PE hot melting joint ultrasonic detection system and method based on hardness distribution

Abstract

The invention relates to the technical field of nondestructive testing, in particular to a PE (polyethylene) hot melting joint ultrasonic testing system and method based on hardness distribution. Hardness distribution data are collected, an internal microstructure is dynamically deduced, an ultrasonic anisotropy model is constructed, and a self-adaptive learning module is used for predicting local anisotropy tensor and sound velocity distribution; and performing predictive dynamic path compensation and multi-mode focusing on the ultrasonic waves. And finally, deeply fusing ultrasonic features, hardness distribution, a microstructure and a defect evolution physical model, and intelligently identifying defects through a neural network. The invention solves or at least relieves the bottleneck of precise recognition of complex defects of a microstructure of the PE hot-melt joint and deep fusion detection of multi-modal data, and provides the ultrasonic detection system and method for the PE hot-melt joint based on hardness distribution, so that precise evaluation of internal defect types, severity and material performance is realized, and the service safety and reliability of a PE pipeline are improved.

Description

Technical Field

[0001] This invention belongs to the field of nondestructive testing technology, specifically an ultrasonic testing system and method for PE hot melt joints based on hardness distribution. Background Technology

[0002] With the widespread application of polyethylene (PE) pipes in critical areas such as gas pipelines, urban water supply pipelines, and chemical media transmission pipelines, the quality of PE heat fusion joints directly affects the safety, stability, and long-term service life of the entire project. PE heat fusion joints are formed by heating and pressurizing the pipe ends to melt them, then butt-jointing and cooling to solidify, creating a homogeneous connection. Its homogeneous internal structure and mechanical properties are essential to withstand external environmental stresses and media pressures. Therefore, non-destructive testing of PE heat fusion joints, using appropriate testing methods and efficient testing processes, to ensure that the connection quality meets design requirements, is a crucial step in ensuring the safe operation of the entire pipeline system.

[0003] Currently, most methods for inspecting PE hot-melt joints rely on their macroscopic geometric parameters and visual observation of the surface. For example, patent application number CN115597471B discloses a method for rapidly inspecting the flanges of PE pipe hot-melt butt joints. The technical concept is to provide a method that can assess the symmetry of the flanges and the alignment of the joints, thereby improving the inspection efficiency during construction and helping to initially ensure construction quality. This is achieved by designing a special inspection ruler and using a method to inspect the geometric shape of the flanges of PE pipe hot-melt butt joints. However, the principle of this technical solution dictates that it can only inspect the geometric shape of the external flanges of the joint. Therefore, it cannot delve into the internal structure of the joint to analyze and understand the physical parameters that play a decisive role in the macroscopic properties of the material during the hot-melt process, such as chain orientation and crystallinity. These complex internal defects caused by thermal history, such as microscopic pores, cold welding, lack of fusion, or initial cracks, cannot be distinguished by the aforementioned method. Furthermore, the patent does not analyze other advanced material data besides geometric dimensions and surface visual inspection, such as hardness distribution. Therefore, the results obtained from its testing method lack comprehensiveness and accuracy, making it impossible to accurately judge the internal integrity of the connector.

[0004] Similarly, patent document CN115371567B discloses a quality inspection device for PE gas-fired hot-melt pipe fittings. Its technical solution employs visual inspection to detect the dimensional deviations at the three ends of T-shaped PE gas-fired hot-melt pipe fittings, aiming to improve the quality inspection accuracy of these fittings. However, this technical solution also uses visual inspection, primarily employing image processing technology to detect external dimensional deviations. Its detection accuracy can only reflect microscopic dimensional deviations on the pipe's surface, and it cannot detect the internal microstructure or acoustic parameters of the fitting. Furthermore, it cannot detect important physical parameters such as localized material hardness distribution and ultrasonic wave propagation in complex media. Therefore, it cannot determine the type of internal defects in the joint, nor can it simultaneously assess material degradation, thus lacking in-depth quality control capabilities.

[0005] While the above examples have improved the control of the macroscopic quality of PE hot-melt joints to some extent, they all share the common principle of observing and judging external geometric shapes or surface appearance characteristics. These methods, based on external observation, are insufficient to meet the increasingly stringent performance requirements of modern engineering regarding the integrity of the internal microstructure and long-term reliability of PE hot-melt joints. Their inherent flaws in principle prevent the resolution of fundamental contradictions. In fact, the formation of a PE hot-melt joint is a complex non-equilibrium thermodynamic melting and crystallization process. The internal molecular chain orientation, crystallinity distribution, residual stress, and the appearance of micro-voids or cold-welding flow regions are closely related not only to welding process parameters and cooling rates but also to the properties of the material itself. The inherent differences in the above examples determine the mechanical properties, resistance to environmental stress cracking, and service life of PE hot-melt joints. However, because they rely solely on external geometric dimensions or appearance, they cannot quickly and accurately reflect the internal microstructural information and its evolutionary patterns. For example, different thermal histories can lead to flanges with the same external shape, but their internal crystallinity gradient or molecular chain orientation may be very different, resulting in the thermofusion joint exhibiting completely different failure modes over the same service life. This is the essential contradiction that the above examples fail to reveal.

[0006] Furthermore, traditional external non-destructive testing methods struggle to detect some non-macroscopic defects, complexly distributed defects, or defects with inconspicuous macroscopic geometric deformations, such as micropores caused by uneven cooling, localized lack of fusion, or initial microcracks in high stress concentration areas. These are often the beginnings of potentially catastrophic failures. Current industrial testing methods generally lack the ability to fuse and analyze information from different physical modes, such as separating and studying hardness distribution information reflecting local mechanical properties from ultrasonic propagation information sensitive to internal structure. Especially for viscoelastic, anisotropic polymer materials like PE, whose microstructure is strongly dependent on thermal history, there is a complex nonlinear relationship between changes in the material properties inside PE hot melt joints (such as hardness, elastic modulus, sound velocity, etc.) and microscopic defects (such as cold welding, local lack of fusion, orientation grain boundaries, etc.). Different changes in material properties will affect the propagation path, attenuation, scattering, and mode conversion of ultrasonic waves. Existing methods cannot acquire and analyze multi-dimensional and heterogeneous physical information, leading to misjudgments, missed detections, and misassessments when faced with complex internal defects. This is especially true when it is necessary to distinguish different types of defects and their severity, and predict their long-term performance impact.

[0007] Therefore, in order to address the shortcomings of existing PE hot melt joint testing technologies in terms of inferring the internal microstructure of the joint, detecting complex defects, and fusion analysis of multimodal information, while ensuring high efficiency and non-destructive testing, this paper proposes an intelligent testing system and method that can comprehensively and accurately infer the type and severity of internal defects and the overall material properties of the joint by introducing key physical quantities (high-resolution hardness distribution) that reflect the internal material properties and advanced ultrasonic testing technology. This is of great significance for helping the industry solve the aforementioned technical bottlenecks and industrial problems. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the prior art, solve or at least alleviate the detection bottlenecks in the prior art regarding the integrity of the internal microstructure of polyethylene (PE) hot melt joints, accurate identification of complex defects, and deep fusion analysis of multimodal data, and to provide an ultrasonic testing system and method for PE hot melt joints based on hardness distribution.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an ultrasonic testing system for PE hot melt joints based on hardness distribution, comprising: The first acquisition unit is used to acquire real-time surface hardness distribution data of the PE hot melt joint detection area. The first acquisition unit includes a micro indenter. The first processing unit is electrically connected to the first acquisition unit and is used to determine whether the hardness distribution non-uniformity or hardness gradient characteristics exceed a preset dynamic threshold, and to trigger or optimize the subsequent processing process based on the determination result. The second processing unit, electrically connected to the first processing unit, is used to dynamically infer the microstructure information inside the PE hot melt joint and construct an ultrasonic anisotropy model; the second processing unit includes a three-dimensional material structure mesh construction module, a hardness data mapping module, a microstructure inference module, and a composite adaptive learning module. The third processing unit, electrically connected to the second processing unit, is used for predictive dynamic acoustic path compensation and multi-mode focusing; the third processing unit includes a propagation behavior prediction module, an acoustic beam focusing and compensation module, an ultrasonic transceiver module, a path feedback module, and a multi-mode adaptive focusing module. The fourth processing unit is electrically connected to the third processing unit and is used for deep defect feature fusion analysis and intelligent recognition. The fourth processing unit includes an ultrasonic feature extraction module, a multimodal data registration module, a defect evolution physical model module, and a multimodal feature fusion neural network module.

[0010] To further realize the present invention, the following technical solutions may be preferred: Preferably, the micro indenter comprises: an indentation head made of diamond material with a specific geometry; a force loading mechanism for applying a controllable load in the range of 0.1 mN to 10 N; and a displacement measuring mechanism based on a capacitive or piezoelectric sensor for real-time monitoring of the indentation depth of the indentation head under the load, with a resolution better than 1 nm; the micro indenter performs a gridded scan on the surface of the PE hot melt joint using a two-dimensional scanning platform driven by a stepper motor, acquiring a series of discrete hardness measurement points with a spatial resolution of 0.5 mm to 2 mm.

[0011] Preferably, the first processing unit is used to determine whether the hardness distribution non-uniformity or hardness gradient characteristics exceed a preset dynamic threshold. Specifically, the determination process includes: calculating the local standard deviation, coefficient of variation, or spatial gradient amplitude of the hardness distribution data; comparing the calculation results with a preset dynamic threshold based on historical data, wherein the dynamic threshold is set by statistical learning methods according to the theoretical hardness range of PE material, known process defect patterns, and their corresponding hardness change characteristics; if the local standard deviation or spatial gradient amplitude exceeds the dynamic threshold, it is identified as a region of hardness non-uniformity or gradient abnormality.

[0012] Preferably, the process by which the microstructure inference module infers and calculates the degree of molecular chain orientation and crystallinity characteristics of the material includes: The first sub-step involves performing hardness gradient field analysis, which involves calculating the gradient tensor of the hardness distribution data in three-dimensional space to identify the anisotropic direction and magnitude of hardness changes. The second sub-step involves measuring the local ultrasonic nonlinear effect. This measurement is performed by transmitting an ultrasonic wave with a center frequency of 5MHz and a bandwidth of 2MHz into the PE hot melt joint, and detecting the amplitude and phase of its second harmonic signal, as well as the change in sound velocity under static and dynamic loads, through a receiver array. The third sub-step involves using a reverse inference algorithm based on finite element analysis, taking the hardness gradient field data and the ultrasonic nonlinear effect measurement results as inputs, and combining the constitutive model and thermodynamic crystallization kinetic model of PE material to iteratively optimize the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes until the hardness gradient and nonlinear ultrasonic response predicted by the model reach the preset convergence threshold with the measured data. The fourth sub-step involves constructing the elastic modulus tensor and anisotropic tensor distribution of the material based on the molecular chain orientation tensor and crystallinity parameters obtained through the iterative optimization.

[0013] Preferably, the machine learning model of the composite adaptive learning module is a convolutional-recurrent neural network based on a spatiotemporal attention mechanism. Its input layer receives multimodal data, including different welding process parameters, environmental parameters, ultrasonic signal features, nonlinear effect parameters, and internal microstructure and macroscopic mechanical properties obtained through destructive testing from historical detection data as training labels. The convolutional layer of the neural network is used to extract the spatial features of ultrasonic signals and hardness distribution, the recurrent layer is used to handle the temporal dependence during the scanning process, and the attention mechanism is used to weight the importance of different modalities and historical data. The model is trained using the Adam optimizer and the mean squared error loss function. In the prediction stage, the model receives real-time hardness distribution data, inferred microstructure information, advanced ultrasonic self-features, and historical learning results, and predicts the local anisotropy tensor and sound velocity distribution gradient of the current scanning area and its subsequent adjacent scanning areas in real time and in advance.

[0014] Preferably, the ultrasonic transceiver module includes a phased array probe, which is a linear array probe with 128 independent controllable piezoelectric crystal units, operating in the frequency range of 1MHz to 10MHz, with each crystal unit having a width of 0.6mm and a spacing of 0.1mm; the adjustment accuracy of the transmission delay parameter is 1ns; the process of the acoustic beam focusing and compensation module adjusting the transmission delay parameter of the phased array probe includes: based on the expected acoustic wave path calculated by the propagation behavior prediction module, using an inverse wave propagation algorithm or an optimization algorithm based on Fermat's principle, calculating the independent transmission delay time of each phased array probe crystal unit; simultaneously, according to the predicted medium attenuation characteristics, weighted adjustment of the transmission excitation voltage of each crystal unit to compensate for acoustic energy loss and optimize the sound field distribution.

[0015] Preferably, the multimodal feature fusion neural network module is a multi-head attention network based on the Transformer architecture. Its input vector includes the ultrasonic time-frequency domain features, the hardness distribution features, the environmental parameters, and the microstructure information identified by the defect evolution physical model module. The output layer of the multimodal feature fusion neural network module adopts the Softmax activation function, and the output includes defect type, severity assessment, and potential risk prediction. The multimodal feature fusion neural network module achieves a comprehensive understanding and collaborative assessment of defects and the overall performance degradation state of materials through deep learning. The training of the neural network adopts the cross-entropy loss function and uses the Adam optimizer for gradient descent. If the defect confidence score output by the neural network is higher than a preset threshold of 0.8, it is marked as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided.

[0016] An ultrasonic testing method for PE hot melt joints based on hardness distribution includes the following steps: S1: Real-time acquisition and preprocessing of material parameters, including acquiring real-time hardness distribution data of the surface of the detection area through a micro indenter, and determining whether the hardness distribution non-uniformity or hardness gradient characteristics exceed a preset dynamic threshold. S2: Dynamic microstructure inference and ultrasonic anisotropy model construction steps, including establishing a three-dimensional material structure mesh containing melt index, molecular chain orientation degree and crystallinity characteristics, accurately mapping the hardness distribution data to the nodes of the three-dimensional mesh, inferring and calculating the molecular chain orientation degree and crystallinity characteristics of the material based on hardness gradient field analysis and local ultrasonic nonlinear effect measurement, and constructing the elastic modulus and anisotropic tensor distribution of the material accordingly, and introducing a composite adaptive learning module to predict the local anisotropic tensor and sound velocity distribution gradient of the current scanning area and its subsequent adjacent scanning areas in real time and ahead of time; S3: Predictive dynamic acoustic path compensation and multi-mode focusing steps, including predicting the propagation behavior of ultrasound in complex anisotropic media based on the advance prediction results of the composite adaptive learning module, performing predictive acoustic beam focusing and compensation to adjust the transmission delay parameters of the phased array probe, transmitting pre-compensated ultrasound and receiving reflected signals, comparing the time difference and beam shape between the predicted acoustic path and the actual received signal, feeding back the comparison results to the composite adaptive learning module for model correction, and executing the multi-mode adaptive focusing algorithm; S4: Deep defect feature fusion analysis and intelligent identification steps, including extracting the time-frequency domain features of the broadband ultrasonic signal after predictive dynamic compensation and multi-mode focusing optimization, accurately spatially registering the composite ultrasonic signal features, predicted sound velocity distribution, inferred microstructure information, hardness distribution and environmental parameters, introducing the defect evolution physical model module for deep physical correlation analysis and acoustic fingerprint matching, and constructing a multi-modal feature fusion neural network for defect type identification, severity assessment and potential risk prediction.

[0017] Preferably, the process of inferring and calculating the molecular chain orientation and crystallinity characteristics of the material based on hardness gradient field analysis and local ultrasonic nonlinear effect measurement in step S2 includes: Perform hardness gradient field analysis to calculate the gradient tensor of the hardness distribution data in three-dimensional space; Local ultrasonic nonlinear effect measurement is performed by transmitting ultrasonic waves with a center frequency of 5MHz and a bandwidth of 2MHz into the PE hot melt joint, and detecting the amplitude and phase of its second harmonic signal, as well as the change of sound velocity under static and dynamic loads, through a receiver array. Using a reverse inference algorithm based on finite element analysis, the hardness gradient field data and the ultrasonic nonlinear effect measurement results are taken as inputs. Combined with the constitutive model and thermodynamic crystallization kinetic model of PE material, the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes are iteratively optimized until the hardness gradient and nonlinear ultrasonic response predicted by the model reach the preset convergence threshold with the measured data. Based on the molecular chain orientation tensor and crystallinity parameters obtained by the iterative optimization, the elastic modulus tensor and anisotropic tensor distribution of the material are constructed.

[0018] Preferably, the process of constructing the multimodal feature fusion neural network in step S4 includes: The extracted composite ultrasonic time-frequency domain features, the hardness distribution and environmental parameters, and the microstructure information identified by the defect evolution physical model module are used as the input vector of the neural network, which is a multi-head attention network based on the Transformer architecture. The neural network uses deep learning to automatically learn the complex implicit correlations and interactions between these multimodal heterogeneous data, so as to achieve a comprehensive understanding and collaborative assessment of defects and the overall performance degradation state of materials. Based on the output of the neural network, defect type identification, severity assessment, and potential risk prediction are performed. If the defect confidence level output by the neural network is higher than a preset threshold of 0.8, it is marked as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided.

[0019] The beneficial effects of this invention are: This invention deeply integrates the local hardness distribution of PE hot-melt joints with the propagation behavior of ultrasonic waves in complex media, enabling precise assessment of the type and severity of internal defects and the overall performance of the material. It can provide comprehensive information on the material's microstructure, defect formation mechanism, and macroscopic performance impact that is unavailable through traditional methods, thus providing more reliable technical support for the safe operation of PE pipeline systems. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the detection system of the present invention.

[0021] Figure 2 This is a schematic diagram of the second processing unit of the present invention.

[0022] Figure 3 This is a schematic diagram of the third processing unit of the present invention.

[0023] Figure 4 This is a schematic diagram of the fourth processing unit of the present invention.

[0024] Figure 5 This is a schematic flowchart of the detection method of the present invention.

[0025] Figure 6 This is a schematic diagram of the hardness data of the present invention mapped to a three-dimensional material structure mesh.

[0026] Figure 7 This is a schematic diagram of the predictive dynamic acoustic path compensation and multi-mode focusing of the present invention. Detailed Implementation

[0027] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0029] This embodiment discloses a non-destructive testing system for PE hot-melt joints based on an ultrasonic phased array, which includes multiple cooperating functional units. For details, please refer to [link / reference needed]. Figure 1 As shown, the system mainly consists of a first acquisition unit, a first processing unit, a second processing unit, a third processing unit, and a fourth processing unit that are electrically connected. These units work together to achieve full-chain non-destructive testing of PE hot melt joints, from surface hardness acquisition, internal microstructure inference, ultrasonic propagation behavior prediction, adaptive compensation focusing, to final defect depth fusion analysis and intelligent identification.

[0030] Specifically, the first acquisition unit obtains real-time surface hardness distribution data of the PE hot-melt joint testing area. The first acquisition unit includes at least one micro indenter. In a preferred embodiment, the micro indenter's core components include an indentation head, a force loading mechanism, and a displacement measuring mechanism. The indentation head can be made of a high-hardness material such as diamond, and its tip has a specific geometry, for example, a standard shape such as a Vickers, Knoop, or Berkovich pyramid to ensure the accuracy and repeatability of the indentation process. The force loading mechanism can accurately apply a force in the range of 0.1 mN to 10 N, which covers the surface hardness testing within the hardness range of PE materials without causing macroscopic plastic deformation of the material. The displacement measuring mechanism can use a capacitive or piezoelectric sensor to monitor the indentation depth of the indentation head under load, with a resolution better than 1 nm to capture details of material deformation. The micro indenter uses a stepper motor-driven two-dimensional scanning platform to perform a grid scan on the surface of the PE hot-melt joint to acquire the hardness distribution of the testing area. This scanning platform acquires a series of hardness measurement points with a spatial resolution of 0.5mm to 2mm, thereby obtaining a fine surface hardness distribution map. For example, when testing a PE100 pipe heat fusion joint, hardness test points can be placed at 10mm on each side of the joint centerline, with axial hardness test points every 1mm and circumferential hardness test points every 0.5mm, forming a 20mm × circumference hardness matrix.

[0031] The first processing unit is electrically connected to the first acquisition unit. Using the surface hardness distribution data acquired by the first acquisition unit, it determines whether the value of the hardness distribution non-uniformity or hardness gradient exceeds the dynamic threshold. If so, it determines that the region is a region of hardness non-uniformity or hardness gradient abnormality. The step of the first processing unit determining whether the value of the hardness distribution non-uniformity or hardness gradient exceeds the dynamic threshold includes the following steps: calculating the local standard deviation, coefficient of variation (the coefficient of variation is the ratio of the standard deviation of the hardness distribution data to the average value of the hardness distribution data) or spatial gradient amplitude of the hardness distribution data using the first processing unit; comparing the result obtained in the first step with the dynamic threshold, wherein the dynamic threshold is not a fixed threshold, but a threshold obtained by dynamically training and adjusting using statistical learning methods based on the theoretical hardness range of the PE material, known or common hot melt welding process defect types (e.g., cold welding, lack of fusion, local oxidation, etc.) and the corresponding hardness distribution non-uniformity or hardness gradient values ​​through experiments or simulations. Preferably, the first processing unit uses statistical learning methods including training a large amount of normal and defective sample data based on Support Vector Machine (SVM) or Random Forest to obtain the dynamic threshold, and collecting data from the first acquisition unit. The dynamic threshold is dynamically adjusted based on the new hardness distribution data. If the local standard deviation or spatial gradient amplitude obtained in the previous step is greater than the dynamic threshold, the first processing unit determines that the region is a region of hardness non-uniformity or hardness gradient abnormality. The specific steps for the first processing unit to determine that the region is a region of hardness non-uniformity or hardness gradient abnormality preferably include: if the local hardness standard deviation of the PE hot-melt joint under test is detected to be greater than twice the normal hardness standard deviation of the PE hot-melt joint under test, or the hardness gradient amplitude is greater than 5 MPa / mm, the first processing unit issues a hardness non-uniformity or hardness gradient abnormality signal; the determination result triggers or optimizes the second processing unit to dynamically infer the microstructure information inside the PE hot-melt joint and construct an ultrasonic anisotropy model, or updates the ultrasonic anisotropy model and adaptive learning model, so that the ultrasonic anisotropy model and adaptive learning model constantly reflect the local microstructure information and state changes of the PE hot-melt joint under test.

[0032] After collecting and initially processing the surface hardness data, this invention also uses a second processing unit to dynamically infer the microstructure information inside the PE hot melt joint and establish an ultrasonic anisotropy model. Its module block diagram is shown below. Figure 2 As shown, the second processing unit includes a three-dimensional material structure mesh construction module, a hardness data mapping module, a microstructure inference module, and a composite adaptive learning module.

[0033] The three-dimensional material structure mesh construction module is first used to establish a three-dimensional material structure mesh that includes melt index, molecular chain orientation degree, and crystallinity characteristics. This process begins by generating a precise geometric model of the PE hot-melt joint based on a CAD model of the joint or geometric data obtained through high-precision 3D scanning (e.g., laser triangulation or structured light scanning). Subsequently, an unstructured tetrahedral or hexahedral mesh generation method is used to discretize the geometric model into a finite element mesh containing millions of nodes, each node representing a micro-region. For example, for a PE hot-melt pipe with a diameter of 160 mm, the joint region can be divided into approximately 5 million tetrahedral elements, each typically 0.2 mm in size. For each mesh node, the melt index, molecular chain orientation degree, and crystallinity characteristics are initially defined as initial unknown variables. The melt index is initially estimated by combining the rheological model of the PE material (e.g., power-law fluid model) and hot-melt process parameters (e.g., welding temperature, heating time, holding pressure), providing a basis for subsequent microstructure inference.

[0034] Next, the hardness data mapping module accurately maps the hardness distribution data acquired by the first acquisition unit to the nodes of the three-dimensional material structure mesh, such as... Figure 6 As shown. This mapping process first uses Kriging interpolation or Inverse Distance Weighting (IDW) interpolation algorithms to interpolate the discrete surface hardness measurement point data into the surface nodes of the three-dimensional material structure mesh. For example, when using Kriging interpolation, a semivariance function is constructed to describe the spatial autocorrelation of the hardness data, thereby providing an optimal linear unbiased estimate of the hardness at unknown points. Furthermore, based on empirical or physical models of surface hardness and internal material properties, such as combining the effects of density, crystallinity, and molecular weight distribution on the elastic modulus, the surface hardness information is extended and corrected to the internal nodes to provide preliminary information on internal mechanical properties. This extension process can be simulated using diffusion equations or solved inversely based on a coupled model of heat conduction and crystallization kinetics to correlate the surface hardness with the gradient relationship of internal crystallinity and density.

[0035] The microstructure inference module is key to the non-contact assessment of the internal microstructure of this invention. Based on the hardness gradient field analysis and local ultrasonic nonlinear effect measurement, this module infers and calculates the degree of molecular chain orientation and crystallinity characteristics of the material, and constructs the elastic modulus and anisotropic tensor distribution of the material accordingly.

[0036] This inference calculation process includes four sub-steps: The first sub-step involves performing hardness gradient field analysis. By calculating the gradient tensor of the hardness distribution data in three-dimensional space, the anisotropic direction and magnitude of hardness changes are identified. This hardness gradient tensor, as a second-order tensor, directly reflects the trends of local stress concentration, density changes, and molecular chain orientation along its principal axes. For example, a significant hardness gradient may appear at the edge of the welded area, indicating a higher cooling rate and shrinkage stress.

[0037] The second sub-step involves measuring localized ultrasonic nonlinear effects. This measurement is performed by transmitting ultrasonic pulses with a center frequency of 5 MHz and a bandwidth of 2 MHz into the PE thermofusion joint. Due to the nonlinear properties of the PE material, the ultrasonic waves generate second harmonics during propagation. By detecting the amplitude and phase of the second harmonic signal using a receiver array, the nonlinear parameters of the material (e.g., the β value) can be evaluated. Furthermore, this measurement also includes measuring the change in sound velocity under static and dynamic loads (i.e., the acoustoelastic effect), for example, by applying an axial preload of 200 N to 1000 N to the PE joint, measuring the rate of change of sound velocity with stress to reflect the elastic nonlinearity of the material.

[0038] The third sub-step utilizes a finite element analysis (FEA)-based inverse inference algorithm, taking the hardness gradient field data and the ultrasonic nonlinear effect measurement results as inputs. Combining the constitutive model of the PE material (e.g., the Drugan-Rice model or the Necking model, which relates the anisotropy of molecular chain orientation, crystallinity, and elastic modulus) and the thermodynamic crystallization kinetic model (e.g., the Avrami model or the Ozawa model, which describes the effect of cooling rate on crystallinity and crystal morphology), the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes are iteratively optimized. The iterative process continues until the model-predicted hardness gradient and nonlinear ultrasonic response reach a preset convergence threshold with the measured data (e.g., a relative error of less than 5%).

[0039] The fourth sub-step involves constructing the elastic modulus tensor and anisotropic tensor distribution of the material based on the molecular chain orientation tensor and crystallinity parameters obtained through iterative optimization. These tensors clearly characterize the anisotropic properties of sound velocity and attenuation at various points within the PE hot-melt joint. For example, the sound velocity in a specific direction can be calculated using the Christoffel equation, providing an accurate physical model for predicting subsequent ultrasonic wave propagation behavior.

[0040] The composite adaptive learning module plays a crucial role in real-time optimization and proactive prediction throughout the entire detection process. It continuously trains and optimizes the machine learning model using historical detection data. In a preferred embodiment of the invention, the machine learning model is a convolutional recurrent neural network (CNN-LSTM) based on a spatiotemporal attention mechanism. Its input layer receives multimodal data, including different welding process parameters (e.g., welding temperature range 190℃-230℃, holding time 100s-300s, cooling rate 0.5℃ / min-2℃ / min), environmental parameters (e.g., ambient temperature -10℃-40℃, humidity 30%-90%), ultrasonic signal characteristics (e.g., A-scan amplitude peak, TOF, spectral energy distribution characteristics in the 1MHz-10MHz frequency band), nonlinear effect parameters (e.g., the ratio of second harmonic amplitude to fundamental amplitude, acoustoelastic coefficient), and internal microstructures obtained through destructive testing (e.g., crystallinity distribution, molecular chain orientation distribution, defect type and size, such as microcrack length 0.1mm-1mm, void diameter 0.05mm-0.5mm) and macroscopic mechanical properties (e.g., tensile strength 20MPa-28MPa, fracture toughness 3.5MPa·m^0.5-4.5MPa·m^0.5) as training labels. The neural network employs convolutional layers (e.g., 3-5 layers, each with 64-128 kernels, for extracting spatial features of the ultrasound signal and hardness distribution, such as edges and textures) to extract spatial features, recurrent layers (LSTM, e.g., 2-3 layers, each with 128 LSTM units) to handle temporal dependencies during scanning, and an attention mechanism to weight the importance of different modalities and historical data, enabling the model to focus on the information that contributes most to the prediction results. The model is trained using an Adam optimizer (learning rate 0.001) and a mean squared error (MSE) loss function to minimize the difference between predicted and true values. During the prediction phase, the model receives real-time hardness distribution data, microstructure information (provided by a microstructure inference module), and advanced ultrasound self-features (e.g., multi-band attenuation spectral coefficients, ultrasound velocity dispersion curves, and scattering intensity versus angle distribution) to predict the local anisotropic tensor and velocity distribution gradient of the current scanning region and its subsequent adjacent scanning regions in real-time and proactively. For example, the anisotropic tensor and sound velocity gradient within a future 5mm scanning area can be predicted, and the prediction results can be used to guide subsequent ultrasonic wave emission.

[0041] The third processing unit is electrically connected to the second processing unit. Its main function is predictive dynamic acoustic path compensation and multi-mode focusing to ensure the accuracy of acoustic wave transmission in anisotropic media. Its structure is as follows: Figure 3As shown, it includes a propagation behavior prediction module, a sound beam focusing and compensation module, an ultrasound transceiver module, a path feedback module, and a multi-mode adaptive focusing module.

[0042] The ultrasonic transceiver module includes a phased array probe. The phased array probe consists of 128 independently controllable piezoelectric crystal units, operating in the frequency range of 1MHz-10MHz. The width of each crystal unit is 0.6mm, and the spacing between crystal units is 0.1mm. This allows for a high range of acoustic field control and frequency range. The control accuracy of the transmission delay parameter is 1ns, which is crucial for focusing in anisotropic media.

[0043] The propagation behavior prediction module acquires the local anisotropic tensor and sound velocity distribution gradient predicted in advance by the composite adaptive learning module, and predicts the propagation behavior of ultrasound in complex anisotropic media. The method involves using either the Fast Marching Method (FMM) or Finite Difference Time Domain (FDTD) numerical simulation within the prediction region. The FMM is suitable for qualitatively predicting propagation time, while the FDTD provides a more accurate sound field distribution, thereby predicting the propagation path, sound velocity distribution, attenuation, and the expected location and shape of the focal point of the ultrasound within the prediction region. For example, the FDTD simulation discretizes the anisotropic tensor of the PE material and calculates the propagation trajectory and energy distribution of the wave packet at a specific incident angle.

[0044] Before ultrasonic wave emission, the beam focusing and compensation module proactively controls the emission delay parameters of the phased array probe based on the prediction results of the propagation behavior prediction module, ensuring the sound beam is pre-focused at the target depth and lateral position. The beam focusing and compensation module calculates the emission delay time of each phased array probe wafer unit using either a reverse propagation algorithm or a Fermat-based optimization algorithm, based on the predicted sound beam path calculated by the propagation behavior prediction module. The reverse propagation algorithm determines the equivalent emission time of each wafer unit by tracing the sound beam propagation at the target focal point in reverse. Furthermore, based on the medium attenuation information predicted by the propagation behavior prediction module (e.g., for 5MHz ultrasonic waves, the attenuation coefficient of PE material is typically 0.1dB / mm-0.5dB / mm), the emission excitation voltage of each wafer unit is weighted and adjusted to compensate for sound energy loss, optimize the sound field distribution, and pre-focus the sound energy at the target depth and lateral position. For example, for sound beams that need to penetrate longer distances or areas with stronger attenuation, the corresponding wafer unit excitation voltage is higher.

[0045] The path feedback module obtains the acoustic wave path predicted by the propagation behavior prediction module and the signal received by the signal receiving module, compares the time difference and beam shape of the two, and sends the comparison result to the composite adaptive learning module to correct the model. The comparison step involves the signal receiving module performing beamforming on the received ultrasonic signal (e.g., Synthetic Aperture Focusing Method SAFT or Total Focusing Method TFM) to reconstruct the actual acoustic beam path and focal position. The reconstructed beam parameters are then compared with the beam parameters predicted by the propagation behavior prediction module, analyzing the time difference deviation, focal offset, and beam distortion rate (e.g., the ratio of the actual focal region's half-width at half-maximum to the predicted value). If the time difference deviation is greater than 10 ns, the focal offset is greater than 0.5 mm, or the beam distortion rate is greater than 15%, within a preset dynamic tolerance range, the path feedback module sends the deviation information to the composite adaptive learning module to correct the model, optimize parameters, and improve the tolerance range for the next prediction.

[0046] The multi-mode adaptive focusing module automatically selects between a wide-area scanning focusing mode and a refined high-resolution focusing mode based on the nature of the current scan test point and the preliminary scan analysis results. In the initial stage of the scan analysis, the system defaults to a wide-area scanning focusing mode, using a lower frequency (e.g., 2MHz-4MHz) and larger aperture (e.g., 16-32 wafer units) ultrasonic beam for scanning to obtain a larger scanning range and preliminary defect information, such as for locating larger defects or suspicious abnormal areas. When the system initially analyzes and determines that there are strong abnormal signals (e.g., significant increase or decrease in local reflection intensity, time-domain waveform distortion of echo, significant changes in signal attenuation, etc.) or "fuzzy" areas, it will automatically switch to a fine high-resolution focusing mode. It will use a higher frequency (e.g., 5MHz-10MHz), a smaller aperture (e.g., 8-16 wafer units), and a fine dynamic focusing method (e.g., multi-angle focusing, synthetic aperture focusing, etc.) to obtain detailed information on the internal structure of the defect and a higher spatial resolution. For example, for small defects that are initially determined to exist, it will automatically perform multi-angle (e.g., ±30°) incident focusing to enhance the defect signal and perform precise focusing and positioning.

[0047] Finally, the fourth processing unit is electrically connected to the third processing unit and is used to perform the final fusion analysis and intelligent identification of several defect features. Its module structure is as follows: Figure 4 As shown, it includes an ultrasonic feature extraction module, a multimodal data registration module, a defect evolution physical model module, and a multimodal feature fusion neural network module.

[0048] The ultrasonic feature extraction module receives and processes broadband ultrasonic signals that have undergone predictive dynamic compensation and multi-mode focusing, calculates their characteristics in the time-frequency domain, and analyzes their multi-band attenuation spectrum, scattering, mode conversion, etc. The ultrasonic feature extraction module first uses short-time Fourier transform or wavelet transform to obtain the time-frequency diagram of the ultrasonic signal, thereby revealing the instantaneous frequency components and energy of the ultrasonic signal. Based on this, the multi-band attenuation spectrum is obtained from the time-frequency diagram. Specifically, the frequency band is divided into sub-bands at 0.5MHz intervals between 1MHz and 10MHz, and the attenuation coefficient in each sub-band is calculated. For example, for a 5MHz ultrasonic signal, the attenuation coefficients of its 18 sub-bands (1-1.5MHz, 1.5-2MHz, 9.5-10MHz) are calculated. The intensity distribution of Rayleigh scattering, Mie scattering, and Bragg scattering is analyzed. Rayleigh scattering, Mie scattering, and Bragg scattering are sensitive to inhomogeneities at different scales within the material. The mode conversion ratios from P to S or S to P at multilayer interfaces or defects are analyzed. The modal transition ratio provides information on the geometric and mechanical properties of the defect interface.

[0049] The multimodal data registration module spatially registers the composite ultrasonic signal parameters extracted by the ultrasonic feature extraction module, the sound velocity distribution predicted by the composite adaptive learning module, the microstructure parameters inferred by the microstructure inference module, and the hardness distribution and environmental parameters acquired by the first acquisition unit. Specifically, a unified three-dimensional Cartesian coordinate system is established, with the central axis of the PE hot-melt joint as the central axis of this three-dimensional Cartesian coordinate system. The scanning positions of the micro-indenter and the ultrasonic phased array probe are synchronized by a high-precision encoder, so that the measurement points of the micro-indenter and the ultrasonic phased array probe are both located in this three-dimensional Cartesian coordinate system. Then, a three-dimensional image registration algorithm (such as a registration algorithm based on maximizing mutual information or the iterative nearest point ICP algorithm) is used to perform voxel matching and fusion of the hardness distribution data (surface data), the ultrasonic imaging data obtained by ultrasonic imaging (volume data), and the inferred microstructure parameters (volume data), so that different multimodal data are geometrically aligned, thereby providing fusion features for subsequent deep learning.

[0050] The defect evolution physical model module establishes a physical equivalence relationship between macroscopic acoustic characteristics and material microstructure parameters. Through fracture mechanics, acoustic scattering theory, and finite element numerical analysis (e.g., using extended finite element method or cohesive model to capture crack propagation process, or using elastic wave scattering theory to analyze acoustic response parameters of various defect morphologies, etc.), it obtains specific scattering signals, absorption coefficients, and mode transition parameters from ultrasonic signals, as well as the acoustic fingerprint patterns corresponding to specific micro-defect types (such as cold welds, microcracks, pores, etc.). Through this physical equivalence relationship, a multi-scale coupled model of macroscopic acoustic parameters and microstructure parameters is established, quantitatively mapping the relationship between microstructure parameters such as crystal size, molecular chain orientation tensor, and porosity and macroscopic elastic parameters such as elastic modulus, absorption coefficient, and sound velocity, and expressing the influence of different microstructure heterogeneities on ultrasonic scattering and absorption behavior. For example, it predicts the equivalent elastic constants of different grain orientations using equivalent media or self-consistent methods, or uses the Geffert model to predict ultrasonic attenuation in porous media.

[0051] The matching process consists of three sub-steps: The first sub-step involves a thorough physical analysis and connection based on the aforementioned macroscopic acoustic characteristics (e.g., ultrasonic attenuation, sound velocity, scattering intensity) and the material's microstructural parameters (e.g., molecular chain orientation, crystallinity, grain size, void size and distribution). This explores the role of the hardness gradient and the ultrasonic nonlinear effect in inferring the physical generation mechanism of microscopic defects, and investigates the sensitivity of the multi-band attenuation spectrum to different microstructural defects. For example, abnormal changes in high- and low-frequency ultrasonic attenuation may indicate micron-level voids or grain boundary scattering, or be related to macroscopic cracks or lack of fusion. Regions with large hardness gradients may exhibit significantly increased ultrasonic nonlinear parameters, indicating localized stress concentration, which may induce microcracks.

[0052] The second sub-step involves using a pre-established acoustic fingerprint database of defects to compare the scattering characteristics (such as the anisotropy ratio of forward scattering intensity to backscattering intensity), attenuation parameters (such as abnormal enhancement of attenuation amplitude within a certain frequency band, and abrupt changes in low-frequency attenuation), and mode conversion parameters (such as abnormal increase in shear wave energy) detected in the ultrasonic signal with the corresponding acoustic fingerprints of micro-defect types (such as cold welds, lack of fusion, microcracks, orientation grain boundaries, and oxide layers) in the database. This database was obtained through numerical simulation (e.g., using ABAQUS or COMSOL software to simulate the propagation of ultrasound in a defective medium using the finite element method) and numerous destructive experiments, and includes the geometry, size, and corresponding acoustic response characteristics of various typical defects.

[0053] The third sub-step examines the relationship between the aforementioned crystallization shrinkage stress (obtained from finite element thermodynamic analysis) and the aforementioned local hardness anomalies (obtained from the aforementioned hardness distribution data) and defects in relation to the formation mechanism and shape of micro-metallurgical structures. For example, simulations show that the crystallization shrinkage stress caused by uneven cooling rates can reach 70% of the PE yield strength in specific regions, and in the aforementioned high hardness gradient regions with local hardness anomalies, this leads to the generation of microcracks, while in locally low hardness regions, it may represent incomplete fusion or high porosity. This step is not limited to the characterization of macroscopic defects but rather reveals the formation mechanism of these defects.

[0054] The multimodal feature fusion neural network module uses all fused features as input to achieve intelligent recognition. In a preferred embodiment of the invention, the multimodal feature fusion neural network module is a multi-head attention network based on the Transformer architecture. This network effectively handles complex dependencies between different modal data through a self-attention mechanism, capturing long-distance dependencies and interactions between different modalities. The input vector of the neural network includes: the ultrasonic time-frequency domain features (e.g., a multi-band attenuation spectrum vector containing 18 attenuation coefficients, a 5x5 scattering intensity distribution matrix, a sequence containing P-wave and S-wave conversion ratios), the hardness distribution features (e.g., the principal direction and amplitude of the hardness gradient tensor, local hardness dispersion), the environmental parameters (e.g., real-time temperature, humidity), and the microstructure information identified by the defect evolution physical model module (e.g., molecular chain orientation tensor, crystallinity distribution, defect formation mechanism feature vector). The output layer of the neural network employs the Softmax activation function, and its output includes: defect type (e.g., cold weld, lack of fusion, microcrack, oxidation, voids, material degradation), severity assessment (e.g., defect size, defect density, predicted propagation trend), confidence score, and, based on the predicted defect type and severity, a quantified assessment of the impact on the material's macroscopic mechanical properties (e.g., percentage loss of tensile strength, factor of fatigue life reduction) and a potential risk level (e.g., low, medium, high). Through deep learning, the neural network automatically learns the complex implicit correlations and interactions between these multimodal heterogeneous data to achieve a comprehensive understanding and collaborative assessment of defects and the overall material performance degradation state. The neural network is trained using a cross-entropy loss function and gradient descent using the Adam optimizer. If the defect confidence score output by the neural network exceeds a preset threshold of 0.8, it is labeled as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided. Example 2

[0055] This embodiment discloses an ultrasonic testing method for PE hot melt joints based on hardness distribution, the flowchart of which is shown below. Figure 5 As shown. The method includes the following steps: S1: Real-time acquisition and preprocessing of material parameters.

[0056] S101: Real-time hardness distribution data of the test area surface is acquired using a micro indenter. The micro indenter comprises an indentation head made of diamond material (e.g., a Vickers diamond indenter), a force loading mechanism (capable of applying a 1N load), and a displacement measurement mechanism (0.5nm resolution). The indenter performs a gridded scan on the PE hot-melt joint surface using a stepper motor-driven two-dimensional scanning platform (e.g., an XY platform with an accuracy of 0.01mm), acquiring a series of discrete hardness measurement points with a spatial resolution of 0.5mm to 2mm (e.g., 1mm selected). In the inspection of a PE100 pipe butt joint (110mm diameter, 10mm wall thickness), one point is measured every 5mm along the joint circumference, and one point is measured every 1mm along the axial direction extending 20mm from the weld center to both sides, for a total of 110*(20+20+1) / 1=4510 hardness points.

[0057] S102: Determine whether the hardness distribution inhomogeneity or hardness gradient characteristics exceed a preset dynamic threshold. This determination process includes: calculating the local standard deviation, coefficient of variation, or spatial gradient amplitude of the hardness distribution data. The calculation result is compared with a preset dynamic threshold based on historical data. For example, the threshold is 1.5 times the standard deviation of the normal hardness fluctuation range of PE100 (e.g., 50-55 Shore D), or the gradient amplitude exceeds 2 MPa / mm. If the local standard deviation or spatial gradient amplitude exceeds the dynamic threshold, it is identified as a region of hardness inhomogeneity or gradient anomalousness, and the subsequent dynamic microstructure inference, ultrasonic anisotropy model, and adaptive learning model update process are triggered or optimized to ensure that the model can reflect the material state in real time.

[0058] S2: Dynamic microstructure inference and construction of ultrasonic anisotropy model.

[0059] S201: Establish a three-dimensional material structure mesh that includes melt flow index, molecular chain orientation degree, and crystallinity characteristics. This three-dimensional material structure mesh is generated by creating a geometric model of the PE hot-melt joint based on a CAD model or 3D scanning data, and then discretizing the geometric model using an unstructured tetrahedral or hexahedral meshing method. For example, finite element software such as ABAQUS or ANSYS is used for meshing, ensuring a finer mesh in the heat-affected zone (HAZ) and weld area. Melt flow index, molecular chain orientation degree, and crystallinity characteristics are defined as initial unknown variables for each mesh node, where the melt flow index can be correlated with welding temperature and pressure using empirical formulas.

[0060] S202: Accurately map the hardness distribution data of S101 onto the nodes of the three-dimensional mesh. The mapping process includes: using Kriging interpolation or inverse distance weighted interpolation algorithms to interpolate the discrete surface hardness measurement point data into the surface nodes of the three-dimensional material structure mesh. Based on empirical or physical models of surface hardness and internal material properties (e.g., linear or nonlinear relationships between hardness and crystallinity, density), the surface hardness information is extended and corrected to the internal nodes, and a diffusion algorithm is used to gradually penetrate the surface information into the interior.

[0061] S203: Based on the hardness gradient field analysis and local ultrasonic nonlinear effect measurement in S101, the molecular chain orientation degree and crystallinity characteristics of the material are inferred and calculated, and the elastic modulus and anisotropic tensor distribution of the material are constructed accordingly. This inference and calculation process includes: performing hardness gradient field analysis to calculate the gradient tensor of the hardness distribution data in three-dimensional space to capture changes in local material properties; and performing local ultrasonic nonlinear effect measurement by emitting ultrasonic waves with a center frequency of 5 MHz and a bandwidth of 2 MHz into the PE thermofusion joint (e.g., using a 128-element phased array probe), and detecting the amplitude and phase of its second harmonic signal through a receiver array (e.g., recording A-scan signals and performing FFT analysis to extract the fundamental and second harmonic components), as well as the change in sound velocity under static and dynamic loads (e.g., by TOF measurement). Using a finite element analysis-based inverse inference algorithm, the hardness gradient field data and the ultrasonic nonlinear effect measurement results are taken as input. Combined with the constitutive model and thermodynamic crystallization kinetic model of the PE material, the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes are iteratively optimized until the model-predicted hardness gradient and nonlinear ultrasonic response reach a preset convergence threshold (e.g., error less than 3%) with the measured data. Based on the molecular chain orientation tensor and crystallinity parameters obtained through iterative optimization, the elastic modulus tensor and anisotropic tensor distribution of the material are constructed.

[0062] S204: Introduces a composite adaptive learning module for real-time and forward prediction.

[0063] S2041: Continuously train and optimize the machine learning model using historical detection data. The machine learning model is a convolutional-recurrent neural network based on a spatiotemporal attention mechanism. Its input layer receives multimodal data, including different welding process parameters (such as clamping force and ambient temperature), environmental parameters, ultrasonic signal characteristics (such as attenuation coefficient and scattering intensity), nonlinear effect parameters, and internal microstructure and macroscopic mechanical properties obtained through destructive testing, as training labels.

[0064] S2042: Based on the real-time hardness distribution data from S101, the microstructure information inferred from S203, and advanced ultrasonic self-features (e.g., multi-band attenuation, sound velocity dispersion, scattering intensity distribution, etc., which are complex responses of ultrasonic waves propagating in a medium) and historical learning results, the machine learning model can predict the local anisotropic tensor and sound velocity distribution gradient of the current scanning area and its subsequent adjacent scanning areas in real time and with anticipation. This prediction directly inverts the medium properties by analyzing the complex response of ultrasonic waves propagating in a medium and considers complex nonlinear mapping relationships, surpassing the limitations of simple empirical formulas. For example, the model can predict the sound velocity distribution gradient and anisotropic principal axis direction along the scanning path in the next 10 mm.

[0065] S3: Predictive dynamic acoustic path compensation and multi-mode focusing.

[0066] S301: Based on the advanced prediction results of the dynamic anisotropic material model and adaptive learning module in S204, predict the propagation behavior of ultrasonic waves in the complex anisotropic medium of the future scanning area. The prediction process includes: using the fast travel method in anisotropic media or the finite difference time-domain numerical simulation method (e.g., through numerical solution of the acoustic wave equation) to calculate the propagation path, sound velocity distribution, attenuation characteristics, and expected location and shape of the focal point of the ultrasonic wave in the prediction area.

[0067] S302: Perform predictive beam focusing and compensation. Based on the material property changes predicted in S301, the emission delay parameters of the phased array probe are proactively adjusted before ultrasonic wave emission. This adjustment aims to optimize the propagation path of the sound beam in complex anisotropic media, ensuring that the acoustic energy is accurately focused to the target depth. The phased array probe is a linear array probe with 128 independently controllable piezoelectric crystal units, operating at a frequency of 3MHz-7MHz. The adjustment accuracy of the emission delay parameters is 1ns. The adjustment process includes: calculating the independent emission delay time of each phased array probe crystal unit based on the expected acoustic wave path (calculated using FMM or FDTD), using an inverse wave propagation algorithm or an optimization algorithm based on Fermat's principle; simultaneously, according to the predicted medium attenuation characteristics, the emission excitation voltage of each crystal unit is weighted and adjusted to compensate for acoustic energy loss and optimize the sound field distribution.

[0068] S303: An ultrasonic phased array probe array emits pre-compensated ultrasonic waves and receives reflected signals. For example, the probe emits at a center frequency of 5 MHz with a pulse width of 2 cycles and receives the echoes through the same array.

[0069] S304: Compare the time difference and beamform between the predicted acoustic path and the actual received signal. The comparison process includes: reconstructing the actual acoustic beam path and focal position by performing beamforming processing on the received ultrasonic signal (e.g., using the TFM algorithm). The reconstructed actual acoustic beam parameters are compared and analyzed with the predicted acoustic beam parameters to calculate the time difference deviation (e.g., the difference between the TOF of the focal point and the predicted TOF), the focal offset (e.g., the Euclidean distance between the actual focal position and the predicted focal position), and the beam distortion rate (e.g., the ratio of the actual sidelobe level to the main lobe level).

[0070] S305: Determine whether the time difference or beam deviation exceeds the preset dynamic tolerance range. If the time difference deviation exceeds 10ns, the focus offset exceeds 0.5mm, or the beam distortion rate exceeds 15% of the preset dynamic tolerance range, then the adaptive learning module in S204 is fed back in real time for model correction, and the phased array transmission delay parameters are optimized again for local compensation, forming a closed-loop control system.

[0071] S306: Execute a multi-mode adaptive focusing algorithm. Based on the characteristics of the current detection area and the preliminary signal analysis results from S305, the system adaptively selects either a wide-area scanning focusing mode or a refined high-resolution focusing mode. For example, when the preliminary analysis identifies potential abnormal signals (such as blurred areas or local signal enhancement in the B-scan image), the system automatically switches to a focusing mode with a wider bandwidth (e.g., 1MHz-10MHz), multiple angles (e.g., -45° to +45°, in 5° increments), or higher spatial resolution (e.g., using synthetic aperture focusing to reduce the beam focusing volume from 5mm³ to 1mm³) to obtain more detailed defect information.

[0072] S4: Deep defect feature fusion analysis and intelligent recognition.

[0073] S401: Extract the time-frequency domain features of the broadband ultrasonic signal after S302 predictive dynamic compensation and S306 multi-mode focusing optimization, and analyze its multi-band attenuation spectrum, scattering, mode conversion, and other characteristics. The extraction process includes: obtaining the time-frequency map of the ultrasonic signal through short-time Fourier transform or wavelet transform (e.g., using Morlet wavelet transform). Extract the multi-band attenuation spectrum from the time-frequency map, analyze the intensity distribution of Rayleigh scattering, Mie scattering, and Bragg scattering, and the P-wave to S-wave or S-wave to P-wave mode conversion ratios occurring at multilayer interfaces or defects. For example, a high P-wave to S-wave conversion ratio is typically associated with an unfused region.

[0074] S402: Precise spatial registration is performed on the composite ultrasonic signal features extracted in S401 (including multi-band attenuation spectrum), the sound velocity distribution predicted in S204, the microstructure information inferred in S203, and the hardness distribution and environmental parameters in S102. The registration process includes: establishing a unified three-dimensional spatial coordinate system (e.g., a Cartesian coordinate system with the origin at the joint center), and synchronizing the scanning positions of the micro-indenter and the ultrasonic phased array probe using a high-precision encoder. A three-dimensional image registration algorithm is used to align and fuse the hardness distribution data, ultrasonic imaging data, and inferred microstructure parameters at the voxel level, ensuring consistency in the geometric positions of different modal data, with each voxel corresponding to a volume of 0.1 mm³.

[0075] S403: Introduce a physical model for defect evolution.

[0076] S4031: Deep physical correlation analysis is conducted by combining macroscopic acoustic characteristics with material microstructure parameters (such as molecular chain orientation, crystallinity, grain size, and void distribution). Particular emphasis is placed on the role of hardness gradients and ultrasonic nonlinear effects in inferring the formation mechanism of micro-defects, as well as the sensitivity of multi-band attenuation spectra to different microstructural defects. For example, high hardness gradients are often associated with stress concentration areas, which are more prone to microcracks; their acoustic fingerprints show a significant enhancement in high-frequency attenuation on a broadband attenuation spectrum.

[0077] S4032: Identifies specific scattering characteristics, attenuation coefficients, and mode transition features in ultrasonic signals and matches them with unique acoustic fingerprints of specific micro-defect types (e.g., cold welds, lack of fusion, microcracks, grain boundaries, etc.). It focuses on matching the unique sensitivities of different defects using multi-band attenuation spectra. For example, cold weld defects may cause abnormal attenuation of ultrasonic waves in the low-frequency range, while internal microvoids may lead to enhanced high-frequency scattering.

[0078] S4033: Analyzes the causal relationships between crystallization shrinkage stress distribution, local hardness anomalies, and the formation mechanisms and morphologies of defects at the microscopic level, going beyond macroscopic defect characterization. For example, the local shrinkage stress distribution map calculated by thermodynamic finite element simulation is highly correlated with the hardness gradient map, which can predict potential stress-induced defect regions.

[0079] S404: Construct a multimodal feature fusion neural network.

[0080] S4041: The composite ultrasonic time-frequency domain features extracted in S401 (especially the multi-band attenuation spectrum), the hardness distribution and environmental parameters in S102, and the microstructure information identified by the physical model in S403 are used as the input vector of the neural network. The neural network is a multi-head attention network based on the Transformer architecture (e.g., a 6-layer encoder, 12 heads per head, with an input feature dimension of 256).

[0081] S4042: The neural network automatically learns the complex implicit correlations and interactions between these multimodal heterogeneous data through deep learning, so as to achieve a comprehensive understanding and collaborative evaluation of defects and the overall performance degradation state of materials.

[0082] S4043: Based on the output of the neural network, perform defect type identification, severity assessment, and potential risk prediction. For example, accurately distinguish between initial microcracks caused by stress concentration and cold welds, and assess their impact on the macroscopic mechanical properties of the material (e.g., tensile strength may decrease by 15%, and fatigue life may decrease by 5 times), while providing a health index of the material's local properties. If the defect confidence score output by the neural network is higher than a preset threshold of 0.8, it is marked as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided. Example 3

[0083] This embodiment focuses on the inspection of a PE100 pipe hot-melt butt joint with a diameter of 160mm and a wall thickness of 15mm. During the fusion process, the heating temperature was slightly lower and the cooling rate was faster, resulting in a typical cold welding defect in the center area of ​​the weld, specifically manifested as a partially unfused area and high local residual stress.

[0084] First, the surface of the PE hot-melt joint was scanned for hardness using a micro indenter. A Berkovich pyramidal diamond indenter was used, with an applied load of 0.5 N and an indentation depth resolution of 0.8 nm. The scanning platform acquired approximately 6500 hardness points with an axial resolution of 1 mm and a circumferential resolution of 0.8 mm. In step S102, the acquired hardness distribution data was analyzed, revealing a hardness gradient amplitude of 3.2 MPa / mm in the weld center region, with a local standard deviation of 2.8 Shore D, significantly exceeding the preset dynamic threshold (normal area hardness gradient < 1.5 MPa / mm, standard deviation < 1.0 Shore D). This anomaly was identified as a potential area of ​​uneven hardness.

[0085] A three-dimensional material structure mesh containing 12 million tetrahedral elements was constructed based on a CAD model, and surface hardness data was mapped to internal nodes using Kriging interpolation and a diffusion algorithm based on a hardness-crystallinity empirical model. A microstructure inference module performed a detailed analysis of the anomalous region. Ultrasonic nonlinear effect measurements showed that the ratio of the second harmonic amplitude to the fundamental amplitude (β value) in this region was as high as 8.5, significantly higher than the 4.0-5.0 in the normal region. Simultaneously, under a 500N axial load, the rate of change of sound velocity indicated a 12% decrease in the local elastic modulus. Combining hardness gradient field data and ultrasonic nonlinear effect measurement results, the molecular chain orientation degree in this region was iteratively calculated to be 0.65 (compared to approximately 0.8-0.9 in the normal region) and the crystallinity to be 58% (compared to approximately 65%-70% in the normal region) using a FEA inverse inference algorithm. The anisotropic tensor constructed based on this showed that the longitudinal wave velocity in this region was reduced by approximately 5% in the direction perpendicular to the weld compared to the direction parallel to the weld.

[0086] The composite adaptive learning module utilizes the aforementioned real-time data, combined with historical detection data, to make predictions on the trained CNN-LSTM model. The model predicts significant deviations in the anisotropic tensor within the anomaly region and its subsequent 5mm range, and predicts a local decrease of approximately 10% in the sound velocity distribution gradient.

[0087] Predictive dynamic acoustic path compensation and multi-mode focusing are initiated. Based on the prediction results of the composite adaptive learning module, the phased array probe (128 crystals, operating frequency 4MHz-8MHz) proactively adjusts the transmission delay parameters before emitting ultrasonic waves using a reverse propagation algorithm and excitation voltage weighting. For example, to focus on the cold-welded area, the delay time of some crystal units is corrected by 25ns, and the excitation voltage is increased by 10%. The ultrasonic transceiver module emits pre-compensated ultrasonic waves. The path feedback module reconstructs the actual acoustic beam path using TFM and finds that the actual acoustic beam focus offset is only 0.2mm, and the time difference deviation is 5ns, both within the preset tolerance range, verifying the effectiveness of predictive compensation. Due to initial signal anomalies, the system automatically switches to a refined high-resolution focusing mode, using an 8MHz center frequency and 12 crystal units for synthetic aperture focusing to obtain a clearer defect image.

[0088] Deep defect feature fusion analysis and intelligent identification were performed. The ultrasonic feature extraction module extracted multi-band attenuation spectra from the optimized ultrasonic signal, finding that the attenuation coefficient in the 2MHz-4MHz band was abnormally increased by 30%, and the mode conversion ratio from P-wave to S-wave reached 0.25 in this region (normal region is less than 0.05). The multimodal data registration module accurately registered these ultrasonic features, predicted sound velocity, inferred microstructure, and hardness distribution data into a unified three-dimensional coordinate system. The defect evolution physical model module combined this data to identify a specific acoustic fingerprint that highly matched cold welding defects (manifested as unfused interfaces, high stress concentration, and low crystallinity) in the database. Simultaneously, this module confirmed that the local high hardness gradient was consistent with the distribution of crystallization shrinkage stress and had a causal relationship with the formation of unfused defects at the microscopic level. Finally, the multimodal feature fusion neural network (Transformer architecture) received all fused feature vectors as input. The neural network output showed that the defect was identified as a "cold welding defect" with a confidence level as high as 0.95. The severity assessment shows that the defect has a lateral dimension of 4 mm and a depth of 30% of the wall thickness. Potential risk prediction indicates that this defect will cause a 20% reduction in the tensile strength of the joint and an 8-fold decrease in fatigue life, classifying it as "high" risk.

[0089] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An ultrasonic testing system for PE hot-melt joints based on hardness distribution, characterized in that, include: The first acquisition unit is used to acquire real-time surface hardness distribution data of the PE hot melt joint detection area, and includes a micro indenter. The first processing unit is used to determine whether the unevenness of hardness distribution or the hardness gradient characteristics exceed the preset dynamic threshold, and to trigger or optimize the subsequent processing process based on the determination result. The second processing unit is used to dynamically infer the microstructure information inside the PE hot melt joint and construct an ultrasonic anisotropy model; it includes a three-dimensional material structure mesh construction module, a hardness data mapping module, a microstructure inference module, and a composite adaptive learning module. The third processing unit is used for predictive dynamic acoustic path compensation and multi-mode focusing; it includes a propagation behavior prediction module, an acoustic beam focusing and compensation module, an ultrasonic transceiver module, a path feedback module, and a multi-mode adaptive focusing module. The fourth processing unit is used for deep defect feature fusion analysis and intelligent recognition; it includes an ultrasonic feature extraction module, a multimodal data registration module, a defect evolution physical model module, and a multimodal feature fusion neural network module.

2. The detection system according to claim 1, characterized in that, The micro indenter includes: an indentation head made of diamond material; a force loading mechanism for applying a controllable load; and a displacement measuring mechanism for real-time monitoring of the indentation depth. The micro indenter performs a gridded scan on the surface of the PE hot melt joint using a two-dimensional scanning platform to obtain hardness measurement points.

3. The detection system according to claim 1, characterized in that, The determination process of the first processing unit includes: calculating the local standard deviation, coefficient of variation, or spatial gradient amplitude of the hardness distribution data; comparing the calculation results with a preset dynamic threshold, which is set according to the theoretical hardness range of PE material, known process defect patterns, and corresponding hardness change characteristics; if the threshold is exceeded, it is identified as an abnormal area.

4. The detection system according to claim 1, characterized in that, The process by which the microstructure inference module infers the degree of molecular chain orientation and crystallinity characteristics includes: Perform hardness gradient field analysis to calculate the gradient tensor of hardness distribution data and identify the anisotropic direction and magnitude of hardness changes. Local ultrasonic nonlinear effects are measured by emitting ultrasonic waves and detecting the amplitude and phase of the second harmonic signal, as well as the change in sound velocity under load. Using a reverse inference algorithm based on finite element analysis, hardness gradient field data and ultrasonic nonlinear effect measurement results are used as inputs. Combined with the constitutive model of PE material and the crystallization kinetic model, the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes are iteratively optimized until the model prediction and the measured data converge. Based on the molecular chain orientation tensor and crystallinity parameters obtained through iterative optimization, the elastic modulus tensor and anisotropic tensor distribution are constructed.

5. The detection system according to claim 1, characterized in that, The machine learning model of the composite adaptive learning module is a convolutional-recurrent neural network based on a spatiotemporal attention mechanism. It receives multimodal data, including welding process parameters, environmental parameters, ultrasonic signal characteristics, nonlinear effect parameters, and destructive detection results from historical detection data as training labels. The machine learning model predicts the local anisotropic tensor and sound velocity distribution gradient of the current scanning area and its subsequent adjacent scanning areas in real time and in advance.

6. The detection system according to claim 1, characterized in that, The ultrasonic transceiver module includes a phased array probe, which is a linear array probe with independent and controllable piezoelectric crystal units, and the transmission delay parameter adjustment accuracy is 1 ns. The process of the sound beam focusing and compensation module adjusting the transmission delay parameter of the phased array probe includes: based on the expected sound wave path calculated by the propagation behavior prediction module, using the reverse wave propagation algorithm or the optimization algorithm based on Fermat's principle, calculating the independent transmission delay time of each crystal unit; and according to the predicted medium attenuation characteristics, weighted adjustment of the transmission excitation voltage of each crystal unit to compensate for sound energy loss and optimize the sound field distribution.

7. The detection system according to claim 1, characterized in that, The multimodal feature fusion neural network module is a multi-head attention network based on the Transformer architecture. Its input vector includes ultrasonic time-frequency domain features, hardness distribution features, environmental parameters, and microstructure information identified by the defect evolution physical model module. The output layer of the multimodal feature fusion neural network module outputs defect type, severity assessment, and potential risk prediction. Through deep learning, the multimodal feature fusion neural network module achieves a comprehensive understanding and collaborative assessment of defects and the overall performance degradation state of materials. If the defect confidence level output by the multimodal feature fusion neural network module is higher than a preset threshold, it is marked as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided.

8. An ultrasonic testing method for PE hot-melt joints based on hardness distribution, applicable to the testing system described in any one of claims 1-7, characterized in that, Includes the following steps: S1: Real-time acquisition and preprocessing of material parameters. Real-time hardness distribution data of the surface of the detection area is obtained through a micro indenter, and it is determined whether the hardness distribution non-uniformity or hardness gradient characteristics exceed the preset dynamic threshold. S2: Dynamic microstructure inference and ultrasonic anisotropy model construction, including the establishment of a three-dimensional material structure mesh, mapping hardness distribution data to mesh nodes, inferring the degree of molecular chain orientation and crystallinity characteristics based on hardness gradient field analysis and local ultrasonic nonlinear effect measurement, and constructing elastic modulus and anisotropic tensor distribution. A composite adaptive learning module is introduced to predict the local anisotropic tensor and sound velocity distribution gradient in real time and ahead of time. S3: Predictive dynamic acoustic path compensation and multi-mode focusing, including predicting ultrasonic wave propagation behavior based on the advance prediction results of the composite adaptive learning module, performing predictive acoustic beam focusing and compensation to adjust the phased array probe transmission delay parameters, transmitting pre-compensated ultrasonic waves and receiving reflected signals, comparing the time difference and acoustic beam shape between the predicted acoustic path and the actual received signal, feeding back the comparison results to the composite adaptive learning module for model correction, and executing the multi-mode adaptive focusing algorithm. S4: Deep defect feature fusion analysis and intelligent identification, including extracting time-frequency domain features of broadband ultrasonic signals after compensation and focusing optimization, accurately spatially registering composite ultrasonic signal features, predicted sound velocity distribution, inferred microstructure information, hardness distribution and environmental parameters, introducing a defect evolution physical model module for deep physical correlation analysis and acoustic fingerprint matching, and constructing a multimodal feature fusion neural network for defect type identification, severity assessment and potential risk prediction.

9. The detection method according to claim 8, characterized in that, The process of inferring the degree of molecular chain orientation and crystallinity characteristics based on hardness gradient field analysis and local ultrasonic nonlinear effect measurement in step S2 includes: Perform hardness gradient field analysis to calculate the gradient tensor of hardness distribution data in three-dimensional space; Local ultrasonic nonlinear effects are measured by emitting ultrasonic waves and detecting the amplitude and phase of the second harmonic signal, as well as the change in sound velocity under load. Using a reverse inference algorithm based on finite element analysis, hardness gradient field data and ultrasonic nonlinear effect measurement results are used as inputs. Combined with the constitutive model of PE material and the crystallization kinetic model, the molecular chain orientation tensor and crystallinity parameters on the three-dimensional mesh nodes are iteratively optimized until the model prediction and the measured data converge. Based on the molecular chain orientation tensor and crystallinity parameters obtained through iterative optimization, the elastic modulus tensor and anisotropic tensor distribution are constructed.

10. The detection method according to claim 8, characterized in that, The process of constructing the multimodal feature fusion neural network in step S4 includes: The extracted composite ultrasonic time-frequency domain features, hardness distribution and environmental parameters, and microstructure information identified by the defect evolution physical model module are used as the input vector of the neural network, which is a multi-head attention network based on the Transformer architecture. The neural network uses deep learning to automatically learn the complex implicit relationships between multimodal heterogeneous data, enabling a comprehensive understanding and collaborative assessment of defects and the overall performance degradation state of materials. Based on the neural network output, defect type identification, severity assessment, and potential risk prediction are performed. If the defect confidence level is higher than a preset threshold, it is marked as a specific structural defect, and its potential mechanical performance impact assessment and risk level are provided.

Images