A cloud production process intelligent monitoring method and system for steel plate cutting

Abstract

This invention provides a cloud-based intelligent monitoring method and system for steel plate cutting, relating to the field of steel plate processing control technology. The method includes: deploying a lightweight image recognition network at the edge nodes of the steel plate cutting process to identify the cutting head position in real time; comparing the results with a preset path to generate preliminary analysis results including positional deviation and state features; uploading the results to a cloud-based digital twin via a highly reliable, low-latency network for high-precision dynamic simulation; generating a globally optimized correction control strategy; and performing microsecond-level dynamic compensation with local real-time sensor feedback; finally generating a driving command to drive the cutting head to complete adaptive correction. This solves the technical problem of existing steel plate cutting methods relying on preset paths and fixed strategies, resulting in poor dynamic response to the real-time cutting process, leading to large cutting deviations and low accuracy. It achieves the technical effect of improving the accuracy and quality of steel plate cutting through multimodal real-time monitoring and dynamic adaptive correction of the cutting process.

Description

Technical Field

[0001] This invention relates to the field of steel plate processing control technology, specifically to a cloud-based intelligent monitoring method and system for steel plate cutting production processes. Background Technology

[0002] In complex manufacturing fields such as steel plate cutting, ensuring high-precision cutting paths and excellent cut quality is a core challenge. Traditional methods mainly rely on offline programming and preset paths, with fixed control strategies that struggle to respond in real time to sudden disturbances during processing, such as internal material stress release, thermal deformation, and equipment vibration. These unpredictable dynamic factors often cause the actual cutting trajectory to deviate from the preset path, leading to reduced cut quality and scrapped parts, severely restricting production efficiency and yield. Existing real-time control schemes based on local sensors, while possessing some responsiveness, have limited decision-making perspectives and often lack global and predictive handling of complex coupled disturbances, making it difficult to achieve truly adaptive optimal control and impacting the accuracy and quality of steel plate cutting. Summary of the Invention

[0003] The purpose of this application is to provide a cloud-based intelligent monitoring method and system for steel plate cutting production processes, which solves the technical problem that existing steel plate cutting relies on preset paths and fixed strategies, resulting in poor dynamic response capabilities to real-time cutting processes.

[0004] In view of the above problems, this application provides a cloud-based intelligent monitoring method and system for steel plate cutting production processes.

[0005] The first aspect of this application provides a cloud-based intelligent monitoring method for steel plate cutting production processes. This method includes: deploying a lightweight image recognition network at the edge nodes of a steel plate cutting device and acquiring image data sequences during the cutting process in real time using industrial vision sensors; constructing a digital twin of the steel plate cutting process on a cloud platform and establishing a real-time synchronous mapping relationship between the digital twin and the physical cutting device; processing the image data sequences in real time using the lightweight image recognition network at the edge nodes to identify the cutting head position and compare it with a preset cutting path to generate preliminary analysis results including cutting head position deviation and cutting state characteristics; uploading the preliminary analysis results to the cloud platform via a highly reliable, low-latency communication network, and performing high-precision dynamic simulation of the current cutting process based on the digital twin to generate a globally optimized correction control strategy; distributing the correction control strategy to the edge nodes, combining local real-time sensor feedback, performing microsecond-level dynamic compensation of control commands, generating final drive commands, and driving the cutting head to complete adaptive correction.

[0006] A second aspect of this application provides a cloud-based intelligent monitoring system for steel plate cutting production processes. This system includes: a monitoring network construction module for deploying a lightweight image recognition network at the edge nodes of the steel plate cutting equipment and acquiring image data sequences during the cutting process in real time using industrial vision sensors; a data mapping module for constructing a digital twin of the steel plate cutting process on a cloud platform and establishing a real-time synchronous mapping relationship between the digital twin and the physical cutting equipment; and a position change trend generation module for processing the image data through the lightweight image recognition network at the edge nodes. The sequence is processed in real time to identify the position of the cutting head and compare it with the preset cutting path to generate preliminary analysis results including the position deviation of the cutting head and the characteristics of the cutting state. The motion prediction module is used to upload the preliminary analysis results to the cloud platform through a high-reliability, low-latency communication network, and based on the digital twin, to perform high-precision dynamic simulation of the current cutting process and generate a globally optimized correction control strategy. The cutting control module is used to send the correction control strategy to the edge nodes, and combined with local real-time sensor feedback, to perform microsecond-level dynamic compensation of the control commands, generate the final drive command, and drive the cutting head to complete adaptive correction.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0008] The method provided in this application deploys a lightweight image recognition network at the edge nodes of a steel plate cutting equipment and collects image data sequences during the cutting process in real time using industrial vision sensors. A digital twin of the steel plate cutting process is constructed on a cloud platform, and a real-time synchronous mapping relationship is established between the digital twin and the physical cutting equipment. The lightweight image recognition network at the edge nodes processes the image data sequences in real time, identifies the cutting head position, compares it with a preset cutting path, and generates preliminary analysis results including cutting head position deviation and cutting state characteristics. These preliminary analysis results are uploaded to the cloud platform via a highly reliable, low-latency communication network. Based on the digital twin, the current cutting process is dynamically simulated with high precision to generate a globally optimized deviation correction control strategy. The deviation correction control strategy is then distributed to the edge nodes, and combined with local real-time sensor feedback, microsecond-level dynamic compensation of control commands is performed to generate final drive commands that drive the cutting head to complete adaptive deviation correction. This achieves the technical effect of improving the accuracy and quality of steel plate cutting through multimodal real-time monitoring and dynamic adaptive deviation correction of the cutting process.

[0009] The above description is merely an overview of the technical solution of this application. To better understand the technical means of this application and to facilitate its implementation according to the description, and to make the above and other objects, features, and advantages of this application more apparent, specific embodiments of this application are described below. It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent through the following description. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0011] Figure 1 This application provides a flowchart illustrating a cloud-based intelligent monitoring method for steel plate cutting production processes.

[0012] Figure 2 This application provides a schematic diagram of the structure of a cloud-based intelligent monitoring system for steel plate cutting production processes.

[0013] Figure labeling: 11 Monitoring network construction module, 12 Data mapping module, 13 Position change trend generation module, 14 Motion prediction module, 15 Cutting and control module. Detailed Implementation

[0014] This application provides a cloud-based intelligent monitoring method and system for steel plate cutting production processes, addressing the technical problem that existing steel plate cutting methods rely on preset paths and fixed strategies, resulting in poor dynamic response capabilities to real-time cutting processes. It achieves the technical effect of improving the accuracy and quality of steel plate cutting through multimodal real-time monitoring and dynamic adaptive correction of the cutting process.

[0015] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. It should be understood that the present invention is not limited to the exemplary embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. It should also be noted that, for ease of description, only the parts related to the present invention are shown in the accompanying drawings, not all of them.

[0016] Example 1, as Figure 1As shown, this application provides a cloud-based intelligent monitoring method for steel plate cutting production processes, which includes:

[0017] A lightweight image recognition network is deployed at the edge nodes of the steel plate cutting equipment, and image data sequences during the cutting process are acquired in real time through industrial vision sensors.

[0018] Furthermore, a lightweight image recognition network is deployed at the edge nodes of the steel plate cutting equipment, including: constructing an image processing model with a multi-layer feature extraction structure, wherein the input of the image processing model is an image data sequence and the output is the image features of the cutting head; performing compression and acceleration processing on the image processing model, removing redundant computational structures, and forming a lightweight image recognition network suitable for edge computing resources; deploying the lightweight image recognition network on the edge computing device and establishing a communication connection with the industrial vision sensor to directly receive and process the image data sequence acquired by the industrial vision sensor.

[0019] Specifically, to achieve high-frequency, real-time visual monitoring of the steel plate cutting process, the primary task is to deploy edge computing nodes next to the steel plate cutting equipment on the production site. This edge node typically consists of an industrial computer with certain graphics processing capabilities or a dedicated edge computing gateway, and its physical location is close to the cutting equipment to minimize data transmission latency.

[0020] The core software module deployed on this edge node is a lightweight image recognition network. The development of this network follows a standardized process from design and optimization to deployment.

[0021] First, the network design phase requires building an image processing model with a multi-layered feature extraction structure. A mature deep convolutional neural network architecture, such as a version of ResNet or MobileNet, can be chosen as the base image processing model. This model is designed to receive raw video streams acquired in real-time by an industrial vision sensor—a high-resolution industrial camera—at a frequency of tens to hundreds of frames per second. This video stream is processed into a temporally ordered sequence of image data within the system. The training objective of the model is to accurately locate and extract visual features representing the cutting head and its key components, such as the cutting nozzle or beam focus, from each frame. These features are ultimately encoded into a high-dimensional feature vector within the model, collectively referred to as the cutting head image features.

[0022] Next, the model optimization phase begins, which involves compressing and accelerating the image processing model. Since the original deep learning model typically contains millions or even billions of parameters and complex computational graphs, its inference speed is slow and resource-intensive, making it unsuitable for real-time operation on edge devices with limited computing power. Therefore, a series of model compression techniques must be employed to remove redundant computational structures. Specifically, network pruning is performed first, by analyzing the importance of neurons or convolutional kernels in each layer of the model and removing those that contribute minimally to the final output, significantly reducing the total number of parameters and computational load. Next, weight quantization is performed, converting the model parameters from 32-bit floating-point numbers to 8-bit integers or even lower precision, significantly reducing the model's storage space and memory bandwidth requirements. The hardware's acceleration capabilities for integer operations are also utilized to improve inference speed. Finally, knowledge distillation is performed, using a large but accurate teacher model to guide the training of a small student model, allowing the student model to approximate the performance of the teacher model while maintaining a smaller size. After this series of optimization operations, the original massive model is transformed into a computationally efficient, memory-efficient, lightweight image recognition network, thus becoming suitable for environments where edge computing resources are typically limited.

[0023] Finally, the system is integrated and deployed. The optimized lightweight image recognition network model file is converted into an inference engine format suitable for the target edge computing hardware, such as a specific model of GPU or neural processing unit, using a specific deployment toolchain. This inference engine is then integrated into the monitoring application deployed on the edge nodes. Simultaneously, a robust communication connection must be established both physically and logically: at the physical level, high-quality Ethernet cables or dedicated camera cables are used to connect the industrial vision sensors to the corresponding data interfaces of the edge computing devices; at the software level, appropriate device drivers and software development kits are configured to ensure that the application can directly and stably receive real-time image data sequence streams from the sensors. After deployment, the entire system forms a closed data loop from sensor acquisition to real-time processing at the edge nodes. The industrial camera continuously captures images of the cutting area, and the resulting image data sequence is instantly pushed to the lightweight network on the edge nodes for inference. The network continuously outputs real-time image features of the cutting head, providing a millisecond-level latency data foundation for subsequent real-time analysis, deviation calculation, and control decisions—a prerequisite for achieving accurate dynamic correction.

[0024] A digital twin of the steel plate cutting process is constructed on a cloud platform, and a real-time synchronous mapping relationship is established between the digital twin and the physical cutting equipment.

[0025] Specifically, a digital twin of steel plate cutting is built on a cloud platform. This digital twin is a high-fidelity virtual model running on a cloud server, with the goal of providing a one-to-one digital mapping and real-time simulation of the steel plate cutting equipment, processing procedures, and workpieces in the physical world—that is, the physical cutting equipment.

[0026] The first step in building a digital twin is model creation. This requires integrating multi-source data and using 3D modeling software or a dedicated digital twin development platform in the cloud to build a composite model that includes geometric, physical, and behavioral rules. The geometric model is built based on the design drawings of the cutting equipment, 3D scanned point cloud data, and the CAD model of the steel plate, accurately reproducing the equipment structure, cutting head, worktable, and the shape, size, and assembly relationships of the steel plate itself. The physical model injects physical properties, such as assigning material properties to the steel plate, defining its density, elastic modulus, and coefficient of thermal expansion; and defining thermodynamic models and material removal models for the cutting process. The behavioral model defines the kinematic and dynamic rules of each moving part of the equipment, as well as the response logic of the control system to drive commands, through programming. The combination of these three elements constitutes a digital twin capable of simulating the physical laws and logical interactions of the real world.

[0027] Next, a real-time synchronous mapping is established between this virtual digital twin and the real physical cutting equipment. This mapping is not static, but rather a dynamic and continuous data flow interaction. To achieve synchronization, downlink data injection is first required: in the cloud, preliminary analysis results—namely, the cutting head position deviation and cutting state characteristics—are transmitted from edge nodes via a highly reliable, low-latency communication network, along with real-time process parameters obtained from the equipment controller. These are continuously and in real-time injected into the corresponding virtual entity in the digital twin as driving signals. For example, the cutting head position deviation data is mapped to the virtual cutting head's position offset in the model, and vibration characteristic data is mapped to the excitation input of the virtual equipment. This process ensures that the state of the virtual model can follow the actual state changes of the physical entity.

[0028] Secondly, state synchronization and calibration are performed. The multiphysics simulation model inside the digital twin runs on a high-precision virtual timeline based on injected real-time data, simulating multi-field coupling effects such as cutting thermal stress, material deformation, and equipment vibration. This process, known as high-precision dynamic extrapolation, predicts the future trajectory of the virtual cutting head and the real-time deformation of the steel plate. Simultaneously, the system continuously compares the simulation predictions of the digital twin with the latest measured data fed back from the physical world. If deviations occur, the system uses this deviation data, employing algorithms such as parameter identification and state estimation, to reverse-calibrate and correct the parameters in the digital twin model. This ensures that the simulated behavior continuously approximates the real behavior of the physical entity, maintaining and optimizing the accuracy of the real-time synchronization mapping relationship.

[0029] Through this bidirectional, dynamic, real-time synchronous mapping relationship, the digital twin on the cloud platform is no longer an offline, static model, but an intelligent mirror that operates in parallel with the physical production line and blends the virtual and real worlds. This allows operators to observe the entire cutting process non-destructively and proactively in virtual space, predict potential problems, and test and optimize control strategies based on this highly simulated environment.

[0030] The image data sequence is processed in real time by the lightweight image recognition network of the edge nodes to identify the position of the cutting head and compare it with the preset cutting path to generate preliminary analysis results including the position deviation of the cutting head and the cutting state characteristics.

[0031] Furthermore, generating preliminary analysis results including cutting head position deviation and cutting state features also includes: extracting the current image frame from the image data sequence, inputting it into the lightweight image recognition network to obtain the image coordinates of the cutting head image in the current image frame; combining a pre-calibrated coordinate transformation relationship to transform the image coordinates of the cutting head image to the physical working coordinate system of the cutting device to obtain the real-time physical coordinates of the cutting head; obtaining the target coordinates of the cutting path corresponding to the current moment from the preset cutting path; calculating the difference between the real-time physical coordinates and the target coordinates of the cutting path to generate the cutting head position deviation; extracting the feature vector of the intermediate layer of the lightweight image recognition network as the cutting state feature representing the current cutting scene; and combining the cutting head position deviation and the cutting state feature to form the preliminary analysis results.

[0032] It should be understood that at the edge node of the steel plate cutting equipment, the lightweight image recognition network is activated and enters a continuously running high-frequency processing loop. Its core task is to process the image data sequence continuously generated by the industrial vision sensor in real time in order to complete the accurate identification and status analysis of the cutting head position.

[0033] The processing begins with frame capture of the image data sequence. The system extracts the image frame corresponding to the current moment from the sensor data stream, which is typically a two-dimensional digital image with a millisecond-level timestamp. This current image frame is immediately fed into a pre-deployed lightweight image recognition network. During forward inference, the network's deep convolutional layers perform layer-by-layer feature extraction and abstraction on the image. Finally, the network output layer provides the location information of the cutting head in the image, typically represented by the coordinates of the center point of a bounding box. These coordinates are defined based on pixel position, i.e., the image coordinates of the cutting head image in the current image frame.

[0034] However, image coordinates exist only in a two-dimensional pixel plane and cannot be directly used for motion control in three-dimensional space. Therefore, they must be transformed to a real-world physical scale. This requires a crucial prior information—a pre-calibrated coordinate transformation relationship. This relationship is obtained through joint calibration of the camera and cutting equipment workspaces, establishing a precise mathematical mapping between the image pixel coordinate system and the physical working coordinate system of the cutting equipment. By applying this transformation relationship, for example using a homography matrix or camera projection model obtained through calibration, the image coordinates of the cutting head image are calculated and converted into the real-time physical coordinates of the cutting head in real three-dimensional space, expressed in units of length. These coordinates precisely describe the instantaneous position of the cutting head in the equipment's workspace.

[0035] Simultaneously, the system retrieves the preset cutting path from the pre-stored process file based on the current processing time. The preset cutting path is a continuous trajectory composed of a series of ordered spatial coordinate points, defining the ideal movement route the cutting head should theoretically follow. The system interpolates or searches within this path to obtain the target coordinates of the cutting path that perfectly correspond to the current image frame acquisition time.

[0036] Subsequently, the deviation calculation module compares the real-time physical coordinates of the cutting head with the target coordinates of the cutting path obtained from the preset path. By calculating the Euclidean distance between the two in three-dimensional space or the difference decomposed onto each coordinate axis, the system generates a quantified cutting head position deviation with orientation and magnitude information. This deviation value intuitively reflects the instantaneous degree of deviation between the actual position and the theoretical position of the cutting head.

[0037] In addition to location information, to more comprehensively perceive the state of the cutting process, the system also extracts deeper semantic information from the lightweight image recognition network. Specifically, it extracts a multi-dimensional feature vector from the output of a certain intermediate layer of the network, such as after the last convolutional layer and before the fully connected layer. This feature vector is a distributed representation of the input image after the network performs a high-level abstraction and understanding. It encodes rich visual information of the current cutting scene, such as the shape of the cutting sparks, the brightness and texture of the molten pool, and the clarity of the cut edges. This information comprehensively characterizes the stability and quality of the cutting process. Therefore, this feature vector is defined as the cutting state feature.

[0038] Finally, the calculated cutting head position deviation and the extracted cutting state features are packaged and combined. The deviation value is processed into a low-dimensional vector, while the state feature is a high-dimensional vector. These are encapsulated together with information such as the current frame's timestamp and device identifier to form a structured data packet. This data packet represents the preliminary analysis result, condensing the edge node's perception and judgment at the moment of cutting, providing the most crucial and concise input data for subsequent uploading to the cloud for in-depth analysis and decision-making. The entire process runs in a closed loop on the edge device with millisecond-level latency, achieving real-time conversion from raw images to preliminary information directly usable for intelligent decision-making.

[0039] The preliminary analysis results are uploaded to the cloud platform through a highly reliable, low-latency communication network. Based on the digital twin, the current cutting process is dynamically simulated with high precision to generate a globally optimized correction control strategy.

[0040] Furthermore, based on the digital twin, a high-precision dynamic simulation of the current cutting process is performed to generate a globally optimized correction control strategy. This includes: establishing a virtual cutting environment in the digital twin that includes material properties, process parameters, and geometric structure; mapping the cutting state characteristics in the preliminary analysis results to the input states of corresponding virtual sensors in the virtual cutting environment; activating a multiphysics simulation model to perform a high-precision dynamic simulation in the virtual cutting environment, predicting the possible evolution of the cutting path under different intervention measures, and generating dynamic simulation prediction results; and calculating and generating an optimal set of control actions based on the dynamic simulation prediction results, with cutting quality and path accuracy as optimization objectives, as the globally optimized correction control strategy.

[0041] Optionally, on the cloud platform, after receiving the preliminary analysis results from the edge nodes through a highly reliable low-latency communication network, the powerful computing power and digital twin simulation capabilities of the cloud can be used to perform high-precision dynamic simulation of the current cutting process and ultimately generate a globally optimized correction control strategy.

[0042] First, the virtual cutting environment within the digital twin is constructed and its state synchronized. The digital twin is not a static model, but a runnable virtual cutting environment containing material properties, process parameters, and geometric structures. Material properties define the physical characteristics of the virtual steel plate, such as density, specific heat capacity, and thermal conductivity; process parameters set operating conditions such as cutting speed, gas pressure, and laser power; and the geometric structure accurately reproduces the physical form of the cutting equipment, cutting head, and steel plate. When new preliminary analysis results are received, the cloud system immediately parses the cutting state features—a high-dimensional vector representing the visual semantics of the current cutting scene—and maps them to the input signals of corresponding virtual sensors in the virtual cutting environment. For example, information encoded in the feature vector, such as high molten pool brightness and numerous burrs on the cut edge, may be mapped to high-temperature readings from a virtual thermal imager or texture anomaly signals from a virtual vision sensor, thus keeping the state of the virtual cutting environment synchronized with the real-time state of the physical world.

[0043] Next, the core multiphysics simulation model is activated. This is a complex computational engine that couples thermodynamics, fluid mechanics, solid mechanics, and even plasma physics. Using the synchronized virtual cutting environment as initial conditions, it accelerates its operation on a virtual timeline, simulating the evolution of steel plate deformation, cutting head vibration, and cutting trajectory under the coupled effects of multiple fields such as thermal stress, recoil pressure, and material phase transformation over a future period. This simulation process is a high-precision dynamic extrapolation. The extrapolation does not perform a single prediction but simulates various intervention measures, such as slightly reducing the cutting speed, fine-tuning the focus position, or increasing the auxiliary gas flow rate. By executing multiple sets of hypothesis-simulation experiments in parallel, the system can proactively predict the possible evolution of the cutting path under different control strategies and evaluate the impact of each evolution on indicators such as final cut quality, perpendicularity, and slag amount. All simulation results are summarized into a dynamic extrapolation prediction result, which includes the predicted path, quality score, and potential defect risks corresponding to different control actions at multiple future time points.

[0044] Finally, the optimization decision-making stage begins. Based on comprehensive dynamic simulation and prediction results, it sets clear optimization objectives: maximizing cutting quality (e.g., smooth cut, no slag) and path accuracy (i.e., minimizing the deviation between the actual trajectory and the preset path). By invoking high-performance optimization algorithms in the cloud, such as model predictive control or gradient-based optimizers, all simulated control action sequences are evaluated, compared, and synthesized within the set process constraints. The algorithm calculates a series of time-coherent control commands that optimize the predicted cutting result; this set of commands is called the optimal control action set. This set precisely specifies how much horizontal and vertical displacement the cutting head should adjust and at what speed in each step over several control cycles. Ultimately, this optimal control action set, fully validated and optimized in the virtual cutting environment, is formally encapsulated as a globally optimized correction control strategy that can be deployed and executed, ready to be distributed to edge nodes via the communication network to drive the cutting head in the physical world for precise correction.

[0045] Furthermore, a multiphysics simulation model is initiated to perform high-precision dynamic simulation in the virtual cutting environment, predicting the possible evolution of the cutting path under different intervention measures and generating dynamic simulation prediction results. This includes: in the multiphysics simulation model, using the cutting head position deviation in the preliminary analysis results as the initial disturbance of the current simulation; setting multiple alternative future control input sequences, each sequence representing a possible intervention measure; for each alternative future control input sequence, driving the multiphysics simulation model to run, calculating the change in the motion trajectory of the virtual cutting head within a preset time window, and generating multiple predicted motion trajectories; comparing the differences between the multiple predicted motion trajectories and the virtual preset path to form a path evolution prediction set.

[0046] Specifically, after the multiphysics simulation model is launched on the cloud platform, a core, forward-looking computational process is activated. Its purpose is to accurately predict the future in the virtual segmentation environment constructed by the digital twin and generate dynamic inference and prediction results, providing a data foundation for optimized decision-making.

[0047] This process begins with the precise setting of the initial simulation conditions. The system extracts the cutting head position deviation from the received preliminary analysis results. This deviation is a real-time error measurement in the physical world, a quantitative representation of how the current cutting state deviates from the ideal state. At the start of the simulation, this cutting head position deviation is directly injected into the virtual model as the initial perturbation for the current simulation. This means that the virtual cutting head in the digital twin does not start moving from the ideal position, but rather from an offset position completely consistent with the physical world. This ensures that the starting point of the virtual simulation is strictly aligned with the real state, a crucial prerequisite for high-precision dynamic simulation.

[0048] Next, the system needs to evaluate various possible future correction schemes. To do this, multiple alternative future control input sequences need to be defined in the control strategy space. This is typically done by a strategy generator. Each such sequence represents a possible intervention, specifying in detail the horizontal and vertical speed or position adjustment commands that the virtual cutting head servo system should receive in each control cycle within a preset time window, such as 0.5 seconds in the future. Different sequences may represent different control logics such as aggressive correction, gentle compensation, or early deceleration.

[0049] Subsequently, the system enters a parallel, large-scale simulation phase. For each pre-defined candidate future control input sequence, the system drives the multiphysics simulation model to run independently once. This model couples the calculations of physical processes such as heat conduction, molten pool flow, and material stress and deformation. During the model's operation, it is subjected to disturbances caused by initial position deviations and continuous vibrations of the virtual equipment, while simultaneously executing the input future control sequences. By solving a system of partial differential equations, the model calculates the trajectory changes of the virtual cutting head under the combined action of the physical field and control forces within the entire preset time window. Each complete simulation run outputs a trajectory line containing spatial coordinates and temporal information from the current moment to a future moment, i.e., generating a predicted motion trajectory. Due to the multiple candidate control sequences, multiple predicted motion trajectories are generated in parallel.

[0050] Finally, these prediction results are systematically evaluated and analyzed. The differences between the predicted motion trajectories and the virtual preset paths—that is, the ideal processing trajectories carried in the twin—are compared. The comparison is quantitative, calculating indicators such as positional deviation and cumulative error at various time points for each predicted trajectory and the ideal path. All the results of these comparative analyses, including the complete coordinate sequence of each trajectory, corresponding deviation indicators, and other quality assessment parameters such as the predicted cut width and heat-affected zone size, are systematically organized, correlated, and stored. This data structure, integrating all scenarios and their results, forms the path evolution prediction set. This set is the core component of the dynamic extrapolation prediction results, clearly demonstrating in data form the different futures led by different intervention measures. It provides direct decision-making basis for the cloud-based intelligent agent to calculate the optimal set of control actions and generate globally optimized corrective control strategies.

[0051] Furthermore, based on the dynamic simulation prediction results, with cutting quality and path accuracy as optimization objectives, an optimal set of control actions is calculated and generated, including: defining a target evaluation function that integrates cutting quality and path accuracy; selecting the optimal predicted motion trajectory from the path evolution prediction set that makes the target evaluation function within a preset time window; generating a virtual control input sequence corresponding to the optimal predicted motion trajectory through reverse analysis, and converting the virtual control input sequence into control commands that the physical actuator can respond to, thus forming the optimal set of control actions.

[0052] Specifically, after completing multiphysics simulation in the cloud and generating a set of predicted path evolutions containing various possibilities, the core task of the system shifts to the decision-making and optimization phase. The purpose of this phase is to calculate the optimal solution that can directly guide the actions of physical devices based on the dynamic inference and prediction results, i.e., to generate the optimal set of control actions.

[0053] The entire process begins with the mathematical definition of the optimization objective. The system must transform the abstract process requirements of improving cutting quality and ensuring path accuracy into calculable criteria. To this end, a comprehensive objective evaluation function integrating cutting quality and path accuracy needs to be defined. This function is a mathematical formula whose inputs are the various quantitative indicators contained in the dynamically extrapolated prediction results, and whose output is a single evaluation value (score). For example, the function might use the maximum deviation between the predicted trajectory and the ideal path (representing path accuracy) and the predicted cut width variance (representing cutting quality uniformity) as negative indicators, and the predicted cutting efficiency as a positive indicator, and then perform a weighted sum after assigning different weighting coefficients.

[0054] With clear evaluation criteria, the search for the optimal solution among numerous possibilities can begin. From the path evolution prediction set, for each predicted trajectory, the previously defined objective evaluation function is called to calculate a comprehensive score for each trajectory within a preset future time window. Then, using a sorting or search algorithm, such as quicksort or traversal comparison, the trajectory that maximizes (or minimizes, depending on the function definition) the objective evaluation function value is selected. This trajectory is considered the optimal predicted trajectory. It represents the trajectory that theoretically delivers the best overall processing effect across all simulation scenarios.

[0055] Next, reverse analysis generates the virtual control input sequence corresponding to the optimal predicted motion trajectory. Since each predicted trajectory in the previous simulation was generated by a specific virtual control input sequence (i.e., a set of hypothetical drive commands), the system maintains this strict causal mapping. Now, by querying the internal mapping table or reverse tracing, the unique virtual control input sequence that generated the optimal predicted motion trajectory is found and extracted. This sequence contains the ideal control quantity for each control cycle on the virtual timeline in response to the virtual hair cutting.

[0056] Finally, the ideal instructions from the virtual world are transformed into executable commands in the physical world. The cloud system performs format conversion and quantization on the virtual control input sequence. Specifically, abstract values ​​such as displacement and velocity in the virtual model are converted into control commands that the physical actuators (such as servo motors and linear motors) can respond to, based on the physical characteristics (such as encoder resolution, maximum speed, and acceleration limits) and control interface protocols (such as pulse commands, analog voltages, and EtherCAT bus commands). For example, millimeter-level displacement increments are converted into the number of position command pulses required by the servo driver. Ultimately, this series of converted, time-sequential control commands constitutes the optimal set of control actions to be sent to the edge nodes. It precisely plans how the cutting head should move at each step over a period of time to achieve the optimal processing path predicted and selected by the cloud.

[0057] The aforementioned correction control strategy is sent to the edge node, and combined with local real-time sensor feedback, microsecond-level dynamic compensation of the control command is performed to generate the final drive command, which drives the cutting head to complete adaptive correction.

[0058] Furthermore, the correction control strategy is distributed to the edge node, and combined with local real-time sensor feedback, microsecond-level dynamic compensation of the control command is performed. This also includes: parsing the globally optimized correction control strategy at the edge node to obtain the reference control command for the current and near-future moments; synchronously acquiring higher-frequency real-time sensor feedback from the edge-side servo system; comparing the reference control command with the real-time feedback data to calculate the high-frequency instantaneous deviation; generating a dynamic compensation command based on the high-frequency instantaneous deviation through a fast compensation unit deployed on the edge side; and fusing the reference control command and the dynamic compensation command to generate the final drive command for driving the cutting head servo system.

[0059] It should be understood that once the globally optimized correction control strategy generated in the cloud is successfully distributed to the edge nodes via a highly reliable, low-latency communication network, the system's control closed loop transitions from cloud-edge collaboration to the final edge-end execution phase. The core task of this phase is to combine local real-time sensor feedback to perform microsecond-level dynamic compensation on the cloud strategy, thereby generating the final drive command that can directly and accurately drive the cutting head to complete adaptive correction.

[0060] After receiving the globally optimized correction control strategy from the cloud, the edge node first parses it. This strategy is typically encapsulated as a data packet containing target actions for multiple control cycles within a future time sequence. The parsing module obtains the current control cycle and the baseline control commands for the next few near-future moments. These baseline control commands are ideal control quantities calculated by the cloud based on macroscopic models and future predictions, such as the target position or velocity of the cutting head within the next 10 milliseconds.

[0061] However, the update frequency and delivery latency of cloud-based strategies make it difficult to respond to sudden disturbances on the device's local side, which occur at the tens of microsecond level. To address this, while parsing cloud commands, edge nodes simultaneously acquire higher-frequency real-time sensor feedback from encoders, current sensors, and potentially additional high-frequency inertial measurement units built into the edge-side servo system. This feedback data (such as the motor's actual position, speed, and current) is acquired at frequencies of several kilohertz or even higher, providing high-resolution status information for real-time compensation.

[0062] Next, the core compensation logic of the edge node is activated. It continuously compares the baseline control commands (as target values) extracted from the cloud strategy with the real-time feedback data (as actual values) synchronously collected from the servo system within the same control cycle. By calculating the difference between the two, the system obtains a high-frequency instantaneous deviation. This deviation is a real-time error caused by local factors such as instantaneous equipment vibration, sudden load changes, and transmission backlash, which the cloud strategy model failed to predict or compensate for.

[0063] To suppress this high-frequency transient deviation, a fast compensation unit deployed at the edge node, operating with extremely low latency, is activated. This unit, for example, implements a lightweight control algorithm within a high-speed digital signal processor or field-programmable gate array. Based on the high-frequency transient deviation and a preset compensation law, such as a proportional-derivative control law, this unit generates a dynamic compensation command. The purpose of this command is not to achieve long-term trajectory tracking, but rather to quickly and effectively cancel out the transient disturbance within a single or several microsecond-level control cycle.

[0064] Finally, within each high-frequency local control cycle, command fusion is performed. The baseline control command parsed from the cloud strategy, applicable to the current moment, is vector-fused with the dynamic compensation command just calculated by the local fast compensation unit. Simple superposition or a more intelligent weighted fusion algorithm ensures command coordination. The result of the fusion is the generated final drive command. This command is the culmination of cloud-based global optimization intelligence and edge-based rapid response, adhering to long-term optimal trajectory planning while possessing robustness against instantaneous disturbances. The final drive command is immediately sent to the driver of the cutting head servo system, converted into drive current, and precisely controls the movement of the servo motor, thereby driving the cutting head to achieve precise adaptive correction in physical space, firmly locking the real-time position of the cutting seam within the error range allowed by the preset path.

[0065] Furthermore, a rapid compensation unit deployed at the edge generates a dynamic compensation command based on the high-frequency instantaneous deviation, including: setting a compensation rule based on real-time deviation adjustment in the rapid compensation unit; the compensation rule calculates the required adjustment amount in real time according to the magnitude and trend of the high-frequency instantaneous deviation; and outputting the required adjustment amount in the form of a control signal as the dynamic compensation command.

[0066] Optionally, at edge nodes, the fast compensation unit is the key execution core for achieving microsecond-level dynamic compensation. It is a dedicated hardware module or a deeply optimized software thread designed specifically for achieving extremely low latency control. Its core function is to generate dynamic compensation commands in real time based on the high-frequency instantaneous deviation calculated from the feedback of the local servo system.

[0067] The implementation of this unit begins with the precise setting of the compensation rules. During the design phase of the fast compensation unit, engineers set compensation rules based on real-time deviation adjustment within the unit, according to the dynamic characteristics of the cutting equipment. This typically means embedding a lightweight yet efficient control algorithm into the unit, most commonly in the form of a proportional-derivative controller rule. This rule is not a complex model, but rather a well-defined mathematical mapping function: its input is the real-time deviation, and its output is the compensation amount. Parameters in the rule, such as proportional gain and derivative gain, are pre-tuned through debugging and system identification to ensure that the unit can respond quickly while maintaining stability.

[0068] Once operational, the compensation unit processes data periodically. Within each extremely short control cycle, such as 100 microseconds, it receives the latest high-frequency instantaneous deviation from the comparator. This deviation value contains two key pieces of information: first, the magnitude of the current deviation, i.e., the absolute amount of deviation from the target value; and second, the trend of the deviation, i.e., the rate of change of the current deviation relative to the previous cycle, which can be obtained by calculating the difference between the current deviation and historical deviations. The compensation rule calculates the required adjustment based on these two pieces of information. The specific calculation logic is as follows: the proportional component is directly proportional to the magnitude of the deviation; the larger the deviation, the greater the compensation force. The derivative component is proportional to the trend of the deviation, predicting its future trend and applying a reverse suppressive force when the deviation grows rapidly, acting as a "damper" to prevent system overshoot or oscillation. These two components are added instantaneously, and the result is the adjustment required for the current control cycle to most effectively offset the instantaneous disturbance.

[0069] After the calculations are completed, the adjustments in the digital world need to be converted into executable commands in the physical world. Therefore, the required adjustments are output as control signals via a digital signal processor or dedicated hardware controller. These control signals strictly adhere to servo drive communication protocols, such as standardized instructions for analog voltage signals, pulse / direction signals, or fieldbus messages. This real-time generated signal carrying the compensation instructions is the dynamic compensation instruction. It is used as the direct output of the fast compensation unit and immediately fed into the subsequent instruction fusion stage, combining with the reference control instructions from the cloud to ultimately form the final drive instructions that can accurately drive the actuators. This effectively suppresses high-frequency, random disturbances, ensuring the robustness and accuracy of the cloud-based macro-optimization strategy at the micro-execution level.

[0070] Furthermore, the method also includes adaptive updates of the edge-cloud model:

[0071] On the cloud platform, preliminary analysis results of cutting tasks, final cutting quality reports, and global optimization correction control strategies from different edge nodes are continuously collected and stored as optimization sample packages. According to a preset cycle, the multiphysics simulation model in the digital twin is back-optimized using the optimization sample packages to update its model parameters.

[0072] Specifically, on the cloud platform, this application is a continuously evolving intelligent system. One of its core mechanisms is edge-cloud model adaptive updating. This mechanism aims to enable the system's core model to learn from long-term operational experience and continuously optimize itself, thereby adapting to different materials, equipment wear, and changes in operating conditions, and gradually improving the generation accuracy of its globally optimized correction and control strategies.

[0073] This process begins with the continuous collection and storage of massive amounts of data. The cloud acts as a data hub, recording the preliminary analysis results of each cutting task from different edge nodes, the final cutting quality report after completion (such as quantitative reports on cut quality and dimensional accuracy generated by quality inspection equipment or manual evaluation), and the global optimization correction and control strategies calculated and distributed by the cloud at that time. These three elements constitute a complete state-action-result triple, encapsulated as an optimization sample package. These sample packages are systematically stored in the cloud's historical database according to timestamps and task numbers, providing rich and realistic training data for subsequent model optimization.

[0074] Next, according to a preset cycle, such as daily, weekly, or after a certain number of cutting tasks are completed cumulatively, the model update process is triggered. The system extracts recent optimization sample packages from the historical database. Using these optimization sample packages, the core of the digital twin, namely the multiphysics simulation model, is back-optimized. For example, this can be done using data-driven model correction or parameter identification methods. Specifically, the system uses the preliminary analysis results from the sample package and the global optimization correction control strategy issued at that time as input to the simulation model, driving the model to perform a replay simulation. Then, the simulation prediction results of the model, such as the predicted final cutting path and cut morphology, are compared with the actual results reflected in the final cutting quality report in the sample package. The difference between the two, i.e., the prediction error, is calculated.

[0075] This prediction error is the core signal driving model optimization. The system can employ optimization algorithms, such as gradient descent and Bayesian optimization, to automatically adjust the model parameters within the multiphysics simulation model, aiming to minimize this prediction error. These parameters may include material thermophysical property coefficients, heat transfer boundary condition coefficients, and efficiency parameters of the material removal model. Through iterative calculations, the system finds a new set of model parameters that makes the simulation output closer to historical real data, thus completing the update.

[0076] Through this periodic adaptive update of the edge-cloud model, the digital twin deployed in the cloud gradually evolves from a model built on initial theories and general data into an expert model that has absorbed massive amounts of real-world experience. This ensures that the correction strategies generated by the system become increasingly accurate and reliable over time, enabling the intelligent manufacturing system to make an intelligent leap from static presets to dynamic self-learning.

[0077] In summary, the embodiments of this application have at least the following technical effects:

[0078] This application achieves real-time and accurate monitoring of the cutting seam position of steel plates by deploying a lightweight image recognition network at the edge nodes of the steel plate cutting equipment, combined with industrial vision sensors and three-dimensional spatial perception devices. Based on the synchronous acquisition and correlation analysis of multimodal real-time monitoring data, the cutting seam trajectory can be predicted and compared with the preset cutting path to generate dynamic cutting control commands, ensuring cutting accuracy and quality. Through high-precision dynamic simulation using a digital twin and multiphysics simulation model on a cloud platform, a globally optimized correction control strategy is generated, thereby improving the accuracy and quality of the cutting process, maximizing cutting path accuracy and reducing waste, improving production efficiency and yield, and solving the accuracy deviation problem caused by preset paths and fixed strategies in traditional steel plate cutting.

[0079] This technology achieves the goal of improving the accuracy and quality of steel plate cutting through real-time multimodal monitoring and dynamic adaptive correction of the cutting process.

[0080] Example 2, based on the same inventive concept as the cloud-based intelligent monitoring method for steel plate cutting in the foregoing examples, such as... Figure 2 As shown, this application provides a cloud-based intelligent monitoring system for steel plate cutting production processes, wherein the cloud-based intelligent monitoring system for steel plate cutting production processes includes:

[0081] The monitoring network construction module 11 is used to deploy a lightweight image recognition network at the edge nodes of the steel plate cutting equipment and to collect image data sequences during the cutting process in real time through industrial vision sensors. The data mapping module 12 is used to construct a digital twin of the steel plate cutting on the cloud platform and establish a real-time synchronous mapping relationship between the digital twin and the physical cutting equipment. The position change trend generation module 13 is used to process the image data sequence in real time through the lightweight image recognition network of the edge nodes, identify the position of the cutting head, compare it with the preset cutting path, and generate preliminary analysis results including the position deviation of the cutting head and the cutting state characteristics. The motion prediction module 14 is used to upload the preliminary analysis results to the cloud platform through a high-reliability, low-latency communication network, and perform high-precision dynamic simulation of the current cutting process based on the digital twin to generate a globally optimized correction control strategy. The cutting control module 15 is used to send the correction control strategy to the edge nodes, combine local real-time sensor feedback, perform microsecond-level dynamic compensation of the control commands, generate the final drive command, and drive the cutting head to complete adaptive correction.

[0082] Furthermore, the monitoring network construction module 11 is also used to: construct an image processing model with a multi-layer feature extraction structure, wherein the input of the image processing model is an image data sequence and the output is a cutting head image feature; perform compression and acceleration processing on the image processing model, remove redundant computing structures, and form a lightweight image recognition network suitable for edge computing resources; deploy the lightweight image recognition network on an edge computing device and establish a communication connection with the industrial vision sensor to directly receive and process the image data sequence collected by the industrial vision sensor.

[0083] Furthermore, the position change trend generation module 13 is also used to: extract the current image frame from the image data sequence, input it into the lightweight image recognition network to obtain the image coordinates of the cutting head image in the current image frame; combine the pre-calibrated coordinate transformation relationship to transform the image coordinates of the cutting head image to the physical working coordinate system of the cutting device to obtain the real-time physical coordinates of the cutting head; obtain the cutting path target coordinates corresponding to the current time from the preset cutting path; calculate the difference between the real-time physical coordinates and the cutting path target coordinates to generate the cutting head position deviation; extract the feature vector of the intermediate layer of the lightweight image recognition network as the cutting state feature representing the current cutting scene; and combine the cutting head position deviation and the cutting state feature to form a preliminary analysis result.

[0084] Furthermore, the motion prediction module 14 is also used to: establish a virtual cutting environment in the digital twin that includes material properties, process parameters, and geometric structure, and map the cutting state characteristics in the preliminary analysis results to the input states of the corresponding virtual sensors in the virtual cutting environment; start a multiphysics simulation model, perform high-precision dynamic simulation in the virtual cutting environment, predict the possible evolution of the cutting path under different intervention measures, and generate dynamic simulation prediction results; based on the dynamic simulation prediction results, calculate and generate the optimal set of control actions with cutting quality and path accuracy as optimization objectives, as the global optimization correction control strategy.

[0085] Furthermore, the motion prediction module 14 is also used to: in the multiphysics simulation model, take the cutting head position deviation in the preliminary analysis results as the initial disturbance of the current simulation; set multiple alternative future control input sequences, each sequence representing a possible intervention measure; for each alternative future control input sequence, drive the multiphysics simulation model to run, calculate the change in the motion trajectory of the virtual cutting head within a preset time window, and generate multiple predicted motion trajectories; compare the differences between the multiple predicted motion trajectories and the virtual preset path respectively to form a path evolution prediction set.

[0086] Furthermore, the cutting control module 15 is also used to: define a target evaluation function that integrates cutting quality and path accuracy; select the optimal predicted motion trajectory from the path evolution prediction set that makes the target evaluation function within a preset time window; generate the virtual control input sequence corresponding to the optimal predicted motion trajectory through reverse parsing, and convert the virtual control input sequence into control commands that the physical actuator can respond to, thereby forming the optimal control action set.

[0087] Furthermore, the cutting control module 15 is also used to: parse the globally optimized correction control strategy at the edge node to obtain the reference control commands for the current and near-future moments; synchronously collect higher-frequency real-time sensor feedback from the edge-side servo system; compare the reference control commands with the real-time feedback data to calculate the high-frequency instantaneous deviation; generate dynamic compensation commands based on the high-frequency instantaneous deviation through a fast compensation unit deployed on the edge side; and fuse the reference control commands with the dynamic compensation commands to generate the final drive commands for driving the cutting head servo system.

[0088] Furthermore, the cutting control module 15 is also used to: set a compensation rule based on real-time deviation adjustment in the rapid compensation unit; the compensation rule calculates the required adjustment amount in real time according to the magnitude and trend of the high-frequency instantaneous deviation; and outputs the required adjustment amount in the form of a control signal as the dynamic compensation instruction.

[0089] Furthermore, the cutting control module 15 is also used to: continuously collect and store preliminary analysis results of cutting tasks, final cutting quality reports and global optimization correction control strategies from different edge nodes on the cloud platform as optimization sample packages; and, according to a preset period, use the optimization sample packages to perform reverse optimization on the multiphysics simulation model in the digital twin and update its model parameters.

[0090] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The cloud-based intelligent monitoring method and specific examples for steel plate cutting described in the foregoing embodiment one are also applicable to the cloud-based intelligent monitoring system for steel plate cutting in this embodiment. Through the foregoing detailed description of the cloud-based intelligent monitoring method for steel plate cutting, those skilled in the art can clearly understand the cloud-based intelligent monitoring system for steel plate cutting in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0091] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0092] Obviously, those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A cloud-based intelligent monitoring method for steel plate cutting production processes, characterized in that, The method includes: A lightweight image recognition network is deployed at the edge nodes of the steel plate cutting equipment, and image data sequences during the cutting process are acquired in real time through industrial vision sensors. A digital twin of steel plate cutting is constructed on a cloud platform, and a real-time synchronous mapping relationship is established between the digital twin and the physical cutting equipment; The image data sequence is processed in real time by the lightweight image recognition network of the edge nodes to identify the position of the cutting head and compare it with the preset cutting path to generate preliminary analysis results including the cutting head position deviation and cutting state features. The preliminary analysis results are uploaded to the cloud platform through a highly reliable, low-latency communication network, and based on the digital twin, the current cutting process is dynamically simulated with high precision to generate a globally optimized correction control strategy. The aforementioned correction control strategy is sent to the edge node, and combined with local real-time sensor feedback, microsecond-level dynamic compensation of the control command is performed to generate the final drive command, which drives the cutting head to complete adaptive correction.

2. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 1, characterized in that, Deploying a lightweight image recognition network at the edge nodes of a steel plate cutting equipment includes: An image processing model with a multi-layer feature extraction structure is constructed. The input of the image processing model is an image data sequence, and the output is the cut head image features. The image processing model is compressed and accelerated to remove redundant computational structures, forming a lightweight image recognition network suitable for edge computing resources; The lightweight image recognition network is deployed on an edge computing device and establishes a communication connection with the industrial vision sensor to directly receive and process image data sequences acquired by the industrial vision sensor.

3. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 1, characterized in that, The image data sequence is processed in real time using a lightweight image recognition network at the edge nodes to identify the cutting head position. This position is then compared with a preset cutting path to generate preliminary analysis results, including cutting head position deviation and cutting state features. Extract the current image frame from the image data sequence, input it into the lightweight image recognition network, and obtain the image coordinates of the cut-head image in the current image frame; By combining the pre-calibrated coordinate transformation relationship, the image coordinates of the cutting head image are transformed to the physical working coordinate system of the cutting equipment to obtain the real-time physical coordinates of the cutting head; Obtain the target coordinates of the cutting path corresponding to the current moment from the preset cutting path; Calculate the difference between the real-time physical coordinates and the target coordinates of the cutting path to generate the cutting head position deviation; The feature vectors of the intermediate layer of the lightweight image recognition network are extracted as features representing the cutting state of the current cutting scene; The cutting head position deviation and the cutting state characteristics are combined to form preliminary analysis results.

4. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 3, characterized in that, Based on the digital twin, the current cutting process is dynamically simulated with high precision to generate a globally optimized correction control strategy, including: In the digital twin, a virtual cutting environment containing material properties, process parameters, and geometric structure is established, and the cutting state characteristics in the preliminary analysis results are mapped to the input state of the corresponding virtual sensors in the virtual cutting environment. A multiphysics simulation model is launched, and a high-precision dynamic simulation is performed in the virtual cutting environment to predict the possible evolution of the cutting path under different intervention measures and generate dynamic simulation prediction results. Based on the dynamic simulation and prediction results, with cutting quality and path accuracy as optimization objectives, an optimal set of control actions is calculated and generated as the global optimization correction control strategy.

5. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 4, characterized in that, A multiphysics simulation model is launched, and high-precision dynamic simulation is performed in the virtual cutting environment to predict the possible evolution of the cutting path under different intervention measures, generating dynamic simulation prediction results, including: In the multiphysics simulation model, the cutting head position deviation in the preliminary analysis results is used as the initial perturbation for the current simulation. Multiple alternative future control input sequences are defined, each sequence representing a possible intervention; For each candidate future control input sequence, drive the multiphysics simulation model to run, calculate the change in the motion trajectory of the virtual cutting head within a preset time window, and generate multiple predicted motion trajectories; The differences between the multiple predicted motion trajectories and the virtual preset paths are compared to form a path evolution prediction set.

6. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 5, characterized in that, Based on the dynamic simulation and prediction results, with cutting quality and path accuracy as optimization objectives, an optimal set of control actions is calculated and generated, including: Define a target evaluation function that integrates cutting quality and path accuracy; From the path evolution prediction set, select the optimal predicted trajectory that makes the target evaluation function within a preset time window; The reverse analysis generates the virtual control input sequence corresponding to the optimal predicted motion trajectory, and converts the virtual control input sequence into control commands that the physical actuator can respond to, thus forming the optimal control action set.

7. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 1, characterized in that, The aforementioned deviation correction control strategy is sent to the edge nodes, and combined with local real-time sensor feedback, microsecond-level dynamic compensation of the control commands is performed to generate the final drive command, driving the cutting head to complete adaptive deviation correction, including: The globally optimized correction control strategy is parsed at the edge nodes to obtain the reference control commands for the current and near-future moments; Synchronously acquire higher frequency real-time sensor feedback from the edge-side servo system; The high-frequency instantaneous deviation is calculated by comparing the reference control command with the real-time feedback data. A rapid compensation unit deployed at the edge generates dynamic compensation commands based on the high-frequency instantaneous deviation. The reference control command and the dynamic compensation command are fused together to generate the final drive command for the cutting head servo system.

8. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 7, characterized in that, A fast compensation unit deployed at the edge generates dynamic compensation commands based on the high-frequency instantaneous deviation, including: In the rapid compensation unit, a compensation rule based on real-time deviation adjustment is set; The compensation rule calculates the required adjustment amount in real time based on the magnitude and trend of the high-frequency instantaneous deviation. The required adjustment amount is output as a control signal, serving as the dynamic compensation instruction.

9. The cloud-based intelligent monitoring method for steel plate cutting production process as described in claim 8, characterized in that, The method also includes adaptive updates of the edge-cloud model: On the cloud platform, preliminary analysis results of cutting tasks, final cutting quality reports, and global optimization correction control strategies are continuously collected and stored from different edge nodes as optimization sample packages. According to a preset cycle, the multiphysics simulation model in the digital twin is back-optimized using the optimized sample package, and its model parameters are updated.

10. A cloud-based intelligent monitoring system for steel plate cutting production processes, characterized in that, The steps for implementing the cloud-based intelligent monitoring method for steel plate cutting as described in any one of claims 1 to 9, wherein the cloud-based intelligent monitoring system for steel plate cutting comprises: The monitoring network construction module is used to deploy a lightweight image recognition network at the edge nodes of the steel plate cutting equipment and to collect image data sequences in real time during the cutting process through industrial vision sensors. The data mapping module is used to construct a digital twin of steel plate cutting on the cloud platform and establish a real-time synchronous mapping relationship between the digital twin and the physical cutting equipment; The position change trend generation module is used to process the image data sequence in real time through the lightweight image recognition network of the edge nodes, identify the position of the cutting head, compare it with the preset cutting path, and generate preliminary analysis results including the cutting head position deviation and cutting state features. The motion prediction module is used to upload the preliminary analysis results to the cloud platform through a highly reliable low-latency communication network, and based on the digital twin, to perform high-precision dynamic simulation of the current cutting process and generate a globally optimized correction control strategy. The cutting control module is used to send the correction control strategy to the edge node, combine it with local real-time sensor feedback, perform microsecond-level dynamic compensation of the control command, generate the final drive command, and drive the cutting head to complete adaptive correction.

Images