Friction stir welding digital twin system based on multi-sensing data fusion and method thereof

Abstract

The invention provides a friction stir welding digital twinning system and method based on multi-sensing data fusion, and belongs to the technical field of advanced welding manufacturing and industrial digitization crossing. The twinning system comprises a physical sensing layer, a data acquisition and fusion layer, a digital twinning model layer and a twinning interaction and service platform; the physical perception layer is connected with the data acquisition and fusion layer, the data acquisition and fusion layer is connected with the digital twinborn model layer, and the digital twinborn model layer is connected with the twinborn interaction and service platform. According to the method, heat-force-material flow multi-dimensional information is covered on the basis of multi-sensor fusion, and the defect detection rate is increased to 98% and is increased by 13% compared with a traditional method; the error of the mixed twin model is reduced by 40% compared with that of a pure data driving model, and the real-time performance reaches a millisecond level; the transformation from post detection to online prediction-real-time intervention is realized, and the rejection rate is reduced by 60%.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of advanced welding manufacturing and industrial digitization, and in particular to a friction stir welding digital twin system and method based on multi-sensor data fusion. Background Art

[0002] As a green solid-phase joining technology, friction stir welding has become the preferred process for welding light alloy materials such as aluminum alloys due to its advantages such as low heat input, small deformation and no smoke. However, its welding process involves complex multi-physical field coupling (heat-force-material flow), and the quality of the weld is highly dependent on the matching of process parameters (such as rotation speed, welding speed, and downforce). Traditional quality control methods have the following limitations: 1. Process invisibility: The welding area is completely covered by the stirring head shoulder, and the plastic flow and defect formation process of the material cannot be directly observed; 2. Offline detection lag: It relies on non-destructive testing such as post-weld X-ray and ultrasonic testing and cannot intervene in real time; 3. Inefficient parameter optimization: Process development relies on "trial and error", which is time-consuming and material-consuming. At present, the monitoring of the friction stir welding process mainly uses a single or a small number of sensors (such as temperature and axial force), which has significant shortcomings such as one-sided information, noise interference, and display data islands.

[0003] In recent years, digital twin technology has shown great potential in intelligent manufacturing, but its application in the field of friction stir welding still faces challenges. Existing twin models mostly rely on simplified mechanisms or static data-driven models, which are difficult to reflect the dynamic welding process in real time. Moreover, pure data-driven models require a large number of labeled samples and have poor physical interpretability. More importantly, most systems do not achieve efficient closed-loop interaction between multi-source sensor data and twin models. For example: Patent CN114012345A proposes a FSW monitoring system based on temperature and force signals, but only realizes threshold alarms and lacks multi-data fusion and prediction capabilities; CN115678901B uses a finite element model to simulate the welding process, but does not integrate real-time sensor data and cannot be dynamically updated; CN116543210A attempts to use digital twins for welding training, but does not solve the quality control problems in actual production. Summary of the Invention

[0004] The present invention aims to provide a friction stir welding digital twin system and method based on multi-sensor data fusion, addressing the technical problem of existing systems' difficulty in reflecting the dynamic welding process in real time. This system aims to overcome the one-sided nature of single-sensor information, generate a more accurate fused feature dataset, and achieve high-fidelity virtual mapping and real-time synchronous updates of the welding process, thereby improving welding quality and production efficiency.

[0005] Through a multi-sensor deep fusion architecture based on the complementary integration of multimodal data such as force, heat, vibration, and acoustic emission, the limitations of single sensor information are overcome. Furthermore, the twin is dynamically modeled by combining a mechanism model (physical constraints) with a data-driven model (real-time adaptability) to construct a high-fidelity virtual welding process. Finally, the twin is used to achieve intelligent decision-making throughout the entire chain, from monitoring and prediction to optimization, to build a closed-loop service system.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] A friction stir welding digital twin system based on multi-sensor data fusion, comprising a physical perception layer, a data acquisition and fusion layer, a digital twin model layer, and a twin interaction and service platform. The physical perception layer is connected to the data acquisition and fusion layer, which is then connected to the digital twin model layer, which is then connected to the twin interaction and service platform.

[0008] The physical perception layer includes a multi-source heterogeneous sensor network deployed on the friction stir welding equipment to collect dynamic physical quantity signals during the welding process in real time;

[0009] The data acquisition and fusion layer is connected to the sensor network to pre-process the collected physical quantity signals and generate a fusion feature data set using data fusion methods;

[0010] The digital twin model layer includes a physical mechanism-based welding process model and / or a data-driven model, which is driven and dynamically updated in real time according to the fused feature data set to build a virtual twin synchronized with the physical welding process;

[0011] The twin interaction and service platform is used to realize two-way data interaction between the virtual twin and the actual welding process, and provide at least one service function of welding process monitoring, quality prediction, process optimization and defect diagnosis.

[0012] Furthermore, the multi-source heterogeneous sensor network includes an axial pressure sensor, a temperature sensor, a vibration sensor, an acoustic emission sensor and a visual sensor. The axial pressure sensor is used to monitor the axial pressure of the stirring head, the spindle torque sensor is used to monitor the rotational torque of the stirring head, the temperature sensor is used to monitor the temperature distribution in the welding area, the vibration sensor is used to collect mechanical vibration signals during the welding process, the acoustic emission sensor is used to detect material micro-deformation and defect signals during the welding process, and the visual sensor is used to obtain the surface morphology of the weld or thermal radiation image.

[0013] Furthermore, the data fusion method is one or more of a multi-sensor data weighted fusion method based on Kalman filtering, a heterogeneous data decision-level fusion method based on DS evidence theory, and an end-to-end multimodal feature fusion model method based on deep learning.

[0014] Furthermore, the digital twin model layer includes mechanism models, data-driven models and hybrid models. The mechanism model simulates the thermal material flow coupling behavior in the welding process based on finite element analysis or computational fluid dynamics. The data-driven model is a quality prediction model or defect classification model constructed based on machine learning or deep learning algorithms. The hybrid model is a modeling method that combines physical mechanism constraints with data-driven correction.

[0015] Furthermore, the functions of the twin interaction and service platform include: real-time monitoring and visualization: dynamic display of the three-dimensional simulation of force, heat and material flow and key parameter curves of the welding process; online quality prediction: based on the fusion of feature data sets and twin model output, real-time prediction of the probability of internal defects or mechanical performance indicators of the weld; process parameter optimization: through virtual twin simulation of welding results with different parameter combinations, output of optimal process parameter recommendations; defect root cause analysis: when an anomaly is detected, the cause of the defect is traced back by combining historical data and the intermediate state of the model.

[0016] Furthermore, online quality prediction outputs the defect probability of the future welding section by inputting real-time fusion features into a pre-trained LSTM neural network model, and judges the potential defect area based on the stress field and temperature field distribution simulated by the mechanism model and the preset threshold.

[0017] Furthermore, the above system also includes a digital archive module for storing full-cycle data of the welding process, including raw sensor data, fusion features, twin status and service output results, to support process traceability and data analysis.

[0018] The system includes building a multi-source heterogeneous sensor network; collecting and fusing multi-source data; building a digital twin model layer; and developing a twin interaction and service platform.

[0019] The multi-source heterogeneous sensor network includes the layout and development of various sensors used in friction stir welding. The mechanical sensors include an axial pressure sensor and a spindle torque sensor, respectively monitoring downforce and rotational resistance. Thermal sensors include thermocouples and infrared thermal imagers for contact and non-contact temperature field data acquisition. Dynamic signal sensors include vibration accelerometers for detecting tool wear or material flow anomalies and acoustic emission sensors for capturing microscopic defect signals. Visual sensors include a high-speed camera and a structured light 3D scanner for observing material surface morphology and acquiring weld geometry, respectively.

[0020] Determine the sensor installation location and sampling frequency based on the distribution of key physical fields in the welding process, such as the plastic flow zone and heat-affected zone. Optimize the sensor layout.

[0021] The data acquisition and fusion layer includes signal preprocessing, feature extraction, and data fusion algorithms. Signal preprocessing primarily uses wavelet thresholding to denoise vibration signals and Kalman filtering to preprocess temperature signals. Time-domain features such as force and torque are extracted using mean, variance, and peak values. Fast Fourier transform energy is used for vibration signals, and wavelet packet energy entropy is used for acoustic signal features. The data fusion algorithm operates at three levels: the data layer, the feature layer, and the decision layer. At the data layer, sensors of the same type are fused using weighted averaging or Kalman filtering. The feature layer often uses convolutional neural networks to fuse the spatiotemporal and frequency-domain features of vibration and acoustic emissions. The decision layer, based on DS evidence theory, outputs defect probabilities based on a comprehensive analysis of force, heat, and vibration signals.

[0022] The digital twin model layer includes: mechanistic models, such as the thermal-mechanical coupled finite element model, to simulate welding temperature and stress-strain fields; computational fluid dynamics models to predict material flow behavior; data-driven models, such as long-short-term memory networks, to map feature sequences to defect probabilities; and residual convolutional neural networks (ResNet-18) to classify defects based on infrared thermal images and vibration signatures.

[0023] The twin interaction and service platform includes: real-time monitoring, quality prediction, process optimization, and defect diagnosis; it uses a three-dimensional visual digital middle platform to dynamically display temperature fields, material flow simulation, and key parameter curves; based on fusion features and twins, it predicts defects such as incomplete penetration and holes 3 seconds in advance; through genetic effects, it simulates parameter combinations in the twins and recommends the optimal rotation speed and welding speed; finally, it combines historical data with the output of the model middle layer to locate the cause of the defect.

[0024] A method for a friction stir welding digital twin system based on multi-sensor data fusion, the method comprising the following steps:

[0025] S1: Real-time acquisition of physical quantity signals during friction stir welding through a multi-source heterogeneous sensor network;

[0026] S2: Preprocessing the physical quantity signal and generating a fusion feature data set using a data fusion method;

[0027] S3: using the fused feature dataset to drive the digital twin model and construct a virtual twin synchronized with the physical welding process;

[0028] S4: Based on the virtual twin, perform at least one service function of real-time monitoring, quality prediction, process optimization or defect diagnosis.

[0029] Furthermore, the data fusion method described in step S2 includes: performing spatiotemporal alignment and weighted fusion of the temperature sensor and infrared thermal imager data to generate a high-precision temperature field distribution; extracting time-frequency domain features of the vibration signal and the acoustic emission signal, and outputting a defect feature vector after fusing them through a convolutional neural network.

[0030] Furthermore, the process optimization function described in step S4 is achieved by simulating the welding process under different combinations of rotation speed, welding speed and downward pressure in the virtual twin; using the genetic algorithm to solve the optimal process parameter combination with the prediction of weld quality and energy consumption as the optimization goal.

[0031] The present invention has the following beneficial effects due to the adoption of the above technical solution:

[0032] This method, based on multi-sensor fusion covering multi-dimensional information from heat, force, and material flow, has increased the defect detection rate to 98%, a 13 percentage point improvement over traditional methods. The hybrid twin model reduces error by 40% compared to purely data-driven models, achieving real-time performance in milliseconds. By shifting from post-detection to online prediction and real-time intervention, the scrap rate has dropped by 60%. The parameter optimization cycle has been shortened from two weeks of traditional trial-and-error to two hours. This approach achieves comprehensive information, high-fidelity models, proactive quality control, and efficient process optimization, significantly improving the efficiency of intelligent quality control for friction stir welding. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a system block diagram of the present invention;

[0034] Figure 2 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0035] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and by way of preferred embodiments. However, it should be noted that many of the details listed in this specification are merely provided to help the reader gain a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be practiced even without these specific details.

[0036] Example 1: Friction stir welding quality control of aluminum alloy cabin for aerospace

[0037] The longitudinal seam welding of a spacecraft fuel tank (material: AA2219 aluminum alloy, thickness 8mm) requires the weld to be defect-free and the tensile strength to be ≥90% of the parent material.

[0038] The core of the present invention is to provide a friction stir welding digital twin system based on multi-source sensor data fusion, such as Figure 1-2As shown, it is used for the quality control of friction stir welding of aluminum alloy cabins for aerospace. According to the invention content, the technical solution is clearly and completely described in detail. The establishment of the system includes the following steps:

[0039] S101 builds a multi-source heterogeneous sensor network, including sensor layout and data acquisition system.

[0040] The sensor consists of three units: mechanical monitoring, temperature monitoring, and dynamic signal monitoring. The mechanical monitoring unit primarily uses the Kistler 9272A for axial pressure monitoring and the HBM T40B for spindle torque monitoring. This unit is installed in a pre-installed mounting cavity within the stirrer. The temperature monitoring unit primarily uses OMEGATJ36-CASS-116G K-type thermocouples, arranged at six points on either side of the weld centerline for contact temperature measurement; a FLIRA655sc infrared thermal imager, mounted in the friction welding chamber for non-contact temperature measurement; and a PCB 352C03 triaxial accelerometer, installed on the rigid connection surface of the stirrer support. Vibration signals are collected using a Physical Acoustics R15α acoustic emission sensor, mounted on the workbench, 20 mm from the weld center on the back of the workpiece, to collect acoustic emission signals during the welding process.

[0041] The data acquisition system includes acquisition equipment and data synchronization methods. It uses the NI cDAQ-9188XT chassis equipped with NI9234, NI9213, and NI9239 modules for data acquisition, and uses the PXI-6683H timing module to achieve multi-channel μs-level synchronization.

[0042] S102 builds the data acquisition and fusion layer, including signal preprocessing and multi-source data fusion.

[0043] Signal preprocessing primarily targets temperature and vibration signals. The temperature signal uses an ADS1118 module to acquire ambient temperature in real time for cold-junction compensation of thermocouples. A second-order Butterworth filter with a cutoff frequency of 10 Hz is used to filter out temperature noise. Finally, a two-dimensional temperature field is reconstructed using a thin-plate spline function. The vibration signal is filtered using a 4-8 kHz bandpass filter, and the wavelet energy entropy is calculated within each 0.1 s time window.

[0044] Among them, multi-source data fusion mainly processes the data layer and feature layer; for multiple temperature sensors, an adaptive weighted fusion algorithm is used. The weight calculation method is shown in the following formula:

[0045]

[0046] Among them, σ iis the standard deviation of the data of the i-th sensor in the recent 10 seconds.

[0047] During the feature layer fusion process, the input feature dimensions are first defined. Temperature field features are defined as 5-dimensional inputs, including maximum temperature, gradient, and asymmetry; mechanical features are defined as 4-dimensional coefficients, including axial force mean and torque fluctuation coefficient; and dynamic features are defined as 6-dimensional coefficients, including vibration energy entropy and acoustic emission RSM. Finally, these inputs are fed into a CNN-LSTM hybrid network structure, consisting of 1-dimensional convolutional layers, max pooling layers, long short-term memory modules, and fully connected layers.

[0048] S103 builds a digital twin model, including mechanism model and data-driven model construction.

[0049] The control equations are solved using COMSOL Multiphysics 6.0 software. See the following equation:

[0050]

[0051] The boundary conditions are the heat input distribution based on the inverse fitting of the measured temperature field and the measured axial force as the mechanical constraint.

[0052] The construction of the data-driven model focuses on a quality prediction model. The network structure of this model consists of a three-layer long short-term memory network with 64, 32, and 16 hidden layers respectively. The training data of the model is the fusion feature sequence of 356 sets of historical welding processes. The output is the probability of hole defects in the next 5 seconds P. defect ∈[0,1).

[0053] In order to achieve real-time data synchronization, the fusion features are input into the mechanism model every 0.5 s to update the heat source efficiency η = η0 + Δη and the friction coefficient μ = μ0 + k·Torque_fluctuation.

[0054] S104 twin interaction and service platform, including: real-time graphical interactive interface, system interaction, and performance statistics;

[0055] Among them, the system main interface includes: three-dimensional temperature field cloud map, material example tracing animation, and key parameter trend curve.

[0056] Among them, system interaction includes: working condition records and system responses.

[0057] Among them, performance statistics include: monitoring indicators, traditional methods, and the system of the present invention.

[0058] Table 1 shows the main technical parameters of the system

[0059]

[0060] In response to the problems of existing stir friction welding process monitoring methods being single, data being isolated, lacking real-time performance, and making it difficult to accurately predict welding quality, the present invention uses a multi-source heterogeneous sensor network to collect key physical quantities of the welding process in real time, and adopts a data fusion algorithm to fuse the multi-source data at the feature level or decision level to generate a high-precision, highly robust fusion feature data set. Based on the fused data, a digital twin model of the stir friction welding process is constructed and updated in real time, including a mechanism model and a data-driven model, to achieve high-fidelity virtual mapping of the welding process. The system supports real-time visual monitoring of welding status, online prediction of weld quality, dynamic optimization of process parameters, and intelligent diagnosis of defects, significantly improving the transparency and quality control capabilities of the welding process, reducing the defect rate, and improving production efficiency and process stability.

[0061] Matters not covered by the present invention are known technologies.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A friction stir welding digital twin system based on multi-sensor data fusion, characterized by: Physical perception layer, data acquisition and fusion layer, digital twin model layer and twin interaction and service platform. The physical perception layer is connected to the data acquisition and fusion layer, the data acquisition and fusion layer is connected to the digital twin model layer, and the digital twin model layer is connected to the twin interaction and service platform. The physical perception layer includes a multi-source heterogeneous sensor network deployed on the friction stir welding equipment to collect dynamic physical quantity signals during the welding process in real time; The data acquisition and fusion layer is connected to the sensor network to pre-process the collected physical quantity signals and generate a fusion feature data set using data fusion methods; The digital twin model layer includes a physical mechanism-based welding process model and / or a data-driven model, which is driven and dynamically updated in real time according to the fused feature data set to build a virtual twin synchronized with the physical welding process; The twin interaction and service platform is used to realize two-way data interaction between the virtual twin and the actual welding process, and provide at least one service function of welding process monitoring, quality prediction, process optimization and defect diagnosis.

2. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 1 is characterized in that: The multi-source heterogeneous sensor network includes axial pressure sensors, temperature sensors, vibration sensors, acoustic emission sensors and vision sensors. The axial pressure sensor is used to monitor the axial pressure of the stirring head, the spindle torque sensor is used to monitor the rotational torque of the stirring head, the temperature sensor is used to monitor the temperature distribution in the welding area, the vibration sensor is used to collect mechanical vibration signals during the welding process, the acoustic emission sensor is used to detect material micro-deformation and defect signals during the welding process, and the vision sensor is used to obtain the surface morphology of the weld or thermal radiation image.

3. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 1 is characterized in that: The data fusion method is one or more of a multi-sensor data weighted fusion method based on Kalman filtering, a heterogeneous data decision-level fusion method based on DS evidence theory, and an end-to-end multimodal feature fusion model method based on deep learning.

4. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 1, characterized in that: The digital twin model layer includes mechanism models, data-driven models and hybrid models. The mechanism model simulates the thermal material flow coupling behavior in the welding process based on finite element analysis or computational fluid dynamics. The data-driven model is a quality prediction model or defect classification model built based on machine learning or deep learning algorithms. The hybrid model is a modeling method that combines physical mechanism constraints with data-driven correction.

5. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 1, characterized in that: The functions of the twin interaction and service platform include: real-time monitoring and visualization: dynamic display of the three-dimensional simulation of force, heat and material flow and key parameter curves of the welding process; online quality prediction: based on the fusion of feature data sets and twin model output, real-time prediction of the probability of internal defects or mechanical performance indicators of the weld; process parameter optimization: through virtual twin simulation of welding results with different parameter combinations, output of optimal process parameter recommendations; defect root cause analysis: when an anomaly is detected, the cause of the defect is traced back by combining historical data and the intermediate state of the model.

6. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 5, characterized in that: Online quality prediction inputs real-time fusion features into a pre-trained LSTM neural network model to output the defect probability of the future welding section. Based on the stress and temperature field distribution simulated by the mechanism model and the preset threshold, the potential defect area is judged.

7. The friction stir welding digital twin system based on multi-sensor data fusion according to claim 1, characterized in that: It also includes a digital archive module for storing full-cycle data of the welding process, including raw sensor data, fusion features, twin status and service output results, to support process traceability and data analysis.

8. The method of a friction stir welding digital twin system based on multi-sensor data fusion according to any one of claims 1 to 7, characterized in that: The method comprises the following steps: S1: Real-time acquisition of physical quantity signals during friction stir welding through a multi-source heterogeneous sensor network; S2: Preprocessing the physical quantity signal and generating a fusion feature data set using a data fusion method; S3: using the fused feature dataset to drive the digital twin model and construct a virtual twin synchronized with the physical welding process; S4: Based on the virtual twin, perform at least one service function of real-time monitoring, quality prediction, process optimization or defect diagnosis.

9. The method of a friction stir welding digital twin system based on multi-sensor data fusion according to claim 8, characterized in that: The data fusion method described in step S2 includes: performing spatiotemporal alignment and weighted fusion of the temperature sensor and infrared thermal imager data to generate a high-precision temperature field distribution; extracting time-frequency domain features of the vibration signal and the acoustic emission signal, and outputting a defect feature vector after fusing them through a convolutional neural network.

10. The method of a friction stir welding digital twin system based on multi-sensor data fusion according to claim 8, characterized in that: The process optimization function described in step S4 is achieved by simulating the welding process under different combinations of rotation speed, welding speed and downward pressure in the virtual twin; using the genetic algorithm to solve the optimal process parameter combination with the prediction of weld quality and energy consumption as the optimization goal.

Images