model uses a Transformer architecture to implement a set function, each output of the base model is an element in the set, since the Transformer supports input of any sequence length, the number of base models is arbitrary, and the fusion coefficients of different models learned by the Transformer can dynamically adjust the effect of model fusion, suppress the models that are poor for the current sample, and increase the attention to the results of the models that are good for the current sample, thereby improving the model performance.

[0067] 4. The dispatching system establishes a data management architecture based on an object storage service OBS, automatically calculates the particle size distribution of the coke or sinter, and determines whether the warning threshold is reached.

[0068] The OBS adopts a flat object storage structure, which is naturally suitable for managing unstructured image data. In terms of data organization, the system ensures one-to-one correspondence between images and labels through the same file method, facilitating automatic reading during model training; at the same time, each batch of data is accompanied by timestamp and version information, supporting incremental management of the training set, effectively improving the flexibility of data updating and iteration. The data management architecture of the object storage service OBS realizes the automatic management of data uploading and downloading by using the browser tool obs-browser-plus and the command line tool obsutil provided by OBS, obs-browser-plus supports concurrent uploading of more than 500 images, meeting the batch return demand after local labeling; obsutil supports breakpoint resume and incremental synchronization through script configuration, adapting to the data processing scene of long-period tasks; the entire data uploading process is recorded to a log file, and combined with the MD5 check mechanism;

[0069] The data acquisition link adopts a Hikvision industrial camera to collect blast furnace belt images at a certain frequency, and the obtained image data has high frame rate and high definition. In order to ensure the representativeness and diversity of model training, the samples cover various working conditions, including uneven material layer accumulation, material edge blur and the like. In the data labeling stage, a polygon instance segmentation labeling mechanism is introduced, the boundary of each piece of coke or sinter is accurately outlined, and the VOC format is used to organize the image and XML label file, so that the model learns the accurate semantic contour. In order to improve the data quality, an annotation auditing process is constructed to eliminate ambiguous and misjudged samples. The training model selects the instance segmentation network architecture built in the Pangu platform, and the complete training process is realized through the "visual large model" workflow provided by the Pangu platform, which supports online management of model definition, training scheduling, log monitoring and model output. The core process includes the following steps: first, create an "image instance segmentation" data set on the platform, after uploading the data path in OBS, the system will automatically parse the image and label structure; second, select the preset segmentation architecture in the model initialization stage, and support loading existing model weights to improve the initial accuracy of training; third, configure the computing resources according to the task size to ensure that the model has good convergence efficiency; fourth, set the maximum number of iterations (default 300 rounds) and start the early stopping mechanism to effectively prevent overfitting during training; and finally, the platform provides multi-version evaluation after training, and outputs core indicators including loss curve, validation accuracy, recall rate and the like, which are used to select the optimal model version. After actual training and optimization, the final model shows excellent segmentation edge quality and the ability to distinguish occluded areas.

[0070] In the AI application of the Pangu platform, the model training can be published as a cloud service one key after completion, and the HTTP API interface is opened for external calling. The deployment process strictly follows the three-stage approval mechanism of "test → test → online", to ensure the stability and reliability of the service. Each model version has an independent ID, which facilitates model effect evaluation and smooth upgrade. In addition, the deployed service can access the observability platform to real-time view the calling frequency, response time and abnormal alarm and other key indicators, to realize all-round operation monitoring and operation and maintenance support.

[0071] The data management architecture automatically calculates the particle size distribution of coke or sinter according to the segmentation result, and judges whether the early warning threshold is reached, whether the proportion of coke particle size less than 25 mm is less than 15%, and the analysis result is output in a structured form and pushed to the production scheduling system in real time.

[0072] The present application is not limited to the above embodiments, and anyone should know that any structural changes made under the inspiration of the present application fall within the protection scope of the present application.

[0073] The technology, shape, and structural parts not described in detail in the present application are all known technology.

Claims

1. A collaborative method for intelligent factory modeling, optimization, and adaptive adjustment of metallurgical process parameters, characterized in that, Includes the following steps: S1. The outline of the material pile is scanned in real time by a three-dimensional laser device and uploaded to the digital material yard module of the scheduling system. A real-time three-dimensional digital model of the material pile is established in the digital material yard module. S2, the digital material yard module collects equipment parameters for stacking and retrieving, automatically identifies the material level and quality scheduling data of stacking and retrieving equipment and material feeding equipment, uses the best path algorithm to match the operation priority, and performs automatic compensation for material mixing. S3. The infrared camera identifies image data and temperature data and uploads them to the scheduling system. The infrared camera integrates an infrared imager as a non-contact temperature measurement device to capture the temperature distribution information of the sintering machine tail section in real time. The fusion model of the scheduling system uses the Transformer architecture to implement set functions. S4. The scheduling system establishes a data management architecture based on the object storage service OBS, automatically calculates the particle size distribution of coke or sinter, and determines whether the warning threshold has been reached.

2. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters as described in claim 1, characterized in that: The digital material yard module in step S1 employs image processing technology and high-precision 3D image reconstruction technology to perform vector modeling of the entire material yard. The 3D laser equipment includes a 3D laser imager, a laser rangefinder, and a machine attitude detector. The 3D laser imager is mounted on the stacker-reclaimer. The onboard front-end computer system of the stacker-reclaimer activates the 3D laser imager to scan the material yard. The 3D radar performs high-speed, high-density measurements of the objects and outputs 3D point clouds. After preprocessing by the onboard front-end computer system, the point cloud data is transmitted via network to the graphics workstation in the central control room, where it undergoes further processing including filtering, noise reduction, geometric fitting, and data refinement. The simplified algorithm further processes the point cloud data. Finally, the processed point cloud data is registered, stitched together, and then processed through surface reconstruction before being updated on the material yard map of the scheduling system. After processing by the graphics workstation, the scheduling system obtains data on the volume, length, and feature point coordinates of the material pile. This data is used in the scheduling system to guide the stacker-reclaimer in automatic operation. During the material reclaiming process, the scheduling system automatically identifies the boundary data of the material pile and adjusts the rotation and inching momentum in a timely manner; it calculates the height and length of each layer, providing a basis for layer opening and changing strategies; and it provides a basis for determining the entry point for each layer's material reclaiming operation.

3. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters as described in claim 2, characterized in that: The scheduling system uses multi-sensor data fusion, attitude positioning, and automatic driving, based on material pile simulation and material flow simulation, for the automatic operation of the stacker-reclaimer. The automatic operation of the stacker-reclaimer includes material flow simulation, automatic stacking, automatic reclaiming, and collision avoidance.

4. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters as described in claim 3, characterized in that, The optimal path algorithm in step S2 matches the task priority. The image identified by the 3D laser equipment is analyzed by computer vision algorithms to determine the travel path of the stacker-reclaimer. The optimal path algorithm includes the following steps: S21. Color space conversion: Use the HSV color space to process images with large changes in lighting. S22. Grayscale and binarization processing: Grayscale is the process of converting a color image into a grayscale image, while binarization is the process of converting a grayscale image into a black and white image. These two processing methods simplify image data. S23. Morphological processing is usually used for the extraction and separation of shapes in images. It employs two operations: dilation and erosion. Dilation is used to fill small holes inside objects, while erosion removes small objects. Gaussian filtering and median filtering are used for image smoothing to remove noise from images. Gaussian filtering is based on a weighted average of the image according to the Gaussian distribution, while median filtering replaces the center pixel value with the median value of the pixel's neighborhood. S24. The processed image is used to calculate the optimal path using the SPFA algorithm. The shortest path estimate of each node is recorded in array d. The adjacency list is used to store the graph G. A first-in-first-out queue is set up using the dynamic approximation method to store the nodes to be optimized. During optimization, the head node u is taken out each time, and the current shortest path estimate of u is used to relax the nodes pointed to by u. If the shortest path estimate of v is adjusted and v is not in the current queue, v is put into the tail of the queue. Nodes are continuously taken out from the queue for relaxation until the queue is empty. Let Dist represent the current shortest distance from S to I, and Fa represent the index of the point preceding I in the current shortest path from S to I; initially, all values ​​of Dist are +∞, except for Dist[S] = 0, and all values ​​of Fa are 0. Maintain a queue that stores all the points that need to be iterated. Initially, there is only one point S in the queue. Use a boolean array to record whether each point is in the queue. In each iteration, take the head point v from the queue, and enumerate the edges v->u starting from v in turn. Let the length of the edge be len. Check if Dist[v] + len is less than Dist[u]. If it is less, improve Dist[u]. Record Fa[u] as v. Since the shortest distance from S to u has become smaller, check if u improves other points. So if u is not in the queue, put u at the tail of the queue. Continue iterating until the queue becomes empty, that is, all the shortest distances from S to u are determined, and the algorithm ends. If a point is enqueued more than n times, then set a negative weight cycle.

5. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters according to claim 1, characterized in that, In step S3, an infrared camera is installed on the sintering machine to collect data on changes in the FeO content of the sintered ore and the thermal field evolution in key sintering areas. The acquisition, storage, and processing mechanism of infrared camera images and temperature data enables FeO content prediction. An infrared thermal imager with a thermal imaging resolution of 384*288 pixels and a frame rate of over 30Hz is deployed at the tail of the sintering machine. It supports continuous shooting and outputs raw infrared images and corresponding temperature matrices. The acquisition frequency is set to 2 frames per second. The point data generated by the PLC control system is time-stamped with the infrared camera. The image data and process parameters, namely sintering machine speed, material thickness, and bellows pressure, are aligned in time sequence to facilitate subsequent fusion analysis. The temperature matrix is ​​designed as a structured storage format. Each frame of infrared image contains a two-dimensional temperature matrix of 384*288 pixels, which is converted into a structured numerical matrix and appended with metadata, timestamps, and temperature values. The system uses Apache Parquet for compressed storage and deploys a MinIO server as a relay station to build an object storage service based on MinIO. This service receives raw temperature data packets from infrared cameras, receives and caches infrared images and temperature data in real time, and provides access control and logging functions. The storage cycle is one sub-file per minute.

6. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters as described in claim 5, characterized in that, The infrared image is a structured numerical matrix, specifically calculated by calibrating the feature points of the RGB image and the infrared image to obtain a transformation relationship matrix. The matrix transformation formula is as follows: (x', y') are the coordinates of the infrared image, (x, y) are the coordinates of the RGB image, and a 1-4 The conversion factor t between the two images was obtained. x t y This represents the translation variable between two images. At least 3 pairs of matching points are required to calculate 6 variables, and at least 10 pairs of matching points are set for the images.

7. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters according to claim 1, characterized in that, The Transformer architecture in step S3 includes the following steps: S31. Image segmentation: ViT first divides the input image into multiple small blocks, each of which is a small region of 16×16 pixels; S32, Flattening and Vectorization: Each small block is flattened into a vector for easier subsequent processing; S33. Position encoding: In order to preserve the spatial information in the original image, position encoding is added to the vector; S34. Global Self-Attention Calculation: The vector with positional encoding information is fed into the standard Transformer encoder and processed through the global self-attention mechanism to capture long-distance dependencies in the entire image. S35, Classification Output: Use the last vector for image classification or other tasks.

8. The intelligent factory modeling, optimization, and collaborative method for adaptive adjustment of metallurgical process parameters according to claim 1, characterized in that, The data management architecture of the Object Storage Service (OBS) in step S4 utilizes the browser tool obs-browser-plus and the command-line tool obsutil provided by OBS to automate data upload and download management. obs-browser-plus supports concurrent upload of more than 500 images, meeting the batch upload requirements after local annotation; obsutil supports breakpoint resume and incremental synchronization through script configuration, adapting to data processing scenarios with long cycle tasks; the entire data upload process is recorded to a log file and combined with an MD5 verification mechanism. The data management architecture automatically calculates the particle size distribution of coke or sinter based on the segmentation results and determines whether the warning threshold has been reached, and whether the proportion of coke particles smaller than 25mm is less than 15%. The analysis results are output in a structured form and pushed to the production scheduling system in real time.

Images