Intelligent control method and system for double-wire submerged arc welding quality of stainless steel reaction kettle

Abstract

The invention discloses an intelligent control method and system for double-wire submerged arc welding quality of a stainless steel reaction kettle, molten pool form and temperature distribution data are obtained through an optical visual sensor and a thermal imaging device, a multi-source data set is formed in combination with arc signals, stability characteristics and defect indexes are analyzed and extracted by using signal entropy, and the quality of the double-wire submerged arc welding quality of the stainless steel reaction kettle is obtained. And when the defect prejudgment exceeds a threshold value, the defect type is determined through machine learning classification, then a swarm intelligence optimization prediction model is input to calculate a welding parameter adjustment value, an optimized current ratio and spacing combination is output, the wire feeding speed is adjusted through intelligent control, and an electric arc stability index is monitored in real time. And if the welding seam offset exceeds the limit, the multi-axis positioning mechanism is used for automatically correcting the deviation, and finally the welding seam forming consistency is achieved. According to the method, through multi-source data fusion, defect intelligent pre-judgment and parameter dynamic optimization, the welding quality and stability are remarkably improved, and the defect occurrence rate is reduced.

Description

Technical Field

[0001] This invention relates to the field of stainless steel reactor welding control technology, and in particular discloses an intelligent control method and system for the quality of double-wire submerged arc welding of stainless steel reactors. Background Technology

[0002] As core pressure vessels in chemical, pharmaceutical, and energy industries, the quality of welds in stainless steel reactors directly determines the equipment's safety, corrosion resistance, and service life. In welding thick-walled stainless steel materials, double-wire submerged arc welding is widely used due to its high deposition efficiency and controllable heat input, making it the preferred process for manufacturing longitudinal and circumferential seams in large reactors. However, as the vessel wall thickness increases to tens or even hundreds of millimeters, the weld requires multiple layers and passes. Ensuring that the molten pool formed between each weld pass and between the two welding wires remains stable and controllable becomes a crucial challenge in the manufacturing process.

[0003] Most existing twin-wire submerged arc welding equipment relies on operators setting fixed parameters based on experience, or employs simple open-loop control methods for welding. This approach reveals significant shortcomings when dealing with the characteristics of stainless steel, such as low thermal conductivity, high coefficient of linear expansion, and poor molten pool fluidity. During welding, strong magnetic field interference and thermal coupling occur between the two arc wires. The arc of the rear wire enters before the molten pool of the front wire has fully solidified, causing disturbance to the molten pool and resulting in uneven ripples, undercut, or localized overheating on the surface. Simultaneously, slight misalignment or thermal deformation is unavoidable in long circumferential or longitudinal welds. Even slight deviations in the welding torch position can cause the arc to deviate from the center of the bevel, resulting in inconsistent penetration depth or incomplete fusion defects. These phenomena do not occur in isolation but are interconnected: minute changes in arc spacing and current ratio directly amplify the disturbance in the molten pool, while instability in the molten pool morphology further exacerbates the arc position deviation, ultimately leading to unpredictable internal weld defects in thick plates with deep bevels.

[0004] The core technical challenge lies in the lack of effective control over the dynamic coupling characteristics of the dual-wire arc. The front and rear wire arcs are not simply parallel; rather, they interact strongly through the molten pool as a common medium: the front arc heats and melts the base material to form the initial molten pool, while the rear arc continues heating on the still-warmed pool. The distance between them, the current magnitude, and the wire feed speed all alter the temperature gradient, flow direction, and surface tension of the molten pool. When this coupling relationship becomes unbalanced, the molten pool is prone to violent oscillations or even overflow, resulting in inconsistent weld width, fluctuating weld reinforcement, and locally shallow or deep weld penetration. For example, when welding the circumferential weld of a reactor vessel, if the rear wire current is too high, the front wire molten pool is subjected to excessive impact, pushing the metal at the tail of the pool to the sides, forming irregular fish-scale patterns and slag inclusion channels. These defects are often hidden within multiple weld layers, making them difficult to detect with conventional non-destructive testing, directly threatening the vessel's ability to withstand high pressure and corrosive media over the long term.

[0005] Therefore, how to perceive and precisely control the dynamic coupling relationship between the front and rear wire arcs and the molten pool in real time during the twin-wire submerged arc welding process, and maintain the stability of the molten pool and the consistency of weld formation, has become a key issue to ensure the high-quality manufacturing of thick-walled welds in stainless steel reactors. Summary of the Invention

[0006] This invention provides an intelligent control method and system for the quality of double-wire submerged arc welding of stainless steel reactors, aiming to solve at least one defect in the prior art.

[0007] One aspect of the present invention relates to an intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors, comprising the following steps: S100: Acquires molten pool morphology and temperature distribution information from the double-wire submerged arc welding process of stainless steel reactor through optical vision sensors and thermal imaging devices, and forms a multi-source fusion dataset by combining arc voltage and current signals. S200. The signal entropy analysis method is used to process the multi-source fusion dataset, extract the stability characteristics of the molten pool and potential defect indicators, and obtain the basic values ​​for defect prediction. S300. If the basic value of defect prediction exceeds the preset threshold, the extracted melt pool stability features are classified using machine learning classification methods to determine the specific defect type. S400: Input the specific defect type into the swarm intelligence optimization prediction model, calculate the adjustment value of the welding parameters, and output the optimized welding parameter set, which includes the current ratio and spacing combination. S500 uses intelligent control methods to adjust the wire feeding speed of the wire feeding module based on the welding parameter set, and obtains the arc stability index after real-time adjustment. S600: Determine the degree of weld deviation from the acquired arc stability index. If the deviation exceeds the threshold, the multi-axis positioning mechanism will automatically correct the deviation to obtain the final weld formation consistency data.

[0008] Further, step S100 includes: S110. Acquire optical image sequences, thermal infrared image sequences, arc voltage signals, and welding current signals collected by optical vision sensors and thermal imaging devices. S120. Extract the geometric feature data of the molten pool morphology from the optical image sequence, and map the geometric feature data of the molten pool morphology onto the thermal infrared image sequence to obtain the temperature distribution matrix corresponding to the spatial data of the geometric feature data of the molten pool morphology. S130. The temperature distribution matrix, molten pool morphology geometric feature data, arc voltage signal and welding current signal are time-aligned to form a multi-source fusion dataset containing molten pool morphology information, temperature distribution information and arc voltage and current signals.

[0009] Further, step S200 includes: S210. Obtain a multidimensional discrete time-series signal sequence. The multidimensional discrete time-series signal sequence is generated by reconstructing the multidimensional acoustic-optical-thermal sensing values ​​in the multi-source fusion dataset. S220. The signal entropy value analysis method is used to process the multidimensional discrete time series signal sequence to obtain the signal entropy value evolution sequence. S230. If the fluctuation amplitude of the signal entropy value evolution sequence exceeds the steady-state reference range, then extract the first molten pool stability feature vector. S240. Generate a potential defect probability distribution matrix based on the stability feature vector of the first molten pool, and analyze the potential defect probability distribution matrix to obtain the basic numerical value for defect prediction.

[0010] Further, step S300 includes: S310. Obtain the basic value of defect prediction. If the basic value of defect prediction exceeds the preset abnormal trigger threshold, extract the second melt pool stability feature vector. S320. Input the second molten pool stability feature vector into the support vector machine classification model, and use the radial basis kernel function to map it to a high-dimensional feature space to obtain feature points; S330. Calculate the geometric distance from the feature point to the optimal hyperplane, and determine the spatial position and assignment probability value relative to the decision boundary based on the geometric distance; S340. Select the category label corresponding to the largest item in the attribution probability values, and determine the specific defect type based on the category label.

[0011] Further, step S400 includes: S410. Match the optimization boundary conditions according to the determined specific defect type, and construct the initial population distribution of the particle swarm optimization algorithm; S420. Import the initial population distribution into the particle swarm optimization model, calculate the fitness value using the ideal melt pool features associated with specific defect types, and iteratively search for the global optimal solution vector based on the fitness value. S430. Decode the global optimal solution vector to obtain the current ratio adjustment step size and the spacing combination adjustment step size; S440. Adjust the step size and spacing combination according to the current ratio to generate an optimized welding parameter set, which includes the current ratio and spacing combination.

[0012] Further, step S500 includes: S510. Obtain real-time welding current and arc voltage values ​​based on the welding parameter set, and obtain the instantaneous arc power sequence and short-circuit transition period duration based on the real-time welding current and arc voltage values. S520. If the standard deviation of the instantaneous arc power sequence exceeds the threshold, the wire feeding speed compensation increment shall be determined according to the short-circuit transition period duration. S530: Generate real-time wire feeding speed control commands based on the wire feeding speed compensation increment to adjust the wire feeding speed of the wire feeding module; S540: Collects arc voltage feedback waveform data after responding to real-time wire feeding speed control command, calculates the distribution variance of arc voltage feedback waveform data, and obtains the arc stability index after real-time adjustment.

[0013] Further, step S600 includes: S610. Construct a weld deviation feature vector by combining the real-time adjusted arc stability index with the welding torch spatial coordinate data, and calculate the weld offset degree and trajectory deviation direction based on the weld deviation feature vector. S620. If the weld offset value exceeds the preset threshold, a trajectory compensation matrix is ​​generated based on the trajectory deviation direction. The trajectory compensation matrix is ​​converted into a correction control command to drive the multi-axis positioning mechanism to adjust the welding torch posture. S630: Collect feedback data after responding to the correction control command, calculate the dynamic balance coefficient of the molten pool, and perform statistical analysis on the dynamic balance coefficient of the molten pool to obtain the final weld formation consistency data.

[0014] Another aspect of the present invention relates to an intelligent control system for the quality of double-wire submerged arc welding of stainless steel reactors, used to execute the above-described intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors, comprising: The multi-source fusion dataset generation module is used to acquire molten pool morphology and temperature distribution information from the double-wire submerged arc welding process of stainless steel reactor through optical vision sensors and thermal imaging devices, and combine it with arc voltage and current signals to form a multi-source fusion dataset. The defect prediction basis acquisition module is used to process multi-source fusion datasets using signal entropy analysis methods, extract melt pool stability characteristics and potential defect indicators, and obtain the defect prediction basis. The specific defect type determination module is used to classify the extracted melt pool stability features using machine learning classification methods to determine the specific defect type if the defect prediction basis exceeds a preset threshold. The welding parameter set output module is used to input specific defect types into the swarm intelligence optimization prediction model, calculate the adjustment values ​​of welding parameters, and output the optimized welding parameter set, which includes current ratio and spacing combination. The arc stability index acquisition module is used to adjust the wire feeding speed of the wire feeding module based on the welding parameter set through intelligent control methods, and to obtain the arc stability index after real-time adjustment. The final weld formation consistency data acquisition module is used to determine the degree of weld deviation from the acquired arc stability index. If the deviation exceeds the threshold, it is automatically corrected by the multi-axis positioning mechanism to obtain the final weld formation consistency data.

[0015] The beneficial effects achieved by this invention are as follows: This invention provides an intelligent control method and system for the quality of double-wire submerged arc welding of stainless steel reactors. Addressing issues such as insufficient molten pool stability, difficulty in real-time defect prediction, and weld misalignment during welding, the method acquires molten pool morphology and temperature distribution data using optical vision sensors and thermal imaging devices. This data, combined with arc signals, forms a multi-source dataset. Stability features and defect indicators are extracted using signal entropy analysis. When a defect prediction exceeds a threshold, machine learning is used to classify and determine the defect type. The data is then input into a swarm intelligence optimization prediction model to calculate welding parameter adjustment values, outputting optimized current ratios and spacing combinations. Intelligent control adjusts the wire feeding speed, and arc stability indicators are monitored in real-time. If weld misalignment exceeds limits, a multi-axis positioning mechanism automatically corrects the deviation, ultimately achieving consistent weld formation. This invention significantly improves welding quality and stability and reduces the defect incidence rate through multi-source data fusion, intelligent defect prediction, and dynamic parameter optimization. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to the present invention. Figure 2 This is a functional block diagram of an embodiment of the intelligent control system for the quality of double-wire submerged arc welding of stainless steel reactors according to the present invention.

[0017] Explanation of icon numbers: 10. Multi-source fusion dataset formation module; 20. Defect prediction basis acquisition module; 30. Specific defect type determination module; 40. Welding parameter set output module; 50. Arc stability index acquisition module; 60. Final weld formation consistency data acquisition module. Detailed Implementation

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] like Figure 1 As shown, the first embodiment of the present invention proposes an intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors, comprising the following steps: Step S100: Obtain molten pool morphology and temperature distribution information from the double-wire submerged arc welding process in a stainless steel reactor using optical vision sensors and thermal imaging devices, and combine them with arc voltage and current signals to form a multi-source fusion dataset.

[0020] The equipment uses optical vision sensors and thermal imaging devices to capture real-time information on the molten pool morphology (such as size, shape, and boundary clarity) and temperature distribution (such as temperature gradient in the molten pool area and temperature range in the heat-affected zone) during the welding process. Simultaneously, it collects electrical signals such as arc voltage and current. Through data alignment and format standardization, the three types of heterogeneous data are fused into a unified multi-source fusion dataset, which comprehensively covers the "geometric-thermal-electrical" characteristics of the welding process. This provides complete data support for subsequent defect prediction and parameter optimization, and is the source guarantee for realizing intelligent control of welding quality.

[0021] Step S200: The multi-source fusion dataset is processed using the signal entropy analysis method to extract the stability characteristics of the molten pool and potential defect indicators, and to obtain the basic values ​​for defect prediction.

[0022] Using the multi-source fusion dataset generated in step S100 as input, various signals are quantified and processed using signal entropy analysis methods (such as Shannon entropy and approximate entropy). The stability features of the weld pool (such as the entropy value of weld pool morphology fluctuation and the entropy value of temperature distribution uniformity) that characterize welding stability are accurately extracted, as well as defect prediction indicators that reflect potential quality problems (such as the entropy value of arc voltage and current fluctuation and the coefficient of sudden change of weld pool temperature). Through feature integration and quantification calculation, the basic numerical values ​​for defect prediction are output, realizing the preliminary quantitative assessment of the quality status of the welding process, and providing triggering basis and feature basis for subsequent defect type determination.

[0023] Step S300: If the defect prediction baseline value exceeds the preset threshold, the extracted melt pool stability features are classified using machine learning classification methods to determine the specific defect type.

[0024] The defect prediction baseline value output in step S200 is used as the basis for judgment and compared with the preset quality qualification threshold. If the baseline value exceeds the threshold, it indicates that there is a quality risk in the welding process. Then, machine learning classification methods (such as random forest, SVM (Support Vector Machine)) are called to classify, train and identify the previously extracted molten pool stability features, accurately determine the specific defect type (such as incomplete penetration, slag inclusion, porosity, cracks, etc.), clarify the core cause of the quality problem, provide a targeted target for subsequent welding parameter optimization, and realize the transformation from "predicting anomalies" to "locating defects".

[0025] Step S400: Input the specific defect type into the swarm intelligence optimization prediction model, calculate the adjustment value of the welding parameters, and output the optimized welding parameter set, which includes the current ratio and spacing combination.

[0026] The specific defect type determined in step S300 is input into a preset swarm intelligence optimization prediction model (such as particle swarm optimization, genetic algorithm-optimized BP (Back Propagation) neural network). Based on the correspondence between historical welding quality data and defects, and combined with the process requirements of stainless steel reactor welding, the model accurately calculates the adjustment values ​​of welding parameters, focusing on optimizing the current ratio (the ratio of the current magnitude of the two wires) and spacing combination (the spatial distance between the two wires) of the two wires, generating an optimized welding parameter set adapted to the current defect type. This provides executable parameter basis for the precise control of the subsequent welding process, achieving targeted matching of "defect type → parameter optimization".

[0027] Step S500: Adjust the wire feeding speed of the wire feeding module based on the welding parameter set using an intelligent control method to obtain the real-time adjusted arc stability index.

[0028] Intelligent control algorithms (such as PID (Proportional-Integral-Derivative) control) are adopted to adjust the wire feeding speed of the wire feeding module in real time based on the optimized welding parameter set output in step S400, so that the wire feeding rhythm is adapted to the optimized current and spacing parameters. At the same time, the adjusted arc stability indicators (such as arc voltage and current fluctuation amplitude and arc combustion continuity) are collected synchronously to quantitatively evaluate the improvement effect of parameter adjustment on welding stability, forming an instant feedback of "parameter adjustment → effect monitoring", which provides process data support for subsequent weld deviation judgment.

[0029] Step S600: Determine the degree of weld offset from the obtained arc stability index. If the offset exceeds the threshold, the multi-axis positioning mechanism will automatically correct the offset to obtain the final weld formation consistency data.

[0030] From the arc stability index obtained in step S500, the relevant features of weld position offset are extracted to determine the degree of offset between the actual position of the weld and the preset benchmark. If the offset exceeds the preset threshold, the multi-axis positioning mechanism is activated to achieve automatic correction of the welding torch head through precise mechanical adjustment, thereby correcting the weld position deviation. Finally, the weld forming data after correction is collected, and the weld forming consistency results (such as weld width uniformity, reinforcement height consistency, and forming flatness) are output. This completes the closed-loop control of the entire welding quality process from "prediction-optimization-control-correction", ensuring the stability and reliability of the stainless steel reactor welding quality.

[0031] Furthermore, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S100 as follows: Step S110: Acquire the optical image sequence, thermal infrared image sequence, arc voltage signal and welding current signal collected by the optical vision sensor and thermal imaging device.

[0032] In welding monitoring systems, acquiring various signal data is crucial. Specifically, high-speed optical cameras and thermal infrared cameras mounted on the welding robot simultaneously acquire optical and thermal infrared image sequences, all with unified timestamps—for example, the timestamp for each frame is accurate to the millisecond level—to ensure subsequent synchronization. Simultaneously, voltage and current sensors record arc voltage and welding current signals in real time, also labeled with the same timestamps. In one implementation, the optical camera captures visible light images of the molten pool area at 100 frames per second, capturing changes in the pool's contour, while the thermal infrared camera records temperature distribution images at the same frame rate. Voltage and current signals are acquired via a data acquisition card at a higher sampling rate, such as 1 kHz, ensuring all data is aligned on the timeline. The principle behind this acquisition method is that unified timestamps prevent data offset, thus providing a foundation for multi-source fusion.

[0033] Step S120: Extract the geometric feature data of the molten pool morphology based on the optical image sequence, and map the geometric feature data of the molten pool morphology onto the thermal infrared image sequence to obtain the temperature distribution matrix corresponding to the spatial data of the geometric feature data of the molten pool morphology.

[0034] The temperature distribution matrix is ​​obtained using the following formula: (1) In formula (1), Represents the spatial grid points after mapping Temperature distribution matrix elements at that location, Represents the spatial grid points after mapping Temperature value in thermal infrared image at that location. Represents the spatial grid points after mapping The geometric feature space indicator mask is used to obtain the temperature distribution matrix corresponding to the geometric feature data space of the molten pool morphology. The control logic of formula (1) is "thermal infrared temperature data + geometric feature mask filtering + spatial alignment temperature matrix generation". Through the binary mask of the molten pool geometric features, the temperature value belonging only to the molten pool area is accurately extracted from the original thermal infrared image, the background interference is filtered, and a temperature distribution matrix strictly aligned with the geometric feature space of the molten pool morphology is generated, providing an accurate thermal data foundation for subsequent multi-source feature fusion and welding quality analysis.

[0035] Mapped spatial grid points Temperature value of thermal infrared image at the location This can be derived from the following formula: (2) In formula (2), and This represents the normalized coordinates of a grid point within a rectangle of adjacent pixels. , , , The temperature values ​​of four adjacent pixels are represented. The control logic of formula (2) is "sub-pixel coordinate decomposition + bilinear weight allocation + four-neighbor pixel weighted fusion + sub-pixel temperature interpolation". By decomposing the normalized coordinates of the grid points into horizontal and vertical weights, the temperature of the four adjacent pixels is weighted and averaged to obtain the accurate temperature value of the sub-pixel position, which solves the problem of spatial resolution mismatch between thermal infrared image and molten pool geometric features and achieves more accurate temperature mapping.

[0036] Mapped spatial grid points Geometric feature space indicator mask This can be derived from the following formula: (3) In formula (3), Represents the spatial grid points after mapping coordinate vector, The center coordinates represent the geometric features of the molten pool. This represents the Gaussian kernel width parameter. The squared Euclidean distance from the grid point to the center is represented by . The control logic of formula (3) is "distance-driven Gaussian weighting + soft boundary mask generation + smooth transition of the molten pool region". By calculating the distance from the grid point to the geometric center of the molten pool, the distance is converted into a continuous weight between 0 and 1, generating a soft mask with smooth transition characteristics, replacing the traditional hard binary mask, which is more in line with the natural gradient characteristics of the molten pool geometric region.

[0037] When extracting molten pool morphology geometric feature data from optical image sequences, image processing algorithms are employed for analysis. For example, the optical images are first preprocessed, such as through grayscale conversion and noise filtering. Then, edge detection methods, such as the Canny algorithm, are used to identify the molten pool boundaries, thereby calculating geometric features, including the length, width, area, and shape factor of the molten pool. Specifically, in an actual welding operation, the system extracts feature data from the image showing a molten pool length of 5.2 mm and a width of 2.8 mm. This data reflects the dynamic morphological changes of the molten pool. This extraction not only quantifies the geometric properties of the molten pool but also provides a spatial reference for subsequent mapping, thereby improving the accuracy of welding quality monitoring. For instance, mapping the molten pool morphology geometric feature data onto a thermal infrared image sequence to obtain the temperature distribution matrix requires spatial registration. In one implementation, image registration techniques, such as affine transformation based on feature points, are used to map the molten pool boundary coordinates in the optical image to the corresponding positions in the thermal infrared image. Specifically, assuming the coordinates of the center point of the molten pool in the optical image are (x=200, y=150), the corresponding point in the thermal infrared image (x'=205, y'=148) is calculated using a registration matrix. Then, the temperature value is extracted within this region, forming a 10×10 temperature distribution matrix, where each element represents the temperature of a pixel, such as 1500℃ in the central region. This mapping associates geometric features with temperature information, achieving spatially corresponding data fusion and bringing higher accuracy to welding defect detection, such as timely identification of overheated areas to prevent crack formation.

[0038] Step S130: Time-align the temperature distribution matrix, molten pool morphology geometric feature data, arc voltage signal and welding current signal to form a multi-source fusion dataset containing molten pool morphology information, temperature distribution information and arc voltage and current signals.

[0039] The multi-source fusion dataset is derived using the following formula: (4) In formula (4), This represents a multi-source fusion dataset. The temperature distribution matrix, This refers to the geometric features of the molten pool morphology. This is the arc voltage signal. This is the welding current signal. The formula (4) represents the fusion after time alignment. It is used to align the temperature distribution matrix, the geometric feature data of the molten pool morphology, the arc voltage signal and the welding current signal and splice them to form a multi-source fusion dataset. The control logic of formula (4) is "multi-source data time alignment + heterogeneous feature structured splicing + unified spatiotemporal benchmark construction". By using a unified timestamp to eliminate the sampling rate difference of different sensors, the heterogeneous features such as temperature, geometry, voltage and current are integrated into a structured dataset according to time steps, providing a spatiotemporally consistent input basis for subsequent multimodal modeling of welding quality.

[0040] When aligning the temperature distribution matrix, molten pool morphology geometry, arc voltage signal, and welding current signal to form a multi-source fusion dataset, timestamps are used as a reference for interpolation synchronization. For example, signals with different sampling rates, such as a 1kHz current signal and a 100Hz image, are aligned to a unified time grid using linear interpolation. Specifically, at a timestamp t=1.5s, the fusion dataset includes a molten pool length of 2.9mm, an average temperature matrix value of 1200℃, a voltage of 25V, and a current of 150A. This alignment ensures the temporal consistency of the dataset, thereby supporting subsequent machine learning model training for predicting welding quality.

[0041] Preferably, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S200: Step S210: Obtain a multidimensional discrete time-series signal sequence. The multidimensional discrete time-series signal sequence is generated by reconstructing the acoustic, optical, and thermal multidimensional sensing values ​​in the multi-source fusion dataset.

[0042] Multidimensional discrete-time signals are derived using the following formula: (5) In formula (5), This represents the reconstructed multidimensional discrete-time signal. Describes a left singular vector matrix. Represents a singular value matrix. This represents the transpose of a right singular vector. The formula (5) represents a multi-source fusion dataset and is used to reconstruct multidimensional discrete time-series signals through singular value decomposition. The control logic of formula (5) is "SVD decomposition of high-dimensional heterogeneous data + key energy retention + low-dimensional robust reconstruction". Through singular value decomposition (SVD), the original multi-source fusion data is projected onto a low-dimensional orthogonal space to remove noise and redundant information, retain the main energy features, and finally reconstruct a simpler and more robust multidimensional discrete time-series signal, providing a more effective input for subsequent welding quality analysis.

[0043] In exploring the acquisition and processing of multidimensional discrete time-series signal sequences, we begin with data reconstruction and gradually delve into subsequent analysis and applications. For example, in the field of welding monitoring, the generation of multidimensional discrete time-series signal sequences is typically based on multi-source fusion datasets, which contain raw data from various sensors such as acoustic, optical, and thermal sensors. In one implementation, the system first preprocesses the acquired acoustic signals, optical image data, and thermal imaging data, unifying signals with different sampling frequencies onto a single time grid through time-series reconstruction. Assuming that during a welding process, the acoustic sensor records sound vibration data at 500Hz, the optical sensor records molten pool morphology changes at 100Hz, and the thermal sensor also collects temperature distribution information at 100Hz, the system will reconstruct these data and align them to a unified 100Hz time grid using interpolation methods, forming a time-series signal sequence containing multidimensional features, laying the foundation for subsequent analysis.

[0044] Step S220: Process the multidimensional discrete time-series signal sequence using the signal entropy analysis method to obtain the signal entropy evolution sequence.

[0045] The signal entropy evolution sequence is obtained using the following formula: (6) In formula (6), This represents the sequence of signal entropy evolution. Represents a multidimensional discrete-time signal at time 10:00. The signal entropy value, The total length of the multidimensional discrete time-series signal is represented by the formula (6). The control logic of formula (6) is "time window division + window-by-window entropy calculation + entropy value time-series splicing". By calculating the entropy value of the reconstructed multidimensional discrete time-series signal according to the time window, the disorder and fluctuation complexity of the signal in each time period are quantified, and the entropy value evolution sequence is generated to capture the dynamic changes and abnormal precursors of the molten pool / arc state during the welding process.

[0046] Multidimensional discrete-time signals at time joint entropy This can be derived from the following formula: (7) In formula (7), Indicates time After quantization of multidimensional discrete-time signals, the first The probability of each state. The total number of quantized states is represented. The control logic of formula (7) is "multidimensional signal state quantization + probability distribution modeling + entropy quantization of joint uncertainty". By discretizing and encoding the state of the multidimensional discrete time sequence signal, the probability of occurrence of each state is calculated. Finally, the entropy value is used to quantify the joint disorder of the multimodal signal at that moment. The higher the entropy value, the more dispersed the signal state distribution and the stronger the uncertainty, which may correspond to a more unstable welding process.

[0047] time After quantization of the multidimensional discrete signal, the first The probability of each state This can be derived from the following formula: (8) In formula (8), Indicates time Multidimensional discrete signals Indicates the first The quantization region of each state, Indicates the first The volume of the quantization region corresponding to each state This represents the Dirac delta function. Indicates the first The quantization region of each state. The control logic of formula (8) is "multidimensional feature space quantization + Delta function precise point positioning + integral determination of region attribution + volume normalization probability calculation". By using the positioning characteristics of the Dirac delta function, combined with integral and volume normalization, the continuous multidimensional discrete signal is mapped to the probability value of discrete state, so as to realize the quantization probability modeling of multimodal welding signal.

[0048] Next, when processing multidimensional discrete-time signal sequences, signal entropy analysis is used to quantify the complexity and trends of the signal. Specifically, signal entropy analysis is a tool for measuring the degree of disorder in a signal; it reflects the dynamic characteristics of the system by calculating the information entropy of the signal within a time window. In one embodiment, assuming the system segments the multidimensional signal sequence during the welding process, with each segment containing 1 second of data, its entropy value is calculated to form an entropy evolution sequence. By observing the change in entropy value over time, abnormal fluctuations occurring during the welding process are captured, providing a basis for further feature extraction.

[0049] Step S230: If the fluctuation amplitude of the signal entropy value evolution sequence exceeds the steady-state reference range, then extract the first melt pool stability feature vector.

[0050] The fluctuation amplitude of the signal entropy evolution sequence is obtained by the following formula: (9) In formula (9), Indicates the fluctuation range. The sequence representing the evolution of signal entropy values. The value at time, This represents the maximum value in the sequence of signal entropy evolution. This represents the minimum value in the sequence of signal entropy evolution. The control logic of formula (9) is "extreme value extraction + amplitude quantization + stability threshold triggering". By calculating the difference between the maximum and minimum values ​​in the sequence of signal entropy evolution, the fluctuation range of the entire entropy sequence is directly quantified. This is used as the basis for judging whether the welding process deviates from the steady state. The larger the fluctuation amplitude, the more violent the fluctuation of the entropy value, and the worse the stability of the corresponding welding process.

[0051] The following formula is used to define the criteria for extracting the stability feature vector of the first molten pool: (10) In formula (10), Indicates the extraction flag. Indicates the steady-state reference range value. Indicates other situations. The control logic of formula (10) is "threshold-triggered binary judgment + precise control of extraction action". By comparing the fluctuation amplitude with the preset steady-state benchmark threshold, the extraction flag of 0 / 1 is output to realize the precise decision of "whether to extract the first molten pool stability feature vector". It is the key logic for triggering welding stability anomalies.

[0052] The stability eigenvector of the first molten pool is obtained by the following formula: (11) In formula (11), Represents the eigenvectors of molten pool stability. Represents the mean of the signal entropy sequence. The standard deviation of the signal entropy sequence is represented by... This represents the range of values ​​for the signal entropy sequence. This indicates the transpose operation. The control logic of formula (11) is "multi-dimensional statistical feature extraction + stability quantization + standardized vector output". By calculating the mean, standard deviation and range (range value) of the signal entropy sequence, the dynamic stability of the entropy sequence is transformed into a structured feature vector, providing interpretable quantitative input for subsequent welding quality assessment and anomaly warning.

[0053] When the fluctuation amplitude of the signal entropy evolution sequence exceeds the preset steady-state reference range, the system triggers the extraction process of the first molten pool stability feature vector. For example, in actual welding operations, the steady-state reference range may be an entropy fluctuation range derived from historical data statistics, assumed to be 0.5 to 1.5. If the entropy fluctuation reaches 2.0 within a certain period, the system will automatically extract the first molten pool stability feature vector. These first molten pool stability feature vectors include the rate of change of geometric parameters of the molten pool morphology, the gradient change of temperature distribution, and the frequency distribution characteristics of the acoustic signal, forming a multi-dimensional vector for subsequent analysis.

[0054] Step S240: Generate a potential defect probability distribution matrix based on the first molten pool stability feature vector, and analyze the potential defect probability distribution matrix to obtain the basic values ​​for defect prediction.

[0055] The basic values ​​for defect prediction are derived using the following formula: (12) In formula (12), This represents the basic numerical value for defect prediction. The first element of the potential defect probability distribution matrix represents the... Line number Column elements, Indicates the number of rows in the matrix. The formula represents the number of columns in the matrix. This formula obtains the basic value for defect prediction by calculating the average value of the probability distribution matrix. The control logic of formula (12) is "global probability average + local fluctuation elimination + comprehensive quantification of defect risk". By calculating the arithmetic average of all elements of the potential defect probability distribution matrix, the probability information of the two-dimensional matrix is ​​compressed into a single scalar to obtain the global comprehensive defect risk benchmark value, which provides a concise and interpretable quantitative basis for subsequent defect prediction.

[0056] The first part of the potential defect probability distribution matrix Line number Column elements This can be derived from the following formula: (13) In formula (13), Indicates the first The and the first The dot product of the eigenvectors of the molten pool stability. Indicates temperature parameter, This represents the total number of samples in the molten pool stability feature vector. Indicates the first The formula sums the similarity weights of each sample with all samples, and generates a probability distribution matrix based on the similarity between feature vectors. The control logic of formula (13) is "feature vector similarity quantification + temperature parameter scaling + Softmax normalization + probability distribution generation". By calculating the dot product similarity between the feature vectors of melt pool stability, and adjusting the relative difference of similarity in combination with the temperature parameter, the similarity is finally converted into a probability value using the Softmax function, generating a probability distribution matrix that reflects the similarity relationship between feature vectors, providing a quantitative similarity basis for defect prediction.

[0057] Generating a potential defect probability distribution matrix based on the first molten pool stability feature vector is a process of mapping the feature vector to defect probability assessment. In one implementation, the system uses a pre-trained classification model to input the feature vector and output a probability distribution matrix, where each element represents the probability of a certain defect type. For example, the probability distribution matrix shows that the probability of cracks caused by molten pool overheating is 0.7, while the probability of porosity defects is 0.3, providing a quantitative basis for subsequent defect prediction. Finally, analyzing the potential defect probability distribution matrix to obtain the basis for defect prediction is a process of transforming probability data into actionable information. For example, in a welding quality monitoring system, the system generates a specific defect prediction report based on the defect type with the highest probability in the matrix, combined with a historical case database. This defect prediction report indicates that welding parameters need to be adjusted to reduce the risk of overheating, thus providing a reference for process optimization. Through this series of steps, from signal acquisition to defect prediction, a complete data processing and application chain is formed.

[0058] Furthermore, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S300 as follows: Step S310: Obtain the defect prediction base value. If the defect prediction base value exceeds the preset abnormal trigger threshold, then extract the second melt pool stability feature vector.

[0059] The following formula is used to define the criteria for extracting the eigenvectors of molten pool stability: (14) In formula (14), Indicates an exception trigger indicator. This represents the basic numerical value for defect prediction. This indicates the preset abnormal trigger threshold. The control logic of formula (14) is "defect risk threshold trigger + binary action control + abnormal response precise decision". By comparing the basic value of defect prediction with the preset abnormal trigger threshold, an abnormal trigger indicator of 0 / 1 is output, realizing the precise control of "whether to extract the second molten pool stability feature vector", which is the key node from the quantification of welding defect risk to action response.

[0060] The second molten pool stability eigenvector is obtained through the following formula: (15) In formula (15), This represents the second molten pool stability eigenvector. Indicates the standard deviation of the molten pool temperature. This represents the average width of the molten pool. The control logic of formula (15) is "multi-dimensional quantification of physical characteristics of the molten pool + structured integration of stability indicators + precise location of defect root causes". By integrating the statistics of the three core physical characteristics of the molten pool temperature fluctuation, width consistency and height fluctuation into a column vector, a stability description that is more focused on the direct state of the molten pool than the first feature vector is formed. This is used to analyze the specific sources of molten pool instability in depth after the defect risk is triggered.

[0061] High volatility of the molten pool This can be derived from the following formula: (16) In formula (16), This represents the average height of the molten pool. Indicates the maximum height. The minimum height is represented by this formula, which calculates the relative volatility of the peak-to-valley difference relative to the average value. The control logic of formula (16) is "absolute volatility range extraction + average value normalization + relative volatility intensity quantification". By calculating the peak-to-valley difference of the molten pool height (the difference between the maximum and minimum height) and normalizing it based on the average height, the dimensionless relative volatility is obtained. This fairly quantifies the intensity of the molten pool height fluctuation under different average heights, and directly reflects the uniformity of the weld reinforcement height.

[0062] The process of obtaining the baseline value for defect prediction typically involves the quantitative evaluation of multi-source signals during the welding process. For example, the system first calculates a comprehensive index from the temperature distribution of the molten pool and optical image data. This comprehensive index integrates multiple parameters, such as the average temperature gradient and the image grayscale change rate, using a weighted averaging method to form a value representing the risk level of potential defects. Assuming the system detects a molten pool temperature gradient of 15 degrees per second and an image grayscale change rate of 0.2, these parameters are input into a preset formula to generate a baseline value for defect prediction, such as 1.8. This value reflects the overall stability of the molten pool dynamics, providing a basis for subsequent decisions. For instance, if this baseline value for defect prediction exceeds a preset anomaly trigger threshold, such as 1.5, the system will initiate the step of extracting a second molten pool stability feature vector. For example, the extraction process includes obtaining frequency peaks from acoustic signals, extracting temperature peak distributions from thermal imaging, and calculating the edge curvature of the molten pool from optical data. These features are combined into a second molten pool stability feature vector, such as [frequency peak: 300 Hz, temperature peak: 1200 degrees, curvature: 0.05]. This second molten pool stability feature vector captures the multidimensional changes of the molten pool over time, helping to identify potential unstable modes.

[0063] Step S320: Input the second melt pool stability feature vector into the support vector machine classification model, and use the radial basis kernel function to map it to the high-dimensional feature space to obtain feature points.

[0064] The feature points implicitly mapped from the radial basis function kernel to the high-dimensional feature space are obtained by the following formula: (17) In formula (17), This represents the feature points implicitly mapped from the radial basis kernel to the high-dimensional feature space. Indicates the first One central point, Indicates the width parameter. The formula (17) represents the high-dimensional basis vector and is used to obtain the feature points in the high-dimensional space. The control logic of the formula (17) is "low-dimensional feature similarity quantification + radial basis kernel implicit mapping + high-dimensional space feature point generation". By calculating the similarity between the second molten pool stability feature vector and the preset center point, and combining the exponential transformation of the radial basis function and the weighting of the high-dimensional basis vector, the low-dimensional molten pool features are implicitly mapped to the high-dimensional feature space, so that the low-dimensional features that were originally linearly inseparable become linearly separable in the high-dimensional space, so that the support vector machine can classify welding defects more accurately.

[0065] When the second molten pool stability feature vector is input into the Support Vector Machine (SVM) classification model, SVM is a supervised learning algorithm that achieves classification by finding the optimal separating hyperplane between data points. In welding defect detection, this SVM classification model is trained to distinguish different defect types. After inputting the second molten pool stability feature vector, the system uses a radial basis function (RBF) kernel to map the second molten pool stability feature vector to a high-dimensional feature space. The RBF kernel is a kernel method that calculates the similarity between data points based on a Gaussian distribution, thereby handling nonlinear problems, such as extending two-dimensional features to three or higher dimensions to obtain new feature points that are more easily linearly separated in high-dimensional space.

[0066] Step S330: Calculate the geometric distance from the feature point to the optimal hyperplane, and determine the spatial position and assignment probability value relative to the decision boundary based on the geometric distance.

[0067] The geometric distance from a feature point to the optimal hyperplane is obtained using the following formula: (18) In formula (18), This represents the geometric distance from the feature point to the optimal hyperplane. This represents the transpose of the normal vector of the hyperplane. The bias term represents the hyperplane. The norm of the normal vector is represented. The control logic of formula (18) is "low-dimensional feature similarity quantification + radial basis kernel implicit mapping + high-dimensional space feature point generation". By calculating the similarity between the second molten pool stability feature vector and the preset center point, combined with the exponential transformation of the radial basis function and the weighting of the high-dimensional basis vector, the low-dimensional molten pool features are implicitly mapped to the high-dimensional feature space, so that the originally linearly inseparable low-dimensional features become linearly separable in the high-dimensional space, so that the support vector machine can classify welding defects more accurately.

[0068] The spatial location relative to the decision boundary is obtained by the following formula: (19) In formula (19), Indicates the spatial location relative to the decision boundary. The sign function is represented by a positive value representing one side and a negative value representing the other side. The control logic of formula (19) is "decision function sign extraction + binary space position determination + direct output of classification". The continuous decision function value of the support vector machine (SVM) is converted into a binary indicator through the sign function, and the spatial position of the high-dimensional feature point relative to the optimal hyperplane is directly determined, thereby completing the classification of welding status (such as "normal / defect").

[0069] The probability of attribution is calculated using the following formula: (20) In formula (20), The value represents the probability of belonging to the target class (range (0, 1)): the closer the value is to 1, the higher the confidence that the feature point belongs to the target class; the closer it is to 0, the lower the confidence; when it is equal to 0.5, it means that it is completely uncertain (the feature point is on the hyperplane). This represents the scaling parameter, which controls the steepness of the Sigmoid curve. The larger the value, the more drastic the change in probability when the distance changes (more sensitive to distance). The smaller the value, the smoother the change (and the less sensitive it is to distance). It represents the absolute value of the geometric distance from the feature point to the optimal hyperplane. It only reflects the distance and does not distinguish between positive and negative (i.e., it does not care which side of the hyperplane it is on, but only its distance from the boundary). The closer the value is to 1, the stronger the classification is to the side closer to the decision boundary. The control logic of formula (20) is "probabilistic mapping of geometric distance + smooth transformation of Sigmoid + quantitative output of confidence". The absolute value of the geometric distance from the feature point to the hyperplane is mapped to the (0,1) interval through the Sigmoid function to generate the classification probability value, which quantifies the confidence of the feature point belonging to a certain class (such as "defect" or "normal"). The farther the distance, the closer the classification probability is to 1, and the more reliable the classification result is.

[0070] Calculating the geometric distance from a feature point to the optimal hyperplane involves measuring the Euclidean distance between each point and the decision boundary. The optimal hyperplane is the classification boundary determined after model training. The distance calculation helps quantify which side a point is located on. If the distance is positive, it belongs to one class; if it is negative, it belongs to another class. Based on this distance, the spatial position relative to the decision boundary and the probability of belonging to the class are further determined. The probability is obtained by transforming the distance value through a logistic function. For example, if the distance of a feature point is 2.3, its probability is 0.85, belonging to the crack class.

[0071] Step S340: Select the category label corresponding to the largest item in the attribution probability values, and determine the specific defect type based on the category label.

[0072] The category label corresponding to the largest item in the attribution probability values ​​is obtained using the following formula: (twenty one) In formula (21), The selected category label indicates the final predicted defect type (such as "porosity", "lack of fusion", "normal welding", etc.), which directly corresponds to the interpretable defect category in production. This represents the category index; the formula determines the corresponding category label by selecting the item with the highest belonging probability value. The control logic of formula (21) is "multi-class probability comparison + maximum probability decision + precise defect type determination". The argmax operator finds the category index corresponding to the largest item among all the belonging probabilities, thereby determining the final defect type label. It is the most intuitive and statistically inferential decision rule in multi-class scenarios.

[0073] The specific defect type is determined by the following formula: (twenty two) In formula (22), It indicates the specific defect type, and the final interpretable output, such as "normal welding", "porosity defect", "lack of fusion defect", etc., is a conclusion that engineers can directly understand and take action on. This represents a mapping function, a pre-defined one-to-one correspondence (usually a dictionary or lookup table operation) that maps abstract category labels (such as indices or codes) within the algorithm to actual defect names or states. This formula is used to map category labels... The specific defect type is determined by the mapping function. The control logic of formula (22) is "mapping of abstract labels to industrial semantics + interpretability conversion + adaptation to production scenarios", which is achieved through a preset mapping function. The abstract category labels output by the algorithm The results are converted into specific defect types that can be directly understood in production, completing the final step from algorithm results to industrially interpretable conclusions, allowing defect determination results to directly guide on-site process adjustments.

[0074] When selecting the category label corresponding to the largest item in the probability values, the system compares the probabilities of all categories. For example, in the process of welding electronic components, if the probability distribution is [crack: 0.7, porosity: 0.2, lack of fusion: 0.1], then the crack label corresponding to 0.7 is selected. Based on this crack label, the specific defect type is determined, such as the crack originating from excessive laser power. This provides direct guidance for adjusting welding parameters and ensures continuous optimization of the process.

[0075] Preferably, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S400 as follows: Step S410: Match the optimization boundary conditions according to the determined specific defect type, and construct the initial population distribution of the particle swarm optimization algorithm.

[0076] The optimal boundary conditions are derived using the following formula: (twenty three) In formula (23), Indicates the specific defect type The corresponding optimization boundary conditions, the parameter search range of the particle swarm optimization algorithm (such as the upper and lower limits of welding current / voltage / gas flow rate), and different optimization boundaries for different defects ensure that the algorithm focuses on the key parameters for solving the current defect. The matching function is essentially a pre-defined rule base / lookup table function. It takes the defect type and matching parameters as input and outputs the corresponding process parameter optimization boundary. It serves as a bridge between factory process experience and algorithms. The matching parameters represent the empirical thresholds for auxiliary matching (such as the default process range of materials, plate thickness, and welding methods), which are used to dynamically adjust the accuracy of the boundary. The control logic of formula (23) is "defect type-driven boundary matching + empirical rule lookup + precise locking of optimization range". Through the preset matching function, the specific defect type is associated with the process experience accumulated by the factory to generate the parameter optimization boundary of the Particle Swarm Optimization (PSO) algorithm, so that the algorithm searches only within the range of process parameters directly related to the defect, which greatly improves the optimization efficiency and accuracy.

[0077] The initial population distribution of the particle swarm optimization algorithm is obtained by the following formula: (twenty four) In formula (24), This represents the initial population position distribution in the particle swarm optimization algorithm. This represents the initial number of particles in the population. This represents the Dirac function. Indicates the specific defect type. Indicates the first The initial position of each particle. This represents the parameter optimization boundary interval based on defect type matching. The control logic of formula (24) is "defect-driven boundary constraint + precise positioning of Dirac function + uniform probability distribution initialization". Through the positioning characteristics and averaging of Dirac function, it is ensured that all initial particles of the particle swarm optimization algorithm fall within the optimization boundary matching the specific defect, and a uniform initial position probability distribution is generated, so that the particle swarm optimization algorithm focuses on solving the parameter space of the current defect from the beginning, which greatly improves the optimization efficiency and convergence speed.

[0078] No. The initial position of each particle The initial position is obtained using the following formula: (25) In formula (25), In the particle swarm optimization algorithm, the first... The initial position of each particle. This indicates the lower boundary of parameter optimization. This indicates the upper boundary of parameter optimization. This represents element-wise multiplication. This represents a random vector that follows a uniform distribution from 0 to 1. The initial number of particles is represented by . The control logic of formula (25) is "uniform random sampling under boundary constraints + independent generation of each element + multi-parameter collaborative initialization". By generating uniformly distributed random offsets within the optimization boundary, it ensures that the initial position of each particle strictly falls within the range of process parameters corresponding to the defect, while ensuring the diversity of the initial population, thus laying the foundation for efficient optimization of the particle swarm algorithm.

[0079] When setting the parameter optimization boundary conditions for the particle swarm optimization algorithm based on the specific defect type, it is necessary to define the value range of variables for different defect characteristics. For example, when the defect type is determined to be porosity, the optimization boundary usually limits the welding current to between 80A and 160A, and the spacing combination is set to 0.5mm to 2.0mm to avoid the expansion of porosity defects caused by bubble aggregation. In actual welding production, the system first reads the defect classification result as porosity, and then automatically loads boundary conditions strongly related to porosity, that is, the lower limit of the current is biased towards a lower value to reduce excessive heat input, and the upper limit is to prevent the shielding gas from failing due to overcooling of the molten pool. Specifically, when constructing the initial population distribution of the particle swarm optimization algorithm, a random uniform initialization strategy is used to generate multiple candidate solution vectors, and each particle represents a parameter combination of current ratio and spacing combination. For example, in a welding task, the initial population size is set to 50 particles. The first particle may be encoded as having a current ratio of 80% for the main arc, 20% for the auxiliary arc, and a spacing of 1.2 mm. The second particle is randomly changed to have a current ratio of 75% for the main arc, 25% for the auxiliary arc, and a spacing of 0.8 mm. These initial positions are scattered in a hypercube space formed by pre-set boundary conditions to ensure that the population has sufficient diversity.

[0080] Step S420: Import the initial population distribution into the particle swarm optimization model, calculate the fitness value using the ideal melt pool features associated with specific defect types, and iteratively search for the global optimal solution vector based on the fitness value.

[0081] Fitness values ​​are derived using the following formula: (26) In formula (26), Represents particles The fitness value (range (0, 1]) is such that the closer the value is to 1, the better the process parameters corresponding to the particle can eliminate defects (the closer the molten pool characteristics are to the ideal state); the closer the value is to 0, the worse the effect of the parameters. Represents particles The position vector. Represents particles The corresponding number The characteristics of the molten pool (such as temperature standard deviation, width mean, height fluctuation rate, etc.) reflect the state of the molten pool under these process parameters. The first one indicates the association of specific defect types. The ideal characteristics of the molten pool (such as the ideal temperature fluctuation range without porosity and the ideal average width without incomplete fusion) are the target state for optimization. It represents the total number of feature dimensions of the molten pool (such as temperature, width, and height), comprehensively reflecting the stability of the molten pool. Indicates the first The standard deviation of the multidimensional feature (square of the standard deviation): used to normalize the deviation, eliminate the dimensional differences between different features (such as temperature in °C² and width in mm²), and make the deviations of different features comparable. The control logic of formula (26) is "multidimensional melt pool feature deviation quantification + normalization weighting + exponential smoothing mapping + fitness value generation". By calculating the deviation between the melt pool feature corresponding to the particle and the ideal melt pool feature associated with the defect, the fitness value is obtained after normalization and exponential transformation. This quantifies the ability of the process parameters corresponding to the particle to eliminate defects, and provides a clear optimization target for the iterative optimization of the particle swarm algorithm.

[0082] The global optimal solution vector is obtained by the following formula: (27) In formula (27), This represents the vector of the global optimal solution. The total number of particles is represented by the formula, which selects the optimal solution by the maximum fitness. The control logic of formula (27) is "comparison of fitness of the whole population + positioning of maximum fitness + locking of optimal process parameters". By traversing the fitness values ​​of all particles through the argmax operator, the position vector corresponding to the particle with the highest fitness is found and used as the global optimal solution vector. This vector directly corresponds to the process parameter combination that can most effectively eliminate the current defect. It is the final decision step of the particle swarm optimization algorithm.

[0083] After importing the initial population into the particle swarm optimization model, the fitness value of each particle is calculated using the ideal melt pool characteristics associated with the porosity defect type. Ideal melt pool characteristics typically include quantitative targets such as a stable melt pool width of approximately 4.5 mm, temperature fluctuations of less than 30℃, and a melt pool oscillation frequency below 8 Hz. The system acquires real-time melt pool images and thermal signals, simulates or maps the parameters corresponding to the current particle to historical data, and calculates the Euclidean distance from the ideal characteristics as the fitness; the smaller the distance, the higher the fitness. For example, a particle with simulated melt pool width of 4.2 mm and temperature fluctuations of 45℃ has a relatively poor fitness value, while another particle with a width close to 4.5 mm and a fluctuation of only 22℃ has a significantly better fitness. Preferably, during the iterative search based on the fitness value, the particle continuously updates its velocity and position according to its own historical best position and the global best position, converging after approximately 40 to 60 iterations. For example, after iterative search, the global optimal solution vector eventually converges to a combination of current ratio of 83% for the main arc, 17% for the auxiliary arc, and a spacing of 1.1 mm. The molten pool characteristics corresponding to this solution are closest to the ideal state.

[0084] Step S430: Decode the global optimal solution vector to obtain the current ratio adjustment step size and the spacing combination adjustment step size.

[0085] The current ratio adjustment step size is obtained using the following formula: (28) In formula (28), Indicates the current ratio adjustment step size. The first element of the global optimal solution vector represents the... Quantity, Indicates the number of bits used for decoding. This represents the maximum adjustment value of the current ratio. This formula is used to obtain the current ratio adjustment step size from the global optimal solution vector through binary decoding. The control logic of formula (28) is "binary encoding to decimal + normalization mapping + process parameter decoding". By converting the binary components of the global optimal solution vector into decimal integers and then normalizing and mapping them to the current ratio adjustment range allowed by the process, the current ratio adjustment step size that can be directly used for equipment adjustment is finally obtained, thus completing the "encoded solution → process value" conversion of the particle swarm optimization algorithm.

[0086] The spacing combination adjustment step size is obtained by the following formula: (29) In formula (29), This indicates the step size for adjusting the spacing combination. The first element of the global optimal solution vector represents the... Quantity, Indicates the number of bits used for decoding. The formula represents the maximum adjustment value of the spacing combination. This formula is used to obtain the spacing combination adjustment step size from the global optimal solution vector through binary decoding. The control logic of formula (29) is "binary encoding to decimal + normalization mapping + process parameter decoding", which is completely consistent with the decoding logic of the current ratio adjustment step size. By converting the binary components of the global optimal solution vector into decimal integers and then normalizing and mapping them to the spacing combination adjustment range allowed by the process, the spacing combination adjustment step size that can be directly used for equipment adjustment is finally obtained, thus completing the "encoded solution → process value" conversion of the particle swarm optimization algorithm.

[0087] When decoding the global optimal solution vector, the encoded current ratio is directly mapped to the ratio of the main arc current to the auxiliary arc current, while the spacing combination is converted into the adjustment step value of the relative position of the welding gun and the laser head. For example, after decoding, the current ratio adjustment step is increased or decreased by 2A each time, and the spacing combination adjustment step is moved by 0.1mm each time, thus forming a control command that can be directly issued.

[0088] Step S440: Adjust the step size and spacing combination according to the current ratio to generate an optimized welding parameter set. The welding parameter set includes the current ratio and spacing combination.

[0089] The optimized welding parameter set is obtained through the following formula: (30) In formula (30), This represents the optimized set of welding parameters. Indicates the current ratio. The formula (30) represents the spacing combination. It generates an optimized parameter set that includes the current ratio and spacing combination by adjusting the linear addition. The control logic of the formula (30) is "incremental adjustment of reference parameters + multi-parameter collaborative optimization + industrial-grade parameter set output". By superimposing the optimized adjustment step size on the current welding reference parameters, a welding parameter set that can be directly used for the equipment is generated, realizing the last step from algorithm optimization results to process implementation.

[0090] Based on the decoding results, an optimized welding parameter set is generated, including a main arc current of 85A, an auxiliary arc current of 20A, and a laser-arc distance of 1.1mm. This optimized welding parameter set, when input into the welding equipment, significantly reduces the probability of porosity defects, providing a stable and reliable process window for subsequent batch production.

[0091] Furthermore, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S500 as follows: Step S510: Obtain real-time welding current and arc voltage values ​​based on the welding parameter set, and obtain the instantaneous arc power sequence and short-circuit transition period based on the real-time welding current and arc voltage values.

[0092] The instantaneous arc power is obtained by the following formula: (31) In formula (31), Indicates the first Instantaneous arc power, Indicates the first Real-time arc voltage value. Indicates the first Real-time welding current value The formula (31) represents the sequence length and is used to generate the instantaneous arc power sequence. The control logic of the formula (31) is "synchronous electrical parameter sampling + direct calculation of instantaneous power + dynamic monitoring of time sequence". By synchronously sampling and multiplying the arc voltage and welding current at each moment, the instantaneous arc power is obtained, and then the power sequence is generated in time order to dynamically reflect the changes in heat input during the welding process, providing a core basis for subsequent analysis of molten pool stability and short-circuit transition period.

[0093] The short-circuit transition period is calculated using the following formula: (32) In formula (32), Indicates the duration of the short-circuit transition period. This indicates the number of sampling points under short-circuit conditions. The sampling time interval is represented by the formula used to calculate the short-circuit transition cycle duration based on real-time current and voltage detection. The control logic of formula (32) is "short-circuit state sampling point count + time interval weighting + precise quantization of cycle duration". By counting the number of short-circuit state sampling points in one short-circuit transition cycle and combining the sampling time interval, the actual cycle duration of the short-circuit transition is calculated, thereby quantifying the stability of the welding process and providing a basis for defect early warning.

[0094] The process of acquiring real-time welding current and arc voltage values ​​based on a welding parameter set typically involves a sensor system deployed on the welding equipment. These sensors monitor electrical signals during the welding process in real time. Specifically, the system first reads the optimized current ratio and spacing combination from the parameter set, for example, setting the main arc current to 90A, the auxiliary arc current to 15A, and the spacing to 1.0mm. Then, it collects data from actual operation through current transformers and voltage probes, obtaining a series of current values, such as fluctuations from 85A to 95A, and voltage values, such as changes from 20V to 25V, thus forming time-series data. This acquisition method ensures the timeliness and accuracy of the data, providing a foundation for subsequent calculations. For example, when calculating the instantaneous arc power sequence based on these real-time values, the power sequence is obtained by multiplying point by point using the power formula P=I×V, such as a power sequence from 1700W to 2200W within one cycle. At the same time, the short-circuit transition period is analyzed, that is, the time interval from arc short circuit to re-ignition, which is usually determined by detecting the zero-point crossing of the voltage waveform. For example, the duration of one cycle is 5ms to 10ms, which helps to evaluate the dynamic characteristics of the welding process.

[0095] Step S520: If the standard deviation of the instantaneous arc power sequence exceeds the threshold, the wire feeding speed compensation increment is determined based on the short-circuit transition period duration.

[0096] The standard deviation of the instantaneous arc power sequence is obtained by the following formula: (33) In formula (33), The standard deviation of the instantaneous arc power sequence is represented by the standard deviation of the arc power sequence. Indicates the sequence length. Indicates the first Instantaneous arc power, This represents the average value of the instantaneous arc power sequence. The control logic of formula (33) is "power sequence fluctuation quantification + root mean square deviation calculation + stability index generation". By calculating the standard deviation of the instantaneous arc power sequence, the severity of power fluctuation is quantified, which is used as the core index for judging the stability of the welding process. When the standard deviation exceeds the preset threshold, it indicates that the heat input fluctuation is too large, which will trigger adjustment actions such as wire feeding speed compensation.

[0097] The wire feed speed compensation increment is derived using the following formula: (34) In formula (34), This indicates the incremental compensation for the wire feeding speed. The compensation ratio coefficient is indicated. The control logic of formula (34) is "short circuit cycle deviation quantification + proportional coefficient mapping + wire feeding speed linear compensation". By using the preset compensation ratio coefficient, the short circuit transition cycle duration is directly converted into the compensation increment of the wire feeding speed, thereby dynamically adjusting the wire feeding speed, stabilizing the short circuit transition cycle, reducing arc power fluctuations, and improving the stability of the welding process.

[0098] If the standard deviation of the instantaneous arc power sequence exceeds a threshold, for example, if the calculated standard deviation is 200W when the threshold is set to 150W, then the wire feed speed compensation increment needs to be determined based on the short-circuit transition period duration. Specifically, if the period duration is too long, such as exceeding 8ms, it indicates that the droplet transition is not smooth. The system will calculate the compensation increment through a preset mapping function. For example, for every 1ms increase in duration, the wire feed speed increases by 0.5m / min, thus generating an increment value of 1.0m / min to optimize droplet formation.

[0099] Step S530: Generate a real-time wire feeding speed control command based on the wire feeding speed compensation increment to adjust the wire feeding speed of the wire feeding module.

[0100] The real-time wire feed speed control command is derived from the following formula: (35) In formula (35), This indicates a real-time wire feed speed control command. The formula represents the reference wire feeding speed. This formula is used to directly generate control commands to adjust the wire feeding speed. The control logic of formula (35) is "reference speed base + dynamic incremental compensation + real-time command generation". By superimposing the real-time calculated compensation increment on the preset reference wire feeding speed, a control command that the wire feeder can directly execute is generated, realizing the dynamic closed-loop adjustment of the wire feeding speed, thereby stabilizing the short-circuit transition period and reducing the arc power fluctuation.

[0101] When generating a real-time wire feeding speed control command based on the wire feeding speed compensation increment, the real-time wire feeding speed control command will be directly sent to the servo motor of the wire feeding module, adjusting the speed from the initial 10m / min to 11m / min. This can respond promptly to the instability caused by power fluctuations.

[0102] Step S540: Collect the arc voltage feedback waveform data after responding to the real-time wire feeding speed control command, calculate the distribution variance of the arc voltage feedback waveform data, and obtain the arc stability index after real-time adjustment.

[0103] The variance of the arc voltage feedback waveform data is obtained by the following formula: (36) In formula (36), This represents the variance of the arc voltage feedback waveform data. This indicates the number of sampling points for the arc voltage feedback waveform data. Indicates the first Arc voltage values ​​at each sampling point This represents the average arc voltage at all sampling points. The control logic of formula (36) is "voltage fluctuation quantification + root mean square deviation calculation + stability index generation". By calculating the distribution variance of the arc voltage feedback waveform, the severity of voltage fluctuation is quantified, which serves as a verification index for arc stability after wire feed speed adjustment. The smaller the distribution variance of the arc voltage feedback waveform data, the more stable the voltage and the more stable the welding process; the larger the distribution variance of the arc voltage feedback waveform data, the more severe the fluctuation, requiring further parameter optimization.

[0104] The real-time adjusted arc stability index is obtained through the following formula: (37) In formula (37), This indicates the arc stability index after real-time adjustment. This indicates the number of arc voltage waveform data points collected. Indicates the number of times after responding to the wire feed speed control command. One arc voltage feedback value, Indicates the number of times after responding to the wire feed speed control command. The average value of the arc voltage feedback values. The control logic of formula (37) is "reciprocal mapping of voltage fluctuation + positive correlation quantification of stability + generation of intuitive index". By calculating the reciprocal of the standard deviation of the arc voltage feedback waveform, the fluctuation amplitude is converted into an index positively correlated with stability: the larger the index value, the smaller the voltage fluctuation and the more stable the arc; the smaller the index value, the larger the fluctuation and the worse the stability. This design makes the quantification result of stability more intuitive (large value = stable, small value = unstable).

[0105] After acquiring the arc voltage feedback waveform data following the response control command, the process of calculating the distribution variance involves statistical analysis of the waveform samples. For example, by acquiring data from 100 voltage points, the variance value is calculated to be, say, 2.5V², thereby obtaining the arc stability index after real-time adjustment. Specifically, in actual operation, if the variance is less than the threshold of 1.0V², the index is displayed as "stable." This not only reflects the effectiveness of wire feed adjustment but also provides a quantitative basis for welding quality control, such as reducing spatter rate and improving weld uniformity in continuous production, ensuring the reliability of the overall process.

[0106] Preferably, the intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment includes step S600 as follows: Step S610: Combine the real-time adjusted arc stability index with the welding torch spatial coordinate data to construct a weld deviation feature vector, and calculate the weld offset degree and trajectory deviation direction based on the weld deviation feature vector.

[0107] The weld deviation eigenvector is obtained by the following formula: (38) In formula (38), It represents the weld deviation feature vector, a 4-dimensional column vector that uniformly encodes arc stability and spatial position deviation, and is the core input for subsequent weld trajectory correction. This represents the arc stability index after real-time adjustment. It is derived from the calculation results of the previous steps and reflects the degree of arc fluctuation (the larger the value, the higher the stability), indirectly characterizing the weld formation quality. Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation, Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation, Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation. Indicates the spatial coordinate deviation of the welding torch: the welding torch in The deviation of the direction from the weld reference position (unit: mm) directly quantifies the positional offset of the welding torch (e.g. Indicates the welding torch is in (2mm to the right). This formula is used to construct a weld deviation feature vector by combining the arc stability index and the welding torch spatial coordinate data. The control logic of formula (38) is "multi-source information fusion + structured feature encoding + unified representation of deviation and stability", which integrates the arc index reflecting the dynamic stability of welding quality with the spatial coordinate deviation reflecting the geometric deviation of the welding torch position into a 4-dimensional feature vector, providing a complete input including "quality + position" for subsequent calculation of weld offset degree and determination of trajectory deviation direction.

[0108] The weld offset value is obtained using the following formula: (39) In formula (39), The value represents the degree of weld offset. It is a dimensionless (or unit-based) value that combines quality stability and spatial position. The larger the value, the more serious the overall deviation. It is the core basis for judging the priority of subsequent corrections. This indicates the arc stability index after real-time adjustment. Indicates welding torch Directional trajectory deviation, Indicates welding torch Directional trajectory deviation, Indicates welding torch Directional trajectory deviation. Indicates welding torch Directional trajectory deviation: Geometric deviation component of spatial position (unit: mm), directly quantifying the offset between the welding torch and the weld reference position. This formula calculates the degree of offset based on the weld deviation characteristics. The control logic of formula (39) is "Euclidean distance calculation of multi-dimensional deviation + unified quantification of quality stability and spatial position", treating the arc stability index and three-dimensional spatial position deviation as components of a 4-dimensional vector, and obtaining a comprehensive weld offset value by calculating the Euclidean distance. The larger this value, the more serious the overall deviation of the welding process in "quality stability" and "spatial position", and the higher the priority of correction.

[0109] When constructing a weld deviation feature vector by combining the real-time adjusted arc stability index with the welding torch spatial coordinate data, it typically involves vector fusion of stability indices such as voltage variance and coordinate data such as three-dimensional position points. Specifically, the system first obtains stability indices from the previous step, for example, a variance of 1.5V², and then collects real-time coordinate data of the welding torch, such as x=100mm, y=50mm, z=20mm. This data is then integrated into a multi-dimensional feature vector using vector construction methods. For example, the weld deviation feature vector might be in the form of [1.5, 100, 50, 20]. This helps capture deviation information during the welding process. This construction process ensures the comprehensiveness of the feature vector, providing a reliable foundation for subsequent analysis. For instance, the process of calculating the weld offset degree and trajectory deviation direction based on the weld deviation feature vector involves evaluating the offset degree using the Euclidean distance formula. For example, the distance between the weld deviation feature vector and the ideal trajectory vector yields an offset value of 2.0mm. Simultaneously, vector direction difference analysis determines the deviation direction, such as a 30-degree rightward offset, reflecting the deviation of the actual welding path.

[0110] Step S620: If the weld offset value exceeds the preset threshold, a trajectory compensation matrix is ​​generated based on the trajectory deviation direction, and the trajectory compensation matrix is ​​converted into a correction control command to drive the multi-axis positioning mechanism to adjust the welding torch posture.

[0111] The following formula is used to define the criteria for generating the trajectory compensation matrix: (40) In formula (40), This represents the compensation trigger flag, a binary variable (0 or 1), which serves as the switch signal for generating the trajectory compensation matrix; 1 indicates that compensation needs to be triggered, and 0 indicates that compensation is not required. The threshold value represents the maximum allowable deviation in the process (e.g., 2.0). If this value is exceeded, the welding quality risk is considered to have increased and compensation is required. The control logic of formula (40) is "threshold comparison + binary switch trigger". By comparing the current weld deviation with the safety threshold preset in the process, a binary compensation trigger flag is output to determine whether to generate a trajectory compensation matrix, thereby realizing the on-demand correction of weld deviation.

[0112] The trajectory compensation matrix is ​​obtained through the following formula: (41) In formula (41), This represents the trajectory compensation matrix, a 3×3 rotation transformation matrix used to correct the trajectory direction deviation of the welding torch in the xy plane. The direction angle of the trajectory deviation is represented by the characteristic vector of the weld deviation. Calculated (e.g.) ), indicating the angle at which the welding torch deviates from the reference trajectory of the weld. , Indicates direction angle The trigonometric function values ​​are used to construct a two-dimensional rotation matrix to achieve angle compensation in the xy plane. The control logic of formula (41) is "two-dimensional plane rotation compensation + z-axis coordinate preservation + precise trajectory direction correction". By constructing a 3×3 rotation transformation matrix, the angle deviation of the welding torch in the xy plane is corrected, while keeping the z-axis (welding torch height) coordinate unchanged, so as to achieve precise compensation of the weld trajectory.

[0113] The corrective control command is derived from the following formula: (42) In formula (42), The correction control command is a drive signal output to a multi-axis positioning mechanism (such as an industrial robot or a cross slide) to directly control the motor's movement and adjust the welding torch's xy-plane rotation angle and posture. This represents the conversion gain matrix, typically with a dimension of "control command dimension × 9" (e.g., a 3×9 matrix for a 3-axis mechanism). Its functions are: 1. Unit conversion: converting the dimensionless rotation information of the matrix into the number of motor pulses / voltage values; 2. Gain amplification: adapting to the power characteristics of the equipment and adjusting the response speed of the correction action; 3. Dimension matching: mapping the 9-dimensional matrix vector to the control command dimension of the multi-axis mechanism. The vectorized form of the trajectory compensation matrix is ​​represented by expanding the 3×3 matrix into a 9-dimensional column vector by columns (or rows), encoding the two-dimensional rotation information into a one-dimensional vector, which facilitates multiplication with the gain matrix. This formula is used to drive the multi-axis positioning mechanism to adjust the welding torch posture. The control logic of formula (42) is "matrix vectorization encoding + gain matrix mapping + multi-axis drive command generation", which converts the rotation transformation information of the trajectory compensation matrix into drive commands that the multi-axis positioning mechanism can directly execute, thereby achieving precise correction of the welding torch posture.

[0114] If the weld offset exceeds a preset threshold, such as 1.5mm, the system will construct a transformation matrix, such as a 3×3 matrix containing rotation and translation components, to correct the offset direction when generating the trajectory compensation matrix based on the trajectory deviation direction. Specifically, if the offset value is 2.5mm and the direction is upward, the system will calculate a compensation amount, such as a rotation angle of -30 degrees and a translation distance of -2.0mm, thereby generating an accurate trajectory compensation matrix. The process of converting the trajectory compensation matrix into correction control commands to drive the multi-axis positioning mechanism to adjust the welding torch posture involves mapping the trajectory compensation matrix parameters to control signals. For example, the trajectory compensation matrix is ​​converted into a command sequence such as "rotation axis A -30 degrees, translation axis X -2.0mm", which is then sent to the multi-axis robot arm to adjust the welding torch from its current position to the corrected posture. This enables real-time optimization of the welding trajectory.

[0115] Step S630: Collect feedback data after responding to the correction control command, calculate the dynamic balance coefficient of the molten pool, and perform statistical analysis on the dynamic balance coefficient of the molten pool to obtain the final weld formation consistency data.

[0116] The final weld formation consistency data is obtained using the following formula: (43) In formula (43), The final weld formation consistency data is represented by the value range (0, 1]. The closer the value is to 1, the more uniform the weld formation; the closer it is to 0, the worse the formation consistency. The number of samples for the equilibrium coefficient indicates the number of samples of dynamic equilibrium coefficients of the molten pool. The larger the number of samples, the more stable and reliable the results. Indicates the first The dynamic equilibrium coefficient of the molten pool reflects the stability of the molten pool at each moment (such as the degree of balance of temperature and shape), and is the basic data for forming consistency. The average value of the dynamic equilibrium coefficient of the molten pool is used as the benchmark for fluctuation calculation to eliminate the influence of the overall equilibrium level on the fluctuation. The control logic of formula (43) is "relative fluctuation quantification + reverse mapping + consistency index generation". By calculating the coefficient of variation (standard deviation / mean) of the dynamic equilibrium coefficient of the molten pool, and then subtracting the coefficient of variation from 1, a weld formation consistency index with a value between (0, 1) is obtained. The closer the index value is to 1, the smaller the relative fluctuation of the dynamic equilibrium coefficient of the molten pool and the more uniform the weld formation; the closer it is to 0, the larger the fluctuation and the worse the formation consistency.

[0117] No. Dynamic equilibrium coefficient of each molten pool This can be derived from the following formula: (44) In formula (44), This indicates the number of feedback data points after correction, the number of sampling points for molten pool parameters (such as temperature, width, and area). The more points, the more stable the result. Indicates the first Feedback data values, and real-time parameters of the molten pool (such as temperature and width) collected after correction. The reference molten pool parameter value is the ideal molten pool parameter preset by the process (such as ideal temperature and width), which serves as the benchmark for the degree of balance. The formula calculates the dynamic balance degree by averaging the relative deviations of the feedback data. The control logic of formula (44) is "relative deviation quantification + reverse mapping to matching degree + averaging of multiple data to calculate balance". By calculating the relative deviation between the feedback data of the molten pool after correction and the reference value, it is reverse mapped to "matching degree", and then the matching degree of multiple data is averaged to obtain the dynamic balance coefficient of the molten pool. The closer the coefficient is to 1, the closer the molten pool is to the ideal state and the higher the degree of dynamic balance; the closer it is to 0, the greater the deviation and the lower the degree of balance.

[0118] Average value of dynamic equilibrium coefficient of molten pool This can be derived from the following formula: (45) In formula (45), Indicates the first The formula is used to perform average statistical analysis on multiple molten pool dynamic balance coefficients. The control logic of formula (45) is "multi-sample arithmetic mean + overall balance level quantification + statistical stability improvement". By performing arithmetic mean on multiple molten pool dynamic balance coefficients, the average dynamic balance level of the molten pool in the entire welding process is obtained, providing a stable statistical benchmark for subsequent calculation of weld formation consistency.

[0119] When calculating the dynamic balance coefficient of the molten pool using feedback data collected after receiving the corrective control command, visual sensors are used to acquire molten pool image data. For example, the variation in molten pool width, such as from 5mm to 6mm, is analyzed. Then, the balance formula, such as the average width divided by the standard deviation, yields a coefficient value of 0.8, which quantifies the dynamic stability of the molten pool. Specifically, the process of statistically analyzing the dynamic balance coefficient of the molten pool to obtain the final weld formation consistency data includes calculating the mean and variance of multiple coefficient samples. For example, the mean of the sample sequence [0.8, 0.85, 0.9] is 0.85, and the variance is 0.0025, thus obtaining consistency data such as a 95% uniformity rate. This provides a quantitative indicator for welding quality assessment, ensuring the reliability of the production process.

[0120] Please see Figure 2 This embodiment provides an intelligent control system for the quality of double-wire submerged arc welding of stainless steel reactors, used to execute the aforementioned intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors. It includes a multi-source fusion dataset formation module 10, a defect prediction basis acquisition module 20, a specific defect type determination module 30, a welding parameter set output module 40, an arc stability index acquisition module 50, and a final weld formation consistency data acquisition module 60. The multi-source fusion dataset formation module 10 is used to acquire molten pool morphology and temperature distribution information from the double-wire submerged arc welding process in a stainless steel reactor using optical vision sensors and thermal imaging devices, and combine this information with arc voltage and current signals to form a multi-source fusion dataset. The defect prediction basis acquisition module 20 is used to process the multi-source fusion dataset using signal entropy analysis to extract molten pool stability features and potential defect indicators, thus obtaining the defect prediction basis. The specific defect type determination module 30 is used to classify the extracted molten pool stability features using machine learning classification methods if the defect prediction basis exceeds a preset threshold, thereby determining the specific defect type. The system includes: a welding parameter set output module 40, which inputs specific defect types into a swarm intelligent optimization prediction model, calculates the adjustment values ​​of welding parameters, and outputs an optimized welding parameter set, which includes current ratio and spacing combination; an arc stability index acquisition module 50, which adjusts the wire feeding speed of the wire feeding module based on the welding parameter set using an intelligent control method, and acquires the real-time adjusted arc stability index; and a final weld formation consistency data acquisition module 60, which judges the degree of weld deviation from the acquired arc stability index. If the deviation exceeds a threshold, it is automatically corrected by a multi-axis positioning mechanism to obtain the final weld formation consistency data.

[0121] The intelligent control method and system for the quality of double-wire submerged arc welding of stainless steel reactors provided in this embodiment, compared with the prior art, acquires molten pool morphology and temperature distribution data through optical vision sensors and thermal imaging devices, combines them with arc signals to form a multi-source dataset, uses signal entropy analysis to extract stability features and defect indicators, and determines the defect type through machine learning when the defect prediction exceeds the threshold. Subsequently, the data is input into a swarm intelligence optimization prediction model to calculate welding parameter adjustment values, outputting optimized current ratios and spacing combinations. Intelligent control is then used to adjust the wire feeding speed and monitor arc stability indicators in real time. If the weld deviation exceeds the limit, a multi-axis positioning mechanism automatically corrects the deviation, ultimately achieving consistent weld formation. This embodiment significantly improves welding quality and stability and reduces the defect incidence rate through multi-source data fusion, intelligent defect prediction, and dynamic parameter optimization.

[0122] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for intelligent quality control of double-wire submerged arc welding in stainless steel reactors, characterized in that, Includes the following steps: S100: Acquires molten pool morphology and temperature distribution information from the double-wire submerged arc welding process of stainless steel reactor through optical vision sensors and thermal imaging devices, and forms a multi-source fusion dataset by combining arc voltage and current signals. S200. The multi-source fusion dataset is processed using the signal entropy analysis method to extract the stability characteristics of the molten pool and potential defect indicators, and to obtain the basic values ​​for defect prediction. S300. If the defect prediction baseline value exceeds a preset threshold, the extracted melt pool stability features are classified using a machine learning classification method to determine the specific defect type. S400. Input the specific defect type into the swarm intelligence optimization prediction model, calculate the adjustment value of the welding parameters, and output the optimized welding parameter set, which includes the current ratio and spacing combination. S500: The wire feeding speed of the wire feeding module is adjusted based on the welding parameter set using an intelligent control method to obtain the real-time adjusted arc stability index. S600. Determine the degree of weld deviation from the obtained arc stability index. If the deviation exceeds the threshold, the multi-axis positioning mechanism will automatically correct the deviation to obtain the final weld formation consistency data.

2. The intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to claim 1, characterized in that, Step S100 includes: S110. Acquire optical image sequences, thermal infrared image sequences, arc voltage signals, and welding current signals collected by optical vision sensors and thermal imaging devices. S120. Extract the geometric feature data of the molten pool morphology based on the optical image sequence, and map the geometric feature data of the molten pool morphology onto the thermal infrared image sequence to obtain a temperature distribution matrix corresponding to the spatial data of the geometric feature data of the molten pool morphology. S130. The temperature distribution matrix, the geometric feature data of the molten pool morphology, the arc voltage signal and the welding current signal are time-aligned to form a multi-source fusion dataset containing molten pool morphology information, temperature distribution information and arc voltage and current signals.

3. The intelligent quality control method for double-wire submerged arc welding of stainless steel reactors according to claim 1, characterized in that, Step S200 includes: S210. Obtain a multidimensional discrete time-series signal sequence, wherein the multidimensional discrete time-series signal sequence is generated by reconstructing the acoustic-optical-thermal multidimensional sensing values ​​in the multi-source fusion dataset. S220. The multidimensional discrete time-series signal sequence is processed using the signal entropy analysis method to obtain the signal entropy evolution sequence. S230. If the fluctuation amplitude of the signal entropy value evolution sequence exceeds the steady-state reference range value, then extract the first melt pool stability feature vector. S240. Generate a potential defect probability distribution matrix based on the first molten pool stability feature vector, and analyze the potential defect probability distribution matrix to obtain the basic value for defect prediction.

4. The intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to claim 3, characterized in that, Step S300 includes: S310. Obtain the defect prediction base value. If the defect prediction base value exceeds the preset abnormal trigger threshold, extract the second melt pool stability feature vector. S320. Input the second molten pool stability feature vector into the support vector machine classification model, and use the radial basis kernel function to map it to a high-dimensional feature space to obtain feature points; S330. Calculate the geometric distance from the feature point to the optimal hyperplane, and determine the spatial position and belonging probability value relative to the decision boundary based on the geometric distance; S340. Select the category label corresponding to the largest item in the attribution probability values, and determine the specific defect type based on the category label.

5. The intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to claim 1, characterized in that, Step S400 includes: S410. Match the optimization boundary conditions according to the determined specific defect type, and construct the initial population distribution of the particle swarm optimization algorithm; S420. The initial population distribution is imported into the particle swarm optimization model, and the fitness value is calculated using the ideal melt pool features associated with the specific defect type. The global optimal solution vector is obtained by iterative search based on the fitness value. S430. Decode the global optimal solution vector to obtain the current ratio adjustment step size and the spacing combination adjustment step size; S440. Adjust the step size according to the current ratio and spacing combination to generate an optimized welding parameter set, wherein the welding parameter set includes the current ratio and spacing combination.

6. The intelligent quality control method for double-wire submerged arc welding of stainless steel reactors according to claim 1, characterized in that, Step S500 includes: S510. Based on the welding parameter set, obtain the real-time welding current value and arc voltage value, and obtain the instantaneous arc power sequence and short-circuit transition period duration according to the real-time welding current value and arc voltage value. S520. If the standard deviation of the instantaneous arc power sequence exceeds the threshold, the wire feed speed compensation increment is determined according to the short-circuit transition period duration. S530. Generate a real-time wire feeding speed control command based on the wire feeding speed compensation increment to adjust the wire feeding speed of the wire feeding module. S540. Collect the arc voltage feedback waveform data after responding to the real-time wire feeding speed control command, calculate the distribution variance of the arc voltage feedback waveform data, and obtain the arc stability index after real-time adjustment.

7. The intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to claim 1, characterized in that, Step S600 includes: S610. Construct a weld deviation feature vector by combining the real-time adjusted arc stability index and the welding torch spatial coordinate data, and calculate the weld offset degree and trajectory deviation direction based on the weld deviation feature vector. The weld deviation eigenvector is obtained by the following formula: ； in, This represents the characteristic vector of weld deviation. This indicates the arc stability index after real-time adjustment. Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation, Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation, Represents the spatial coordinates of the welding torch Components and weld reference positions Component deviation; The weld offset value is obtained using the following formula: ； in, This indicates the numerical value representing the degree of weld offset. Indicates welding torch Directional trajectory deviation, Indicates welding torch Directional trajectory deviation, Indicates welding torch Directional trajectory deviation; S620. If the weld offset value exceeds a preset threshold, a trajectory compensation matrix is ​​generated based on the trajectory deviation direction, and the trajectory compensation matrix is ​​converted into a correction control command to drive the multi-axis positioning mechanism to adjust the welding torch posture. S630. Collect feedback data after responding to the correction control command, calculate the dynamic balance coefficient of the molten pool, and perform statistical analysis on the dynamic balance coefficient of the molten pool to obtain the final weld formation consistency data.

8. The intelligent control method for the quality of double-wire submerged arc welding of stainless steel reactors according to claim 7, characterized in that, In step S620, the following formula is used to define the judgment condition for generating the trajectory compensation matrix: ； in, Indicates the compensation trigger flag. This indicates the preset offset threshold. The trajectory compensation matrix is ​​obtained through the following formula: ； in, Represents the trajectory compensation matrix. Indicates the direction angle of trajectory deviation; The corrective control command is derived from the following formula: ； in, This indicates a correction control command. Represents the transformation gain matrix. This represents the vectorized form of the trajectory compensation matrix.

9. The intelligent quality control method for double-wire submerged arc welding of stainless steel reactors according to claim 8, characterized in that, In step S630, the final weld formation consistency data is obtained using the following formula: ； in, This indicates the final weld formation consistency data. Indicates the number of samples for the balance coefficient. Indicates the first The dynamic balance coefficient of the molten pool This represents the average value of the dynamic equilibrium coefficient of the molten pool.

10. A smart control system for the quality of double-wire submerged arc welding of stainless steel reactors, used to execute the smart control method for the quality of double-wire submerged arc welding of stainless steel reactors as described in any one of claims 1 to 9, characterized in that, include: The multi-source fusion dataset forming module (10) is used to obtain molten pool morphology information and temperature distribution information from the double-wire submerged arc welding process of stainless steel reactor through optical vision sensors and thermal imaging devices, and combine them with arc voltage and current signals to form a multi-source fusion dataset. The defect prediction basis acquisition module (20) is used to process the multi-source fusion dataset using the signal entropy analysis method, extract the melt pool stability characteristics and potential defect indicators, and obtain the defect prediction basis; The specific defect type determination module (30) is used to classify the extracted melt pool stability features using a machine learning classification method to determine the specific defect type if the defect prediction basis exceeds a preset threshold. The welding parameter set output module (40) is used to input the defect type into the prediction model of the group intelligent optimization, calculate the adjustment value of the welding parameters, and output the optimized welding parameter set, which includes the current ratio and spacing combination. The arc stability index acquisition module (50) is used to adjust the wire feeding speed of the wire feeding module based on the welding parameter set through an intelligent control method, and to acquire the arc stability index after real-time adjustment. The final weld formation consistency data acquisition module (60) is used to determine the degree of weld deviation from the acquired arc stability index. If the deviation exceeds the threshold, the multi-axis positioning mechanism will automatically correct the deviation to obtain the final weld formation consistency data.

Images