Real-time closed-loop regulation and control method and system for stainless steel welded pipe forming process

Abstract

The invention discloses a real-time closed-loop regulation and control method and system for the forming process of a stainless steel welded pipe. The method comprises the steps that multi-source heterogeneous data including a pipe shape image, a laser ranging sequence, temperature field distribution and roller pressure time sequence data in the forming process are synchronously collected through a distributed multi-source sensor network; performing space-time alignment and feature level fusion on the multi-source heterogeneous data, and constructing a multi-dimensional dynamic digital twinborn body in the forming process; on the basis of the multi-dimensional dynamic digital twins, an online rolling prediction model is adopted to deduce the development trend of weld forming quality and pipe diameter size deviation in real time; and according to the development trend, a cooperative regulation and control instruction set of the roller gap and the welding power is generated through a self-adaptive fuzzy inference system. By means of the embodiment of the invention, multivariable look-ahead perception and intelligent collaborative closed-loop regulation and control of technological parameters in the stainless steel welded pipe forming process can be achieved, and the stability and consistency of the forming quality are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of welded pipe forming process, in particular to a real-time closed-loop regulation method and system for stainless steel welded pipe forming process. BACKGROUND

[0002] The forming process of stainless steel welded pipe involves multiple processes such as rolling, welding, sizing, etc., and its quality control highly depends on the dynamic stability of process parameters. Traditional regulation methods mainly use independent control strategies based on single sensors or fixed models, which are difficult to comprehensively perceive the complex interaction and real-time changes of multiple variables such as weld formation, pipe diameter size, temperature field, etc. in the forming process. Especially when dealing with material fluctuations and working condition disturbances, the existing methods generally lack global and forward-looking cognition of the process state, resulting in lagging regulation, limited precision, and easy occurrence of size out-of-tolerance, welding defects, etc. With the development of industrial intelligence, introducing digital twinning and multi-sensor fusion technology to achieve more accurate process regulation has become an urgent need in the industry. However, existing technologies still lack systematic solutions in efficient fusion of multi-source heterogeneous data, online rolling prediction based on dynamic digital models, and multi-parameter collaborative adaptive closed-loop regulation. SUMMARY

[0003] The purpose of the present application is to provide a real-time closed-loop regulation method and system for stainless steel welded pipe forming process to solve the problems in the prior art and achieve multi-variable forward-looking perception and intelligent collaborative closed-loop regulation of process parameters for stainless steel welded pipe forming process, thereby improving the stability and consistency of forming quality.

[0004] One embodiment of the present application provides a real-time closed-loop regulation method for stainless steel welded pipe forming process, which comprises: synchronously collecting multi-source heterogeneous data including pipe morphology images, laser ranging sequences, temperature field distribution, and rolling pressure time series data in the forming process through a distributed multi-source sensor network; performing spatio-temporal alignment and feature-level fusion on the multi-source heterogeneous data to construct a multi-dimensional dynamic digital twin of the forming process; based on the multi-dimensional dynamic digital twin, using an online rolling prediction model to real-time deduce the development trend of weld formation quality and pipe diameter size deviation; according to the development trend, generating a collaborative regulation instruction set for roll gap and welding power through an adaptive fuzzy reasoning system to realize online closed-loop regulation of forming process parameters.

[0005] Another embodiment of the present application provides a real-time closed-loop regulation system for stainless steel welded pipe forming process, which comprises: The acquisition module is used for synchronously acquiring multi-source heterogeneous data including pipe material appearance images, laser ranging sequences, temperature field distribution and rolling pressure time sequence data in the forming process through a distributed multi-source sensing network; The construction module is used for performing time-space alignment and feature-level fusion on the multi-source heterogeneous data, and constructing a multi-dimensional dynamic digital twin of the forming process; The deduction module is used for deducing the development trend of the weld forming quality and the pipe diameter size deviation in real time based on the multi-dimensional dynamic digital twin and by using an online rolling prediction model; The generation module is used for generating a collaborative control instruction set of the roll gap and the welding power according to the development trend and by using an adaptive fuzzy inference system, so as to realize online closed-loop control of the forming process parameters.

[0006] Another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any one of the above embodiments when running.

[0007] Another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method described in any one of the above embodiments.

[0008] Compared with the prior art, the real-time closed-loop control method for the stainless steel welded pipe forming process provided by the present application can realize multi-variable forward-looking perception and intelligent collaborative closed-loop control of process parameters for the stainless steel welded pipe forming process, and improve the stability and consistency of the forming quality. BRIEF DESCRIPTION OF DRAWINGS

[0009] Figure 1 The present application provides a hardware structure block diagram of a computer terminal for the real-time closed-loop control method for the stainless steel welded pipe forming process; Figure 2 The present application provides a flowchart of the real-time closed-loop control method for the stainless steel welded pipe forming process; Figure 3 The present application provides a structure diagram of a real-time closed-loop control system for the stainless steel welded pipe forming process. DETAILED DESCRIPTION

[0010] The embodiments described below with reference to the drawings are exemplary and are only used to explain the present application, and cannot be explained as a limitation of the present application.

[0011] The present application provides a real-time closed-loop control method for the stainless steel welded pipe forming process, which can be applied to an electronic device, such as a computer terminal, specifically, a general computer, etc.

[0012] The following will be described in detail taking the computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal for providing a stainless steel welded pipe forming process real-time closed-loop regulation method is provided in the embodiments of the present application. As shown in the figure, Figure 1 The computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0013] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions which, when executed, can cause the processor to execute any stainless steel welded pipe forming process real-time closed-loop regulation method.

[0014] The processor is used to provide computing and control capabilities to support the operation of the entire computer device.

[0015] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any stainless steel welded pipe forming process real-time closed-loop regulation method.

[0016] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art can understand that, Figure 1 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0017] It should be understood that the processor can be a central processing unit (CPU), and the processor can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0018] Referring to Figure 2 The embodiments of the present application provide a stainless steel welded pipe forming process real-time closed-loop regulation method, which can include the following steps: S201, synchronously collecting multi-source heterogeneous data including pipe morphology image, laser ranging sequence, temperature field distribution and rolling pressure time series data in the distributed multi-source sensing network synchronous acquisition integrated process; Specifically, a multi-sensor array can be designed and deployed, including a high-speed industrial camera, a laser displacement sensor array, an infrared thermal imager and a distributed pressure sensor, which are respectively installed at the inlet, welding area, sizing area and outlet station of the forming machine to generate a sensor spatial layout scheme. The core of this step is to realize the global coverage collection of key physical quantities in the forming process by scientifically laying out multiple types of sensors, to provide a comprehensive data source for subsequent data fusion and regulation. The specific implementation is as follows: The design of the sensor array takes the principle of “workstation requirement-sensor characteristic” matching as the principle. For the four core workstations of the inlet, welding area, sizing area and outlet of the forming machine, the appropriate sensor type is configured respectively. The core requirement of the inlet workstation is to monitor the initial morphology and outer diameter reference of the pipe. One high-speed industrial camera and four laser displacement sensors are deployed. The industrial camera is installed 0.8 m above the inlet, with the lens vertically downward aiming at the center of the pipe. The parameter settings are resolution 2048x1536 pixels, frame rate 100 fps (to ensure capturing clear images of fast-moving pipes), and exposure time 50 μs (to reduce motion blur). The four laser displacement sensors are evenly distributed around the circumference of the pipe (at an interval of 90°), installed at a height of 0.5 m, with a measurement range of 50-300 mm, an accuracy of ±0.01 mm, and real-time scanning of the outer diameter profile of the pipe inlet end.

[0019] The welding area needs to monitor the weld morphology, temperature field and welding area pressure. Two high-speed industrial cameras are deployed (installed 0.6 m on both sides of the welding area at an angle of 45°, aiming at the weld, capturing the front and back surface morphology of the weld), one infrared thermal imager (installed 1.0 m above the welding area, measuring temperature range -20℃~1500℃, measuring temperature accuracy ±2%, resolution 640x480 pixels, sampling frequency 50Hz, real-time acquisition of weld and heat affected zone temperature distribution), and four distributed pressure sensors (embedded in the upper and lower left and right roller surfaces of the welding area, measuring range 0-50MPa, accuracy ±0.1MPa, sampling frequency 1kHz, monitoring the contact pressure of the roller and the pipe).

[0020] The core requirement of the sizing area is to control the pipe diameter size accuracy. Eight laser displacement sensors are deployed (evenly distributed around the circumference at an interval of 45°, installed at a height of 0.4 m in the middle of the sizing area, measuring range 80-250 mm, accuracy ±0.005 mm, scanning frequency 200 Hz, densely collecting pipe diameter profile point cloud). The outlet station is used for final quality verification, deploying one high-speed industrial camera (0.7 m above the outlet) and four laser displacement sensors (distributed around the circumference), with the same parameters as the inlet station, forming a “inlet-outlet” full-process size comparison.

[0021] The sensor spatial layout scheme is solidified in the form of engineering drawings, which clearly defines the installation coordinates of each sensor (based on the mechanical coordinate system of the molding machine, with the origin set as the center of the molding machine), the installation angle, the fixing method (such as bracket welding, bolt fastening), and the protection measures (such as adding a dust cover to the camera, and providing a high-temperature resistant sheath for the pressure sensor), for example, "welding area left industrial camera: coordinates (X=1.2m, Y=0.6m, Z=1.5m), installation angle 45° with pipe axis, protection level IP67", to ensure that the layout scheme can directly guide the on-site installation.

[0022] Based on the sensor spatial layout scheme, IEEE1588 precise time protocol is used to synchronize the hardware clocks of all sensor nodes, ensuring microsecond-level time consistency of data acquisition, and generating a synchronous clock network; The core of this step is to eliminate the time deviation of multi-sensor acquisition through standardized time synchronization protocol, laying the foundation for subsequent space-time alignment. The specific implementation is as follows: The deployment of IEEE1588 precise time protocol (PTP) adopts a "master clock-slave clock" architecture. The master clock selects a high-precision time server (clock accuracy ≤10ns / day) in the control cabinet of the molding machine, which is connected to an industrial Ethernet switch and serves as the time reference for the entire synchronization network. All sensor nodes are equipped with hardware clock modules that support the PTP protocol and are connected to the switch as slave clocks, forming a star-shaped topology of the synchronous clock network.

[0023] The configuration process of clock synchronization includes protocol parameter setting and hardware calibration: the PTP protocol working mode is set to "master-slave mode", the synchronization period is 1 second (i.e. the master clock sends a synchronization message to the slave clock once every second), and the delay measurement uses a "two-way delay mechanism" (the master and slave clocks send delay request and response messages to each other, calculate the link delay and compensate). In the hardware calibration stage, a special calibration tool is used to fine-tune the frequency of the crystal oscillator of each slave clock, ensuring that the frequency deviation of the free-running clock is ≤1ppm (parts per million). During synchronization, the synchronization message sent by the master clock contains an accurate timestamp (accurate to nanoseconds), which is received by the slave clock. After calculating the link delay and the local clock deviation, the local clock is automatically adjusted to ensure that the clocks of all sensor nodes are consistent with the master clock.

[0024] Synchronization accuracy verification is achieved by comparing the same event timestamps collected by different sensors. For example, a high-speed flash is used to trigger a synchronous collection, and the collection timestamps of industrial cameras, laser displacement sensors, and pressure sensors are read respectively to calculate the time deviation. The time deviation of all sensors is required to be ≤5 μs (microsecond-level consistency). If the deviation exceeds the threshold, it is corrected by adjusting the synchronization period (such as shortening to 0.5 seconds) or optimizing the network link (reducing switch forwarding delay). The final generated synchronization clock network ensures that all sensors trigger collection at the same physical time, and the timestamp error of collected data is controlled within microseconds, providing a guarantee for time alignment of multi-source data.

[0025] According to the synchronization clock network, each sensor node is triggered to collect synchronously. A high-speed industrial camera captures the surface image of the pipe, a laser displacement sensor array scans the pipe diameter profile point cloud, an infrared thermal imager collects the temperature field distribution of the weld, and a pressure sensor records the roller pressure time waveform to generate a raw multi-source data stream. The core of this step is to trigger multi-sensor cooperative collection based on the synchronization clock to obtain multi-type raw data covering the whole forming process. The specific implementation is as follows: The collection trigger mechanism is uniformly controlled by the master clock of the synchronization clock network. The master clock sends a trigger signal at a preset collection frequency (matching the sampling frequency of each sensor, such as 100 Hz), and all slave clock sensor nodes start collection synchronously after receiving the signal, ensuring that different types of data are captured at the same time point. During the collection process, each sensor generates raw data according to the preset parameters: the surface image of the pipe captured by the high-speed industrial camera is in RAW format, with a single frame size of about 6 MB, containing details such as pipe surface scratches and weld forming. For example, the image captured by the weld area camera can clearly distinguish defects such as weld crown and undercut. The pipe diameter profile point cloud generated by the laser displacement sensor array contains 100 three-dimensional coordinate points (X, Y, Z) for each sensor each time, and 800 points are output by 8 sensors at a time, forming a complete profile of the pipe diameter. The point cloud data is in ASCII code, containing coordinate values and sensor identifiers. The temperature field distribution data collected by the infrared thermal imager is a two-dimensional matrix (640x480), with each element being the temperature value (unit: °C) of the corresponding pixel point. For example, the temperature of the center pixel of the weld is 1200°C, and the temperature of the heat-affected zone pixel is 800°C. The roller pressure time waveform recorded by the pressure sensor is a continuous voltage signal conversion value (through analog-to-digital conversion, resolution 16 bits), which reflects the dynamic changes of the roller pressure in real time. For example, the waveform curve of the pressure from 30 MPa to 35 MPa and then to 32 MPa during the welding process.

[0026] The original multi-source data stream is stored in a structured format of "sensor identification-time stamp-data content", transmitted in real time to the data processing server through industrial Ethernet, and the transmission protocol uses TCP / IP to ensure data is not lost. For example, a certain original data entry is "laser sensor-003_2025-04-0314:30:00.123456_X=5.2m, Y=0.3m, Z=0.85m_ temperature=25℃", wherein the time stamp is accurate to microseconds, consistent with the synchronous clock network. The data stream is also cached locally (cache capacity ≥ 100 GB) to prevent data loss due to network interruption, and the generated original multi-source data stream completely covers the key physical quantities of the forming process, such as morphology, size, temperature, and pressure.

[0027] The original multi-source data stream is preprocessed in real time, and a wavelet denoising algorithm is used to filter out pressure signal noise, adaptive histogram equalization is applied to enhance image contrast, and outliers are removed from laser point cloud to generate a preprocessed multi-source heterogeneous data set.

[0028] The core of this step is to improve the quality of the original data through targeted preprocessing algorithms, eliminate noise and interference, and provide clean data for subsequent spatio-temporal alignment and feature fusion. The specific implementation is as follows: The wavelet denoising algorithm for pressure signal selects db4 wavelet basis (considering both denoising effect and computational efficiency), and the processing procedure is as follows: first, the original pressure time series signal is decomposed into 3 layers of wavelet, obtaining the approximate coefficient (low frequency signal, containing the real change of pressure) and the detail coefficient (high frequency noise, such as sensor vibration interference); then the soft threshold function is used to process the detail coefficient, the threshold λ = σ × √(2 × ln(N)), where σ is the noise standard deviation (estimated by the detail coefficient after decomposition), N is the signal length, for example, a certain segment of pressure signal length N = 1000, σ = 0.5MPa, then λ = 0.5 × √(2 × ln(1000)) ≈ 0.5 × 4.6 ≈ 2.3MPa, the detail coefficient less than λ is set to zero, and the coefficient greater than λ is shrunk according to the soft threshold rule; finally, the signal is reconstructed by inverse wavelet transform to obtain the denoised pressure time series waveform, the signal-to-noise ratio of the denoised signal is improved to more than 30dB, for example, the 2MPa high frequency noise contained in the original pressure signal is effectively filtered out, and the waveform is smoother.

[0029] Adaptive Histogram Equalization (CLAHE) is used to enhance the contrast of industrial camera images, especially addressing the issue of low contrast between weld seams and the pipe background. During processing, the image is divided into 8×8 local blocks (block size adjusted according to image resolution). Each local block has its histogram calculated independently and is then equalized. A contrast limit parameter is set (typically 2.0 to avoid areas being too bright or too dark). Bilinear interpolation eliminates grayscale jumps at block boundaries. For example, blurry weld seam edges in the original weld seam image become clearer and have richer grayscale levels after processing, facilitating subsequent edge feature extraction.

[0030] Outlier removal in laser point clouds employs a statistical filtering algorithm. First, the neighborhood of each point cloud data point is analyzed, with a set number of 15 neighborhood points (15 adjacent points for each point). The average distance and standard deviation σ of the neighborhood points are calculated. Then, points whose deviation from the average distance exceeds twice the standard deviation are identified as outliers (e.g., isolated points caused by sensor interference or false points due to dust in the air) and removed. For example, if the average neighborhood distance of a laser point is 0.05m and the standard deviation σ = 0.01m, and the deviation from the average distance is 0.03m (exceeding 2 × 0.01m = 0.02m), it is identified as an outlier. After removal, the data is downsampled to retain key contour points, and the point cloud density is controlled within a reasonable range (e.g., 3 points per millimeter) to ensure efficient subsequent processing.

[0031] The preprocessed multi-source heterogeneous dataset integrates all processed data and stores them in categories such as "timestamp-sensor type-data content", for example, "2025-04-03 14:30:00.123456_industrial camera_processed weld image_laser point cloud_noise-reduced pressure signal_temperature field matrix". The data format is unified into a standardized binary format, which facilitates subsequent spatiotemporal alignment and feature extraction.

[0032] S202, perform spatiotemporal alignment and feature-level fusion on the multi-source heterogeneous data to construct a multidimensional dynamic digital twin of the forming process; Specifically, timestamps from each data source can be extracted from the preprocessed multi-source heterogeneous dataset, and cubic spline interpolation algorithm can be used to unify all data sequences to the same millisecond-level time base to generate time-strictly aligned multi-source data sequences; The core of this step is to eliminate the time asynchrony problem caused by the difference in sampling frequency among multiple sensors. Precise interpolation is used to synchronize data in the time dimension, laying a time-consistent foundation for subsequent spatial registration and feature fusion. The specific implementation method is as follows: Firstly, the timestamp corresponding to each data entry in the pre-processed multi-source heterogeneous data set is extracted. The timestamp format is "YYYY-MM-DDHH:MM:SS.ssssss", accurate to microseconds, derived from the unified time granted by the synchronous clock network. The sampling frequencies of various data sources differ: the high-speed industrial camera has a sampling frequency of 100 fps (1 data point every 10 milliseconds), the laser displacement sensor array has a sampling frequency of 200 Hz (1 data point every 5 milliseconds), the infrared thermal imager has a sampling frequency of 50 Hz (1 data point every 20 milliseconds), and the distributed pressure sensor has a sampling frequency of 1 kHz (1 data point every 1 millisecond). To achieve time unification, the millisecond-level time reference is set to 1 millisecond, i.e., each time step is 1 millisecond, covering the sampling period of all data sources to ensure no data is missed.

[0033] The application core of the cubic spline interpolation algorithm is to fill in the data gaps through smooth curve fitting, ensuring that the interpolated data maintains the trend and characteristics of the original sequence and avoids distortion. The algorithm principle is as follows: for any two adjacent original data points t_i and t_j (t_j > t_i), a cubic polynomial S(x) is constructed in the interval [t_i, t_j] that satisfies S(t_i) = y_i, S(t_j) = y_j (y is the data value), and the first and second derivatives of the polynomials in adjacent intervals are continuous at the junction points to ensure smooth curves. For example, the temperature value of the infrared thermal imager at t = 100 milliseconds is 800°C, and at t = 120 milliseconds it is 850°C. The 19 time steps between 101-119 milliseconds need to be interpolated, and the cubic polynomial S(x) = ax 3 + bx 2 + cx + d is constructed. By solving the coefficients a, b, c, d through boundary conditions, the temperature value at each millisecond-level time point is calculated, such as 825°C at 105 milliseconds, which not only conforms to the temperature rising trend but also maintains smooth transition.

[0034] Interpolation processing for different types of data requires targeted adaptation: image data interpolation focuses on key feature points (such as pixel coordinates of weld edges), and calculates the feature point position at each millisecond time point through cubic spline interpolation to avoid overall image stretching and distortion; laser point cloud data interpolation interpolates the three-dimensional coordinates of each point separately to ensure the continuity of the point cloud contour; temperature field data interpolation calculates pixel by pixel to maintain the spatial consistency of temperature distribution; pressure signal interpolation retains key features such as peaks and valleys of the original waveform, and the signal-to-noise ratio of the interpolated signal is not lower than that of the original preprocessed signal. After interpolation, all data sources generate continuous data sequences with a time step of 1 millisecond. For example, the original 1kHz data from the pressure sensor (1 point per millisecond) is directly retained, and the 100fps data from the industrial camera (1 point per 10 milliseconds) is supplemented with 9 intermediate points through interpolation. Finally, the time axes of all data sequences are completely aligned, and the corresponding data from all sensors can be extracted at any millisecond time point, generating a multi-source data sequence with strict time alignment.

[0035] Based on sensor calibration parameters and mechanical structure coordinate system, time-strictly aligned multi-source data sequences are mapped to a unified three-dimensional spatial coordinate system with the rolling centerline as the reference. In this process, rigid body transformation is used to align the laser point cloud with the image pixel coordinates to generate a multi-source data field with spatial coordinate registration. The core of this step is to eliminate the spatial heterogeneity problem caused by the different installation locations of multiple sensors. By using coordinate mapping and rigid body transformation, data is unified in the spatial dimension, and a globally consistent multi-source data field is constructed. The specific implementation method is as follows: First, a unified three-dimensional coordinate system is defined, with the rolling centerline of the forming machine as the Z-axis (along the tube forming direction), the midpoint of the rolling centerline as the origin O(0,0,0), the horizontal axis perpendicular to the Z-axis and pointing to the left of the forming machine as the X-axis, and the vertical axis pointing upwards as the Y-axis, with units in millimeters. The calibration of the mechanical structure coordinate system is completed using a laser tracker, measuring the coordinates of key components of the forming machine (such as the roll center and sensor mounting bracket) in the unified coordinate system, with the error controlled within ±0.01 mm, providing a reference for sensor coordinate transformation.

[0036] The sensor calibration parameters include the intrinsic and extrinsic parameters of each sensor: the intrinsic parameters such as the focal length f = 16 mm, the principal point coordinates (u0 = 1024 pixels, v0 = 768 pixels), the distortion coefficients k1 = -0.01, k2 = 0.005 of the industrial camera, which are used to convert the image pixel coordinates into three-dimensional coordinates in the camera coordinate system; the extrinsic parameters such as the translation vector and rotation matrix of the sensor installation position, which are used to convert the camera coordinate system, the laser sensor local coordinate system into a unified three-dimensional space coordinate system. For example, the extrinsic parameters of the high-speed industrial camera are: the translation vector T = (500, 800, 1200) (indicating that the position of the camera in the unified coordinate system is X = 500 mm, Y = 800 mm, Z = 1200 mm), and the rotation matrix R represents the angle relationship between the camera lens and the unified coordinate system, which is used to convert the point cloud in the camera coordinate system to the unified coordinate system.

[0037] The rigid body transformation of the laser point cloud and the image pixel coordinates is the key to spatial registration, which includes two steps of translation and rotation, without changing the shape and size of the point cloud, only adjusting the spatial position and attitude. First, the hand-eye calibration of the laser sensor and the industrial camera is completed through the calibration board, and the transformation matrix T_trans = [R, t] is obtained, where R is a 3x3 rotation matrix and t is a 3x1 translation vector. For example, the coordinates of a certain point of the laser point cloud in the laser local coordinate system are P_laser = (10, 20, 30), after rigid body transformation, the coordinates in the unified coordinate system are P_unified = R x P_laser + t, assuming that R is a unit matrix (no rotation) and t = (490, 780, 1200), then P_unified = (10 + 490, 20 + 780, 30 + 1200) = (500, 800, 1230). For the image pixel coordinates, first convert the pixel coordinates (u, v) to the ray direction in the camera coordinate system through the camera intrinsic parameters, and combine the depth information (distance from the Z axis of the camera) of the laser point cloud to calculate the three-dimensional space coordinates corresponding to the pixel, and then map it to the unified coordinate system through the rigid body transformation, to realize the spatial alignment of the laser point cloud and the image feature points.

[0038] After the coordinate mapping and rigid body transformation of all time-strictly-aligned data sequences, they are all unified into a three-dimensional space coordinate system with the rolling center line as the reference, and the multi-source data at any millisecond time point correspond to the same spatial position, for example, at t = 1000 milliseconds, the weld edge pixels captured by the industrial camera, the pipe diameter contour points scanned by the laser sensor, the temperature pixels collected by the infrared thermal imager, and the roller contact points recorded by the pressure sensor can all find corresponding spatial coordinates in the unified coordinate system, generating a multi-source data field with spatial coordinate registration, and each spatial point in the data field contains multi-dimensional information such as topography, size, temperature, and pressure.

[0039] Multi-modal features are extracted from the multi-source data field registered in space coordinates, including weld edge contour features based on a Canny operator, heat-affected zone features based on a temperature gradient, rolling force distribution features based on a pressure waveform, and pipe diameter ellipticity features based on a point cloud, to generate a multi-modal feature vector set; The core of this step is to mine key features that can represent the forming quality from the unified multi-source data field. Multi-modal feature extraction realizes data dimension reduction and information condensation, providing core input for subsequent digital twin construction and quality prediction. The specific implementation is as follows: The weld edge contour feature extraction based on the Canny operator is aimed at the pipe surface topography image collected by an industrial camera. The Canny operator is a classic algorithm for edge detection, which accurately locates the weld edge through multi-stage processing. First, Gaussian filtering is performed on the image, with a Gaussian kernel size of 5x5 and a standard deviation σ=1.4, to smooth the image noise while preserving edge details. Then, the image gradient amplitude and direction are calculated. The Sobel operator is used to calculate the X-axis (horizontal) and Y-axis (vertical) direction gradients, respectively. The gradient amplitude is calculated by √(Gx 2 +Gy 2 ), and the gradient direction is determined by arctan(Gy / Gx). Next, non-maximum suppression is performed to eliminate pixels with local maximum gradient amplitude and refine the edge. Finally, a double threshold (low threshold 50, high threshold 150) is set. Pixels above the high threshold are determined as strong edges, pixels below the low threshold are discarded, and pixels between the two thresholds that are connected to strong edges are determined as weak edges, finally forming a continuous weld edge contour. For example, the processed weld edge contour is composed of a series of three-dimensional coordinate points: (502.1, 801.3, 1235.6), (502.3, 800.9, 1236.2), …, (503.5, 799.8, 1240.1). The number of edge points, average spacing, and curvature parameters are extracted to form a 12-dimensional weld edge contour feature vector, such as [128, 0.08mm, 0.03rad / mm,...].

[0040] The heat-affected zone feature extraction based on the temperature gradient is aimed at the temperature field distribution data collected by an infrared thermal imager. The temperature gradient reflects the rate of change of temperature in space and is the core basis for dividing the heat-affected zone. The central difference method is used to calculate the temperature gradient. For any pixel point (i, j) in the temperature field, the X-direction gradient Gx=(T(i+1, j)-T(i-1, j)) / (2Δx), the Y-direction gradient Gy=(T(i, j+1)-T(i, j-1)) / (2Δy), where Δx, Δy are the actual spatial distances corresponding to the pixels (infrared thermal imager pixel resolution 0.5mm / pixel). The temperature gradient amplitude G=√(Gx 2 +Gy 2). The heat-affected zone is defined as the area with a temperature gradient amplitude between 5-20℃ / mm. The area, perimeter, center coordinates, average temperature, maximum temperature, and distance from the weld center are extracted to form a 10-dimensional heat-affected zone feature vector. For example, at a certain time point, the heat-affected zone area is 120mm 2 , the perimeter is 45mm, the center coordinates are (505.2, 802.5, 1238.0), the average temperature is 650℃, the maximum temperature is 8mm away from the weld center, and the feature vector is [120, 45, 505.2, 802.5, 1238.0, 650, 8,...].

[0041] The rolling force distribution feature extraction based on pressure waveform is aimed at the time series pressure data recorded by the distributed pressure sensor. The pressure waveform contains the dynamic contact force information between the roller and the pipe during the rolling process. The extracted features include peak value, valley value, mean value, variance, peak value occurrence time, rising slope, falling slope, and waveform distortion rate, etc. The waveform distortion rate is calculated by the mean square error of the actual waveform and the ideal sine waveform. For example, in a 100ms time window, the pressure data: peak value 38.5MPa, valley value 29.2MPa, mean value 33.8MPa, variance 2.6MPa 2 , peak value occurrence time at 45ms in the window, rising slope 0.32MPa / ms, falling slope 0.28MPa / ms, waveform distortion rate 0.15, forming an 8-dimensional rolling force distribution feature vector [38.5, 29.2, 33.8, 2.6, 45, 0.32, 0.28, 0.15].

[0042] The pipe diameter ellipticity feature extraction based on point cloud is aimed at the pipe diameter profile point cloud scanned by the laser displacement sensor array. Ellipticity is a key indicator to measure the circularity of the pipe diameter, reflecting the roundness accuracy of the pipe forming. First, the pipe diameter point cloud at each time point is fitted with an ellipse, and the least squares method is used to construct the ellipse equation Ax 2 +Bxy+Cy 2+Dx+Ey+F=0, the length of the long axis a and the length of the short axis b of the ellipse are calculated by eigenvalue decomposition, and the ellipticity calculation formula is e=(a-b) / a x 100% (the smaller the value, the better the roundness). At the same time, the average diameter, maximum diameter, minimum diameter, diameter standard deviation and other parameters of the point cloud are extracted to form a 6-dimensional pipe diameter ellipticity feature vector. For example, after fitting, the ellipse long axis a=114.2mm, the short axis b=113.8mm, the ellipticity e=(114.2-113.8) / 114.2 x 100%≈0.35%, the average diameter is 114.0mm, the maximum diameter is 114.3mm, the minimum diameter is 113.7mm, the standard deviation is 0.2mm, and the feature vector is [114.2, 113.8, 0.35%, 114.0, 114.3, 113.7, 0.2] (Note: the actual vector dimension is integrated with 6 key parameters).

[0043] The above four types of feature vectors are spliced in time sequence, and each millisecond time point corresponds to a multi-modal feature vector, and the total dimension of the vector is 12+10+8+6=36, generating a multi-modal feature vector set, for example, the feature vector corresponding to t=1000 milliseconds is [128, 0.08,..., 120, 45,..., 38.5, 29.2,..., 0.35%, 114.0,...], which fully represents the multi-dimensional state of the molding process at this time point.

[0044] The multi-modal feature vector set is input into the physics-based modeling engine, combined with the finite element analysis model and the material rheology model, to update the state parameters of the digital twin in real time, and to build a multi-dimensional dynamic digital twin reflecting the dynamic changes of the molding process.

[0045] The core of this step is to combine multi-modal features with physical laws of the molding process through physical modeling to drive the dynamic evolution of the digital twin, and to realize accurate digital replication of the molding process. The specific implementation is as follows: The physics-based modeling engine is a core module that integrates multi-physical field simulation and data-driven, supporting functions such as finite element analysis, material rheology calculation, state parameter updating, etc., with an operation delay of ≤10 milliseconds, meeting the real-time control requirements. The input of the engine is the multi-modal feature vector set, and the output is the state parameters of the digital twin, including geometric state (weld width, reinforcement, pipe diameter size, ellipticity), physical state (temperature distribution, rolling force distribution, stress and strain distribution), material state (grain size, phase change ratio), etc., realizing dynamic mapping of multi-dimensional state.

[0046] The finite element analysis model is constructed to analyze the mechanical and thermal properties of stainless steel welded pipe forming process. Quadrilateral plane strain elements are used as the element type, and the mesh size is adaptively adjusted according to key areas: 0.5mm mesh size for weld seam and heat-affected zone (for refined calculation), and 2mm mesh size for other areas (balancing efficiency and accuracy), with a total of approximately 8000 elements. The model boundary conditions are dynamically updated based on multimodal characteristics: temperature boundary conditions are derived from the average temperature and temperature gradient of the heat-affected zone; mechanical boundary conditions are derived from the pressure peak and distribution pattern of the rolling force distribution; and displacement boundary conditions restrict the degrees of freedom of the fixed components of the forming machine. For example, when the average temperature of the heat-affected zone in the multimodal eigenvector is 650℃ and the peak rolling force is 38.5MPa, the finite element model applies this temperature value to the weld area and the pressure value to the contact surface between the roll and the tube. By solving the heat conduction equation and the elasticity equation, the temperature field distribution (e.g., weld center temperature 1200℃, heat-affected zone temperature 600-800℃) and stress distribution (e.g., weld edge tensile stress 250MPa) of the tube are calculated.

[0047] The Johnson-Cook model was selected as the material rheology model. This model can accurately describe the plastic rheological behavior of stainless steel under high temperature and high strain rate. The model parameters were calibrated experimentally: Yield strength σ_y = A + Bε^n(1 + Cln) )(1-T^m), where A=700MPa (initial yield strength), B=500MPa (hardening coefficient), n=0.3 (hardening index), C=0.05 (strain rate sensitivity coefficient), m=0.8 (temperature sensitivity coefficient), and ε is the equivalent plastic strain. The equivalent plastic strain rate is T, where T is the homologous temperature ((T_actual-T_room) / (T_melt-T_room), T_actual is the actual temperature, T_room=25℃, and T_melt=1450℃ is the melting point of stainless steel). For example, when the temperature of the weld area in the eigenvector is 1200℃, the strain rate during the rolling process is 10s^-1, and the plastic strain is 0.2, the calculated yield strength σ_y = 700 + 500 × 0.2^0.3 × (1 + 0.05 × ln10) × (1 - ((1200-25) / (1450-25))^0.8) ≈ 700 + 500 × 0.58 × 1.12 × (1 - 0.85^0.8) ≈ 700 + 324.8 × 0.13 ≈ 742.2 MPa, this parameter is used for stress calculation in the finite element model to ensure that the mechanical properties of the material are consistent with reality.

[0048] The state parameter real-time updating mechanism of the digital twin is as follows: every time a multi-modal feature vector is received (with an interval of 1 millisecond), the modeling engine immediately drives the finite element analysis model and the material rheology model to calculate, uses the geometric features (such as the weld edge profile and the pipe diameter ovality) in the feature vector as the initial condition correction of the model, inputs the physical features (such as the temperature gradient and the rolling force) as the boundary condition, solves to obtain the temperature, stress, strain, material state and other parameters at the current time, and updates the corresponding state of the digital twin. For example, when the number of weld edge profile points in the feature vector increases to 150 (indicating that the weld width increases), the weld width parameter in the digital twin is updated from 6.2 mm to 6.5 mm; when the rolling force peak value decreases from 38.5 MPa to 36.8 MPa, the roller contact pressure distribution in the digital twin is adjusted, and the stress value in the stress concentration area is also reduced synchronously.

[0049] The construction of the multi-dimensional dynamic digital twin is based on a three-dimensional geometric model, which is created based on the actual sizes of the forming machine and the pipe material, with a scale of 1:1, and contains core components such as pipe material, rollers, and welding equipment. Through state parameter updating, the digital twin can not only visually present the geometric shape changes of the pipe material (such as weld forming and pipe diameter size fluctuation), but also can display the temperature field (rendered with a color gradient, with red indicating a high temperature zone), the stress field (represented by arrows indicating the stress direction and size), and the rolling force distribution (represented by a pressure cloud chart) in real time. It also supports cross-section viewing and real-time query of key parameters, such as clicking on the weld area of the digital twin to display the current weld width of 6.5 mm, the crown of 1.2 mm, the temperature of 1200°C, and the tensile stress of 245 MPa, etc. It fully reflects the dynamic changes of the forming process and provides a precise digital carrier for subsequent quality prediction and control.

[0050] S203, based on the multi-dimensional dynamic digital twin, using an online rolling prediction model to real-time deduce the development trend of the weld forming quality and the pipe diameter size deviation; Specifically, the key state variables at the current time can be extracted from the multi-dimensional dynamic digital twin, including the weld width, the crown, the heat-affected zone width, the pipe diameter size distribution, and the rolling force distribution, to form the input feature vector of the prediction model; The core of this step is to select the key variables that play a decisive role in the forming quality from the multi-dimensional state of the digital twin, and form a standardized input through feature integration to provide a precise and efficient data source for the subsequent prediction model. The specific implementation is as follows: The extraction of key state variables is based on the real-time update data of the digital twin, and each variable has a clear physical definition and extraction logic. The weld width refers to the maximum horizontal distance between the fusion lines in the weld cross section, which is extracted from the geometric state parameters of the digital twin, calculated based on the fitting results of the weld edge contour features, such as the weld edge point cloud extracted by the Canny operator, and the horizontal distance between the two edges is measured after fitting as a straight line, the current extraction value is 6.2mm, and the unit is unified as millimeter. The weld reinforcement refers to the maximum vertical height of the weld surface above the base material surface, which is calculated by the vertical distance between the base material surface reference line and the highest point of the weld in the three-dimensional geometric model of the weld area in the digital twin, and the current extraction value is 1.1mm.

[0051] The heat-affected zone width refers to the width of the region around the weld that has undergone organizational changes due to welding heat but has not melted, which is extracted from the material state parameters of the digital twin, based on temperature gradient distribution and material phase change threshold, such as setting the phase change temperature of stainless steel to 800℃, extracting the width of the region between 600-800℃, the current extraction value is 8.5mm. The pipe diameter size distribution refers to the diameter statistical characteristics of the cross section of the pipe diameter area, which is extracted from the ellipse parameters fitted by the laser point cloud, including the average diameter, the maximum diameter, the minimum diameter and the diameter standard deviation, the current extraction values are 114.0mm, 114.3mm, 113.7mm and 0.2mm respectively.

[0052] The rolling force distribution refers to the pressure distribution characteristics on the contact surface between the roll and the pipe, which is extracted from the physical state parameters of the digital twin, including the average pressure, peak pressure and pressure uniformity of the upper, lower, left and right four groups of rolls, the pressure uniformity is calculated by the ratio of pressure standard deviation and average pressure, the current average pressure of the upper roll is 33.8MPa, the peak pressure is 38.5MPa, and the uniformity is 0.08, the corresponding values of the lower roll are 32.6MPa, 37.2MPa and 0.07, the left roll are 34.2MPa, 39.1MPa and 0.09, and the right roll are 33.5MPa, 38.0MPa and 0.08.

[0053] The construction of the input feature vector adopts the "variable normalization + sequential splicing" method. First, each key state variable is normalized by Min-Max normalization to map to the [0, 1] interval. The normalization formula is x' = (x - x_min) / (x_max - x_min), where x_min and x_max are the historical extreme values of the variable (determined by statistics of the last 3 months of forming data). For example, the x_min of the weld width is 5.0 mm, the x_max is 7.0 mm, and the current value 6.2 mm is normalized to (6.2-5.0) / (7.0-5.0)=0.6; the x_min of the weld height is 0.5 mm, the x_max is 1.5 mm, and the current value 1.1 mm is normalized to (1.1-0.5) / (1.5-0.5)=0.6. After normalization of all variables, they are sequentially spliced in the order of "weld width - height - HAZ width - average diameter - maximum diameter - minimum diameter - diameter standard deviation - upper roller average pressure - upper roller peak pressure - upper roller uniformity - lower roller average pressure - lower roller peak pressure - lower roller uniformity - left roller average pressure - left roller peak pressure - left roller uniformity - right roller average pressure - right roller peak pressure - right roller uniformity" to form a 19-dimensional input feature vector. For example, the current feature vector is [0.6, 0.6, 0.425, 0.5, 0.6, 0.4, 0.3, 0.38, 0.45, 0.2, 0.36, 0.43, 0.17, 0.4, 0.47, 0.25, 0.37, 0.44, 0.18]. The vector dimension strictly matches the input layer dimension of the prediction model.

[0054] The input feature vector is input into the pre-trained online rolling prediction model, which adopts a structure combining gated recurrent unit network and attention mechanism, and outputs the weld forming quality index and pipe diameter size prediction sequence for the next ten sampling periods; The core of this step is to use a hybrid model that combines time series modeling capability and key information focusing capability to infer future quality changes based on the current state, achieving ultra-short-term trend prediction. The specific implementation is as follows: The pre-training process of the online rolling prediction model is based on historical forming data. The training data set contains nearly 3 months of key state variable sequences and corresponding forming quality results, with a total of 1 million samples, divided into training set, validation set and test set in the ratio of 7:2:1. The model structure consists of input layer, GRU layer, attention layer, fully connected layer and output layer. The input layer dimension is 19 (matching the input feature vector dimension), the GRU layer contains 2 layers of hidden layer, each layer contains 128 hidden units, the activation function uses tanh, and the weights of the reset gate and update gate are optimized through training to capture the time correlation of state variables, such as forgetting irrelevant historical data through the reset gate and retaining long-term information useful for current prediction through the update gate.

[0055] The attention mechanism layer, located after the GRU layer, plays a crucial role in dynamically weighting the importance of features at different historical time steps, thus preventing irrelevant historical information from interfering with the prediction results. Its implementation logic is as follows: the hidden state sequence output by the GRU layer is used as keys and values, and the input feature vector at the current time step is used as the query. The similarity between the query and each key is calculated using a dot product, normalized by a softmax function to obtain the attention weights, and then the weights are summed with the values ​​to obtain the attention-enhanced feature representation. For example, if the similarity between the hidden state at a certain historical time step and the current query is 0.8, significantly higher than the 0.1-0.3 similarities at other time steps, the corresponding attention weight is 0.6. Features at this time step are given priority, effectively improving the model's ability to capture key historical information.

[0056] The fully connected layer consists of two layers. The first layer has an output dimension of 64 and uses ReLU as the activation function. The second layer has an output dimension of 20 (corresponding to two quality indicators for the next ten sampling periods: weld formation quality and pipe diameter, with 10 predicted values ​​for each indicator). The output layer has no activation function and directly outputs continuous predicted values. The model training uses mean squared error (MSE) as the loss function, with Adam as the optimizer. The initial learning rate is set to 1e-3, and it decreases to 1e-4 when the validation set loss does not decrease for 10 consecutive rounds. The training iterations are 200 rounds. The final test set shows a weld formation quality prediction error ≤0.03mm and a pipe diameter prediction error ≤0.05mm, meeting the accuracy requirements for real-time control.

[0057] The 19-dimensional input feature vector at the current moment is input into the pre-trained model. The model mines the temporal evolution of variables through the GRU layer, strengthens key historical information through the attention layer, and finally outputs the prediction sequence for the next ten sampling periods (each sampling period is 1 millisecond, for a total of 10 milliseconds). For example, the predicted sequence for weld formation quality indicators (with weld width as the core characterization) is [6.2mm, 6.3mm, 6.4mm, 6.5mm, 6.6mm, 6.7mm, 6.8mm, 6.9mm, 7.0mm, 7.1mm], and the predicted sequence for pipe diameter is [114.0mm, 114.1mm, 114.2mm, 114.3mm, 114.4mm, 114.5mm, 114.6mm, 114.7mm, 114.8mm, 114.9mm]. Each predicted value is accompanied by a confidence level (calculated based on the variance of the model output). For example, the confidence level of the first predicted value is 0.95, and the confidence level gradually decreases as the prediction time step increases. The confidence level of the tenth predicted value is 0.82, reflecting the change in the uncertainty of the prediction.

[0058] The deviation from the process standard value is calculated based on the predicted sequence, and a sliding time window mechanism is adopted. The input feature vector is updated and the rolling prediction is re-performed every time new sensor data is received, realizing online adaptive update of model parameters. The core of this step is to quantify the difference between the prediction quality and the standard, and through a dynamic updating mechanism, the model is adapted to the drift characteristics of the forming process, ensuring the long-term accuracy of the prediction. The specific implementation is as follows: The process standard value is determined by industry specifications and production requirements, which clearly defines the allowable range of each quality indicator. For example, the process standard value for weld width is 6.0 ± 0.3 mm (i.e., the acceptable range is 5.7-6.3 mm), and the process standard value for pipe diameter size is 114.0 ± 0.5 mm (the acceptable range is 113.5-114.5 mm). The deviation calculation uses the absolute value of "predicted value - standard center value". If the predicted value exceeds the acceptable range, the deviation is the actual excess; if it is within the acceptable range, the deviation is 0, and the deviation unit is consistent with the predicted value (mm). For example, in the weld width prediction sequence, the first three values (6.2 mm, 6.3 mm, 6.4 mm) correspond to deviations of 0.2 mm (6.2-6.0), 0.3 mm (6.3-6.0), and 0.4 mm (6.4-6.0, exceeding the upper limit of the acceptable range by 0.1 mm), respectively. In the pipe diameter size prediction sequence, the first five values are within the acceptable range, with a deviation of 0, the sixth value is 114.5 mm with a deviation of 0.5 mm, and the seventh value is 114.6 mm with a deviation of 0.6 mm (exceeding the upper limit of the acceptable range by 0.1 mm).

[0059] The sliding time window mechanism is used to dynamically update the input feature vector. The window size is set to 20 sampling periods (20 milliseconds), i.e., the input feature vector not only contains 19-dimensional key state variables at the current time, but also integrates variable data from the previous 19 sampling periods, forming a 19x20=380-dimensional time series feature vector, enhancing the model's ability to capture trend changes. The window sliding method is "first-in, first-out". Every time new sensor data is received (at an interval of 1 millisecond), the window removes the data from the earliest sampling period and adds the latest data, reconstructs the time series feature vector and inputs it into the model, triggering a new round of rolling prediction, realizing a closed-loop process of "real-time data - update input - re-prediction".

[0060] The online adaptive update of model parameters adopts an incremental learning mechanism. After each rolling prediction, the error between the prediction deviation and the actual deviation (the difference between the actual quality data collected subsequently and the standard value) is calculated, and the weight parameters of the model are fine-tuned through the stochastic gradient descent (SGD) algorithm. The learning rate is set to 5e-5 (much smaller than the pre-training learning rate to avoid parameter mutation), and only the weights of the attention layer and the fully connected layer are adjusted, while the GRU layer parameters remain stable (to reduce computational load). For example, when the predicted weld width deviation is 0.4 mm, the actual deviation is 0.35 mm, and the error is 0.05 mm, the weight of the corresponding historical time step in the attention layer is adjusted by -0.002 through gradient descent, and the weight of the fully connected layer is adjusted by +0.001, so that the model gradually adapts to the characteristics of the current forming process. The frequency of online update is consistent with rolling prediction (once every 1 millisecond), ensuring that the model can quickly respond to the drift of data distribution and maintain prediction accuracy.

[0061] According to the rolling prediction results, the time-space evolution curve of the weld forming quality and the pipe diameter size deviation is drawn, the critical point of accelerated increase and the trend inflection point of the deviation are identified, and a development trend report containing warning levels is generated.

[0062] The core of this step is to visually present the change law of quality deviation through visualization and trend analysis, accurately identify risk nodes, and provide clear decision basis for subsequent regulation. The specific implementation is as follows: The time-space evolution curve of the weld forming quality and the pipe diameter size deviation takes time as the horizontal coordinate (unit: millisecond) and deviation value as the vertical coordinate (unit: millisecond), uses a smooth curve for drawing, and the curve colors are red (weld deviation) and blue (pipe diameter deviation), and the curve thickness is 2 pixels for easy distinction. The horizontal coordinate covers the current time to 10 milliseconds (ten sampling periods) in the future, and the vertical coordinate range is set according to the allowable deviation of the process standard, for example, the weld deviation vertical coordinate range is 0-1.0 mm, and the pipe diameter deviation vertical coordinate range is 0-1.0 mm. A horizontal warning line (weld deviation 0.3 mm, pipe diameter deviation 0.5 mm) is drawn to intuitively mark the boundaries of qualified and unqualified. For example, the weld deviation evolution curve starts from 0.2 mm at the current time and gradually rises with time, breaking the warning line (0.3 mm) at the 3rd millisecond and reaching 0.6 mm at the 10th millisecond; the pipe diameter deviation curve maintains 0 mm before the 5th millisecond, reaches 0.5 mm (warning line) at the 6th millisecond, and reaches 0.9 mm at the 10th millisecond, clearly showing the accelerating upward trend of the deviation.

[0063] The identification of the critical point and the trend inflection point is based on the deviation rate of change and the second derivative analysis. The deviation rate of change calculation formula is Δe / Δt (Δe is the deviation difference value of adjacent sampling periods, Δt=1 millisecond), when the rate of change exceeds the preset threshold value (weld deviation rate of change threshold value 0.05 mm / millisecond, pipe diameter deviation rate of change threshold value 0.08 mm / millisecond) for 2 consecutive sampling periods, it is determined that it is a critical point of accelerated increase of deviation; the trend inflection point refers to the time when the deviation rate of change changes from increasing to decreasing or from decreasing to increasing, by calculating the second derivative of the deviation sequence, when the second derivative changes from positive to negative or from negative to positive and the absolute value exceeds 0.02 mm / millisecond 2 , it is determined that it is an inflection point. For example, the weld deviation rate of change in the 2-3 millisecond is 0.05 mm / millisecond and 0.06 mm / millisecond respectively, which continuously exceeds the threshold value, and the 2 millisecond is determined as a critical point; the second derivative of the pipe diameter deviation in the 7 millisecond changes from positive to negative (from 0.03 mm / millisecond 2 to-0.02 mm / millisecond 2 ), the 7 millisecond is determined as a trend inflection point, indicating that the deviation growth rate will slow down.

[0064] The development trend report containing the warning level integrates the curve analysis result and the risk assessment, and the warning level is divided into three levels: mild warning (the deviation breaks through the warning line but the rate of change does not reach the critical point), moderate warning (the deviation breaks through the warning line and reaches the critical point), and severe warning (the deviation greatly exceeds the warning line, the rate of change continuously increases or an unfavorable inflection point appears). The report contains four core parts: trend overview (briefly describes the future trend of the weld and pipe diameter deviation), key node identification (lists the specific time and deviation value of the critical point and the inflection point), warning level determination (gives the warning based on the severity of the trend), and adjustment suggestion (preliminarily prompts the direction of the process parameters that need to be adjusted). For example, a development trend report fragment is: "trend overview: in the next 10 milliseconds, the weld width deviation will continuously rise from 0.2 mm to 0.6 mm, and the pipe diameter size deviation will rise from 0 mm to 0.9 mm, both of which will break through the upper limit of the process standard; key nodes: the weld deviation critical point is at the 2 millisecond (deviation 0.25 mm, rate of change 0.05 mm / millisecond), the pipe diameter deviation critical point is at the 5 millisecond (deviation 0.4 mm, rate of change 0.08 mm / millisecond), and the pipe diameter deviation inflection point is at the 7 millisecond (deviation 0.6 mm); warning level: severe warning; adjustment suggestion: the weld power needs to be immediately reduced to suppress the increase of the weld width, and the roll gap needs to be adjusted to control the pipe diameter expansion", the report is output in the form of structured text, which is convenient for subsequent adaptive fuzzy reasoning system calling.

[0065] S204, according to the development trend, a set of coordinated control instructions of the roll gap and the welding power is generated by an adaptive fuzzy reasoning system to realize online closed-loop control of the forming process parameters.

[0066] Specifically, key input variables can be extracted from the development trend report, including weld width deviation, pipe diameter ovality deviation, temperature gradient deviation, and deviation change rate, fuzzy processing is performed using a triangular membership function to generate a fuzzy input variable set; The core of this step is to convert the quantized deviation data into fuzzy language variables to provide an adaptive input form for fuzzy reasoning. The fuzzy characteristics of the variables are accurately described through a triangular membership function to ensure the rationality and robustness of the reasoning. The specific implementation is as follows: The extraction of key input variables is directly related to the core conclusions of the development trend report. Each variable is a real-time quantized value with a unified unit and clear physical meaning. The weld width deviation e_w is the difference between the predicted weld width and the process standard center value, taking the absolute value, with a unit of mm. For example, the weld width deviation in the next 5 milliseconds is 0.5 mm. The pipe diameter ovality deviation e_d is the difference between the actual pipe diameter ovality and the standard ovality (≤0.3%), also taking the absolute value, with a unit of %, and the example value is 0.25%. The temperature gradient deviation e_t is the difference between the actual temperature gradient in the weld area and the ideal temperature gradient (5℃ / mm), taking the absolute value, with a unit of ℃ / mm, and the example value is 1.8℃ / mm. The deviation change rate r is the ratio of the difference between the deviation values of two adjacent sampling periods (1 millisecond) to the time interval, reflecting the growth or decay rate of the deviation, with a unit of mm / ms (or % / ms, ℃ / (mm・ms)), and in the example, the weld width deviation change rate r_w = 0.06 mm / ms, indicating that the deviation is accelerating.

[0067] The triangular membership function is the core tool for fuzzy processing, with the characteristics of simple calculation and clear physical meaning. It is defined by three key parameters (a, b, c), where a is the left boundary with a membership degree of 0, b is the peak point with a membership degree of 1, and c is the right boundary with a membership degree of 0. The function expression is: when x≤a or x≥c, μ(x)=0; when a<x≤b, μ(x)=(x-a) / (b-a); when b<x<c, μ(x)=(c-x) / (c-b). Each input variable defines three fuzzy subsets: small (S), medium (M), and large (L), covering the entire domain range of the variable. The domain range is determined based on historical data statistics to ensure that 99% of the actual variable values fall within the domain.

[0068] The fuzzification of a specific variable is shown as follows: the domain of the weld width deviation e_w is [0, 1.0 mm], and the parameters of the three fuzzy subsets are: small (S) (0, 0.2, 0.4), medium (M) (0.3, 0.5, 0.7), and large (L) (0.6, 0.8, 1.0). If the current e_w = 0.5 mm, the function is calculated as: μ_S(0.5) = (0.4-0.5) / (0.4-0.2) = 0 (since 0.5 > 0.4, the membership degree of the small subset is 0); μ_M(0.5) = (0.5-0.3) / (0.5-0.3) = 1.0 (peak point, membership degree is 1); μ_L(0.5) = (0.8-0.5) / (0.8-0.6) = 1.5 (exceeds the medium set range, membership degree is 0), and the final fuzzification result of the weld width deviation is "medium (M)", and the membership degree vector is [0, 1.0, 0].

[0069] The domain of the pipe diameter ovality deviation e_d is [0, 0.5%], and the parameters of the fuzzy subsets are: small (S) (0, 0.1, 0.2), medium (M) (0.15, 0.25, 0.35), and large (L) (0.3, 0.4, 0.5). The current e_d = 0.25 mm, and the fuzzification result is "medium (M)", and the membership degree vector is [0, 1.0, 0]. The domain of the temperature gradient deviation e_t is [0, 3 ℃ / mm], and the parameters of the fuzzy subsets are: small (S) (0, 0.8, 1.6), medium (M) (1.2, 2.0, 2.8), and large (L) (2.4, 2.7, 3.0). The current e_t = 1.8 ℃ / mm, and the calculation results are μ_S = 0, μ_M = (2.0-1.8) / (2.0-1.2) = 0.25, and μ_L = 0. The fuzzification result is "medium (M)" (membership degree 0.25). The domain of the deviation change rate r_w is [0, 0.1 mm / ms], and the parameters of the fuzzy subsets are: small (S) (0, 0.02, 0.04), medium (M) (0.03, 0.06, 0.09), and large (L) (0.08, 0.09, 0.1). The current r_w = 0.06 mm / ms, and the fuzzification result is "medium (M)" (membership degree 1.0). After the fuzzification of all variables, the fuzzy input variable set is integrated: {e_w = M(1.0), e_d = M(1.0), e_t = M(0.25), r_w = M(1.0)}, which provides input for subsequent fuzzy reasoning.

[0070] An adaptive fuzzy rule base is constructed, the rule weights of which are dynamically adjusted according to historical control effect data through reinforcement learning, the fuzzy input variable set is subjected to Mamdani reasoning with the fuzzy rules, and a fuzzy output set representing the roll gap adjustment amount and the welding power adjustment amount is generated; The core of this step is to map the fuzzy input to the fuzzy output through dynamic rule base and classical fuzzy reasoning, and the self-adaptive adjustment of rule weight ensures that the reasoning result adapts to the dynamic changes of the forming process. The specific implementation is as follows: The adaptive fuzzy rule base adopts the production rule structure of "If-Then", the antecedent of the rule is the fuzzy subset combination of the four input variables, and the consequent is the fuzzy subset of the two output variables (roller gap adjustment amount Δg, welding power adjustment amount ΔP). The output variables also define three fuzzy subsets: negative large (NL, corresponding to adjustment amount reduction), zero (Z, corresponding to no adjustment), and positive large (PL, corresponding to adjustment amount increase), which correspond to the control direction of "reduce parameter", "maintain parameter" and "increase parameter" respectively. The rule base initially contains 81 complete rules (3^4=81, three fuzzy subsets for each of the four input variables), covering all possible input combinations. An example rule is as follows: Rule 1: If e_w=S and e_d=S and e_t=S and r=S, Then Δg=Z and ΔP=Z; Rule 2: If e_w=M and e_d=M and e_t=M and r=M, Then Δg=PL and ΔP=NL; Rule 3: If e_w=L and e_d=L and e_t=L and r=L, Then Δg=PL and ΔP=NL; Rule 4: If e_w=M and e_d=S and e_t=L and r=M, Then Δg=PL and ΔP=NL; The initial value of the rule weight ω_i (i=1 to 81) is set to 1.0, representing the initial importance of all rules being consistent, and the weight range is [0.1, 2.0], rules below 0.1 are considered invalid rules, and rules above 2.0 are considered core rules. The dynamic adjustment of the weight is realized through reinforcement learning, and the reward function R is defined as R=1-|e_new-e_target| / e_old, where e_old is the deviation value before control, e_new is the deviation value after control, and e_target is the target deviation value (0). The value range of R is [-∞, 1], when R>0, positive reward is given, and the weight increases; when R<0, negative reward is given, and the weight decreases. The adjustment formula is ω_i(t+1)=ω_i(t)×(1+η×R), where η=0.05 is the learning rate, which controls the weight adjustment amplitude. For example, after the application of rule 2, the weld width deviation decreases from 0.5mm to 0.3mm, R=1-|0.3-0| / 0.5=0.4, the weight adjustment is 1.0×(1+0.05×0.4)=1.02, and the importance of the rule is increased; if the deviation increases after the application of a certain rule, R=-0.2, the weight adjustment is 1.0×(1+0.05×(-0.2))=0.99, and the importance is reduced.

[0071] Mamdani reasoning is the core algorithm of fuzzy reasoning, which contains three steps: rule matching, activation strength calculation, and fuzzy conclusion synthesis. In the rule matching stage, the fuzzy input variable set is compared with the rule antecedent one by one, and the activation rule whose antecedent completely matches the input is selected, for example, if the current input set is {e_w=M, e_d=M, e_t=M, r=M}, rule 2 is activated; if there are multiple fuzzy subsets with non-zero membership degrees in the input, multiple related rules will be activated. The activation strength calculation uses the "min" operator, that is, the activation strength α_i=min(μ_1,μ_2,μ_3,μ_4), where μ_1-μ_4 are the membership degrees of the input variable corresponding to the antecedent fuzzy subset, for example, the activation strength of rule 2 is α_2=min(1.0,1.0,0.25,1.0)=0.25.

[0072] The fuzzy conclusion synthesis uses the "max" operator, and after the antecedent fuzzy subset of all activated rules is trimmed according to the activation strength, it is merged into the final fuzzy output set. For example, the consequent of rule 2 is Δg=PL, ΔP=NL, and the activation strength is 0.25. The trimmed PL subset membership function is μ_PL'(x)=min(0.25,μ_PL(x)), and the NL subset is μ_NL'(x)=min(0.25,μ_NL(x)). If rule 4 is also activated with an activation strength of 0.15 and the same consequent, the trimmed membership functions are μ_PL''(x)=min(0.15,μ_PL(x)) and μ_NL''(x)=min(0.15,μ_NL(x)). The synthesized fuzzy output set is the PL subset of Δg (max(0.25,0.15)=0.25) and the NL subset of ΔP (max(0.25,0.15)=0.25), which may also contain fuzzy subsets of other activated rules, finally forming a complete fuzzy output set representing the fuzzy intention of "positive large" adjustment of the roll gap and "negative large" adjustment of the welding power.

[0073] The barycentric method is used to solve the defuzzification algorithm to convert the fuzzy output set into precise control quantities, generating preliminary coordinated control instructions including the gap adjustment values of the upper roll, lower roll, left roll, and right roll, as well as the power adjustment value of the high-frequency welding power supply. The core of this step is to convert the qualitative intention of fuzzy output into quantitative precise control quantities, and the barycentric method is used to ensure the rationality of the control quantities. At the same time, according to the deviation distribution characteristics, the adjustment amount of each roll is allocated to form the preliminary control instruction, and the specific implementation is as follows: The barycentric method (Centroid of Area, COA) is a classical algorithm for defuzzification, which calculates the horizontal coordinate value corresponding to the barycenter of the area under the membership function curve of the fuzzy output set as the precise control quantity, and the formula is: where μ_U(x) is the membership function of the fuzzy output set, and the integral range covers the domain of the output variable. The advantage of this algorithm is that it makes full use of all fuzzy information, and the result is smooth and robust, which meets the accuracy requirements of real-time regulation.

[0074] The domain of the output variable and the fuzzy subset parameters need to be defined in advance: the domain of the roll gap adjustment amount Δg is [-0.3mm, 0.3mm], a negative value indicates a decrease in the gap, and a positive value indicates an increase in the gap, and the fuzzy subset parameters are NL(-0.3, -0.2, -0.1), Z(-0.05, 0, 0.05), and PL(0.1, 0.2, 0.3); the domain of the welding power adjustment amount ΔP is [-10kW, 10kW], a negative value indicates a decrease in power, and a positive value indicates an increase in power, and the fuzzy subset parameters are NL(-10, -6, -2), Z(-1, 0, 1), and PL(2, 6, 10).

[0075] De-fuzzification example: for the PL subset of Δg in the fuzzy output set (membership 0.25), the area centroid under the membership function curve is calculated as follows: The effective range of the PL subset in the domain is [0.1, 0.3mm], the integral numerator ∫x·μ_PL(x)dx=∫0.1^0.2x·(x-0.1) / (0.2-0.1)dx+∫0.2^0.3x·(0.3-x) / (0.3-0.2)dx≈0.019; the integral denominator ∫μ_PL(x)dx=∫0.1^0.2(x-0.1) / 0.1dx+∫0.2^0.3(0.3-x) / 0.1dx≈0.02; then (the activation intensity 0.25 weighted), the approximate value is 0.24mm, that is, the roll gap needs to be increased by 0.24mm.

[0076] De-fuzzification of the NL subset of the welding power adjustment amount ΔP (membership 0.25): the effective range of the domain is [-10, -2kW], the integral numerator ∫x·μ_NL(x)dx=∫-10^-6x·(x+10) / (-6+10)dx+∫-6^-2x·(-2-x) / (-2+6)dx≈-128-42.67≈-170.67; the integral denominator ∫μ_NL(x)dx=∫-10^-6(x+10) / 4dx+∫-6^-2(-2-x) / 4dx≈4+4=8; then , the approximate value is -5.3kW, that is, the welding power needs to be reduced by 5.3kW.

[0077] The generation of the preliminary coordinated control instruction needs to distribute the total roll gap adjustment amount Δg to the four rolls, and the distribution principle is based on the deviation distribution of the pipe diameter ovality: if the pipe diameter has no obvious deviation (such as e_d=M and uniform distribution), the adjustment amounts of the upper, lower, left and right rolls are equal, that is, Δg_up=Δg_down=Δg_left=Δg_right=Δg / 4=0.24 / 4=0.06 mm; if there is left deviation (the left diameter is larger), the adjustment amount of the left roll is reduced by 20%, and the adjustment amount of the right roll is increased by 20%, and in the example, there is no deviation, so the adjustment amount is equally distributed. The welding power adjustment amount ΔP* is directly used as the adjustment value of the high-frequency welding power supply, and does not need to be distributed, and the final preliminary coordinated control instruction is: the upper roll gap +0.06 mm, the lower roll gap +0.06 mm, the left roll gap +0.06 mm, the right roll gap +0.06 mm, and the welding power -5.3 kW.

[0078] According to the dynamic response characteristics of the forming machine actuator and the process safety constraints, the preliminary coordinated control instruction is subjected to amplitude limiting and rate limiting optimization to generate the final executable coordinated control instruction set, which is sent to each actuator through the field bus to realize online closed-loop precise control of the process parameters.

[0079] The core of this step is to ensure the feasibility and safety of the control instruction through double limiting optimization to avoid overloading of the actuator or sudden change of the process parameters, and then to realize precise issuance of the instruction through a reliable field bus to complete the last link of the closed-loop control. The specific implementation is as follows: The dynamic response characteristics of the forming machine actuator are an important basis for limiting optimization. The maximum response speed of the roll gap adjustment mechanism is 0.1 mm / ms (i.e. the maximum adjustment is 0.1 mm per millisecond), and the maximum single adjustment amplitude is ±0.3 mm (exceeding this amplitude will cause mechanical impact). The maximum response speed of the high-frequency welding power supply is 1 kW / ms, and the maximum single adjustment amplitude is ±10 kW, and the response delay is ≤2 milliseconds. These parameters are determined by the hardware characteristics of the actuator, which ensures that the instruction is within the physical capability of the mechanism.

[0080] The process safety constraints are another core basis for limiting optimization, which aims to avoid forming defects caused by excessive parameter adjustment: the total adjustment amount of the roll gap should not exceed ±10% of the initial gap (assuming the initial gap is 2.0 mm, the maximum adjustment is ±0.2 mm), to prevent the pipe from being deformed due to too small gap or the pipe diameter from being out of control due to too large gap; the adjustment amount of the welding power should not be lower than the minimum welding power (assuming it is 30 kW) or higher than the maximum welding power (assuming it is 60 kW), to avoid incomplete penetration due to insufficient power or weld burn-through due to excessive power.

[0081] The amplitude clipping optimization is for the single adjustment amount of the preliminary instruction, and if it exceeds the maximum single adjustment amplitude or the process constraint, it is truncated to the boundary value. For example, the single roll adjustment amount of the preliminary instruction is 0.06 mm, the total adjustment amount is 0.24 mm, the initial gap is 2.0 mm, and 10% of 0.2 mm is 0.24 mm, which exceeds the process constraint, and needs to be reduced in proportion: the reduction coefficient k = 0.2 / 0.24 ≈ 0.833, the adjusted single roll adjustment amount = 0.06 x 0.833 ≈ 0.05 mm, the total adjustment amount = 0.05 x 4 = 0.2 mm, which meets the constraint; the welding power preliminary adjustment amount is -5.3 kW, and the current power is assumed to be 40 kW, the adjusted power = 40-5.3 = 34.7 kW, which is within the range of 30-60 kW, and no amplitude clipping is required.

[0082] The rate clipping optimization is for the response speed of the actuator to ensure smooth adjustment process without impact. The roll gap adjustment rate clipping is 0.1 mm / ms, the single roll adjustment amount is 0.05 mm, and the required time = 0.05 / 0.1 = 0.5 ms, which can be completed within the sampling period (1 ms) without adjustment; if the preliminary adjustment amount is 0.12 mm, the rate clipping needs to be completed in 2 sampling periods, the current adjustment is 0.1 mm, and the next period adjustment is 0.02 mm. The welding power adjustment amount is -5.3 kW, the response speed is 1 kW / ms, the required time = 5.3 / 1 = 5.3 ms, which needs to be completed in 6 sampling periods, the current period adjustment is 1 kW, and the subsequent 5 periods are adjusted by 1 kW, 1 kW, 1 kW, 1 kW, 0.3 kW, respectively, to avoid sudden temperature changes caused by too fast single adjustment.

[0083] The final executable cooperative control instruction set integrates the adjusted amount and time sequence after clipping optimization, and the instruction format includes "actuator identification-adjustment amount-adjustment period-priority", for example: upper roll (GR-UP) + 0.05 mm (1 period completed), lower roll (GR-DN) + 0.05 mm (1 period completed), left roll (GR-LT) + 0.05 mm (1 period completed), right roll (GR-RT) + 0.05 mm (1 period completed), welding power (PWR-01) -1 kW (1st period), -1 kW (2nd period), -1 kW (3rd period), -1 kW (4th period), -1 kW (5th period), -0.3 kW (6th period), and the priority is set to 1 (highest priority).

[0084] The field bus selects PROFINET industrial Ethernet, the transmission rate is 100Mbps, the transmission delay is less than or equal to 1ms, and the real-time control demand is met.CRC-32 check is carried out before the instruction is issued, 4 bytes of check code are generated, and the integrity of the instruction transmission is ensured;The process of issuing adopts the mode of "master-slave communication", the control host is the master station, each actuator is the slave station, the master station issues the instruction according to the actuator identifier, and the slave station returns the confirmation frame (including the instruction receiving state and the state of the slave station) after receiving the instruction;If the master station does not receive the confirmation frame, it retries 3 times (each time interval is 2ms), and if the retry fails, it records the alarm log and notifies the administrator.

[0085] After the actuator receives the instruction, the operation is performed according to the adjustment amount and the period: the roll gap adjustment mechanism drives the lead screw through the servo motor, accurately adjusts the gap to the target value, and the position sensor feedbacks the adjustment result in real time, the deviation is less than or equal to 0.005mm;The welding power supply adjusts the output power step by step according to the time sequence, and the power sensor monitors the output value in real time, and the deviation is less than or equal to 0.1kW.After the execution is completed, the multi-sensor network re-collects the integrated data, enters the next round of closed-loop control process, and realizes the real-time and accurate control of the stainless steel welded pipe forming process.

[0086] It can be seen that the distributed multi-source sensor network synchronously collects multi-source heterogeneous data including pipe appearance image, laser ranging sequence, temperature field distribution and roll pressure time sequence data in the forming process;The multi-source heterogeneous data is spatio-temporally aligned and feature-level fused to construct a multi-dimensional dynamic digital twin of the forming process;Based on the multi-dimensional dynamic digital twin, an online rolling prediction model is used to real-time deduce the development trend of the weld forming quality and the pipe diameter size deviation;According to the development trend, the self-adaptive fuzzy reasoning system generates the collaborative control instruction set of the roll gap and the welding power, so that the multivariate forward-looking perception and intelligent collaborative closed-loop control of the process parameters of the stainless steel welded pipe forming process can be realized, and the stability and consistency of the forming quality are improved.

[0087] Another embodiment of the application provides a real-time closed-loop control system for a stainless steel welded pipe forming process, referring to Figure 3 , the system can include: The acquisition module 301 is used for synchronously collecting multi-source heterogeneous data including pipe appearance image, laser ranging sequence, temperature field distribution and roll pressure time sequence data in the forming process through a distributed multi-source sensor network; The construction module 302 is used for spatio-temporally aligning and feature-level fusing the multi-source heterogeneous data to construct a multi-dimensional dynamic digital twin of the forming process; The deduction module 303 is used for based on the multi-dimensional dynamic digital twin, using an online rolling prediction model to real-time deduce the development trend of the weld forming quality and the pipe diameter size deviation; The generating module 304 is configured to generate, by the adaptive fuzzy inference system, a coordinated control instruction set of the roll gap and the welding power according to the development trend, so as to realize online closed-loop control of the forming process parameters.

[0088] The embodiment of the present application further provides a storage medium, and the storage medium has a computer program stored therein, wherein the computer program is arranged to execute the steps in any one of the method embodiments.

[0089] The embodiment of the present application further provides an electronic device, comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is arranged to execute the computer program to execute the steps in any one of the method embodiments.

[0090] Specifically, the electronic device can further comprise a transmission device and an input / output device, wherein the transmission device is connected with the processor, and the input / output device is connected with the processor.

[0091] The above embodiment according to the drawings illustrates the structure, features and effects of the present application, and the above description is only the preferred embodiment of the present application, but the present application is not limited to the drawings, and any change or modification within the scope of the present application, or equivalent embodiments within the scope of the present application, should be within the protection scope of the present application.

Claims

1. A real-time closed-loop control method for the forming process of stainless steel welded pipes, characterized in that, The method includes: Multi-source heterogeneous data, including pipe morphology images, laser ranging sequences, temperature field distribution, and roll pressure time series data, are synchronously collected during the forming process through a distributed multi-source sensor network. The multi-source heterogeneous data is spatiotemporally aligned and feature-level fused to construct a multi-dimensional dynamic digital twin of the forming process. Specifically, timestamps from each data source are extracted from the preprocessed multi-source heterogeneous dataset, and a cubic spline interpolation algorithm is used to unify all data sequences to the same millisecond-level time reference, generating a time-strictly aligned multi-source data sequence. Based on sensor calibration parameters and the mechanical structure coordinate system, the time-strictly aligned multi-source data sequence is mapped to a unified three-dimensional spatial coordinate system with the rolling centerline as the reference. Rigid body transformation is used to align the laser point cloud with the image pixel coordinates, generating a spatially registered multi-source data field. Multimodal features are extracted from the spatially registered multi-source data field, including weld edge contour features based on the Canny operator, heat-affected zone features based on temperature gradients, rolling force distribution features based on pressure waveforms, and pipe diameter ellipticity features based on point clouds, generating a multimodal feature vector set. The multimodal feature vector set is input into a physics-based modeling engine, combined with a finite element analysis model and a material rheology model, to update the state parameters of the digital twin in real time, constructing a multi-dimensional dynamic digital twin reflecting the dynamic changes of the forming process. Based on the aforementioned multidimensional dynamic digital twin, an online rolling prediction model is used to predict in real time the development trend of weld formation quality and pipe diameter deviation. Based on the aforementioned development trend, an adaptive fuzzy inference system is used to generate a set of coordinated control instructions for roll gap and welding power, thereby achieving online closed-loop control of forming process parameters.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous data, including pipe morphology images, laser ranging sequences, temperature field distribution, and roll pressure time series data, which are synchronously collected during the forming process through a distributed multi-source sensor network, includes: Design and deploy a multi-sensor array, including a high-speed industrial camera, a laser displacement sensor array, an infrared thermal imager, and a distributed pressure sensor, which are installed at the inlet, welding area, sizing area, and outlet stations of the forming machine, respectively, to generate a sensor spatial layout scheme. Based on the sensor spatial layout scheme, the IEEE 1588 precise time protocol is used to synchronize the hardware clock of all sensor nodes to ensure the microsecond-level time consistency of data acquisition and generate a synchronous clock network. Based on the synchronous clock network, each sensor node is triggered to collect data synchronously. A high-speed industrial camera captures the surface morphology image of the pipe, a laser displacement sensor array scans the pipe diameter contour point cloud, an infrared thermal imager collects the temperature field distribution of the weld, and a pressure sensor records the timing waveform of the roll pressure, generating a raw multi-source data stream. The original multi-source data stream is preprocessed in real time. Wavelet denoising algorithm is used to filter out pressure signal noise, adaptive histogram equalization is applied to enhance image contrast, and outlier points in the laser point cloud are removed to generate a preprocessed multi-source heterogeneous dataset.

3. The method according to claim 2, characterized in that, The method based on the multidimensional dynamic digital twin, employing an online rolling prediction model to predict in real time the development trend of weld formation quality and pipe diameter deviation, includes: Key state variables at the current moment are extracted from the multidimensional dynamic digital twin, including weld width, weld height, heat-affected zone width, pipe diameter distribution, and rolling force distribution, to form the input feature vector of the prediction model. The input feature vector is fed into a pre-trained online rolling prediction model, which adopts a structure combining a gated recurrent unit network and an attention mechanism to output a sequence of predicted weld formation quality indicators and pipe diameter for the next ten sampling periods. Based on the deviation between the predicted sequence calculation and the process standard value, and using a sliding time window mechanism, the input feature vector is updated and rolling prediction is performed again as soon as new sensor data is received, so as to realize the online adaptive update of model parameters. Based on the rolling prediction results, the spatiotemporal evolution curves of weld formation quality and pipe diameter deviation are plotted, the critical points and trend inflection points of accelerated deviation increase are identified, and a development trend report including early warning levels is generated.

4. The method according to claim 3, characterized in that, Based on the aforementioned development trend, an adaptive fuzzy inference system is used to generate a set of coordinated control instructions for roll gap and welding power, thereby achieving online closed-loop control of forming process parameters, including: Key input variables were extracted from the development trend report, including weld width deviation, pipe diameter ellipticity deviation, temperature gradient deviation, and deviation change rate. The fuzzification process was performed using a triangular membership function to generate a set of fuzzy input variables. An adaptive fuzzy rule base is constructed, and its rule weights are dynamically adjusted through reinforcement learning based on historical control effect data. The fuzzy input variable set and fuzzy rules are subjected to Mamdani inference to generate a fuzzy output set representing the roll gap adjustment amount and welding power adjustment amount. The centroid method defuzzification algorithm is used to convert the fuzzy output set into precise control values, generating preliminary coordinated control commands that include the gap adjustment values ​​of the upper roll, lower roll, left roll, and right roll, as well as the power adjustment value of the high-frequency welding power source. Based on the dynamic response characteristics of the molding machine actuator and process safety constraints, the amplitude and rate limits of the preliminary coordinated control commands are optimized to generate the final executable coordinated control command set, which is then sent to each actuator via fieldbus to achieve online closed-loop precise control of process parameters.

5. A real-time closed-loop control system for the forming process of stainless steel welded pipes, characterized in that, The system includes: The acquisition module is used to synchronously acquire multi-source heterogeneous data, including pipe morphology images, laser ranging sequences, temperature field distribution, and roll pressure time series data, through a distributed multi-source sensor network during the forming process. The construction module is used to perform spatiotemporal alignment and feature-level fusion on the multi-source heterogeneous data to construct a multi-dimensional dynamic digital twin of the forming process. Specifically, it extracts the timestamps of each data source from the preprocessed multi-source heterogeneous dataset, and uses a cubic spline interpolation algorithm to unify all data sequences to the same millisecond-level time reference, generating a time-strictly aligned multi-source data sequence. Based on sensor calibration parameters and the mechanical structure coordinate system, it maps the time-strictly aligned multi-source data sequence to a unified three-dimensional spatial coordinate system with the rolling centerline as the reference. Rigid body transformation is used to connect the laser point cloud with the image... The system aligns primitive coordinates to generate a multi-source data field with spatial coordinate registration. Multimodal features are extracted from this data field, including weld edge contour features based on the Canny operator, heat-affected zone features based on temperature gradients, rolling force distribution features based on pressure waveforms, and pipe diameter ellipticity features based on point clouds, generating a multimodal feature vector set. This multimodal feature vector set is then input into a physics-based modeling engine, combined with a finite element analysis model and a material rheology model, to update the state parameters of the digital twin in real time, constructing a multidimensional dynamic digital twin that reflects the dynamic changes during the forming process. The deduction module is used to deduce the development trend of weld formation quality and pipe diameter deviation in real time based on the multidimensional dynamic digital twin using an online rolling prediction model; The generation module is used to generate a set of coordinated control instructions for roll gap and welding power through an adaptive fuzzy inference system based on the development trend, so as to realize online closed-loop control of forming process parameters.

6. The system according to claim 5, characterized in that, The acquisition module is specifically used for: Design and deploy a multi-sensor array, including a high-speed industrial camera, a laser displacement sensor array, an infrared thermal imager, and a distributed pressure sensor, which are installed at the inlet, welding area, sizing area, and outlet stations of the forming machine, respectively, to generate a sensor spatial layout scheme. Based on the sensor spatial layout scheme, the IEEE 1588 precise time protocol is used to synchronize the hardware clock of all sensor nodes to ensure the microsecond-level time consistency of data acquisition and generate a synchronous clock network. Based on the synchronous clock network, each sensor node is triggered to collect data synchronously. A high-speed industrial camera captures the surface morphology image of the pipe, a laser displacement sensor array scans the pipe diameter contour point cloud, an infrared thermal imager collects the temperature field distribution of the weld, and a pressure sensor records the timing waveform of the roll pressure, generating a raw multi-source data stream. The original multi-source data stream is preprocessed in real time. Wavelet denoising algorithm is used to filter out pressure signal noise, adaptive histogram equalization is applied to enhance image contrast, and outlier points in the laser point cloud are removed to generate a preprocessed multi-source heterogeneous dataset.

7. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-4 when it is run.

8. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-4.

Images