Wire production control system and method based on visual feedback

Abstract

The application discloses a kind of welding wire production control system and method based on visual feedback, belong to precision machining technical field.System includes multimodal intelligent perception subsystem, adaptive model predictive controller and safety and exception handling subsystem, and quality closed loop and two-stage control architecture of execution closed loop are constructed.Method is through multimodal perception real-time acquisition welding wire surface micro roughness, macro geometry and defect etc multidimensional quality characteristics, and fusion is comprehensive quality index;Using model predictive control algorithm of on-line updating parameter is carried out rolling optimization, generates control instruction;While parallel operation trend early warning based on statistical process control and sudden failure diagnosis based on model residual analysis, realize initiative safety protection.The application solves the quality instability caused by incoming material fluctuation and tool time-varying in welding wire planing processing, control lag and safety passive problem, realizes the adaptive accurate control and intelligent safety protection of processing process.

Description

Technical Field

[0001] This invention relates to the field of precision machining technology, and specifically to a welding wire production control system and method based on visual feedback. Background Technology

[0002] Welding wire is a critical consumable material in welding processes such as arc welding and gas shielded welding. Its surface quality is crucial to welding stability, wire feeding smoothness, and the mechanical properties of the weld. The shaving process is a core step in finishing the surface, aiming to remove oxide scale and other defects to obtain a bright, uniform, and dimensionally accurate finished product. Currently, the mainstream control technologies and limitations of the welding wire shaving process are as follows:

[0003] 1. Open-loop fixed-parameter control scheme: The operator presets the process parameters of the planer according to the welding wire parameters and keeps them unchanged. The disadvantage is that it cannot compensate for fluctuations in the incoming material, and improper cutting depth may damage the welding wire or leave residual defects; it also cannot adapt to changes in equipment status, resulting in deterioration of processing quality over time and poor process stability.

[0004] 2. Single-loop control scheme based on simple sensors: This scheme employs online measurement sensor feedback control, typically by installing a sensor after the cutter head exit to measure the diameter and compare it with a set value to adjust the feed speed or pressure. The drawbacks are: control lag, a "remedial" approach, and limited effectiveness in high-speed production control; limited sensing dimension, measuring only the diameter; and limitations inherent in the sensor itself, such as wear and interference.

[0005] 3. First-generation vision-assisted monitoring solution: Some advanced production lines have introduced industrial camera vision inspection, but it is only used for offline spot checks or end-point macroscopic defect alarms. The drawback is that it is not integrated into the real-time control loop, it only "detects" but does not "control", and cannot intervene in real time; the information utilization is superficial and the value of the data is not fully explored.

[0006] In summary, the fundamental problem with existing technologies is the lack of comprehensive, real-time, and accurate perception and intelligent decision-making control strategies for the multi-dimensional states of the processing. Specifically, this manifests as: a one-sided definition of quality, lacking a comprehensive quantitative index system for "surface quality"; delayed control response, unable to provide preventative adjustments; insufficient system adaptability, leading to long-term performance degradation; and passive safety protection, lacking early warning and intelligent diagnostic decision-making capabilities. Therefore, there is an urgent need in this field for an intelligent control system that integrates advanced sensing technologies, possesses online learning and predictive optimization capabilities, and can achieve proactive safety protection, thereby improving the consistency of welding wire planing process quality, production efficiency, and operational reliability. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention discloses a visual feedback-based welding wire production control system and method, which solves problems such as unstable processing quality, lagging control, and passive safety protection caused by time-varying factors such as fluctuations in incoming material size and material, and tool wear in existing welding wire shaving processing technology.

[0008] This invention is achieved through the following technical solution:

[0009] On the one hand, this invention proposes a welding wire production control system based on visual feedback. The system adopts the overall design principle of "perception-decision-execution-safety" integration, which is specifically reflected in the core architecture of "two-level closed loop and three-layer collaboration".

[0010] Two-stage closed loop:

[0011] Quality closed loop (outer loop): The core controlled variable is the comprehensive surface quality index Q of the welding wire, which is extracted and quantified in real time by a high-speed vision system. This closed loop directly tracks the optimal quality target defined by process requirements, including micro-roughness, macro-geometric accuracy, and defect status.

[0012] The closed-loop (inner loop) execution uses physical quantities (such as position, speed, and pressure) of actuators such as servo motors and proportional valves that drive the planer head as controlled variables. The setpoints are dynamically calculated by the outer loop controller based on deviations from overall quality indicators, ensuring that the executed actions accurately serve the final quality objective.

[0013] Three-tier collaboration:

[0014] Intelligent sensing layer: Composed of a multimodal sensor network, responsible for comprehensively and synchronously collecting information on incoming material status, physical signals during processing, and surface images of finished products.

[0015] Collaborative Decision Layer: Serving as the control center, this layer integrates the Adaptive Model Predictive Controller (AMPC) with the safety and anomaly handling subsystems. The AMPC is responsible for feedforward-feedback optimization decisions based on perceived information and online-updated process models; the safety subsystem operates in parallel, providing comprehensive protection from trend warnings to emergency response.

[0016] Precision execution layer: Composed of servo drive units with high dynamic response and pressure control units, it is responsible for executing control commands issued by the decision-making layer with high precision.

[0017] The subsystem of this invention provides comprehensive status information input for control, specifically including:

[0018] Front-mounted vision measurement unit: Located in front of the cutter head inlet, it employs dual-sided telecentric backlighting and a linear array camera, and uses a sub-pixel edge detection algorithm to measure the instantaneous diameter of the incoming welding wire in real time. and ellipticity This provides a basis for feedforward compensation.

[0019] In-situ visual monitoring unit: This is the core quality inspection unit, located near the processing point. It employs a high frame rate area array camera combined with a coaxial light and low-angle ring light illumination scheme, and runs multiple feature extraction algorithms in parallel.

[0020] Micro-roughness characteristics Extraction: Wavelet packet transform is performed on the surface texture image to characterize texture activity using the energy of specific high-frequency subbands. The calculation formula is as follows:

[0021]

[0022] in, These are wavelet packet coefficients. For scale and subband indexing, For a selected set of high-frequency subbands, For its quantity, It is a dimensionless roughness characteristic value.

[0023] Macroscopic geometric features Extraction: The low-angle light stripe image is processed to extract the centerline of the light stripe and calculate its root mean square deviation from the ideal straight line. The calculation formula is:

[0024]

[0025] in, Let i be the coordinates of the extracted center point. To fit the coordinates of corresponding points on the straight line. For points, The unit is pixels (which can be calibrated as a length unit).

[0026] Defect detection features Extraction: A lightweight convolutional neural network is used to classify and segment image blocks in real time, and output the probability of defects and the area ratio as feature values ​​to quantify the severity of defects.

[0027] Process status sensing unit: integrates force sensor and acoustic emission sensor to measure cutting force in real time. With vibration / acoustic emission signals It is used to indirectly assess the health status of cutting tools and process anomalies.

[0028] The AMPC of this invention is the intelligent decision-making core of the system, realizing the precise mapping from quality objectives to control commands.

[0029] Online updated predictive models: built to control input increments As input, with comprehensive quality indicators This is the output discrete state-space model. Model parameter vector. The system uses a recursive least squares method with a forgetting factor to identify and update data in real time online, adapting to slow time-varying factors such as tool wear.

[0030] Comprehensive quality indicators Defined as a weighted combination of multi-dimensional features:

[0031]

[0032] in, For the target roughness, These are non-negative weighting coefficients.

[0033] The model expression is:

[0034]

[0035] in, Let be the system state vector. For the system matrix, It is noise.

[0036] Rolling optimization and feedback correction: in each control cycle Based on the current state and parameter estimates Solve the following finite-time optimization problem:

[0037]

[0038] in, These are the prediction time domain and the control time domain, respectively. This is the weight matrix. This represents the lower limit for quality and safety. After solving, only the first control increment in the optimal sequence is considered. It acts on the system.

[0039] The subsystem of this invention provides layered, proactive security protection and works in conjunction with AMPC.

[0040] Trend early warning layer (based on SPC): for key variables (such as...) Apply the index-weighted moving average control chart. Calculate the statistic. ,in is the smoothing factor. When Exceeding the control limits indicates a slow deterioration trend in the process. The system issues a warning and triggers AMPC to enter conservative mode or prompts for maintenance.

[0041] Emergency Diagnosis Layer (Based on Residual Analysis): Real-time calculation of quality characteristic observations Compared with model predictions residual The residual energy was monitored using the chi-square test. ,when When the threshold is exceeded, it is quickly identified as a sudden failure such as wire breakage or blade chipping.

[0042] Active constraint layer (SOAinMPC): This layer sets physical safety limits (such as maximum pressure) within the MPC framework. The rolling optimization problem of AMPC is rigidly embedded in the form of inequality constraints to eliminate unsafe operations from the source of decision-making.

[0043] Emergency Execution Layer: When the emergency diagnostic layer confirms a serious fault, it immediately triggers the highest priority emergency shutdown command and locks the equipment.

[0044] On the other hand, the present invention provides a welding wire production control method based on visual feedback, comprising the following steps:

[0045] Step S1: Through the multimodal intelligent sensing subsystem, the diameter and ellipticity information of the incoming welding wire, the multi-dimensional visual image of the surface of the processed welding wire, and the status signal of the cutting process are collected in real time.

[0046] Step S2: Process the acquired visual image to extract micro-roughness features. Macroscopic geometric features and defect detection features ;

[0047] Step S3: Build and update online a system that uses incremental control inputs as input and a comprehensive quality index. The output is an adaptive prediction model;

[0048] Step S4: In each control cycle, minimize With target value Taking the future deviation as the objective and considering the safety operating area constraint, the rolling optimization problem is solved to calculate the optimal control input increment sequence, and the first increment is sent to the actuator.

[0049] Step S5: Parallel execution of safety monitoring, including: providing early warning of process trends based on statistical process control charts, diagnosing sudden failures based on model prediction residuals, and dynamically adjusting control strategies or performing safety operations based on the diagnosis results;

[0050] Step S6: The actuator drives the planer head to move according to the received control command, so as to realize closed-loop precise control and adaptive adjustment of the surface quality of the welding wire.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention directly uses multidimensional visual quality features as controlled variables, and the system is directly responsible for the final processing effect, achieving submicron-level surface quality control. The process capability index CpK is significantly better than traditional methods, ensuring extremely high consistency.

[0053] This invention employs AMPC with online parameter identification, which can track and compensate for time-varying and random disturbances such as tool wear and material fluctuations in real time, automatically maintaining the optimal machining state over a tool life cycle of up to several hours, reducing the need for object changes and manual intervention.

[0054] This invention combines statistical process control, model residual diagnosis, and model predictive control constraint optimization to construct a four-layer linkage safety mechanism of "early warning-constraint-diagnosis-emergency stop," achieving full coverage from preventive maintenance to instantaneous safety response, and elevating safety from passive response to proactive protection.

[0055] This invention significantly reduces scrap rates such as overcutting and undercutting due to its high-quality stability; adaptive optimization reduces unnecessary energy consumption; and predictive maintenance reduces unplanned downtime, thereby improving overall equipment efficiency (OEE) and production cost-effectiveness.

[0056] The system architecture of "multimodal perception + adaptive predictive control + hierarchical security" proposed in this invention is not only applicable to wire cutting, but its methodology can also be extended to other continuous processing scenarios that require precise control of surface quality, such as precision wire drawing and strip rolling, and has good prospects for industrial promotion. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is a schematic diagram of the principle of a welding wire production control system based on visual feedback;

[0059] Figure 2 This is a simplified block diagram of a welding wire production control method based on visual feedback. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Reference Figure 1As shown, this embodiment provides a complete and preferred specific implementation of the technical solution of the present invention, demonstrating the entire process from system construction, algorithm implementation to effect verification.

[0062] This embodiment provides a complete implementation system including a sensing unit, a computing control unit, and an execution unit.

[0063] Sensing unit:

[0064] The front-end dimension measurement module consists of a high-resolution line scan camera (such as the Basler L801k, with a resolution of 4096 pixels) and a set of dual-sided telecentric parallel backlit LED light sources (such as the SICK OT12 series). This module is installed in front of the blade cutter head inlet to acquire a high-contrast projected image of the incoming welding wire.

[0065] The in-situ surface monitoring module consists of two high-frame-rate area array cameras (such as the Basler acA2040) and a matching light source system. One camera, equipped with a coaxial diffuse light source (such as the Schott GEO series), is used to acquire surface texture; the other camera, equipped with a low-angle ring light source (incident angle of approximately 15°), is used to highlight macroscopic surface undulations. The two cameras are aligned at a certain angle with the welding wire surface near the cutter head exit.

[0066] A high frame rate area scan camera combined with a coaxial light and low-angle ring light illumination scheme is used to run multiple feature extraction algorithms in parallel:

[0067] Micro-roughness characteristics Extraction: Wavelet packet transform is performed on the surface texture image to characterize texture activity using the energy of specific high-frequency subbands. The calculation formula is as follows:

[0068]

[0069] in, These are wavelet packet coefficients. For scale and subband indexing, For a selected set of high-frequency subbands, For its quantity, It is a dimensionless roughness characteristic value.

[0070] Macroscopic geometric features Extraction: The low-angle light stripe image is processed to extract the centerline of the light stripe and calculate its root mean square deviation from the ideal straight line. The calculation formula is:

[0071]

[0072] in, Let i be the coordinates of the extracted center point. To fit the coordinates of corresponding points on the straight line. For points, The unit is pixels (which can be calibrated as a length unit).

[0073] Defect detection features Extraction: A lightweight convolutional neural network is used to classify and segment image blocks in real time, and output the probability of defects and the area ratio as feature values ​​to quantify the severity of defects.

[0074] Process status sensing unit: integrates force sensor and acoustic emission sensor to measure cutting force in real time. With vibration / acoustic emission signals It is used to indirectly assess the health status of cutting tools and process anomalies.

[0075] The AMPC of this invention is the intelligent decision-making core of the system, realizing the precise mapping from quality objectives to control commands.

[0076] Online updated predictive models: built to control input increments As input, with comprehensive quality indicators This is the output discrete state-space model. Model parameter vector. The system uses a recursive least squares method with a forgetting factor to identify and update data in real time online, adapting to slow time-varying factors such as tool wear.

[0077] Comprehensive quality indicators Defined as a weighted combination of multi-dimensional features:

[0078]

[0079] in, For the target roughness, These are non-negative weighting coefficients.

[0080] The model expression is:

[0081]

[0082] in, Let be the system state vector. For the system matrix, It is noise.

[0083] Rolling optimization and feedback correction: in each control cycle Based on the current state and parameter estimates Solve the following finite-time optimization problem:

[0084]

[0085] in, These are the prediction time domain and the control time domain, respectively. This is the weight matrix. This represents the lower limit for quality and safety. After solving, only the first control increment in the optimal sequence is considered. It acts on the system.

[0086] Process status sensing module: including a rotary piezoelectric force gauge (such as Kistler 9129AA) integrated in the tool holder for measuring cutting force; and an acoustic emission sensor (such as Physical Acoustics Nano30) mounted on the machine tool bed for acquiring high-frequency vibration signals during machining.

[0087] Computational Control Unit: This unit employs an industrial control computer equipped with a multi-core processor (such as an Intel Core i7 series), ample memory, and an interface card supporting real-time communication. It runs the core control algorithm.

[0088] The execution unit includes a servo motor drive system (such as the Yaskawa Σ-7 series) for precise control of the wire feeding speed, a pneumatic / hydraulic cylinder (such as the Festo VPPM series) controlled by a proportional valve for adjusting the cutter head pressure, and a spindle servo motor for driving the cutter head.

[0089] All hardware is synchronized and communicated via high-speed industrial networks (such as EtherCAT) to ensure the timing accuracy of data acquisition and control.

[0090] This embodiment's subsystem provides layered, proactive security protection and works in conjunction with AMPC.

[0091] Trend early warning layer (based on SPC): for key variables (such as...) Apply the index-weighted moving average control chart. Calculate the statistic. ,in is the smoothing factor. When Exceeding the control limits indicates a slow deterioration trend in the process. The system issues a warning and triggers AMPC to enter conservative mode or prompts for maintenance.

[0092] Emergency Diagnosis Layer (Based on Residual Analysis): Real-time calculation of quality characteristic observations Compared with model predictions residual The residual energy was monitored using the chi-square test. ,when When the threshold is exceeded, it is quickly identified as a sudden failure such as wire breakage or blade chipping.

[0093] Active constraint layer (SOAinMPC): This layer sets physical safety limits (such as maximum pressure) within the MPC framework. The rolling optimization problem of AMPC is rigidly embedded in the form of inequality constraints to eliminate unsafe operations from the source of decision-making.

[0094] Emergency Execution Layer: When the emergency diagnostic layer confirms a serious fault, it immediately triggers the highest priority emergency shutdown command and locks the equipment.

[0095] In one embodiment, reference is made to Figure 2 As shown, the system software is developed on a Robot Operating System (ROS) or similar framework, achieving modularity. The core workflow is as follows:

[0096] Step S1: Synchronous Acquisition and Preprocessing of Multimodal Data

[0097] The system synchronously triggers all sensors with a fixed control cycle (e.g., 10 milliseconds). Linear array images are used to calculate incoming material dimensions; two area array cameras synchronously capture surface images; analog signals from force sensors and acoustic emission sensors are synchronously sampled and digitized.

[0098] Step S2: Real-time extraction of multi-dimensional quality features

[0099] This step extracts quantized features from the collected raw data:

[0100] Incoming material size feature extraction: Subpixel edge detection is performed on the linear array image to calculate the wire width of each row, and then the real-time diameter is obtained statistically. and ellipticity Specifically, ellipticity can be quantified by calculating the ratio of the difference between the maximum and minimum diameters over a period of time to the average diameter.

[0101] Surface micro-roughness characteristics ( Extraction: For surface texture images under coaxial light, the region of interest is first defined. Then, a multi-scale wavelet packet transform is performed on this region. The micro-roughness of the surface is characterized by calculating the energy sum of specific high-frequency subband coefficients and mapping it to a logarithmic scale eigenvalue. The formula for calculating this eigenvalue can be expressed as:

[0102]

[0103] in, This indicates that the coordinates in the k-th selected high-frequency sub-band are... wavelet packet coefficients, It is the set of selected high-frequency subbands. This represents the number of subbands in the set.

[0104] Surface macroscopic geometric features Extraction: For low-angle light stripe images, a light stripe center extraction algorithm (such as the Steger algorithm) is used to obtain the sub-pixel position sequence of the light stripe center. An ideal center line is obtained through linear fitting, and the root mean square error of the vertical distance from each actual center point to the fitted line is calculated as a feature value. The calculation formula is:

[0105]

[0106] in, It is the distance from the i-th center point to the fitted line. It is the total number of center points.

[0107] Surface defect characteristics ( Extraction: Using a pre-trained lightweight convolutional neural network model, the surface image is analyzed and segmented in real time, and the probability of the presence of defects and the pixel area ratio of the defect area are output as quantitative features of defects.

[0108] Process state feature extraction: Filtering and feature extraction (such as calculating the root mean square value) of cutting force and acoustic emission signals are used to indirectly evaluate tool state and process stability.

[0109] Step S3: Comprehensive Quality Assessment and Adaptive Prediction Model Update

[0110] Constructing a comprehensive quality index (Q): This involves fusing the multiple features extracted above into a single scalar index to comprehensively evaluate individual quality points. This index is the weighted sum of squares of the deviations from each feature, for example:

[0111]

[0112] in, It is the desired target value for micro-roughness. , , These are weighting coefficients set based on the importance of the process.

[0113] Online update process prediction model: A dynamic mathematical model (e.g., a state-space model) is established, taking control inputs (such as feed rate v, tool head pressure P, and tool head speed ω) as inputs and a comprehensive quality index Q as output. The parameters of this model are considered time-varying. A recursive least squares method with a forgetting factor is used to update the model parameters online in real time using the latest input and output data streams, enabling the model to adapt to slow time-varying disturbances such as tool wear and material changes.

[0114] Step S4: Model-based rolling optimization control

[0115] In each control cycle, perform the following operations:

[0116] Based on the latest updated process model and the current system state, predict the trajectory of the overall quality index Q under different control input sequences in the future time domain (prediction time domain).

[0117] Solving a finite-time optimization problem involves finding the optimal future control input sequence such that the predicted mass trajectory is as close as possible to the target value, while ensuring that the control variables and their rates of change are not excessive, and satisfying all physical and safety constraints (such as actuator limits, mass safety lower limits, etc.). The mathematical description of this optimization problem typically includes an objective function and constraints.

[0118] The first control increment in the optimal control sequence obtained by solving is immediately applied to the actuator (servo motor, proportional valve, etc.).

[0119] Step S5: Parallel security monitoring and anomaly handling

[0120] Trend early warning (based on statistical process control): for key quality characteristics (such as...) The system monitors time-series data of process variables (such as cutting force) using statistical methods such as exponentially weighted moving average control charts. When the monitored statistics exceed the preset control limits, it indicates that the process may be slowly deteriorating (such as gradual tool wear). The system issues an early warning and can switch the controller to a more conservative operating mode.

[0121] Sudden Fault Diagnosis (Based on Model Residual Analysis): The actual measured values ​​of key quality characteristics are compared with the predicted values ​​output by the internal prediction model in step S3 in real time to obtain the residual sequence. When the residual energy increases significantly in a short period of time and exceeds the threshold obtained from the statistical analysis of historical normal data, it is quickly determined that a sudden fault has occurred (such as wire breakage or tool breakage), and the highest level emergency stop command is triggered.

[0122] Safety constraint embedding: In the rolling optimization problem in step S4, the safe operating range of the actuator and process is directly embedded in the form of inequality constraints to ensure production safety from the source of control commands.

[0123] Step S6: Closed-loop execution and data recording. The actuator responds to the control command and drives the cutter head to complete the machining. Simultaneously, the system continuously records all sensor data, feature values, control commands, and model parameters for process traceability, performance analysis, and continuous model learning.

[0124] 3. Implementation Effect Verification: In an 8-hour continuous production experiment, the above system was used to process ER50-6 welding wire. Experimental results show that compared with traditional open-loop control and PID control based on single diameter feedback, this system exhibits significant advantages:

[0125] Quality stability: The fluctuation range of surface roughness Ra was significantly reduced, and the process capability index Cpk increased to 1.58, indicating that the process is highly stable and under control.

[0126] Adaptive capability: During the tool wear cycle, the system automatically adjusts the pressure parameters to keep the machining quality stable, without the quality degradation that occurs with tool wear in traditional methods.

[0127] Safety and reliability: Successfully detected all man-made simulated wire breakage faults with no false alarms, achieving effective safety protection from trend warning to emergency shutdown.

[0128] Overall benefits: The defect rate of overcut and undercut products is reduced to below 1%, while the energy consumption per unit length of processing is reduced.

[0129] In one embodiment, the process adaptation of aluminum-magnesium alloy soft welding wire is implemented. This embodiment demonstrates that the system can adapt to the processing of welding wires with different material properties by adjusting parameters, reflecting its flexibility and versatility.

[0130] Scenario characteristics: Aluminum-magnesium alloy welding wire is soft and highly viscous, making it prone to built-up edge during processing. It is sensitive to control parameters and has high requirements for surface brightness and defect control.

[0131] Adjustments to be implemented:

[0132] Sensing parameter adjustment: Enhance coaxial light intensity to better capture subtle textures and adhesions on soft metal surfaces; adjust the training set of the defect detection neural network and increase typical defect samples of aluminum-magnesium alloys.

[0133] Control parameter adjustment:

[0134] In terms of comprehensive quality indicators In the weight configuration, the defect features should be appropriately increased. weight This reflects a higher requirement for surface cleanliness.

[0135] Initialize or load the parameter set of the process prediction model for the properties of soft materials.

[0136] In rolling optimization constraints, stricter upper limits for cutter head pressure and feed rate ranges are set to prevent excessive material deformation or tearing.

[0137] Safety monitoring adjustments: Incorporate the spectral characteristics of cutting force signals into trend warning analysis, because the changing patterns of cutting force in soft material processing may differ from those in steel.

[0138] Implementation results: After the above adjustments, the system can effectively suppress the generation of built-up edge, and while ensuring dimensional accuracy, it can achieve a surface finish that meets high-end requirements, proving the system's strong adaptability to different processes.

[0139] In one embodiment, an extended implementation integrated into a digital workshop is presented, demonstrating the application model of the system as an intelligent production unit integrated with a factory-level information system.

[0140] Scenario Description: The system is deployed in a digital welding wire production workshop and needs to interact with manufacturing execution systems, material management systems, etc.

[0141] Implementation method:

[0142] Data uplink interface: The system uploads key data to the workshop MES in real time via standard protocols (such as OPC UA), including: real-time comprehensive quality indicators. Feature extraction values ​​( , The data includes process status (average cutting force), equipment health indicators (such as SPC-based early warning levels), and the changing trends of model parameters obtained through self-learning.

[0143] Command and Recipe Downlink Interface: The MES sends production orders to the system, which include product specifications, material codes, and quality grade codes. Based on these codes, the system automatically retrieves the corresponding process recipe from a local or cloud database. This recipe includes: parameters for the feature extraction algorithm (such as ROI location and light intensity settings), and weighting coefficients for comprehensive quality indicators. ), the initial model parameters of the controller, the set of constraints for the optimization problem, etc.

[0144] Collaboration and Optimization: Based on uploaded real-time quality and equipment health data, MES can optimize production rhythm and trace quality across the entire workshop. When the system warns that a tool needs to be replaced, MES can automatically schedule the logistics system to prepare the tool and plan equipment maintenance windows.

[0145] Implementation Results: The system has transformed from an independent intelligent controller into a key node in the cyber-physical system of a digital factory. It not only achieves high-quality local closed-loop control but also provides precise data support and configurable execution capabilities for flexible production scheduling, predictive maintenance, and global quality optimization at the shop floor level, thereby improving overall production efficiency and management level.

[0146] In summary, by directly using multidimensional visual quality features as controlled variables, the system is directly responsible for the final processing effect, achieving submicron-level surface quality control. The process capability index CpK is significantly better than traditional methods, ensuring extremely high consistency.

[0147] This invention employs AMPC with online parameter identification, which can track and compensate for time-varying and random disturbances such as tool wear and material fluctuations in real time, automatically maintaining the optimal machining state over a tool life cycle of up to several hours, reducing the need for object changes and manual intervention.

[0148] This invention combines statistical process control, model residual diagnosis, and model predictive control constraint optimization to construct a four-layer linkage safety mechanism of "early warning-constraint-diagnosis-emergency stop," achieving full coverage from preventive maintenance to instantaneous safety response, and elevating safety from passive response to proactive protection.

[0149] This invention significantly reduces scrap rates such as overcutting and undercutting due to its high-quality stability; adaptive optimization reduces unnecessary energy consumption; and predictive maintenance reduces unplanned downtime, thereby improving overall equipment efficiency (OEE) and production cost-effectiveness.

[0150] The system architecture of "multimodal perception + adaptive predictive control + hierarchical security" proposed in this invention is not only applicable to wire cutting, but its methodology can also be extended to other continuous processing scenarios that require precise control of surface quality, such as precision wire drawing and strip rolling, and has good prospects for industrial promotion.

[0151] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A visual feedback-based welding wire production control system, applied to the welding wire planing process, characterized in that, The system includes a multimodal intelligent perception subsystem, an adaptive model prediction controller, and a security and anomaly handling subsystem; the system adopts a two-level closed-loop architecture, wherein: The quality closed loop is the outer loop, and the multi-dimensional quality characteristics of the welding wire surface acquired in real time by the multi-modal intelligent sensing subsystem are used as the controlled variable, and its set value is the optimal quality target required by the process. The execution closed loop is the inner loop, and the physical quantity of the actuator that drives the planer head is used as the controlled variable. Its set value is output by the adaptive model predictive controller based on the feedback of the quality closed loop. The adaptive model predictive controller integrates a controlled object prediction model containing time-varying parameters and uses a recursive least squares method with a forgetting factor to update the model parameters online. In each control cycle k, the controller solves a finite-time rolling optimization problem to minimize the tracking error of the comprehensive quality index Q and the change of the control input, and outputs the optimal control increment command. The solution to the rolling optimization problem satisfies control input constraints, input increment constraints, and quality and safety constraints. The safety and anomaly handling subsystem operates in parallel with the adaptive model prediction controller, implementing trend early warning based on statistical process control, sudden fault diagnosis based on model prediction residual analysis, and active protection by embedding safe operating area constraints in the rolling optimization problem. The controlled object prediction model is described in discrete state-space form: in, For discrete-time indexing; Let k be the system state vector at time k, containing internal variables related to processing quality and equipment status; Let be the control increment vector at time k, and its components represent the feed rate increment, the tool head pressure increment, and the tool head speed increment, respectively. This is the time-varying parameter vector at time k, reflecting the dynamic characteristics of the object model; , , A system matrix with corresponding dimensions; The output of the comprehensive quality index at time k; and These represent process noise and measurement noise, respectively. The comprehensive quality indicators It is composed of the weighted average of the aforementioned multi-dimensional quality features, and is defined as follows: in, , , These represent the micro-roughness characteristic value, macro-geometric characteristic value, and defect detection characteristic value at time k, respectively. The target set value for micro-roughness; , , These are non-negative coefficients used to balance the weights of each feature term; The rolling optimization problem is described as follows: in, This represents the control input sequence to be optimized starting from time k; To predict the length of the time domain; To control the length of the time domain; This represents the predicted value of the comprehensive quality index at time k+i from time k. Set a value for its target; This represents the predicted value of the control input at time k+j at time k; This indicates the corresponding control input increment; , , This is the weight matrix for the corresponding terms; The parameter estimates obtained from online identification at time k; , , , These are the lower and upper limits of the control input and its increment, respectively; This represents the lower limit of safety for comprehensive quality indicators.

2. The system according to claim 1, characterized in that, The multimodal intelligent sensing subsystem includes: a front-mounted vision measurement unit, positioned in front of the blade cutter head inlet, equipped with dual-sided telecentric backlighting and a linear array camera, used to measure the real-time diameter of the incoming welding wire. and ellipticity ;in, Represents a time variable. This represents the measured diameter of the incoming material at time t. This represents the calculated ellipticity of the incoming material at time t; The in-situ visual monitoring unit, located in the machining area of ​​the planer head, is equipped with a high frame rate area array camera and a combined illumination source to extract the micro-roughness features of the weld wire surface after machining. Macroscopic geometric features and defect detection features ; The process state sensing unit integrates a force sensor and an acoustic emission sensor for measuring cutting force. With vibration signal And assess the tool wear condition; among which, This represents the measured cutting force at time t. This represents the measured value of the acoustic emission signal at time t.

3. The system according to claim 2, characterized in that, The in-situ visual monitoring unit extracts micro-roughness features. The method is as follows: perform wavelet packet transform on the acquired surface image, and calculate the energy of a specific high-frequency sub-band, the expression of which is: in, This indicates that the wavelet packet coefficient matrix located at scale j and frequency subband k is in the range of... The coefficient value of the location; This represents the total energy of the subband; The number of selected high-frequency sub-bands; The calculated micro-roughness characteristic value.

4. The system according to claim 2, characterized in that, The in-situ visual monitoring unit extracts macroscopic geometric features. The method is as follows: Extract the centerline of the light stripe from the low-angle illumination image and perform curve fitting; calculate the root mean square deviation of this deviation from the ideal straight line; the expression is: in, Here are the vertical coordinates of the center point of the i-th light stripe extracted from the image; The vertical coordinate value corresponding to the i-th point on the ideal straight line obtained by fitting; The total number of center points involved in the calculation; These are the calculated macroscopic geometric eigenvalues.

5. The system according to claim 1, characterized in that, In the safety and anomaly handling subsystem, the trend early warning based on statistical process control specifically refers to: early warning of key quality characteristics or process parameters. The statistical properties of the applied index-weighted moving average control chart. The calculation formula is ,in As a smoothing factor, When the statistic If the preset control limit is exceeded, the system triggers an early warning signal, and the adaptive model predictive controller switches to conservative control mode or generates a maintenance prompt based on this signal.

6. The system according to claim 1, characterized in that, In the aforementioned safety and anomaly handling subsystem, the sudden fault diagnosis based on model prediction residual analysis specifically involves: real-time calculation of observed values ​​of key quality characteristics. The predicted value output by the prediction model of the controlled object residuals between ; Calculate the energy value of the residual within a time window. ;when When the fault threshold preset based on the chi-square test is exceeded, a sudden fault is determined to have occurred, and an emergency shutdown command is immediately triggered.

7. A visual feedback-based method for controlling the production of welding wire, applied to the welding wire planing process, characterized in that, The method is executed based on the visual feedback-based welding wire production control system according to any one of claims 1 to 6, and includes the following steps: Step S1: Through the multimodal intelligent sensing subsystem, the diameter and ellipticity information of the incoming welding wire, the multi-dimensional visual image of the surface of the processed welding wire, and the status signal of the cutting process are collected in real time. Step S2: Process the acquired visual image to extract micro-roughness features. Macroscopic geometric features and defect detection features ; Step S3: Build and update online a system that uses incremental control inputs as input and a comprehensive quality index. The output is an adaptive prediction model; Step S4: In each control cycle, minimize With target value Taking the future deviation as the objective and considering the safety operating area constraint, the rolling optimization problem is solved to calculate the optimal control input increment sequence, and the first increment is sent to the actuator. Step S5: Parallel execution of safety monitoring, including: providing early warning of process trends based on statistical process control charts, diagnosing sudden failures based on model prediction residuals, and dynamically adjusting control strategies or performing safety operations based on the diagnosis results; Step S6: The actuator drives the planer head to move according to the received control command, so as to realize closed-loop precise control and adaptive adjustment of the surface quality of the welding wire.

Images