Full-thickness three-dimensional residual stress field reconstruction method and multi-source detection device

Abstract

A method for reconstructing a three-dimensional residual stress field of full thickness and a multi-source detection device are disclosed, relating to the interdisciplinary field of nondestructive testing and artificial intelligence. The stress field reconstruction method comprises the following steps: acquiring first stress data from the surface of the component under test and second acoustic data from its interior; constructing a physical information-driven neural network model, using the first stress data as spatial boundary conditions and converting the second acoustic data into internal physical field constraints, configuring the composite loss function of the model; iteratively optimizing the solution; and using a multi-source detection device to reconstruct the residual stress field, which includes an X-ray diffraction module for acquiring surface stress data, an ultrasonic detection module for acquiring internal acoustic data, and a data processing control unit for executing the reconstruction method. This invention reconstructs the three-dimensional residual stress field of the component under test by using surface stress data as spatial boundary conditions, converting acoustic data into internal physical field constraints, and introducing mechanical equilibrium equations as physical information loss terms.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of nondestructive testing and machine learning, and in particular to a method for reconstructing a three-dimensional residual stress field of full thickness and a multi-source detection device. Background Technology

[0002] Residual stress is the internal stress that remains in equilibrium within a component when no external force is applied. Its distribution directly determines the fatigue strength, dimensional stability, and service safety of the workpiece. Especially for large metal components such as rolls, the presence of harmful tensile stress concentrations within the component, as well as the stress gradient changes on the surface, are key factors affecting its performance and preventing fracture failure. Therefore, obtaining a high-precision three-dimensional residual stress distribution across the entire thickness of the component from the surface to its core is a core prerequisite for achieving quality control in high-end equipment manufacturing.

[0003] Currently, the industry mainly uses X-ray diffraction (XRD) and ultrasonic methods for non-destructive testing of residual stress, but both have significant limitations in principle: Limitations of X-ray diffraction (XRD): XRD technology, based on the principle of crystal diffraction, calculates stress by measuring lattice strain, offering high accuracy in surface measurements and is a recognized standard method for stress detection. However, due to the extremely shallow penetration depth of X-rays (typically 10–30 μm), this technology can only obtain two-dimensional stress information on the surface of a component, failing to perceive the stress state at the subsurface and deeper internal layers. For thick components, surface data alone cannot represent the overall stress situation and is insufficient to support a full-section safety assessment.

[0004] The ultrasonic testing method has insufficient resolution: While it leverages the acoustoelastic effect and utilizes the change in wave velocity under stress for inversion, possessing strong penetrating power (up to several meters), its drawbacks are: First, low spatial resolution. The signal received by the ultrasonic probe is the superposition of all media interactions along the sound wave propagation path, measuring only the average stress along the propagation path and failing to analyze the stress gradient distribution along the depth direction. Second, poor anti-interference capability. The inhomogeneity of the material's microstructure (such as texture and grain size) significantly affects the sound velocity, resulting in low absolute accuracy of stress inversion, making it difficult to meet high-precision testing requirements.

[0005] In industrial practice, a combined strategy of "XRD calibration + ultrasonic testing" is typically used to balance depth and accuracy. This involves first measuring the stress values ​​at several discrete points on the surface using X-rays, and then using these values ​​to correct the acoustoelastic coefficient measured by ultrasound. However, this combined strategy merely utilizes data superficially and lacks a deep physical model coupling, thus exhibiting the following inherent flaws: Data silos: The discrete stress data obtained by XRD and the path integral data obtained by ultrasound are physically independent (not organically unified) and lack a unified mathematical model for correlation and constraint.

[0006] Limited reconstruction methods: For areas that have not been measured, existing technologies mostly use linear interpolation or empirical formulas for estimation, which cannot construct a three-dimensional stress field that simultaneously satisfies surface boundary conditions and the continuity of the internal physical field.

[0007] Lack of physical consistency: Existing data processing often ignores prior physical knowledge such as the equilibrium equations of elasticity, which may lead to non-physical abrupt changes or discontinuities in the reconstructed stress field, resulting in a lack of physical self-consistency and difficulty in guaranteeing reliability.

[0008] In summary, existing technologies suffer from problems such as "high surface precision but low internal resolution" and "disconnect between physical mechanisms and data-driven approaches," failing to provide a full-thickness three-dimensional residual stress field reconstruction scheme that can integrate multi-source heterogeneous data and strictly adhere to physical conservation laws. Particularly for roll components, key hidden dangers such as abrupt stress changes in the internal quenching transition zone and residual tensile stress in the core cannot be effectively identified and quantified using existing methods. Therefore, a novel detection method and device are urgently needed to achieve accurate and visual assessment of the internal stress state of components. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a full-thickness three-dimensional residual stress field reconstruction method and a multi-source detection device. It solves the problems of shallow detection depth of existing X-ray diffraction method (cannot know the internal stress), low spatial resolution of ultrasonic detection method (difficult to analyze depth gradient), and lack of physical model coupling of multi-source detection data (poor physical self-consistency and low reliability of reconstruction results).

[0010] The technical solution of this invention is: a method for reconstructing a three-dimensional residual stress field of full thickness, the steps of which are as follows: S1. Acquire test data: Acquire the first stress data on the surface of the component under test, and acquire the second acoustic data inside the component under test; S2. Constructing a Prediction Model: Constructing a physical information-driven neural network model, which is a physical information neural network, takes the three-dimensional spatial coordinates of the measured component as input and the stress tensor components of the corresponding coordinates as output; simultaneously, the first stress data is used as the spatial boundary condition, and the second acoustic data is converted into internal physical field constraints. Based on the spatial boundary condition and the internal physical field constraints, a composite loss function of the model is configured; the composite loss function includes at least a data loss term that measures the difference between the model's predicted value and the first stress data and the second acoustic data, and a physical information loss term that measures the residual of the model's predicted value satisfying the mechanical equilibrium equation; S3. Iterative optimization solution: The parameters of the model are adjusted iteratively to minimize the composite loss function until the model converges; S4. Reconstructing the residual stress field: Call the converged model to perform forward inference calculations on the discrete spatial coordinate points within the full thickness range of the tested component, and output the stress tensor components of the corresponding coordinate points, thereby reconstructing the three-dimensional residual stress field of the full thickness of the tested component.

[0011] A further technical solution of the present invention is as follows: In step S1, the first stress data includes lattice strain data obtained by performing sin²ψ measurement on the surface of the component under test by X-ray diffraction, or the surface residual stress value calculated from the lattice strain data; the second acoustic data includes the ultrasonic flight time measured by ultrasonic detection, and / or the rate of change of sound velocity calculated from the ultrasonic flight time, and / or acoustoelastic birefringence data.

[0012] A further technical solution of the present invention is: in step S2, the composite loss function Represented as formulas 1 to 3; Formula 1: ; in, This is a data loss term used to measure the difference between the rate of change of sound velocity corresponding to the stress location predicted by the model and the rate of change of sound velocity in the second acoustic data. This is the physical information loss term, used to measure the residuals of the model's predicted stress field satisfying the elastic equilibrium equations at the internal sampling points; For the weighting coefficients of the data loss term, The weighting coefficients for the physical information loss term; Formula 2: ; in, The rate of change of sound speed in the second acoustic data; The rate of change of sound velocity is obtained from the transformation of the full-thickness three-dimensional residual stress field based on model prediction; Let m be the m-th ultrasound propagation path; m is the index of the ultrasound propagation path; M is the total number of ultrasound propagation paths; Formula 3: ; in, This represents the total number of internal sampling points; Let i be the spatial coordinates of the i-th internal sampling point; This is the predicted value of the stress tensor; The divergence of the stress tensor, where i is the index of the internal sampling point.

[0013] A further technical solution of the present invention is as follows: The operation of "acquiring the first stress data of the surface of the component under test" in step S1 includes: selecting at least 3 measurement areas on the surface of the component under test using an X-ray diffraction module, collecting lattice strain data of no less than 9 points in each area, and calculating the surface residual stress value of each point using the sin²ψ method; The operation of "acquiring the second acoustic data inside the component under test" in step S1 includes: arranging a transmitter and a receiver on the surface of the component under test using an ultrasonic testing module, so that the ultrasonic wave propagation path runs through the full thickness of the component under test; collecting the ultrasonic wave flight time and converting the flight time into the rate of change of sound velocity; collecting the shear wave splitting signal of the ultrasonic wave, calculating the arrival time difference or sound velocity difference between the fast wave and the slow wave, thereby obtaining the acoustoelastic birefringence data.

[0014] A further technical solution of the present invention is: an intermediate step M is provided after step S1 and before step S2; M. Data Preprocessing and Coordinate Registration: The first stress data is denoised, and the second acoustic data is spatially registered so that the two types of data can be processed in a unified spatial coordinate system.

[0015] A further technical solution of the present invention is: the network structure of the Physical Information Neural Network (PINN) in step S2 includes an input layer, a hidden layer and an output layer. The hidden layer adopts a fully connected layer, and a skip connection mechanism is introduced in the fully connected layer to improve the propagation efficiency of the deep network.

[0016] A further technical solution of the present invention is: the "internal residual sampling point" in step S2 The number of “10” 3 ~10 5 The sampling method uses Latin hypercube sampling to ensure uniform spatial distribution.

[0017] A further technical solution of the present invention is that the values ​​of various weight coefficients in step S2 satisfy: α=0.8~1.2, β=0.5~1.0.

[0018] A further technical solution of the present invention is as follows: the "internal physical field constraint" in step S2 further includes: using pre-calibrated acoustoelastic constants as fixed parameters for model training, substituting the acoustoelastic constants into the acoustoelastic constitutive relation to establish a mapping relationship between the model-predicted stress field and the model-predicted rate of change of sound velocity; based on the mapping relationship, converting the rate of change of sound velocity and / or acoustoelastic birefringence data in the second acoustic data into constraints on the model-predicted acoustic characteristics to avoid interference from the inhomogeneity of the material microstructure on the sound velocity-stress conversion.

[0019] The technical solution of the present invention is: a multi-source detection device for full-thickness three-dimensional residual stress field reconstruction, adapted to the full-thickness three-dimensional residual stress field reconstruction method, including an X-ray diffraction module, an ultrasonic detection module and a data processing control unit; The X-ray diffraction module is used to perform sin²ψ measurements on the surface of the component under test to obtain the first stress data; The ultrasonic testing module is used to emit ultrasonic waves that penetrate the entire thickness of the component under test and receive the echo signal to obtain second acoustic data. The data processing control unit is communicatively connected to the X-ray diffraction module and the ultrasonic detection module, respectively, and is configured to execute the full-thickness three-dimensional residual stress field reconstruction method as described in claim 7.

[0020] A further technical solution of the present invention is as follows: the data processing control unit incorporates a physical information-driven attention neural network, which is a physical information neural network that incorporates an attention mechanism module; the attention mechanism module includes a spatial coordinate mask layer, used to identify whether the input coordinates are located on the surface of the measured component during the forward propagation of the model, and when identified as surface coordinates, the network output value is forcibly locked to the corresponding first stress data, which does not participate in the gradient update of the backpropagation; during the model training process, the attention mechanism module is used to directly embed the first stress data as a hard spatial boundary condition into the forward propagation computation graph of the model, forcibly setting the network output value located at the coordinate point on the surface of the measured component to be equal to the first stress data, so as to perform geometric topological constraints on the reconstruction process of the full-thickness three-dimensional residual stress field.

[0021] A further technical solution of the present invention is as follows: the data processing control unit includes a process decision module; the process decision module is configured to: take the reconstructed full-thickness three-dimensional residual stress field as input, firstly perform a risk assessment on the tensile stress concentration area and gradient inside the component based on a preset expert rule base, then calculate the process adjustment amount matching the risk level through a regression algorithm, and finally output the optimized process parameters for the tested component; the expert rule base includes mapping rules between tensile stress threshold and risk level; the regression algorithm is multiple linear regression or support vector regression; the process parameters include at least one of tempering temperature, tempering time, or quenching cooling rate.

[0022] Compared with the prior art, the present invention has the following advantages: 1. Breaking down detection barriers and achieving full-thickness stress field reconstruction: This invention overcomes the physical bottlenecks of traditional X-ray diffraction (surface-constrained) and ultrasonic testing (path averaging), constructing a reconstruction paradigm with a triple coupling of "boundary-data-physics". By using XRD data as a rigid spatial boundary, ultrasonic data as an internal constraint, and mechanical equilibrium equations as regularization terms, it achieves full-thickness three-dimensional stress field reconstruction that simultaneously satisfies surface accuracy, internal continuity, and physical conservation laws, solving the "data silo" problem where multi-source data cannot be organically integrated.

[0023] 2. Introducing an attention mechanism to overcome the problem of boundary information decay: Addressing the boundary accuracy loss defect in deep training of existing Physical Information Neural Networks (PINNs), this invention designs a Physical Information Driven Attention Neural Network (PATN). This model, through a unique attention mechanism module, directly embeds XRD data as an unshakeable hard boundary condition into the forward propagation computation graph, rather than fitting it through soft-weighted loss. "Soft-weighted loss fitting" refers to constraining the network output to approximate the boundary value through the weight coefficients in the loss function (but the network can still deviate), while "hard spatial boundary condition" means directly forcing the output to equal the measured value during forward propagation, without participating in gradient updates. This mechanism ensures that regardless of the network depth, the surface stress always locks into the measured value, thereby guiding the neural network to perform high-fidelity extrapolation across the entire thickness range.

[0024] 3. Dynamic calibration constants to shield against material microstructure interference errors: The main reason for the low accuracy of existing ultrasonic methods is that sound velocity is easily affected by the microstructure of the material (texture, grain size). This invention incorporates an XRD-ultrasonic automatic calibration module, which uses high-precision XRD data to invert the acoustoelastic constants in real time and substitutes them as fixed parameters into the constitutive relation. Since model training cannot "cheat" the fit by adjusting constants, it can only be forced to correct the stress distribution, thus successfully eliminating the interference of material microstructure inhomogeneity and significantly improving the absolute accuracy of internal stress inversion.

[0025] 4. Integrated Decision Model to Construct a Closed-Loop Inspection Process: Unlike traditional equipment that only outputs data charts, this invention incorporates a process decision module based on expert rules. This model can not only reconstruct the three-dimensional stress field but also automatically identify dangerous tensile stress areas and, combined with material characteristics, output optimization suggestions such as tempering temperature, holding time, and cooling rate. This design extends the equipment's function from "passive quality inspection" to "active process optimization," providing decision support for the intelligent manufacturing of large, critical components.

[0026] 5. Optimized hardware configuration for multi-scenario industrial testing: The design fully considers the complexity of industrial environments. The multi-axis oscillation mechanism of the X-ray module effectively overcomes the coarse-grain effect in large forgings. The ultrasonic module can be configured with a contact LCR probe for precision testing, or seamlessly switched to an EMAT electromagnetic ultrasonic probe for online inspection at high temperatures of 200–300°C. This modular design achieves full-scenario coverage from harsh workshops to cloud computing, enhancing the engineering applicability of the device.

[0027] The present invention will be further described below with reference to the figures and embodiments. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of a multi-source detection device.

[0029] Legend: X-ray diffraction module 1; Generating unit 11; Detecting unit 12; Multi-axis oscillation mechanism 13; Signal processing and interface unit 14; Ultrasonic detection module 2; Transmitting probe 21; Receiving probe 22; Ultrasonic signal transceiver host 23; Signal processing and interface unit 24; Data processing and control unit 3; Data acquisition and preprocessing layer 31; Core algorithm and storage layer 32; Application and presentation layer 33. Detailed Implementation

[0030] Example 1: like Figure 1 As shown, the method for reconstructing the three-dimensional residual stress field of full thickness includes the following steps: S1. Acquiring Test Data: Acquiring first stress data of the surface of the tested component and second acoustic data of the interior of the tested component. The first stress data includes lattice strain data obtained by performing sin²ψ measurements on the surface of the tested component using X-ray diffraction, or the surface residual stress value calculated from the lattice strain data. The second acoustic data includes the ultrasonic flight time measured by ultrasonic testing, and / or the rate of change of sound velocity calculated from the ultrasonic flight time, and / or acoustoelastic birefringence data (when ultrasonic waves encounter a stress field, shear waves split into fast and slow waves; this phenomenon is birefringence, and the acoustoelastic birefringence data refers to the difference in sound velocity or time between the fast and slow waves).

[0031] S2. Constructing a Prediction Model: Construct a physical information-driven neural network model, which is a physical information neural network. It takes the three-dimensional spatial coordinates of the measured component as input and the stress tensor components of the corresponding coordinates as output. Simultaneously, the first stress data is used as the spatial boundary condition, and the second acoustic data is converted into internal physical field constraints. Based on the spatial boundary condition and the internal physical field constraints, the composite loss function of the model is configured. The composite loss function includes at least a data loss term that measures the difference between the model's predicted value and the first stress data and the second acoustic data, and a physical information loss term that measures the residual of the model's predicted value satisfying the mechanical equilibrium equation.

[0032] S3. Iterative optimization solution: The parameters of the model are adjusted iteratively to minimize the composite loss function until the model converges.

[0033] S4. Reconstructing the residual stress field: Call the converged model to perform forward inference calculations on the discrete spatial coordinate points within the full thickness range of the tested component, and output the stress tensor components of the corresponding coordinate points, thereby reconstructing the three-dimensional residual stress field of the full thickness (from the surface to the core) of the tested component.

[0034] Preferably, the operation of "obtaining the first stress data of the surface of the component under test" in step S1 includes: selecting at least 3 measurement areas on the surface of the component under test using an X-ray diffraction module, collecting lattice strain data of no less than 9 points in each area, and calculating the surface residual stress value of each point using the sin²ψ method.

[0035] Preferably, the operation of "acquiring the second acoustic data inside the component under test" in step S1 includes: using an ultrasonic testing module to arrange a transmitter and a receiver on the surface of the component under test, so that the ultrasonic wave propagation path runs through the full thickness of the component under test; collecting the flight time of the ultrasonic wave, and converting the flight time into the rate of change of sound speed through the law of acoustic elasticity; collecting the shear wave splitting signal of the ultrasonic wave (split into fast wave and slow wave), and calculating the arrival time difference or sound speed difference between the fast wave and the slow wave to obtain the acoustic elastic birefringence data.

[0036] Preferably, the network structure of the Physical Information Neural Network (PINN) in step S2 includes an input layer, a hidden layer, and an output layer. The hidden layer is a fully connected layer, and a skip connection mechanism is introduced in the fully connected layer to improve the propagation efficiency of the deep network.

[0037] Preferably, in step S2, the composite loss function Represented as formulas 1 to 3; Formula 1: ; in, This is a data loss term used to measure the difference between the rate of change of sound velocity corresponding to the stress location predicted by the model and the rate of change of sound velocity in the second acoustic data. The physical information loss term measures the residuals of the model's predicted stress field satisfying the elastic equilibrium equation at internal sampling points; α is the weighting coefficient of the data loss term, and β is the weighting coefficient of the physical information loss term.

[0038] Formula 1 establishes a mathematical framework for multi-objective optimization. By introducing weighting coefficients α and β, a dynamic balance between data-driven and physics-driven terms is achieved. In the early stages of training, α has a higher weight to quickly fit ultrasonic observation data; in the later stages of training, β has a relatively higher weight to strengthen the mechanical equilibrium constraints. This weighting mechanism avoids overfitting caused by relying solely on experimental data or convergence difficulties caused by relying solely on physical equations, ensuring the model's generalization ability with limited samples.

[0039] Formula 2: ; in, The rate of change of sound speed in the second acoustic data; The rate of change of sound velocity is obtained from the transformation of the full-thickness three-dimensional residual stress field based on model prediction; Let m be the m-th ultrasound propagation path; m is the index of the ultrasound propagation path (m=1,2,3,...,M); M is the total number of ultrasound propagation paths.

[0040] Equation 2 establishes the quantitative correlation between the model's internal representation and the external ultrasonic detection signal. Because... It is derived from the integral of the stress field over the full thickness, and The average value is the path value across the thickness. Equation 2, by minimizing the difference between the two, forces the stress field output by the model to conform to the physical integral characteristics of ultrasonic propagation in the depth direction. At the same time, since the acoustoelastic constant K remains fixed during training, it effectively suppresses the sound velocity drift caused by the inhomogeneity of the material's microstructure and eliminates the interference of non-stress factors on the inversion results.

[0041] Formula 3: ; in, This represents the total number of internal sampling points. Let i be the spatial coordinates of the i-th internal sampling point; This is the predicted value of the stress tensor; The divergence of the stress tensor, where i is the index of the internal sampling point (i=1,2,3,...,N). internal ).

[0042] Equation 3 introduces differential constraints from the equilibrium equations of elasticity. Unlike traditional data interpolation, which only focuses on numerical matching of discrete points, Equation 3, by calculating the divergence residuals of the stress field in the spatial domain, forces the stress distribution generated by the model to be continuous and smooth in geometric topology, and to satisfy the static equilibrium condition under no external force. This solves the problem of non-physical abrupt changes or discontinuity artifacts that easily occur in the reconstructed stress field in existing technologies, ensuring the physical self-consistency of the reconstruction results.

[0043] Preferably, an intermediate step M is provided after step S1 and before step S2; M. Data Preprocessing and Coordinate Registration: The first stress data is denoised, and the second acoustic data is spatially registered so that the two types of data can be processed in a unified spatial coordinate system.

[0044] Preferably, the "internal residual sampling points" in step S2 The number of “ is 10” 3 ~10 5 The sampling method uses Latin hypercube sampling to ensure uniform spatial distribution.

[0045] Preferably, the values ​​of various weight coefficients in step S2 satisfy: α=0.8~1.2, β=0.5~1.0.

[0046] Preferably, the "internal physical field constraint" in step S2 further includes: using pre-calibrated acoustic elastic constants as fixed parameters for model training, and substituting the acoustic elastic constants into the acoustic elastic constitutive relation (the acoustic elastic constitutive relation is the mathematical expression of the acoustic elastic law: ,in, The rate of change of the speed of sound (dimensionless). Acoustic elastic constant (MPa) -1 ), The stress (MPa) is used to establish a mapping relationship between the model-predicted stress field and the model-predicted rate of change of sound velocity. Based on this mapping relationship, the rate of change of sound velocity and / or acoustoelastic birefringence data in the second acoustic data are converted into constraints on the model-predicted acoustic features, avoiding interference from material microstructure inhomogeneities on the sound velocity-stress conversion. The acoustoelastic constitutive relation is used to describe the physical relationship between the stress tensor and acoustic features. Specifically, the model-predicted stress field can be used to calculate the corresponding rate of change of sound velocity (relative change in the velocity of longitudinal or transverse waves) and the corresponding acoustoelastic birefringence data (difference between fast and slow waves in shear wave splitting) through the acoustoelastic constitutive relation. By comparing the rate of change of sound velocity and the acoustoelastic birefringence data in the second acoustic data with the corresponding features predicted by the model, internal physical field constraints are constructed. Since the acoustoelastic constant remains fixed during training, the model cannot fit the data by adjusting material parameters and can only be forced to correct the stress field distribution, thus effectively shielding the interference of material microstructure inhomogeneities.

[0047] like Figure 2 As shown, a multi-source detection device for full-thickness three-dimensional residual stress field reconstruction is adapted to the full-thickness three-dimensional residual stress field reconstruction method, including an X-ray diffraction module 1, an ultrasonic detection module 2, and a data processing and control unit 3.

[0048] X-ray diffraction module 1 is used to perform sin²ψ measurements on the surface of the component under test to obtain first stress data.

[0049] The ultrasonic testing module 2 is used to emit ultrasonic waves that penetrate the entire thickness of the component under test and receive the echo signal to obtain second acoustic data.

[0050] The data processing control unit 3 is communicatively connected to the X-ray diffraction module 1 and the ultrasonic detection module 2, respectively. The data processing control unit 3 is configured to execute the above-described full-thickness three-dimensional residual stress field reconstruction method.

[0051] Preferably, the data processing control unit 3 has a built-in Physical Information Driven Attention Neural Network (PATN), which is a physical information neural network that incorporates an attention mechanism module. During model training, the attention mechanism module is used to directly embed the first stress data as a hard spatial boundary condition into the forward propagation computation graph of the model, and the network output value located at the coordinate point on the surface of the measured component is forced to be equal to the first stress data, so as to perform geometric topological constraints on the reconstruction process of the full-thickness three-dimensional residual stress field.

[0052] Preferably, the data processing control unit 3 includes a process decision module 331; the process decision module 331 is configured to: take the reconstructed full-thickness three-dimensional residual stress field as input, firstly perform a risk assessment on the tensile stress concentration area and gradient inside the component based on a preset expert rule base (e.g., if the core tensile stress is >150MPa, it is recommended to supplement tempering), then calculate the process adjustment amount matching the risk level through a regression algorithm, and finally output the optimized process parameters for the tested component, wherein the process parameters include at least one of tempering temperature, tempering time or quenching cooling rate. Regarding tempering temperature: Based on the temperature-stress curve at the location of maximum tensile stress (e.g., for 45 steel, the temperature needs to be increased by 30℃ for every 50MPa decrease in stress), a recommended temperature (e.g., range 200~600℃) is output; Regarding tempering time: Based on the stress gradient (e.g., when the gradient is >10MPa / mm, the tempering time is extended by 20%), a recommended holding time (1~4h) is output; Regarding quenching cooling rate: If the surface compressive stress is insufficient (e.g., <-200MPa), it is recommended to reduce the cooling rate (e.g., change from water cooling to oil cooling).

[0053] Preferably, the X-ray diffraction module 1 is a portable X-ray stress meter (e.g., the Xstress 3000 manufactured by Stresstech GmbH, Finland) used to acquire the first stress data on the surface of the component under test (e.g., a roll).

[0054] Its specific hardware configuration and working principle are as follows: The X-ray generating and detection unit includes a generating unit 11 and a detection unit 12. The generating unit 11 is equipped with an X-ray tube with a Cr or Mn target to emit characteristic X-rays of a specific wavelength and irradiate the surface of the component under test at a preset incident angle. The detection unit 12 is used to receive the diffracted X-rays from the surface under test and record the intensity and diffraction angle of the diffraction peaks. The multi-axis oscillation mechanism 13 is used to drive the goniometer or sample stage to generate a compound oscillation at the measurement point (driven by a high-precision stepper motor, including unidirectional or lateral oscillation). The compound oscillation includes reciprocating oscillation in the ψ-angle direction, with an oscillation amplitude of -3° to +6° and an oscillation frequency of 2 to 5 Hz. Through this compound oscillation, the X-ray beam can cover a sufficient number of grain orientations during the measurement process, increase the number of effective grains participating in diffraction, eliminate measurement deviations caused by texture, and improve the statistical reliability of stress calculation results (the background of setting up the multi-axis oscillation mechanism is that large rolling mills and other components often have coarse grains or texture effects, which can easily lead to asymmetric diffraction spots or peaks, seriously affecting measurement accuracy). The signal processing and interface unit 14 amplifies, filters, reduces noise, and performs analog-to-digital conversion on the raw photon signal acquired by the detection unit 12, extracting precise information on the interplanar spacing variation. Subsequently, the signal processing and interface unit 14 utilizes its built-in sin...2 The ψ algorithm calculates the surface residual stress value and sends the first stress data (i.e., the surface residual stress value) to the data processing control unit via a high-speed data cable for subsequent model reconstruction.

[0055] Preferably, the ultrasonic testing module 2 is a split-type critical refractive longitudinal wave (LCR) testing module (e.g., Hangzhou Jianwei USG-V2-XX-P1.3 ultrasonic stress meter), equipped with piezoelectric ultrasonic probes with fixed spacing, used to acquire second acoustic data characterizing the internal stress state of the tested component (e.g., a roll).

[0056] Its specific hardware configuration and working principle are as follows: The transmitting probe 21 and the receiving probe 22 are rigidly mounted on a special fixture at a fixed interval. The fixed interval is 20–100 mm, determined specifically based on the thickness of the component being measured. The special fixture is designed to adapt to the curved contour of the roll, ensuring good acoustic coupling between the probe and the measured surface. During measurement, multiple sets of measuring points are arranged along the axial or circumferential direction of the roll, allowing the ultrasonic wave propagation path to penetrate the entire thickness range of the component being measured. The ultrasonic transceiver 23 is used to generate high-frequency high-voltage electrical pulses to excite the transmitting probe 21 to generate an LCR wave; at the same time, it amplifies, filters and converts the weak echo signal captured by the receiving probe 22 to capture the time of flight (ToF) of the LCR wave in the propagation path. The signal processing and interface unit 24 performs envelope detection and cross-correlation calculations on the acquired ultrasonic waveform data to extract time-of-sound variation data. Since the LCR wave is highly sensitive to stress components parallel to the surface, and the beam penetrates the entire thickness, the measured time-of-sound variation data essentially reflects the average stress state along the penetration path. Subsequently, the signal processing and interface unit 24 transmits the calculated time-of-flight (ToF) or rate of change of sound speed, and other second acoustic data, via a data cable to the data processing and control unit to construct internal physical field constraints.

[0057] Preferably, the data processing control unit 3 is an industrial computer (IPC) with a multi-core processor and large-capacity memory, or a remote computing node (i.e., a cloud server) deployed on a cloud server. As the control center of the system, the data processing control unit 3 maintains real-time communication with the X-ray diffraction module 1 and the ultrasonic detection module 2, and is responsible for coordinating the data fusion processing, stress field reconstruction calculation, and process decision generation.

[0058] Its internal functional architecture is divided into three logical layers from top to bottom: data acquisition and preprocessing layer 31, core algorithm and storage layer 32, and application and presentation layer 33.

[0059] The data acquisition and preprocessing layer 31 is primarily responsible for the access and spatiotemporal alignment of multi-source heterogeneous data. Specifically, the XRD data interface module receives and parses the surface stress coordinates and values ​​sent by the X-ray diffraction module via TCP / IP or RS22 communication protocols; the US data interface module receives the raw waveforms from the ultrasonic testing module and formats them into a structured dataset containing "path start point, path end point, and time of flight (ToF)" according to the spatial arrangement of the probe array. Subsequently, the data synchronization module aligns the two types of asynchronously acquired data based on high-precision timestamps, generating a spatiotemporally consistent joint input dataset as input for subsequent reconstruction calculations.

[0060] The core algorithm and storage layer 32 is the core of the device's intelligent reconfiguration. First, the XRD-ultrasonic automatic calibration module performs initial physical calibration: using the surface stress values ​​obtained by the XRD-ultrasonic automatic calibration module, combined with the acoustoelastic equation (ΔV / V=K·σ), it inverts and calculates the acoustoelastic constant K under the current operating condition in real time, and dynamically transmits this parameter to the subsequent neural network model to eliminate systematic errors caused by individual material differences. Subsequently, PATN performs the core three-dimensional stress field reconstruction task. PATN's input is the three-dimensional spatial coordinate grid of the roll. PATN simultaneously receives two physical constraints: one is the internal stress constraint converted by the XRD-ultrasonic automatic calibration module; the other is the XRD surface stress data located on the surface of the coordinate grid (Z=0). A key technical feature of PATN lies in its unique attention mechanism module. This module does not rely on soft weighting of the loss function, but instead directly embeds the XRD surface stress data as a hard spatial boundary condition into the forward propagation computation graph of the network. Specifically, when the network predicts the stress tensor at a certain spatial coordinate point, if it detects that the point is located on the surface of the component under test, the module will force the network output value at that point to be strictly equal to the XRD measured value. This mechanism ensures the absolute geometric accuracy of the reconstruction result at the surface boundary. Finally, driven by composite physical constraints, the network outputs the three-dimensional residual stress field distribution across the entire thickness. In addition, the material-stress database 223 of the measured object is built into the core algorithm and storage layers to store historical test data and calibration parameters of different materials such as MC5, D2, and 45Cr4NiMoV under different heat treatment processes, providing prior knowledge support for the rapid convergence of PATN.

[0061] The application and presentation layer 33 is user-facing, enabling the engineering application of test results. The process decision module 331 receives the three-dimensional stress field output by the Physical Information-Driven Attention Neural Network (PATN), automatically assesses the current service risk of the component based on a built-in expert rule base, and outputs specific process parameters (e.g., warning: +350MPa tensile stress exists 5mm from the surface; recommended tempering temperature is set to 580℃). The full lifecycle "stress fingerprint" management module is responsible for establishing a baseline stress profile for the rolls and periodically comparing it during service to analyze stress evolution trends. Finally, the visualization and report generation module renders the abstract stress field as an interactive color cloud map and supports one-click generation of test reports containing stress assessment, microstructure analysis, and process recommendations.

[0062] Preferably, the model structure of the Physical Information Driven Attention Neural Network (PATN) includes: an input layer (3 nodes: X / Y / Z coordinates) → 8 hidden layers (fully connected layers, 128 neurons per layer, introducing a skip connection mechanism, i.e., the output of the nth layer is added to the output of the (n-2)th layer to alleviate gradient vanishing) → an output layer (6 nodes: output stress tensor component σ). xx ,σ yy ,σ zz ,σ xy ,σ xz ,σ yz The Physical Information Driven Attention Neural Network (PATN) uses a spatial coordinate masking module to force the output value of the surface node (Z=0) to be equal to the first stress data (measured value) during the model's forward propagation, preventing it from participating in the backpropagation gradient update. This mechanism uses XRD data as a "hard spatial boundary condition" to ensure the accuracy of surface stress.

[0063] The workflow of the multi-source detection device for full-thickness three-dimensional residual stress field reconstruction is as follows: 1. Initialization: Based on the material of the component being tested (e.g., 42CrMo), the pre-calibrated acoustoelastic constant K is applied, such as... Longitudinal wave K L =1.2×10 −5 MPa −1 Shear wave K S =2.5×10 −5 MPa −1 ; 2. Data Acquisition: The X-ray diffraction module completes surface stress measurement, and the ultrasonic testing module completes internal acoustic data acquisition; 3. Model Calculation: The data processing and control unit performs PATN model training and stress field reconstruction, generating a three-dimensional stress cloud map (which can be displayed in real time via a touch screen). 3. Decision Output: The process decision module 331 outputs optimization suggestions and automatically generates inspection reports (including stress distribution diagrams, risk levels, and process parameters), which can be wirelessly transmitted to the factory MES system.

[0064] Example 2: The only difference between this embodiment and Embodiment 1 is the selection of the ultrasonic testing module. Ultrasonic testing module 1 uses an electromagnetic ultrasonic transducer (EMAT). EMAT uses the principle of electromagnetic induction to directly generate ultrasonic waves on the surface of the component without the need for a coupling agent. It can achieve non-contact measurement when the temperature of the component being tested is as high as 200-300°C, and is suitable for rapid online inspection of rolls immediately after they come off the production line.

Claims

1. A method for reconstructing a three-dimensional residual stress field of full thickness, characterized by the following steps: as follows: S1. Acquire test data: Acquire the first stress data on the surface of the component under test, and acquire the second acoustic data inside the component under test; S2. Constructing a Prediction Model: Constructing a physical information-driven neural network model, which is a physical information neural network, takes the three-dimensional spatial coordinates of the measured component as input and the stress tensor components of the corresponding coordinates as output; simultaneously, the first stress data is used as the spatial boundary condition, and the second acoustic data is converted into internal physical field constraints. Based on the spatial boundary condition and the internal physical field constraints, a composite loss function of the model is configured; the composite loss function includes at least a data loss term that measures the difference between the model's predicted value and the first stress data and the second acoustic data, and a physical information loss term that measures the residual of the model's predicted value satisfying the mechanical equilibrium equation; S3. Iterative optimization solution: The parameters of the model are adjusted iteratively to minimize the composite loss function until the model converges; S4. Reconstructing the residual stress field: Call the converged model to perform forward inference calculations on the discrete spatial coordinate points within the full thickness range of the tested component, and output the stress tensor components of the corresponding coordinate points, thereby reconstructing the three-dimensional residual stress field of the full thickness of the tested component.

2. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 1, characterized in that: In step S1, the first stress data includes lattice strain data obtained by performing sin²ψ measurement on the surface of the component under test by X-ray diffraction, or the surface residual stress value calculated from the lattice strain data; the second acoustic data includes ultrasonic flight time measured by ultrasonic testing, and / or the rate of change of sound velocity calculated from the ultrasonic flight time, and / or acoustoelastic birefringence data.

3. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 2, characterized in that: S2 In the steps, the composite loss function Represented as formulas 1 to 3; Official 1: ; in, This is a data loss term used to measure the difference between the rate of change of sound velocity corresponding to the stress location predicted by the model and the rate of change of sound velocity in the second acoustic data. This is the physical information loss term, used to measure the residuals of the model's predicted stress field satisfying the elastic equilibrium equations at the internal sampling points; For the weighting coefficients of the data loss term, The weighting coefficients for the physical information loss term; Official 2: ; The rate of change of sound speed in the second acoustic data; The rate of change of sound velocity is obtained from the transformation of the full-thickness three-dimensional residual stress field based on model prediction; Let m be the m-th ultrasound propagation path; m is the index of the ultrasound propagation path (m=1,2,3,...,M); M is the total number of ultrasound propagation paths; Official 3: ; in, This represents the total number of internal sampling points. Let i be the spatial coordinates of the i-th internal sampling point; This is the predicted value of the stress tensor; The divergence of the stress tensor, where i is the index of the internal sampling point (i=1,2,3,...,N). internal ).

4. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 3, characterized in that: The operation of "obtaining the first stress data on the surface of the component under test" in step S1 includes: selecting at least 3 measurement areas on the surface of the component under test using an X-ray diffraction module, collecting lattice strain data at no less than 9 points in each area, and calculating the surface residual stress value at each point using the sin²ψ method; the operation of "obtaining the second acoustic data inside the component under test" in step S1 includes: arranging the transmitting end and receiving end on the surface of the component under test using an ultrasonic testing module, so that the ultrasonic wave propagation path runs through the full thickness of the component under test; collecting the ultrasonic wave flight time and converting the flight time into the rate of change of sound velocity; collecting the shear wave splitting signal of the ultrasonic wave, calculating the arrival time difference or sound velocity difference between the fast wave and the slow wave, thus obtaining the acoustoelastic birefringence data.

5. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 4, characterized in that: There is an intermediate step M between step S1 and step S2; M. Data Preprocessing and Coordinate Registration: The first stress data is denoised, and the second acoustic data is spatially registered so that the two types of data can be processed in a unified spatial coordinate system.

6. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 5, characterized in that: S2 The physical information neural network structure in the step includes an input layer, a hidden layer, and an output layer. The hidden layer is a fully connected layer, and a skip connection mechanism is introduced in the fully connected layer to improve the propagation efficiency of the deep network. "Internal residual sampling points" in step S2 The number of “10” 3 ~10 5 The sampling method uses Latin hypercube sampling to ensure uniform spatial distribution; The values ​​of various weight coefficients in step S2 satisfy: α = 0.8 ~ 1.2, β = 0.5 ~ 1.

0.

7. The full-thickness three-dimensional residual stress field reconstruction method as described in claim 6, characterized in that: S2 The "internal physical field constraint" step also includes: using pre-calibrated acoustoelastic constants as fixed parameters for model training, substituting the acoustoelastic constants into the acoustoelastic constitutive relation to establish a mapping relationship between the model-predicted stress field and the model-predicted rate of change of sound velocity; based on the mapping relationship, converting the rate of change of sound velocity and / or acoustoelastic birefringence data in the second acoustic data into constraints on the model-predicted acoustic characteristics to avoid interference from the inhomogeneity of the material's microstructure on the sound velocity-stress conversion.

8. A multi-source detection device for reconstructing a three-dimensional residual stress field of full thickness, adapted to the three-dimensional residual stress field reconstruction method of full thickness as described in claim 7, characterized in that, It includes an X-ray diffraction module, an ultrasonic testing module, and a data processing control unit; the X-ray diffraction module is used to perform sin²ψ measurement on the surface of the component under test to obtain first stress data; the ultrasonic testing module is used to emit ultrasonic waves that penetrate the entire thickness of the component under test and receive the echo signal to obtain second acoustic data; the data processing control unit is communicatively connected to the X-ray diffraction module and the ultrasonic testing module respectively, and the data processing control unit is configured to execute the full-thickness three-dimensional residual stress field reconstruction method as described in claim 7.

9. The multi-source detection device for full-thickness three-dimensional residual stress field reconstruction as described in claim 8, characterized in that: The data processing and control unit incorporates a physical information-driven attention neural network, which is a physical information neural network that incorporates an attention mechanism module. This attention mechanism module includes a spatial coordinate mask layer, used to identify whether the input coordinates are located on the surface of the measured component during model forward propagation. When the coordinates are identified as surface coordinates, the network output value is forcibly locked to the corresponding first stress data, which does not participate in the gradient update during backpropagation. During model training, the attention mechanism module directly embeds the first stress data as a hard spatial boundary condition into the model's forward propagation computation graph, forcibly setting the network output value at the coordinate point on the surface of the measured component to be equal to the first stress data, thereby providing geometric topological constraints for the reconstruction process of the full-thickness three-dimensional residual stress field.

10. The multi-source detection device for full-thickness three-dimensional residual stress field reconstruction as described in claim 9, characterized in that: The data processing and control unit includes a process decision module; The process decision module is configured to: take the reconstructed full-thickness three-dimensional residual stress field as input, firstly assess the risk of tensile stress concentration areas and gradients inside the component based on a preset expert rule base, then calculate the process adjustment amount matching the risk level through a regression algorithm, and finally output the optimized process parameters for the tested component; the expert rule base contains mapping rules between tensile stress thresholds and risk levels; the regression algorithm is multiple linear regression or support vector regression; the process parameters include at least one of tempering temperature, tempering time, or quenching cooling rate.

Images