An edge-computing-based hot-rolled slab number multi-node collaborative identification system and method

Abstract

The application discloses a hot-rolled slab number multi-node collaborative identification system and method based on edge computing and relates to the technical field of intelligent manufacturing of the iron and steel industry.The system comprises a field acquisition layer, an edge computing layer, a cloud management layer and a collaborative identification scheduling layer; the FPGA preprocessing module in the edge computing layer carries out multistage pipeline denoising and enhancement on the image, the ARM master control module carries out character recognition and outputs the recognition result and confidence; the collaborative identification scheduling layer is connected with multiple edge computing nodes, acquires the recognition result, confidence and load state of each node, triggers a cross-node collaborative verification mechanism when the recognition confidence of any node is lower than a threshold value, schedules another node to carry out secondary identification on the same image, and outputs the final slab number based on the arbitration of the multi-node result. Through the multi-node collaborative arbitration driven by the slab number recognition confidence and the dynamic load redistribution, the application significantly improves the recognition accuracy and system robustness under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of automation and intelligent manufacturing technology in the steel industry, specifically to a multi-node collaborative identification system and method for hot-rolled slab numbers based on edge computing. Background Technology

[0002] In hot-rolled steel production, slab numbers are the core identifier throughout steelmaking, rolling, and finished product warehousing. Identifying and recording characters on the slab surface using industrial cameras is crucial for product traceability and production management. Currently, some existing technologies propose preprocessing and recognizing slab images at the edge, such as using FPGAs or ARMs for image acceleration and centrally managing data in the cloud. Typical examples include CN201210091150.9, which discloses an automatic slab code recognition method; CN118797274A, which discloses a cloud-edge collaborative architecture for the hot rolling process; and a well-known FPGA and ARM combined image processing system.

[0003] However, the above solutions still have shortcomings when facing the extreme working conditions of hot rolling mills: high-temperature oxide scale coverage, drastic changes in lighting, and moisture interference can cause significant fluctuations in the character recognition confidence of a single edge node, leading to an increased error rate. While existing cloud-edge architectures can alleviate network load, they lack an adaptive processing mechanism for declining recognition confidence and fail to utilize the redundant computing power of multiple edge nodes for real-time collaborative verification of low-confidence results, resulting in insufficient robustness of the system under complex interference. Furthermore, traditional load balancing is based solely on node resource utilization and does not incorporate recognition quality indicators into the scheduling criteria, affecting the overall recognition accuracy.

[0004] Therefore, there is an urgent need for a system and method that can automatically trigger multi-node collaborative verification and dynamically schedule tasks based on the recognition quality when the confidence level of the identification decreases at the edge. Summary of the Invention

[0005] This invention aims to address the problems of existing edge-side slab number identification systems lacking multi-node collaborative verification mechanisms and having insufficient robustness when confidence decreases under complex interference. It provides a multi-node collaborative identification system and method for hot-rolled slab numbers based on edge computing.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A multi-node collaborative identification system for hot-rolled slab numbers based on edge computing includes:

[0008] The on-site acquisition layer contains several acquisition stations for acquiring raw image data of the slab surface on the hot rolling production line.

[0009] An edge computing layer comprises at least two edge computing nodes, each connected to the field acquisition layer via an industrial Ethernet network, for edge-side processing of raw image data. Each edge computing node includes an FPGA preprocessing module and an ARM main control module. The FPGA preprocessing module is connected to the ARM main control module via an internal bus to form a hierarchical processing structure. The FPGA preprocessing module is configured to perform multi-stage pipelined processing on the raw image data for denoising and enhancement, while the ARM main control module is configured to perform character recognition on the processed image and output structured recognition results and corresponding recognition confidence scores.

[0010] The cloud management layer connects to the edge computing layer via industrial Ethernet and is used to centrally manage the structured recognition results uploaded by each edge computing node.

[0011] The collaborative identification and scheduling layer is connected to each edge computing node and the cloud management layer via industrial Ethernet. The collaborative identification and scheduling layer is configured to: acquire the working status, load information, and identification confidence of the previous identification output of each edge computing node in real time, and compare the identification confidence with a preset confidence threshold; when the identification confidence output by any edge computing node is lower than the confidence threshold, a cross-node collaborative verification process is triggered, and the original image data or the image data processed by the FPGA preprocessing module is scheduled to at least one other edge computing node for secondary identification, and arbitration is performed based on the identification results of the main edge computing node and the collaborative edge computing node to output the final slab number information.

[0012] Furthermore, the collaborative identification and scheduling layer includes a confidence assessment unit, a task arbitration unit, and a load balancing unit;

[0013] The confidence assessment unit receives the identification confidence uploaded by each edge computing node and compares it with the confidence threshold.

[0014] After triggering the cross-node collaborative verification process, the task arbitration unit obtains the slab number identification results returned by multiple edge computing nodes and determines the final slab number information according to the preset arbitration rules.

[0015] The load balancing unit periodically collects the CPU utilization, memory usage, and queue length of pending tasks from each edge computing node, and dynamically adjusts the task distribution weights based on the recognition confidence level.

[0016] Furthermore, the preset arbitration rules are as follows:

[0017] When multiple edge computing nodes return the same slab number identification result, the consistent result is used as the final slab number information;

[0018] When the identification results are inconsistent, the identification result with the highest confidence level is selected as the final slab number information;

[0019] If two or more recognition results have the same and the highest confidence level, then the image is re-acquired and recognition is performed again.

[0020] Furthermore, the FPGA preprocessing module internally incorporates a multi-stage pipeline structure, which includes at least:

[0021] An adaptive median filter stage is used to remove salt-and-pepper noise and Gaussian noise from the original image;

[0022] A local enhancement level based on limited contrast adaptive histogram equalization is used to enhance low-contrast character areas on the slab surface caused by oxide scale coverage.

[0023] And a guided filtering-based edge-preserving smoothing level is used to suppress image ripple interference caused by thermal radiation and preserve character edges.

[0024] Furthermore, the ARM main control module integrates a character recognition model based on a deep convolutional neural network. The character recognition model is trained with hot-rolled slab sample data containing different oxidation levels and different illumination angles, and the recognition confidence is output through the Softmax function in the model output layer.

[0025] Furthermore, the collaborative identification and scheduling layer is also used to dynamically migrate the identification task corresponding to a certain edge computing node to the edge computing node with the lowest load when a continuous failure or load of a certain edge computing node is detected, and to automatically switch back the task after the failure is recovered.

[0026] Furthermore, the cloud management layer includes an MES server and a database. The MES server receives the final slab number information, timestamp, and workstation information output by the collaborative identification and scheduling layer through a data interface and stores them in the database.

[0027] Furthermore, it also includes a hardware carrier, in which at least a portion of the hardware of the edge computing layer is deployed; the hardware carrier is an industrial-grade ruggedized cabinet with a double-layer metal shielding shell, conductive sealing strips, and a composite heat dissipation system consisting of a heat pipe radiator and an intelligent speed-regulating fan.

[0028] Unlike existing technologies that only perform load balancing based on resource utilization without coordinating optimization of recognition quality, this invention incorporates recognition confidence into collaborative scheduling decisions, constructing a closed-loop mechanism of "single-point recognition - confidence assessment - cross-node collaborative verification - arbitration output." This mechanism utilizes the redundant computing power of multiple nodes at the edge to correct low-quality recognition results, significantly improving recognition accuracy and system robustness under complex operating conditions. Simultaneously, the pipeline structure specifically designed for hot-rolled oxide scale and thermal radiation interference in the FPGA preprocessing module further ensures the reliability of the single-point recognition baseline, creating complementary gains.

[0029] This system is particularly suitable for typical hot-rolled steel environments such as high temperature, strong electromagnetic interference, and uneven oxide scale coverage. Its confidence threshold, pre-processing pipeline, and collaborative arbitration strategy are all custom-designed for the above-mentioned industrial specificities, and have significant industrial applicability.

[0030] A multi-node collaborative identification method for hot-rolled slab numbers based on edge computing, applied to the above system, includes the following steps:

[0031] S1. The original image data of the hot-rolled slab surface is collected by the field acquisition layer and transmitted to the corresponding main edge computing node in the edge computing layer (200);

[0032] S2. The FPGA preprocessing module in the main edge computing node performs multi-stage pipelined denoising and enhancement processing on the original image data.

[0033] S3. The ARM main control module in the main edge computing node performs character recognition on the processed image, generates structured recognition results and recognition confidence, and uploads the recognition confidence to the collaborative recognition scheduling layer.

[0034] S4. The collaborative recognition scheduling layer compares the recognition confidence with the confidence threshold. If it is lower than the threshold, it sends the original image data or preprocessed image data to at least one other edge computing node for secondary recognition.

[0035] S5. The collaborative identification and scheduling layer receives the identification results of multiple edge computing nodes and outputs the final slab number information according to the preset arbitration rules.

[0036] S6. Upload the final slab number information to the cloud management layer for storage and display.

[0037] Furthermore, the arbitration rules preset in S5 are as follows:

[0038] When multiple edge computing nodes return the same slab number identification result, the consistent result is used as the final slab number information;

[0039] When the identification results are inconsistent, the identification result with the highest confidence level is selected as the final slab number information;

[0040] If two or more recognition results have the same and the highest confidence level, then the image is re-acquired and recognition is performed again.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. This invention constructs a closed-loop mechanism of "single-point identification—confidence assessment—cross-node collaborative verification—arbitration output." When the identification confidence of any edge computing node falls below a threshold, it automatically triggers other nodes to perform secondary identification on the same image. Based on the multi-node results, the final slab number is output through arbitration, effectively correcting the low-quality identification results caused by single-node factors such as high-temperature oxide scale, moisture, and strong electromagnetic interference. Actual test data shows that after adopting this collaborative verification mechanism, the slab number identification accuracy increased from 92.3% for single nodes to 99.6%.

[0043] 2. This invention uses the historical average recognition confidence of edge computing nodes as a task distribution weight factor, so that recognition tasks are preferentially allocated to nodes with high recognition quality and sufficient resources, which not only ensures the load balance of each node, but also continuously optimizes the overall recognition quality of the system.

[0044] 3. The multi-stage pipeline built inside the FPGA preprocessing module of this invention, including adaptive median filtering, CLAHE local enhancement, and guided filtering edge preservation smoothing, is specifically designed to address typical problems such as oxide scale coverage and thermal radiation ripple interference on the surface of hot-rolled slabs. It completes image denoising and enhancement within a low latency of 2ms, effectively ensuring the reliability of the single-node recognition baseline and providing high-quality input for collaborative verification.

[0045] 4. The edge computing node hardware carrier of the present invention adopts a composite heat dissipation system consisting of a double-layer metal shielding shell, conductive sealing strip, heat pipe heat dissipation and intelligent speed-regulating fan, which can stably operate in the hot rolling site of steel with high temperature of 55℃, high dust and strong electromagnetic interference, ensuring the long-term reliability of the identification system.

[0046] 5. The collaborative identification scheduling layer of the present invention adopts a primary and backup redundant deployment, and can automatically migrate the identification task to the backup node when a node is detected to have continuous failures or overload. After the fault is recovered, it will automatically switch back, realizing high availability under unattended operation and avoiding production data interruption caused by single point of failure. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention;

[0048] Figure 2 This is a schematic diagram of the edge computing layer structure of the present invention;

[0049] Figure 3 This is a schematic diagram of the functional modules of the collaborative identification and scheduling layer of the present invention;

[0050] Figure 4 This is a flowchart illustrating the method of the present invention.

[0051] In the diagram: 100, Field acquisition layer; 101, Industrial camera; 102, Laser illumination assembly;

[0052] 200. Edge computing layer; 201. FPGA preprocessing module; 202. ARM main control module; 210 / 220 / 230. Edge computing node;

[0053] 300. Cloud management layer; 500. Human-computer interaction terminal;

[0054] 400. Collaborative Identification and Scheduling Layer; 401. Confidence Evaluation Unit; 402. Task Arbitration Unit; 403. Load Balancing Unit; Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0056] like Figure 1 As shown in the figure, this embodiment provides a multi-node collaborative identification system for hot-rolled slab numbers based on edge computing, including a field acquisition layer 100, an edge computing layer 200, a cloud management layer 300, and a collaborative identification and scheduling layer 400.

[0057] The field acquisition layer 100 is set at the entrance of the hot rolling production line and key identification stations, including an industrial camera 101 with a resolution of 2048×1080 and a frame rate of no less than 60fps and a laser fill light component 102 that is triggered synchronously with it. The industrial camera 101 is directly connected to the main edge computing node of the corresponding station through a dedicated gigabit network cable.

[0058] The edge computing layer 200 contains at least two edge computing nodes. Figure 1 As shown in nodes 210, 220, and 230, each node's hardware is deployed in an independent industrial-grade ruggedized cabinet. This cabinet has a double-layer metal shielded shell, with conductive sealing strips filling the seams, and is connected to the factory's grounding grid via grounding terminals. Internally, it is equipped with a composite cooling system consisting of copper heat pipe radiators and PWM intelligent speed-regulating fans, ensuring stable operation in high-temperature and high-dust environments up to 55°C.

[0059] like Figure 2As shown, a single edge computing node integrates an FPGA preprocessing module 201 and an ARM main control module 202. The FPGA preprocessing module 201 uses a Xilinx Kintex-7 series chip and internally constructs a four-stage pipeline: the first stage is adaptive median filtering to filter out salt-and-pepper noise and Gaussian noise; the second stage is local enhancement based on CLAHE (Contrast Limiting Adaptive Histogram Equalization) to improve the character contrast in the oxide-covered area; the third stage is edge-preserving smoothing based on guided filtering to suppress image ripples caused by thermal radiation; and the fourth stage is bilinear interpolation scaling to unify the image to 512×256 pixels. The pipeline achieves pixel-level parallel processing through internal Block RAM cache, with latency controlled within 2ms.

[0060] The ARM main control module 202 uses a multi-core Cortex-A72 processor with a main frequency of 1.8GHz, and is connected to the FPGA preprocessing module 201 via a PCIe 3.0 x4 bus (bandwidth approximately 4GB / s). Internally, the ARM main control module 202 deploys a deep convolutional neural network character recognition model optimized with TensorFlowLite. This model uses MobileNetV3 as its backbone and is trained using 12,000 sample images of hot-rolled slabs with different oxidation levels and illumination angles. The output layer outputs the probability of each category through a Softmax function, and the maximum probability value is taken as the recognition confidence score. The average inference time of the model is 45ms, and the confidence threshold is set to 0.85. This threshold was determined experimentally: in 6,000 test samples, the recognition error rate increases significantly when the confidence score is below 0.85. Using 0.85 balances the false negative rate and the overhead of collaborative communication, ensuring overall efficiency. When the confidence level is ≥0.85, the structured recognition result is directly output, including the slab number string, confidence level, timestamp, and workstation number; when the confidence level is <0.85, in addition to the output result, a "low confidence" status bit is also marked in the uploaded data packet.

[0061] The collaborative identification and scheduling layer 400 is deployed on a dedicated rack-mounted industrial server. To improve system reliability, a primary-backup redundancy deployment is adopted. When the primary scheduling server fails, it automatically switches to the backup scheduling server, ensuring uninterrupted scheduling functionality. The scheduling layer connects to each edge computing node and the cloud management layer 300 via a gigabit industrial Ethernet switch. Its functional modules are as follows: Figure 3 As shown, it includes a confidence assessment unit 401, a task arbitration unit 402, and a load balancing unit 403. The three units work collaboratively, and the specific process is as follows:

[0062] (1) Each edge computing node reports heartbeat information to the collaborative recognition scheduling layer 400 every 200ms. The information payload includes node CPU utilization, memory utilization, length of the image queue to be processed, confidence of the most recent recognition task, and low confidence marker.

[0063] (2) The confidence evaluation unit 401 analyzes the confidence reported by each node and compares it with the threshold 0.85. Once a low confidence marker is detected by a main edge computing node, the cross-node collaborative verification process for the slab image is immediately triggered.

[0064] (3) The task arbitration unit 402 distributes the 512×256 pixel image preprocessed by the master node FPGA (in order to ensure that the image content processed by each collaborative node is completely consistent, the preprocessed normalized image is transmitted first to reduce the network load; if the master node preprocessing stage is abnormal, the original image is transmitted and preprocessed by the collaborative node itself) to the two edge computing nodes with the lowest load in the system other than the master node as collaborative edge computing nodes for secondary identification.

[0065] (4) After the collaborative edge computing nodes complete the identification, they return their respective slab number strings and confidence scores to the task arbitration unit 402, and the task arbitration unit 402 executes the arbitration rules:

[0066] Rule 1: If the slab number returned by the main edge computing node and the two collaborative edge computing nodes are completely consistent, then the consistent result shall be used as the final slab number information.

[0067] Rule 2: If the results are inconsistent, the slab number with the highest confidence level among all results shall be selected as the final result;

[0068] Rule 3: If there are two or more results with the same confidence level and are tied for the highest, the task arbitration unit 402 sends a re-sampling command to the field acquisition layer 100, the industrial camera 101 re-captures the slab image, and starts a new round of recognition process.

[0069] (5) In addition to periodically collecting hardware resource data, the load balancing unit 403 also uses the historical average identification confidence level of the nodes as a scheduling weight factor. A specific weight adjustment algorithm example is as follows:

[0070] Task distribution weight W i =α·(1-U i )+β·C i ;

[0071] Among them U i C is the normalized total load value for node i (a weighted average of CPU utilization and memory utilization). i Let α be the average confidence score of node i in the most recent 100 recognition tasks, and β be the weighting coefficients.

[0072] In this embodiment, α=0.4 and β=0.6 are used to give higher distribution weights to nodes with high confidence and sufficient resources, thus prioritizing high-quality nodes during normal task allocation and maintaining a high overall recognition quality. When a node is detected to have three consecutive heartbeat timeouts or malfunctions, the acquired images of the corresponding workstation of that node are automatically migrated to a preset backup node, and automatic switchback is triggered after the fault is recovered.

[0073] The cloud management layer 300 is a server configured with dual Intel Xeon Silver 4214 processors, 64GB of DDR4 ECC memory, and a RAID5 disk array, running the MES system. The MES server receives the final slab number information, timestamp, workstation information, and collaborative verification flags output by the collaborative identification and scheduling layer 400 via a RESTful API and stores them in a PostgreSQL database. The human-machine interface terminal 500 communicates with the cloud management layer 300 via WebSocket, displaying real-time identification results for each workstation, system topology status, and confidence trend graphs for each node.

[0074] like Figure 4 As shown in the method flow, during operation:

[0075] Step S1: The industrial camera 101 acquires an image of the slab surface under laser illumination synchronization and transmits it to the corresponding main edge computing node;

[0076] Step S2, the FPGA preprocessing module 201 performs four-stage pipeline processing;

[0077] Step S3: The ARM main control module 202 runs the character recognition model and outputs the structured results and confidence scores.

[0078] In step S4, when the collaborative recognition scheduling layer 400 detects a confidence level below 0.85, it distributes the image to at least one collaborative edge computing node; if the confidence level is normal, it directly adopts the result from the master node.

[0079] Step S5: After multiple collaborative edge computing nodes complete the identification, the task arbitration unit 402 executes the arbitration logic to generate the final slab number.

[0080] Step S6: The final slab number is uploaded to the MES system database and displayed as 500 on the human-computer interaction terminal.

[0081] This embodiment underwent 30 days of continuous on-site testing on a hot rolling production line at a steel plant, collecting a total of 5120 slab image samples, covering different oxidation levels, lighting conditions, and moisture conditions. The average accuracy rate of independent single-node recognition was 92.3%. After enabling the collaborative verification mechanism, the accuracy rate of the final output slab number increased to 99.6%. The test results show that this system has extremely high robustness in recognizing complex interference in industrial environments.

[0082] The above description is merely 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 within the protection scope of the present invention.

Claims

1. A multi-node collaborative identification system for hot-rolled slab numbers based on edge computing, characterized in that, include: The field acquisition layer (100) contains several acquisition stations for acquiring raw image data of the slab surface on the hot rolling production line; An edge computing layer (200) includes at least two edge computing nodes. Each edge computing node is connected to the field acquisition layer (100) via an industrial Ethernet and is used to perform edge-side processing on the raw image data. The edge computing node includes an FPGA preprocessing module (201) and an ARM main control module (202). The FPGA preprocessing module (201) is connected to the ARM main control module (202) via an internal bus to form a hierarchical processing structure. The FPGA preprocessing module (201) is configured to perform multi-stage pipeline processing on the raw image data for denoising and enhancement. The ARM main control module (202) is configured to perform character recognition on the processed image and output the structured recognition result and the corresponding recognition confidence level. The cloud management layer (300) is connected to the edge computing layer (200) via industrial Ethernet and is used to centrally manage the structured recognition results uploaded by each edge computing node; The collaborative identification scheduling layer (400) is connected to each edge computing node and the cloud management layer (300) via industrial Ethernet. The collaborative identification scheduling layer (400) is configured to: acquire the working status, load information and identification confidence of the previous identification output of each edge computing node in real time, and compare the identification confidence with the preset confidence threshold; when the identification confidence output by any edge computing node is lower than the confidence threshold, the cross-node collaborative verification process is triggered, the original image data or the image data processed by the FPGA preprocessing module is scheduled to at least one other edge computing node for secondary identification, and arbitration is performed based on the identification results of the main edge computing node and the collaborative edge computing node to output the final slab number information.

2. The system according to claim 1, characterized in that, The collaborative identification and scheduling layer (400) includes a confidence evaluation unit (401), a task arbitration unit (402), and a load balancing unit (403). The confidence assessment unit (401) receives the identification confidence uploaded by each edge computing node and compares it with the confidence threshold. After triggering the cross-node collaborative verification process, the task arbitration unit (402) obtains the slab number identification results returned by multiple edge computing nodes and determines the final slab number information according to the preset arbitration rules. The load balancing unit (403) periodically collects the CPU utilization rate, memory utilization rate and the length of the task queue to be processed of each edge computing node, and dynamically adjusts the task distribution weight in combination with the recognition confidence.

3. The system according to claim 2, characterized in that, The preset arbitration rules are as follows: When multiple edge computing nodes return the same slab number identification result, the consistent result is used as the final slab number information; When the identification results are inconsistent, the identification result with the highest confidence level is selected as the final slab number information; If two or more recognition results have the same and the highest confidence level, then the image is re-acquired and recognition is performed again.

4. The system according to claim 1, characterized in that, The FPGA preprocessing module (201) internally constructs a multi-stage pipeline structure, which includes at least: An adaptive median filter stage is used to remove salt-and-pepper noise and Gaussian noise from the original image; A local enhancement level based on limited contrast adaptive histogram equalization is used to enhance low-contrast character areas on the slab surface caused by oxide scale coverage. And a guided filtering-based edge-preserving smoothing level is used to suppress image ripple interference caused by thermal radiation and preserve character edges.

5. The system according to claim 1, characterized in that, The ARM main control module (202) integrates a character recognition model based on a deep convolutional neural network. The character recognition model is trained with hot-rolled slab sample data containing different oxidation levels and different illumination angles, and the recognition confidence is output through the Softmax function in the model output layer.

6. The system according to claim 1, characterized in that, The collaborative identification scheduling layer (400) is also used to dynamically migrate the identification task corresponding to a certain edge computing node to the edge computing node with the lowest load when a continuous failure or load of a certain edge computing node is detected, and to automatically switch back the task after the failure is recovered.

7. The system according to claim 1, characterized in that, The cloud management layer (300) includes an MES server and a database. The MES server receives the final slab number information, timestamp, and workstation information output by the collaborative identification and scheduling layer (400) through a data interface and stores them in the database.

8. The system according to any one of claims 1 to 7, characterized in that, It also includes a hardware carrier, in which at least a portion of the hardware of the edge computing layer is deployed; the hardware carrier is an industrial-grade ruggedized cabinet with a double-layer metal shielding shell, conductive sealing strips, and a composite heat dissipation system consisting of a heat pipe radiator and an intelligent speed-regulating fan.

9. A multi-node collaborative identification method for hot-rolled slab numbers based on edge computing, applied to the system according to any one of claims 1 to 8, characterized in that, Includes the following steps: S1. The original image data of the hot-rolled slab surface is collected by the field acquisition layer (100) and transmitted to the corresponding main edge computing node in the edge computing layer (200); S2. The FPGA preprocessing module (201) in the main edge computing node performs multi-stage pipelined denoising and enhancement processing on the original image data. S3. The ARM main control module (202) in the main edge computing node performs character recognition on the processed image, generates structured recognition results and recognition confidence, and uploads the recognition confidence to the collaborative recognition scheduling layer (400). S4. The collaborative recognition scheduling layer (400) compares the recognition confidence with the confidence threshold. If it is lower than the threshold, it sends the original image data or preprocessed image data to at least one other edge computing node for secondary recognition. S5. The collaborative identification and scheduling layer (400) receives the identification results of multiple edge computing nodes and outputs the final slab number information according to the preset arbitration rules. S6. Upload the final slab number information to the cloud management layer for storage and display.

10. The method according to claim 9, characterized in that, The arbitration rules preset in S5 are as follows: When multiple edge computing nodes return the same slab number identification result, the consistent result is used as the final slab number information; When the identification results are inconsistent, the identification result with the highest confidence level is selected as the final slab number information; If two or more recognition results have the same and the highest confidence level, then the image is re-acquired and recognition is performed again.

Images