A cloud-based self-learning-based automatic welding pool dual-closed-loop collaborative control method
By adopting a cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method, the problems of insufficient CPU visual computing power and difficulty in data traceability during the welding process are solved, achieving high-precision welding pool control and efficient welding production, and meeting the quality traceability requirements of special equipment.
Patent Information
- Application Number
- CN202611238404.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-15
- Publication Date
- 2026-11-10
AI Technical Summary
Existing welding technologies suffer from problems such as insufficient CPU visual computing power, image transmission frame loss, uneven weld formation quality, and difficulty in data traceability, which cannot meet the requirements for high-precision and high-efficiency welding.
An automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning is adopted. Through the collaborative work of hardware units and remote cloud AI training cluster, the image is directly connected to the NPU module, an independent bandwidth channel and a parallel execution link are realized. Combined with the 5G communication module, real-time welding control and asynchronous model updates are performed, enabling rapid identification of weld pool characteristics and synchronous adjustment of electrical parameters.
It achieves high-precision molten pool control, reduces molten pool fluctuations and weld defects, improves production efficiency and data traceability capabilities, and meets the quality traceability requirements of special equipment.
Smart Images

Figure CN123060200A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent manufacturing automated welding technology, and in particular to an automated welding pool dual-closed-loop collaborative control method based on cloud self-learning. Background Technology
[0002] Welding, as a core fundamental process in equipment manufacturing, is widely used in key industries such as special equipment, rail transportation, automobile manufacturing, and heavy engineering machinery. The quality of weld formation directly determines the structural strength and service safety of components. With the deepening of Industry 4.0 and intelligent manufacturing, the industry has placed higher demands on the control precision, production efficiency, and quality traceability capabilities of the welding process. The existing architecture of "centralized visual processing by industrial control computers + post-processing data analysis in the cloud" has four inherent technical bottlenecks:
[0003] Insufficient CPU visual computing power and frame loss in image transmission: The molten pool image is transmitted through the industrial control relay, and the bus is congested and frame loss is easy; the CPU serial processing of the image has a single frame delay of 100~300ms, and the molten pool size detection is lagging and distorted, which cannot meet the high-precision tracking requirements of ±0.1mm for ship welding.
[0004] Two control loops are executed in steps, causing molten pool oscillation: the weld visual tracking loop and the arc current and voltage regulation loop operate independently and in time. When the workpiece simultaneously exhibits weld deviation and abnormal molten pool width, two control cycles are required for step-by-step correction, which leads to severe fluctuations in the molten pool, uneven weld width, and a high defect rate in multi-layer and multi-pass welding.
[0005] Fragmented quality traceability data: There is no unified index for images, electrical parameters, and dimensional time-series data, making it difficult to retrieve welding data for special equipment throughout the entire process, and failing to meet the industry's digital quality traceability compliance requirements. Summary of the Invention
[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method. Its purpose is to solve problems such as CPU visual computing power bottleneck, image transmission frame loss, step-by-step adjustment of visual electrical parameter control loop, and poor model generalization.
[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method, comprising a local hardware unit and a remote cloud AI training cluster; the local hardware unit includes: a high-speed welding pool camera equipped with a composite optical lens, an independent embedded NPU hardware acceleration module, an industrial control computer with an integrated 5G communication module, a PLC controller, an electrical parameter acquisition module, a welding power supply, and a two-dimensional stepping fine-tuning slide.
[0008] Hardware communication links are divided into four types of independent bandwidth isolated channels:
[0009] (1) Image acquisition channel: The melt pool camera is directly connected to the NPU module, and the raw image is not forwarded by the industrial control computer;
[0010] (2) Local feature control channel: The NPU is connected to the industrial control computer via industrial Ethernet and only outputs the melt pool quantization feature value to the industrial control computer;
[0011] (3) Cloud-based interaction channel: The industrial control computer relies on the 5G module to conduct bidirectional encrypted communication with the cloud-based AI training cluster;
[0012] (4) Bottom execution channel: The industrial computer and PLC communicate bidirectionally. The PLC is connected to the two-dimensional stepper fine-tuning slide, the electrical parameter acquisition module and the welding power supply respectively.
[0013] The method describes two independent execution links that do not compete for bandwidth: a local real-time welding dual-closed-loop collaborative control main line and a 5G cloud-based asynchronous AI self-learning iterative auxiliary line.
[0014] The local real-time welding dual closed-loop collaborative control main line executes S1~S6 in a loop:
[0015] S1: Pre-welding initialization: retrieve the reference weld pool width W0, ±0.1mm offset threshold, reference current I0, and reference voltage U0 from the standard process library and send them to the NPU and PLC. Complete the NPU model verification before arc ignition.
[0016] S2: The molten pool camera acquires images at 60~200FPS and pushes them unidirectionally to the NPU cache. The industrial control computer asynchronously stores the original images for cloud upload after shutdown.
[0017] S3: The NPU hardware performs four levels of parallel computation: hardware image preprocessing, lightweight network inference to extract the actual width W of the melt pool, horizontal offset ΔX, vertical offset ΔY, 60-frame mean filtering to smooth features, and quantization output ΔW=W-W0, which is only transmitted to the industrial control system.
[0018] S4: The industrial control computer synchronously generates two types of instructions: two-dimensional slide table position compensation pulse and arc current voltage adjustment increment, and sends them to the PLC simultaneously.
[0019] S5: The PLC operates two closed loops synchronously within the same hardware clock cycle: the vision position closed loop drives the slide to stably control ΔX within ±0.1mm; the electrical parameter closed loop synchronously acquires I and U at 1kHz, and couples ΔW to adjust the welding power supply; when ΔW>0, I and U are lowered, and when ΔW<0, I and U are raised.
[0020] S6: The industrial control system locally caches the original image of the molten pool, characteristic timing, IU arc curve and operating condition label, and cycles through S2~S6 until the arc is extinguished;
[0021] The 5G cloud-based asynchronous AI self-learning iterative auxiliary line silently executes T1~T5 only during downtime:
[0022] T1: After shutdown, the welding dataset is bound to a unique workpiece code and encrypted. It is then uploaded to the cloud via 5G breakpoint resume transmission. The cloud establishes an image, electrical parameter, and dimension timing association index.
[0023] T2: Cloud-based collection of heterogeneous samples from multiple plants for parallel training of location recognition network, electrical parameter mapping network, and four-dimensional coupled adaptive network; loss function incorporates thermal deformation and melt pool size constraints; updates global process library.
[0024] T3: The cloud performs quantization pruning and lightweighting on the training model, and integrates the lightweight model package and the new version of the global process library into an encrypted distribution file;
[0025] T4: The cloud 5G sends encrypted files to the industrial control system, which verifies them and then sends a model switching command; the NPU is divided into two independent model cache partitions, the new model is loaded in the background, and the new model is switched for inference in the next frame of the image, without any downtime.
[0026] T5: The updated lightweight model is integrated into the S3 solution process. The system automatically outputs the optimal compensation amount and target electrical parameters, and iterates and optimizes continuously by looping from T1 to T5.
[0027] The entire production line's welding data is permanently stored in the cloud. Authorized users can remotely retrieve complete welding data for a single piece by entering the workpiece code, enabling quality traceability for special equipment.
[0028] Furthermore, the two-dimensional stepper fine-tuning slide adopts a ball screw and a closed-loop stepper motor, with a minimum single-axis compensation unit of 0.1mm and a hardware response time of ≤18ms. When the NPU synchronous output exceeds the standard ΔX and ΔW, the PLC outputs position compensation pulses and power adjustment signals in parallel during the same control cycle. The two corrections take effect synchronously, eliminating defects such as molten pool oscillation and uneven weld width caused by step adjustment.
[0029] Furthermore, the composite optical lens is composed of a 940nm narrowband infrared filter and a gradient neutral density filter to filter out interference from electric arc light, spatter, and smoke, and to stably extract the molten pool profile and bevel baseline.
[0030] Furthermore, the NPU hardware parallel computation is compatible with traditional machine vision algorithms; edge detection and threshold segmentation algorithm parameters are synchronously iterated in the cloud, and then distributed to the NPU after being lightweighted, so that melt pool feature extraction can be completed independently when there is no deep learning model.
[0031] Furthermore, a wire feeding mechanism is added to the underlying execution unit; the PLC adjusts the wire feeding speed in real time based on ΔW, forming a three-closed-loop coordinated control of position, electrical parameters, and wire feeding speed, with hardware links and dual parallel execution chains.
[0032] Furthermore, the cloud-based AI training cluster collects heterogeneous samples from multiple plants within the group to build a global welding sample library. The three-layer neural network is trained in parallel to improve the model recognition and generalization accuracy under conditions of bevel fluctuation, thermal deformation, and dust interference, thereby reducing the frequency of on-site manual process adjustments.
[0033] Furthermore, an automatic weld pool adaptive control system is provided for executing the control method described in any one of claims 1 to 5, comprising: an image acquisition module, an NPU hardware acceleration module, an industrial control scheduling module, a low-level execution module, and a remote cloud AI training cluster; the image acquisition module is a high-speed weld pool camera equipped with a composite optical lens; the NPU hardware acceleration module has built-in dual independent model inference cache partitions for hardware parallel weld pool image preprocessing and feature calculation; the industrial control scheduling module integrates a 5G communication module and local solid-state storage; the low-level execution module includes a PLC, a two-dimensional stepping fine-tuning slide, a 1kHz electrical parameter acquisition module, a welding power supply, and an optional wire feeding mechanism; the remote cloud AI training cluster is used for sample collection, parallel training of a three-layer neural network, model lightweighting, remote distribution, and permanent traceability storage of welding data.
[0034] Furthermore, a computer-readable storage medium storing a computer-executable program is provided, characterized in that, when the program is run by a processor, it implements the cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method as described in any one of claims 1 to 6.
[0035] The beneficial effects of this invention are as follows:
[0036] This invention presents a cloud-based self-learning-based automatic welding pool dual-closed-loop collaborative control method. Direct NPU hardware acquisition eliminates image frame drops and significantly reduces detection latency. Direct connection between the welding pool camera and the NPU eliminates the need for an industrial control relay link, preventing bus congestion and image frame drops. The NPU performs full hardware parallel processing, with single-frame image processing time ≤20ms, compared to 100~300ms for traditional CPU solutions. Combined with a composite infrared optical lens, welding pool recognition is stable under harsh conditions of smoke and strong light, and long weld seam offset is stably controlled within ±0.1mm, completely resolving continuous weld deviation defects. The industrial control fully releases image processing computing power, supporting 24-hour uninterrupted continuous production without lag or crashes.
[0037] This invention discloses a cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method. The PLC synchronous dual-closed-loop coupling significantly reduces weld pool fluctuations and welding defects. The same hardware clock cycle outputs two signals in parallel: position correction and power adjustment, enabling dual compensation for composite anomalies in a single operation. The weld pool width fluctuation is reduced by 70% compared to traditional step-by-step adjustment, resulting in uniform multi-layer, multi-pass weld formation. Actual welding defect rejection rate is reduced by over 85%, and the welding defect rate of thick-plate pressure vessels in ships is reduced from 12.7% to 1.78%.
[0038] This invention presents a cloud-based self-learning-based automatic welding pool dual-closed-loop collaborative control method. This method utilizes a cloud-based three-layer neural network with downtime-free hot updates, reducing manual debugging and improving uptime. Cloud-based aggregation of heterogeneous samples from multiple plants trains a coupled network, improving the model's generalization accuracy under bevel fluctuations and thermal deformation. Manual process debugging time in a single workshop is reduced by 60%. The NPU dual-partition model hot updates require no downtime, ensuring continuous production without interruption and significantly improving production line uptime. For older workshops, 5G wireless transmission eliminates the need to re-lay industrial network cables, reducing equipment upgrade costs by 60%.
[0039] This invention provides a cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method for full-chain digital traceability of a single workpiece. It meets the compliance requirements of special equipment, binding all welding data to a unique workpiece code, establishing a multi-dimensional association index in the cloud, and allowing quality inspectors and customers to remotely retrieve complete electronic welding archives. It is compatible with classification societies and pressure vessel industry standards for factory flaw detection and quality auditing, solving the problems of data dispersion and difficulty in obtaining certification in distributed production lines. Attached Figure Description
[0040] Figure 1 This is a flowchart of the hardware and data flow of the automatic welding molten pool monitoring and adaptive control system of the present invention;
[0041] Figure 2 This is an NPU identification diagram with the standard melt pool size of the present invention, where the green outline is the melt pool boundary and the red line is the bevel baseline;
[0042] Figure 3 This is an NPU identification diagram for the excessively wide melt pool defect of the present invention;
[0043] Figure 4 This is an NPU identification diagram of the narrow melt pool defect of the present invention;
[0044] Figures 5 to 7 These are real-world images of the melt pool under different operating conditions based on the NPU lightweight network, as presented in this invention. Detailed Implementation
[0045] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. Identical components are indicated by the same reference numerals.
[0046] It should be noted that the terms “front,” “back,” “left,” “right,” “up,” and “down” used in the following description refer to the directions shown in the attached diagram, while the terms “inside” and “outside” refer to the directions toward or away from the geometric center of a specific component, respectively.
[0047] To make the content of this invention easier to understand, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings.
[0048] like Figures 1 to 7 As shown, the overall hardware architecture of a cloud-based self-learning-based automatic welding pool dual-closed-loop collaborative control method is illustrated in the attached diagram. Figure 1 (Automatic welding pool monitoring and adaptive control flowchart)
[0049] The system is divided into two main modules: local hardware unit and remote cloud AI training cluster. The local hardware unit includes: a high-speed molten pool camera with a composite optical lens, an independent embedded NPU hardware acceleration module, an industrial control computer with integrated 5G communication module, a PLC controller, a 1kHz electrical parameter acquisition module, an industrial welding power supply, and a two-dimensional stepping fine-tuning slide. An optional wire feeding mechanism can be added to expand the three-loop control.
[0050] Category 4 bandwidth-isolated independent communication channels
[0051] To address image congestion and real-time link occupancy issues, the hardware communication link is divided into four independent channels that do not compete for bandwidth:
[0052] (1) Image acquisition channel: The high-speed melt pool camera is directly connected to the NPU module, and the original melt pool image does not go through the industrial control computer for relay, thus eliminating bus frame loss;
[0053] (2) Local feature control channel: The NPU connects to the industrial control computer via industrial Ethernet and only transmits three sets of quantization deviation values of ΔX, ΔY and ΔW to the industrial control computer. It does not transmit large-capacity original images, thus reducing the bus load.
[0054] (3) Cloud interaction channel: The industrial control computer relies on the built-in 5G module to communicate bidirectionally with the cloud AI training cluster. During welding operation, it only receives lightweight model files and uploads complete datasets in batches during downtime.
[0055] (4) Low-level execution channel: the industrial control computer and PLC communicate bidirectionally, the PLC is uniformly connected to the two-dimensional slide table, electrical parameter acquisition module and welding power supply, and the unified hardware clock outputs drive signals synchronously.
[0056] Two parallel isolated execution links
[0057] This method synchronously runs the local real-time welding dual closed-loop collaborative control main line (which runs in real time throughout the welding process with high priority) and the 5G cloud asynchronous AI self-learning iterative auxiliary line (which runs silently only during downtime / material change intervals and does not occupy real-time control bandwidth). The two links have isolated bandwidth and do not compete for resources.
[0058] Link 1: Local real-time welding dual closed-loop collaborative control main line (cyclic execution of S1~S6)
[0059] S1: Before welding, the workpiece material, plate thickness, and bevel parameters are entered. The industrial control system retrieves the standard process library and outputs the reference weld pool width W0, ±0.1mm lateral offset threshold, reference current I0, and reference voltage U0, which are then sent to the NPU and PLC for solidification. Before arc ignition, the integrity of the NPU lightweight model is verified, and the latest inference model is loaded in advance.
[0060] S2: NPU non-intermediate melt pool image direct acquisition cache. The melt pool camera acquires melt pool images at 60~200FPS. The composite optical lens filters out arc light, splashes, and smoke interference. The original image is pushed unidirectionally to the NPU high-speed cache. The industrial control computer asynchronously stores the original image locally, which is only used for cloud upload after shutdown and does not intervene in the real-time image transmission link.
[0061] S3: NPU hardware parallel four-level feature decomposition (full hardware acceleration, single frame ≤20ms).
[0062] Hardware parallel image preprocessing: The dedicated NPU hardware synchronously completes grayscale conversion, overexposure repair, and noise filtering without occupying the industrial control CPU.
[0063] Lightweight network hardware inference: Output bevel centerline and weld pool outline; extract real-time weld pool width W, lateral offset ΔX, and vertical offset ΔY; 60-frame mean smoothing filtering.
[0064] The cached features of 60 consecutive frames are averaged to eliminate numerical jumps caused by momentary occlusion and welding torch jitter; the output is then quantized and simplified.
[0065] Calculate the weld pool width deviation ΔW = W - W0, and transmit only three sets of values, ΔX, ΔY, and ΔW, to the industrial control system.
[0066] S4: Industrial control synchronously generates two-way coordinated adjustment instructions. The industrial control calculates the two-dimensional slide position compensation pulse based on ΔX and ΔY; calculates the arc current and voltage adjustment increment based on ΔW; and simultaneously sends two types of instructions to the PLC without any sequential waiting.
[0067] S5: PLC Same Hardware Clock Synchronization Dual Closed-Loop Coupling Regulation (Core Control Innovation) The PLC has a built-in unified synchronous clock, enabling synchronous calculation of visual position closed-loop and electrical parameter morphology closed-loop, and synchronous output of drive signals. Composite anomalies are corrected in a single operation.
[0068] Visual position closed loop: Read the stepper motor encoder coordinates, drive the two-dimensional slide table to micro-calibrate the welding torch, and stably constrain ΔX within the range of ±0.1mm; the minimum compensation unit of the slide table single axis is 0.1mm, and the hardware response is ≤18ms;
[0069] Electrical parameter closed loop: 1kHz synchronous acquisition of instantaneous I and U, coupled with ΔW linkage adjustment of power supply; ΔW>0 (melt pool too wide) I and U are reduced; ΔW<0 (melt pool too narrow) I and U are increased; the two correction signals take effect in parallel in the same control cycle, avoiding defects such as molten pool oscillation and uneven weld width caused by step adjustment.
[0070] S6: Local data persistent cache, industrial control local solid-state storage of molten pool original image, timing characteristics, IU arc curve, workpiece condition label; cyclically execute S2~S6 until the arc extinguishing signal is triggered.
[0071] Link 2: 5G cloud-based asynchronous AI self-learning iterative auxiliary line (silent execution of T1~T5 during downtime)
[0072] T1: During the welding operation phase, the large image upload channel is closed. After shutdown, the molten pool image, feature timing, arc data, and working condition label are bound to a unique workpiece traceability code, encrypted, and then transmitted to the cloud via 5G breakpoints. The cloud establishes an image-size-electrical parameter association index based on the workpiece code to support accurate traceability of a single piece.
[0073] T2: Cloud-based three-layer coupled neural network parallel training. Cloud-based aggregation of heterogeneous welding samples from multiple plants within the group. Parallel training of three types of dedicated networks. The loss function incorporates thermal deformation and molten pool size constraints, adapting to ship and pressure vessel welding specifications: ① Position recognition network: Optimizes the accuracy of extracting bevel and molten pool contours under strong light and dust; ② Electrical parameter mapping network: Establishes the matching relationship between molten pool geometry and optimal thermal input; ③ Four-dimensional coupled adaptive network: Integrates four-dimensional features of bevel gap, thermal deformation, molten pool size, and arc parameters to generate general adaptive process rules; synchronously updates the global standard process library.
[0074] T3: Lightweight Model Packaging and Compression. The cloud performs quantization and pruning lightweight processing on the trained large model to generate a lightweight inference package adapted to NPU computing power; the lightweight model and the new version of the global process library are integrated into an encrypted distribution file.
[0075] T4: 5G downlink without downtime, NPU hot update, cloud-encrypted packaged files are sent to industrial control computers on each production line; industrial control verification files are tamper-proof before model switching instructions are sent; the NPU is divided into two independent model cache partitions, the old model continues to run in the current frame to ensure uninterrupted welding, and the new model is loaded in parallel in the background; the next frame image directly switches to the new model for inference, without any downtime or workstation restart required.
[0076] T5: The lightweight model after self-learning closed-loop iterative update is connected to the main line S3 feature calculation process; the system can autonomously output the optimal position compensation amount and target arc parameters based on the real-time melt pool status; the production samples are continuously collected and iteratively optimized in cycles T1~T5 to improve the model recognition accuracy.
[0077] It is equipped with a cloud-based end-to-end quality traceability function, which permanently stores all welding datasets of the entire production line in the cloud; authorized users can input a unique workpiece code to remotely retrieve complete molten pool images, dimensional time sequence, arc curves, and working condition information for a single piece, meeting the compliance requirements for digital quality auditing and flaw detection traceability of special equipment.
[0078] Expansion and optimization solutions
[0079] 1) Hardware limitations of the slide table: The two-dimensional stepper micro-adjustment slide table adopts a ball screw + closed-loop stepper motor, with a minimum single-axis compensation of 0.1mm, a response of ≤18ms, and continuous dynamic compensation for thermal deformation of long weld seams.
[0080] 2) Optical lens limitations: The composite optical lens is composed of a 940nm narrowband infrared filter and a gradient neutral density filter to filter out strong arc light and splash interference;
[0081] 3) Compatibility with traditional vision: The NPU supports traditional machine vision algorithms such as edge detection and threshold segmentation. Algorithm parameters are distributed synchronously in the cloud, and melt pool features can still be extracted independently even without a deep learning model.
[0082] 4) Three-closed-loop expansion: A wire feeding mechanism is added to the bottom layer. The PLC adjusts the wire feeding speed synchronously according to ΔW to form a three-closed-loop collaborative control of position + electrical parameters + wire feeding speed, while the core architecture remains unchanged.
[0083] 5) Supporting System: Provides an automatic weld pool adaptive control system for performing the above methods;
[0084] 6) Storage medium: Provides a computer-readable medium for storing the program, which implements all the steps of this method when the program is run.
[0085] Embodiment 1 of this invention patent:
[0086] 30mm thick marine pressure vessel plate V-groove MIG multi-layer multi-pass automatic welding
[0087] Hardware configuration: 200FPS high-speed global shutter molten pool camera, 940nm composite optical lens, independent embedded NPU computing board, integrated industrial 5G SA / NSA module industrial control computer, 16-axis PLC, 0.01mm precision ball screw two-dimensional slide table, 1000A industrial MIG welding machine, 1kHz synchronous current and voltage acquisition module.
[0088] Actual performance indicators
[0089] The NPU single-frame melt pool image and size calculation takes an average of 18ms, the melt pool width recognition error is ≤0.03mm, and the lateral offset is stably controlled within ±0.08mm;
[0090] Synchronous dual closed-loop coupled control reduces molten pool width fluctuation by 72%, and multi-layer, multi-pass welds are free of defects such as incomplete fusion, burn-through, and slag inclusion.
[0091] During daily downtime, 5G batch uploads of thick plate welding samples are performed. Data from three branch plants are collected in the cloud and iterated on a three-layer coupled network. The model adapts to fluctuations in bevel gap from 0 to 2.5 mm, eliminating the need for operators to adjust the process on-site.
[0092] Defect rate comparison: The defect rate of traditional vision welding is 12.7%, while the defect rate of this solution is 1.78%, a reduction of 86%;
[0093] Source verification: The classification society's quality inspectors input the workpiece code and remotely retrieve the entire molten pool width curve, arc timing, and original molten pool image to meet the pressure vessel audit specifications.
[0094] Embodiment 2 of this invention patent:
[0095] 6-12mm thin-plate steel structure submerged arc automatic welding
[0096] Challenges in operation: long welds on thin plates are prone to wavy thermal deformation; the positioning accuracy of tooling varies greatly between workshops; the interference from slag and dust in submerged arc welding is extremely strong; and the outline of the molten pool is easily blurred and distorted.
[0097] Real-world test results
[0098] The 940nm composite optical lens, combined with NPU hardware parallel noise reduction, filters out interference from welding slag and dust, and improves the accuracy of molten pool and bevel centerline recognition by 38% compared to traditional CPU vision.
[0099] The cloud-based aggregation of thin plate samples from various branch factories is used to iteratively couple a network, adaptively compensating for the thermal deformation of thin plate waves, resulting in an 82% reduction in defects such as undercut and lack of fusion.
[0100] By leveraging industrial 5G wireless modules, older workshops no longer need to re-lay industrial wired network cables, reducing equipment upgrade and renovation costs by 60%.
[0101] The cloud-based unified archiving system provides access to all welding sequence data for thin plate workpieces, allowing downstream customers to remotely retrieve complete electronic files for factory acceptance testing.
[0102] Extended implementation of this invention patent:
[0103] In the two embodiments described above, a wire feeding mechanism is added to the underlying execution unit. The PLC adjusts the wire feeding speed in real time according to the molten pool width deviation ΔW, forming a three-way collaborative control of visual position closed loop + electrical parameter closed loop + wire feeding speed closed loop. The NPU hardware acceleration, four types of independent communication channels, dual parallel links, and cloud self-learning core architecture remain unchanged, further optimizing the molten pool filling uniformity, making it suitable for high-speed welding scenarios of thin plates.
[0104] Compatible with traditional machine vision implementation methods:
[0105] The cloud-based edge detection, adaptive threshold segmentation, and contour fitting parameters of traditional visual algorithms are synchronously iterated and then delivered to the NPU for hardware acceleration. When the cloud-based deep learning model fails to be delivered, the NPU can switch to the traditional visual algorithm to independently complete the extraction of melt pool width and offset features, ensuring uninterrupted production of the equipment.
[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A cloud-based self-learning-based automatic welding pool dual-closed-loop collaborative control method, characterized in that, It includes local hardware units and remote cloud AI training clusters; the local hardware units include: a high-speed molten pool camera equipped with a composite optical lens, an independent embedded NPU hardware acceleration module, an industrial control computer with an integrated 5G communication module, a PLC controller, an electrical parameter acquisition module, a welding power supply, and a two-dimensional stepping fine-tuning slide. Hardware communication links are divided into four types of independent bandwidth isolated channels: (1) Image acquisition channel: The melt pool camera is directly connected to the NPU module, and the raw image is not forwarded by the industrial control computer; (2) Local feature control channel: The NPU is connected to the industrial control computer via industrial Ethernet and only outputs the melt pool quantization feature value to the industrial control computer; (3) Cloud-based interaction channel: The industrial control computer relies on the 5G module to conduct bidirectional encrypted communication with the cloud-based AI training cluster; (4) Bottom execution channel: The industrial computer and PLC communicate bidirectionally. The PLC is connected to the two-dimensional stepper fine-tuning slide, the electrical parameter acquisition module and the welding power supply respectively. The method describes two independent execution links that do not compete for bandwidth: a local real-time welding dual-closed-loop collaborative control main line and a 5G cloud-based asynchronous AI self-learning iterative auxiliary line. The local real-time welding dual closed-loop collaborative control main line executes S1~S6 in a loop: S1: Pre-welding initialization: retrieve the reference weld pool width W0, ±0.1mm offset threshold, reference current I0, and reference voltage U0 from the standard process library and send them to the NPU and PLC. Complete the NPU model verification before arc ignition. S2: The molten pool camera acquires images at 60~200FPS and pushes them unidirectionally to the NPU cache. The industrial control computer asynchronously stores the original images for cloud upload after shutdown. S3: The NPU hardware performs four levels of parallel computation: hardware image preprocessing, lightweight network inference to extract the actual width W of the melt pool, horizontal offset ΔX, vertical offset ΔY, 60-frame mean filtering to smooth features, and quantization output ΔW=W-W0, which is only transmitted to the industrial control system. S4: The industrial control computer synchronously generates two types of instructions: two-dimensional slide table position compensation pulse and arc current voltage adjustment increment, and sends them to the PLC simultaneously. S5: The PLC operates two closed loops synchronously within the same hardware clock cycle: the vision position closed loop drives the slide to stably control ΔX within ±0.1mm; the electrical parameter closed loop synchronously acquires I and U at 1kHz, and couples ΔW to adjust the welding power supply; when ΔW>0, I and U are lowered, and when ΔW<0, I and U are raised. S6: The industrial control system locally caches the original image of the molten pool, characteristic timing, IU arc curve and operating condition label, and cycles through S2~S6 until the arc is extinguished; The 5G cloud-based asynchronous AI self-learning iterative auxiliary line silently executes T1~T5 only during downtime: T1: After shutdown, the welding dataset is bound to a unique workpiece code and encrypted. It is then uploaded to the cloud via 5G breakpoint resume transmission. The cloud establishes an image, electrical parameter, and dimension timing association index. T2: Cloud-based collection of heterogeneous samples from multiple plants for parallel training of location recognition network, electrical parameter mapping network, and four-dimensional coupled adaptive network; loss function incorporates thermal deformation and melt pool size constraints; updates global process library. T3: The cloud performs quantization pruning and lightweighting on the training model, and integrates the lightweight model package and the new version of the global process library into an encrypted distribution file; T4: The cloud 5G sends encrypted files to the industrial control system, which verifies them and then sends a model switching command; the NPU is divided into two independent model cache partitions, the new model is loaded in the background, and the new model is switched for inference in the next frame of the image, without any downtime. T5: The updated lightweight model is integrated into the S3 solution process. The system automatically outputs the optimal compensation amount and target electrical parameters, and iterates and optimizes continuously by looping from T1 to T5. The entire production line's welding data is permanently stored in the cloud. Authorized users can remotely retrieve complete welding data for a single piece by entering the workpiece code, enabling quality traceability for special equipment.
2. The automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning as described in claim 1, characterized in that, The two-dimensional stepper fine-tuning slide adopts a ball screw and a closed-loop stepper motor, with a minimum single-axis compensation unit of 0.1mm and a hardware response time of ≤18ms. When the NPU synchronous output exceeds the standard ΔX and ΔW, the PLC outputs position compensation pulses and power adjustment signals in parallel during the same control cycle. The two corrections take effect synchronously, eliminating defects such as molten pool oscillation and uneven weld width caused by step adjustment.
3. The automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning according to claim 1, characterized in that, The composite optical lens is composed of a 940nm narrowband infrared filter and a gradient neutral density filter to filter out interference from electric arc light, spatter, and smoke, and to stably extract the molten pool profile and bevel baseline.
4. The automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning as described in claim 1, characterized in that, The NPU hardware parallel computation is compatible with traditional machine vision algorithms; edge detection and threshold segmentation algorithm parameters are synchronously iterated in the cloud, and then distributed to the NPU after being lightweighted, so that melt pool feature extraction can be completed independently when there is no deep learning model.
5. The automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning as described in claim 1, characterized in that, The underlying execution unit is equipped with a wire feeding mechanism; the PLC adjusts the wire feeding speed in real time according to ΔW, forming a three-closed-loop coordinated control of position, electrical parameters, and wire feeding speed, with hardware links and dual parallel execution chains.
6. The automatic welding pool dual-closed-loop collaborative control method based on cloud self-learning as described in claim 1, characterized in that, The cloud-based AI training cluster collects heterogeneous samples from multiple plants within the group to build a global welding sample library. The three-layer neural network is trained in parallel to improve the model recognition and generalization accuracy under conditions of bevel fluctuation, thermal deformation, and dust interference, thereby reducing the frequency of on-site manual process adjustments.
7. An automatic welding pool adaptive control system, characterized in that, The control method for executing any one of claims 1 to 5 comprises: an image acquisition module, an NPU hardware acceleration module, an industrial control scheduling module, a low-level execution module, and a remote cloud-based AI training cluster; the image acquisition module is a high-speed molten pool camera equipped with a composite optical lens; the NPU hardware acceleration module has built-in dual independent model inference cache partitions for hardware parallel molten pool image preprocessing and feature calculation; the industrial control scheduling module integrates a 5G communication module and local solid-state storage; the low-level execution module includes a PLC, a two-dimensional stepper fine-tuning slide, a 1kHz electrical parameter acquisition module, a welding power supply, and an optional wire feeding mechanism; the remote cloud-based AI training cluster is used for sample collection, parallel training of a three-layer neural network, model lightweighting, remote distribution, and permanent traceability storage of welding data.
8. A computer-readable storage medium having a computer-executable program stored thereon, characterized in that, When the program is run by the processor, it implements the cloud-based self-learning automatic welding pool dual-closed-loop collaborative control method as described in any one of claims 1 to 6.
