A friction stir welding method and system based on ultrasonic testing

Abstract

This application relates to the field of intelligent control technology for welding quality, and discloses a friction stir welding method and system based on ultrasonic detection. The ultrasonic detection-based friction stir welding method includes the following steps: acquiring material density and sound velocity, calculating acoustic impedance and matching wavelet basis functions; performing pre-weld ultrasonic detection based on the function to construct a defect-free reference signal; acquiring real-time ultrasonic signals and temperature and stress information during welding; extracting features and inputting them into a neural network to output a defect probability map; constructing a state vector and inputting it into a reinforcement learning module; outputting and executing welding parameter adjustment actions; and using a physical information neural network to predict performance, triggering parameter reset when the performance falls below a threshold. This invention employs a method of constructing a defect probability map based on ultrasonic data, mapping the ultrasonic detection signal into a spatial probability distribution map, achieving visual perception of internal defects in the weld, and achieving the technical effect of accurately identifying potential defects in the welding area without damaging the sample.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for welding quality, specifically to a friction stir welding method and system based on ultrasonic testing. Background Technology

[0002] In the field of high-end manufacturing and engineering equipment welding, friction stir welding (FSW) has been widely used in aerospace, high-speed trains, and new energy vehicles—applications with extremely high structural reliability requirements—due to its advantages such as low heat input, high joint strength, and no need for filler material. Especially in welding scenarios involving light alloys and dissimilar materials, traditional fusion welding methods face numerous technical challenges, while FSW, with its solid-state bonding mechanism, has become an ideal choice for addressing material heterogeneity and heat sensitivity issues. However, the welding process involves many uncertainties in quality, such as the complex formation mechanism of weld defects and the significant impact of process parameter fluctuations, urgently requiring a more real-time and intelligent method for ensuring welding quality.

[0003] In existing technologies, quality control during friction stir welding primarily employs the following measures: Ultrasonic non-destructive testing is used to identify defects in the post-weld structure, effectively identifying typical internal discontinuities such as pores and cracks; some systems incorporate thermocouples and force sensors to collect physical parameters during the process for preliminary monitoring; in terms of control, some welding systems dynamically adjust process parameters such as rotational speed and feed rate using preset empirical rules combined with PID or fuzzy logic control algorithms to maintain process stability; and at the performance evaluation level, most rely on sampling tensile tests and metallographic analysis to determine weld strength post-construction. These technical approaches can achieve a certain level of quality assurance under standardized welding conditions and possess engineering practicality.

[0004] While the aforementioned technologies have improved welding quality control to some extent, they still suffer from the following unavoidable drawbacks: First, traditional ultrasonic testing methods heavily rely on fixed threshold judgments, failing to construct continuous defect probability distribution maps and limiting identification sensitivity and resolution, thus unable to meet the needs of online defect spatial localization. Second, existing condition monitoring primarily relies on single physical quantity feedback, lacking the ability to jointly model stress, temperature, and defect information, making it difficult for the control system to globally perceive the welding state, resulting in delayed adjustments or insufficient responses. Third, control strategies are typically based on static models or empirical rules, lacking the ability to automatically learn and optimize the process from real-time conditions, leading to instability in complex environments such as abrupt welding path changes and material variations. Finally, existing weld performance evaluation methods are mainly post-inspection, unable to participate in the control process, and have long evaluation cycles and high costs, failing to achieve real-time correction under feedforward drive. Therefore, those skilled in the art propose an ultrasonic testing-based friction stir welding method and system to address these issues. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a friction stir welding method and system based on ultrasonic testing, which solves the problems of welding defect identification relying on fixed thresholds, single state perception, rigid process control rules, and lagging performance evaluation in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a friction stir welding method based on ultrasonic testing, comprising the following steps:

[0007] S1. Obtain the physical parameters of the material to be welded, calculate the acoustic impedance value based on the obtained physical parameters, and match a more adaptable wavelet basis function based on the acoustic impedance value.

[0008] S2. Based on the wavelet basis function, perform pre-welding ultrasonic testing on the area to be welded, acquire the original ultrasonic signal, perform filtering preprocessing, and establish a reference signal template.

[0009] S3. Based on the reference signal template, real-time ultrasonic signals are collected synchronously during the welding process, and combined with the temperature field and stress field information of the welding area, a multi-source data foundation of multi-physical coupling state is formed.

[0010] S4. Based on the real-time ultrasound signal, the wavelet basis function is used to extract time-frequency domain features, and the features are input into a pre-trained neural network model to output a defect probability distribution map.

[0011] S5. Using the defect probability distribution map as the core input, and combining the temperature field and stress field information, construct a state vector and input it into the reinforcement learning control module for policy judgment.

[0012] S6. The reinforcement learning control module outputs the corresponding welding process parameter adjustment action based on the input state vector;

[0013] S7. Combining the welding process parameters and state vectors, a physical information neural network is used to predict and analyze the microstructure and macroscopic mechanical properties of the weld.

[0014] Preferably, the step of obtaining the physical parameters of the material to be welded includes:

[0015] The physical parameter of the material is density. With the speed of sound Obtained through experiments or by looking up tables;

[0016] According to the density With the speed of sound Calculate acoustic impedance ;

[0017] Select an appropriate wavelet basis function based on the calculated acoustic impedance value. .

[0018] Preferably, the pre-welding ultrasonic testing step includes:

[0019] The wavelet basis function is applied to the area to be welded to construct an ultrasonic wave propagation model, and the original ultrasonic signal is acquired and recorded. ;

[0020] For the original signal Bandpass filtering is performed to remove noise with frequencies outside the effective detection range, resulting in the processed signal. ;

[0021] A reference signal template is established by performing time-frequency analysis on the processed signal.

[0022] Preferably, the step of simultaneously acquiring real-time ultrasound signals and forming a multi-source data basis from the multi-physical coupling state includes:

[0023] During the welding process, ultrasonic signals within the welding area are acquired in real time and synchronously. At the same time, temperature sensors are used to collect the temperature field in the welding area. Stress sensors are used to collect stress fields in the welding area. ,in Represents the spatial coordinates of the weld cross-section or surface. Indicates the sampling time point;

[0024] These multi-source data are merged through a data synchronization algorithm to form a comprehensive data foundation under multi-physical coupling conditions.

[0025] Preferably, the steps of time-frequency domain feature extraction and outputting the defect probability distribution map include:

[0026] The real-time ultrasound signal is analyzed using the wavelet basis function. Perform continuous wavelet transform to obtain time-frequency domain features. ,in As a scale factor, For time translation;

[0027] The extracted time-frequency features are input into a pre-trained deep neural network model, which outputs a spatial distribution map reflecting the probability of defects. .

[0028] Preferably, the steps for constructing the state vector include:

[0029] Combining defect probability diagrams Temperature field information and stress field information Construct a multi-physical quantity state vector:

[0030] ;

[0031] in: For at any time The state vector; This represents the defect probability distribution function; This represents the area of ​​the spatial detection region corresponding to the defect probability map. This represents the maximum temperature value in the temperature field at the current time point; This represents the spatial variance of the stress field.

[0032] Preferably, the reinforcement learning control module outputs corresponding welding process parameter adjustment actions based on the input state vector, including:

[0033] The control module uses the input state vector Output welding parameter adjustment action:

[0034] ;

[0035] in: This refers to the action of adjusting welding process parameters; This is the adjustment amount for the stirring head rotation speed; This is the adjustment amount for downforce; This is the adjustment amount for the welding travel speed;

[0036] The parameter adjustment action is used to dynamically correct the welding process.

[0037] Preferably, the welding performance prediction step of the physical information neural network includes:

[0038] Using multi-physical quantity state vector As input, the microstructure and macroscopic mechanical properties of the weld, especially the grain size, are predicted. and tensile strength ;

[0039] By training a physical information neural network, the prediction results are optimized to achieve accurate welding quality assessment.

[0040] Preferably, the method of using a physical information neural network to predict and analyze the microstructure and macroscopic mechanical properties of the weld specifically includes:

[0041] If the tensile strength predicted by the physical information neural network is... Less than the set threshold This triggers a parameter reset, the parameter reset steps of which include:

[0042] ;

[0043] in: This refers to the stirring head rotation speed; The axial pressure applied by the stirring head to the material to be welded; The speed at which the stirring head propels the workpiece; These are the initial process parameters. This is the compensation action calculated by the control module based on the current state.

[0044] A friction stir welding system based on ultrasonic testing, comprising:

[0045] Friction stir welding host, used to perform the welding process and support the adjustment of welding parameters;

[0046] The ultrasonic testing module includes a phased array probe and a signal processing unit, used to acquire and process ultrasonic signals;

[0047] The thermal imaging device and stress sensor module are used to acquire temperature field and stress field information of the welding area;

[0048] The neural network analysis unit is used to extract time-frequency features from the ultrasonic signal and output a defect probability map.

[0049] The reinforcement learning control module is used to adjust welding process parameters based on the state vector of multiple physical quantities.

[0050] Physical information neural network units are used to evaluate the microstructure and macroscopic mechanical properties of welds after welding.

[0051] The control execution unit is used to modify the welding process execution instructions based on the calculated adjustment actions.

[0052] In summary, this application includes at least one of the following beneficial technical effects:

[0053] 1. This invention employs a method for constructing a defect probability map based on ultrasonic data, mapping ultrasonic detection signals into a spatial probability distribution map to achieve visual perception of internal defects in welds. This achieves the technical effect of accurately identifying potential defects in the welding area without damaging the sample. Compared with the existing technology that relies solely on amplitude gating for judgment, this invention solves the problems of poor identification ability and high misjudgment rate for early defects such as microcracks.

[0054] 2. By introducing a multi-physics state quantity fusion mechanism, this invention constructs a unified state vector from the temperature, stress, and defect distribution during the welding process, achieving full information perception of the welding process. This vector can serve as the basis for control strategy decisions. Compared with the existing technical path of single temperature or force sensing feedback, this invention effectively overcomes the problems of insufficient state information dimensions and feedback control lag.

[0055] 3. This invention proposes an online optimization mechanism for process parameters that integrates reinforcement learning algorithms. Based on the current state, the stirring head rotation speed, pressure, and travel speed are intelligently adjusted, thereby achieving the goal of adaptive adjustment of process parameters. Unlike the traditional setpoint control method based on empirical rules, this effectively avoids the technical shortcomings of rigid parameter settings and weak ability to cope with disturbances.

[0056] 4. By employing a physical information neural network to predict the mechanical properties of welds online, this invention can predict the microstructure and tensile strength in real time during the later stages of the welding process, thereby realizing a performance-driven process correction path. Compared with the method that relies on offline test evaluation in the later stages, this invention solves the problems of lag and untimely adjustment, and enhances the feedforward capability of quality control. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0058] Figure 2 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0059] The technical solutions in 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.

[0060] Please see the appendix Figure 1 This invention provides a friction stir welding method based on ultrasonic testing, comprising the following steps:

[0061] S1. Obtain the physical parameters of the material to be welded, calculate the acoustic impedance value based on the obtained physical parameters, and match a more adaptable wavelet basis function based on the acoustic impedance value.

[0062] Specifically, in this invention, step S1 involves welding system initialization and material matching configuration. The purpose of this step is to provide accurate material parameters and suitable wavelet basis functions for the subsequent ultrasonic testing process. By appropriately selecting an ultrasonic signal processing method that matches the material properties, potential defects can be effectively detected and identified during the welding process. First, the system needs to acquire the physical parameters of the material to be welded, especially its density and sound velocity. These parameters are the foundation of ultrasonic testing; they directly affect the propagation characteristics of ultrasonic waves, thus determining the detection accuracy and reliability.

[0063] In this embodiment, the physical parameters of the material to be welded are first obtained, including density. and speed of sound These parameters are typically obtained through experimental testing or by looking up tables. Based on the obtained physical parameters, the acoustic impedance can be calculated. The calculation formula is as follows:

[0064] ;

[0065] in: Density of a material is expressed in kg / m³. 3 ; The speed of sound in a material is expressed in m / s. This represents the acoustic impedance value.

[0066] The acoustic impedance value obtained through the above formula is an important physical quantity describing the impedance characteristics of a material to ultrasound.

[0067] In one possible implementation, the acquired acoustic impedance value is used to select a more adaptable wavelet basis function. Specifically, based on the acoustic impedance matching principle, the wavelet basis function most suitable for the material to be welded can be selected from a wavelet basis function library. These wavelet basis functions will be used for subsequent ultrasonic signal feature extraction to improve the accuracy of time-frequency analysis. Alternatively, the selection of wavelet basis functions may depend not only on the acoustic impedance value but also on other factors such as the material's thermal conductivity and stress state.

[0068] Specifically, the selected wavelet basis functions This signal processing method will be used for pre-weld ultrasonic testing of the area to be welded. By extracting time-frequency features from the ultrasonic signal using the wavelet basis function, important feature information in the signal can be effectively captured, especially signal features related to potential defects in the welding process.

[0069] In some embodiments, the selection of wavelet basis functions is not only based on the acoustic impedance of the material, but may also take into account the actual welding process requirements. Wavelet basis functions have strong time-frequency localization characteristics, which can effectively separate the time-domain and frequency-domain information of ultrasonic signals, which is crucial for real-time defect detection.

[0070] Through the above-mentioned technology, step S1 ensures that during the welding process, the system can dynamically select the most suitable wavelet basis function according to the physical properties of the welding material, thereby improving the accuracy and reliability of ultrasonic testing.

[0071] S2. Based on wavelet basis functions, perform pre-welding ultrasonic testing on the area to be welded, acquire the original ultrasonic signal, perform filtering preprocessing, and establish a reference signal template.

[0072] Specifically, after acquiring the physical parameters of the welding materials and calculating the acoustic impedance, and matching a highly adaptable wavelet basis function, preliminary detection and analysis of the area to be welded is required to achieve pre-weld ultrasonic testing and evaluation and establish a signal template for a defect-free reference state. This step is not only a fundamental preparatory process for the defect identification module, but also provides a benchmark for subsequent real-time signal comparison, thus playing a crucial role in the entire welding control method.

[0073] In this embodiment, based on the wavelet basis function determined in step S1 Pre-welding ultrasonic testing is performed on the area to be welded.

[0074] Generally, during the pre-welding ultrasonic testing stage, a phased array ultrasonic probe is used to perform multi-angle, multi-path coverage scanning of the welding path area and record the raw ultrasonic signal data. .in, This represents the original ultrasound signal in the time domain, expressed in volts (V) or normalized units.

[0075] To improve the accuracy of subsequent signal analysis, this original signal The signal processing requires filtering and preprocessing. Specifically, the signal processing flow includes bandpass filtering to remove environmental and instrument noise that exceeds the detection frequency band. In one possible implementation, the frequency range of the bandpass filter is set based on the principal frequency domain response characteristics of the selected wavelet basis function, and can be expressed as:

[0076] ;

[0077] in: This represents the signal after filtering. BPF stands for Bandpass Filter Operator, and its frequency band range is... It needs to be set in conjunction with the center frequency of the ultrasonic transducer and the attenuation characteristics of the material.

[0078] After signal preprocessing, time-frequency domain features are further extracted based on wavelet transform. Specifically, the processed signal is transformed into a two-dimensional time-frequency space using continuous wavelet transform (CWT), and the calculation formula is as follows:

[0079] ;

[0080] in: These are wavelet coefficients; This is the scaling factor (corresponding to frequency characteristics), and its unit is dimensionless. This is the translation factor (corresponding to the time characteristic), with the unit being seconds (s). This represents the conjugate of the wavelet basis functions.

[0081] As an alternative, to enhance the interpretability of the wavelet coefficient matrix, a two-dimensional feature map can be constructed using scale-energy mapping to identify regions with stable energy density distributions.

[0082] Subsequently, based on the wavelet time-frequency analysis results, a reference signal template for the defect-free state is constructed. This template can be represented as a set of characteristic matrices or vectors, specifically in the form of:

[0083] ;

[0084] in: This represents the baseline template; its value range constitutes the baseline dataset for subsequent real-time signal comparison with standard states.

[0085] In some embodiments, to ensure the stability and representativeness of the reference template, the system can repeatedly collect pre-welding signals and use multiple superposition averaging methods to reduce the impact of environmental disturbances, thereby improving the consistency of the template.

[0086] In addition, in some embodiments, the ultrasonic path simulation calculation involved in the construction of the reference signal template can be incorporated into the modeling process, that is, the frequency domain expression capability of the template can be optimized by ultrasonic propagation path simulation and echo response model.

[0087] S3. Based on the reference signal template, real-time ultrasonic signals are collected synchronously during the welding process. Combined with the temperature field and stress field information of the welding area, a multi-source data foundation with multiple physical coupling states is formed.

[0088] Specifically, after completing the pre-weld ultrasonic testing and constructing the reference signal template, in order to achieve real-time defect identification and adaptive process control during welding, it is necessary to simultaneously acquire dynamic ultrasonic signals from the weld area during welding execution and integrate multiple physical field information to comprehensively characterize the welding state. This step not only expands the temporal dimension of signal detection but also introduces a multi-physics coupling mechanism, laying a multi-source data foundation for subsequent feature extraction and intelligent control modules.

[0089] In this embodiment, a synchronous data acquisition mechanism for the welding process is activated simultaneously with the start of friction stir welding by the welding host.

[0090] Specifically, the same phased array ultrasonic testing device as before welding is used, based on the selected wavelet basis function. Periodic or continuous ultrasonic testing is performed on the weld area to acquire dynamic ultrasonic signals during the welding process in real time. .in, The real-time ultrasonic response signal varies over time, and the unit is volts (V) or normalized value.

[0091] In general, in order to achieve real-time monitoring of multi-dimensional information of the welding process, in addition to collecting ultrasonic signals, temperature field and stress field data need to be introduced simultaneously.

[0092] Alternatively, a thermocouple array or infrared thermal imaging device can be deployed in the weld area to obtain the two-dimensional temperature field distribution that evolves over time during the welding process. ,in: It indicates temperature, and the unit is Kelvin (K) or Celsius (°C). Indicates the spatial coordinates of the weld section or surface, in millimeters (mm). Indicates the sampling time point, in seconds (s).

[0093] At the same time, strain gauges or stress grating sensor arrays are used to acquire stress field information, forming spatiotemporal distribution data of stress within the welding area. ,in: It represents the stress tensor or scalar form, with the unit being megapascals (MPa); other parameters are defined as above.

[0094] In one possible implementation, the acquisition of the aforementioned ultrasonic signals, temperature field information, and stress field information employs a unified timestamp calibration mechanism to ensure the temporal consistency of multi-source data. The data synchronization module can perform data alignment processing using either hardware trigger signals or software time synchronization algorithms.

[0095] The above-mentioned multiple data sources are integrated in a structured manner to form a unified data foundation, denoted as:

[0096] ;

[0097] in: Indicates at a certain moment The multi-source datasets collected below serve as the basic input for subsequent feature extraction and state analysis.

[0098] In some embodiments, to further enhance the spatial coverage of data acquisition, sensor arrays can be symmetrically arranged on both sides of the welding head path to achieve full-domain reconstruction of the physical field of the weld area. In the welding of specific structural components, a three-dimensional reconstruction mechanism can also be introduced to extend the two-dimensional temperature field into a three-dimensional heat distribution body. This enhances the ability to characterize thermal and mechanical fields.

[0099] Furthermore, after multi-source data fusion, algorithms such as linear interpolation and multi-scale resampling can be used to achieve grid unification between different physical quantities, ensuring that when they are input as unified state input vectors into neural networks or reinforcement learning modules, they have consistent spatial and temporal characteristics.

[0100] S4. Based on real-time ultrasonic signals, wavelet basis functions are used to extract time-frequency domain features, which are then input into a pre-trained neural network model to output a defect probability distribution map.

[0101] Specifically, in the aforementioned steps, the system has established a multi-source data foundation, including real-time ultrasonic signals during the welding process, as well as synchronously acquired temperature and stress field information. To effectively identify potential defects during welding and provide a clear basis for subsequent control strategies, it is necessary to perform in-depth time-frequency domain feature extraction on the real-time ultrasonic signals and combine this with a well-trained neural network model to predict and represent the probability distribution of defects. This step plays a crucial role in information transformation and intelligent identification.

[0102] In this embodiment, the wavelet basis functions already matched in step S1 are first used as the basis for the initial implementation. For the real-time acquired ultrasound signals Perform a continuous wavelet transform (CWT) operation. This method can map a one-dimensional time-domain signal into a two-dimensional time-frequency feature space to capture local signal variation characteristics.

[0103] In general, the ultrasound signal is processed using the following transformation formula:

[0104] ;

[0105] in: This represents the wavelet coefficient matrix obtained after the transformation; It is a scale factor that reflects the frequency information of the signal, and its unit is dimensionless; This is the translation factor, corresponding to time information, in seconds (s); This is the complex conjugate form of the wavelet basis functions.

[0106] Specifically, scale The selected range should cover all effective frequencies within the operating frequency band of the ultrasonic transducer, typically selected as multiple discrete scale values ​​to enhance frequency domain resolution. In some embodiments, a logarithmic scaling strategy is employed to balance the distribution of low-frequency and high-frequency information.

[0107] The two-dimensional time-frequency coefficients obtained after wavelet transform The energy distribution of the ultrasound signal at different frequencies and time locations was characterized. Alternatively, the system can... Normalization is performed, and statistical features such as maximum value, mean, and energy density are extracted to construct feature vectors, which are used as input to the neural network model.

[0108] In one possible implementation, the feature vector is fed into a pre-trained offline deep neural network model for defect classification and probability prediction. This neural network model typically includes multiple convolutional and fully connected layers, possessing strong spatial awareness and nonlinear fitting capabilities, enabling it to accurately identify minute defect features in ultrasound signals.

[0109] The output of the neural network is a two-dimensional defect probability distribution map, denoted as:

[0110] ;

[0111] in: Represents coordinates in two-dimensional space The defect probability value is within the range of [value missing]. ; Represents the inference function of a neural network; These are the spatial coordinates of the weld area, in millimeters (mm).

[0112] In some embodiments, the defect probability map can be compared with the feature distribution of a reference signal template to further eliminate misjudgments caused by operating condition disturbances.

[0113] Furthermore, when the structure of a specific welding path is complex or the acoustic properties of the material are non-uniform, an attention mechanism can be introduced into the network structure to enhance the feature representation of key regions, thereby improving prediction accuracy.

[0114] In this invention, the defect probability distribution map serves as core intermediate data, reflecting not only the current health status of the weld but also directly as an important component for subsequent state vector construction and control decision input.

[0115] S5. Using the defect probability distribution map as the core input, and combining temperature field and stress field information, construct a state vector and input it into the reinforcement learning control module for policy judgment.

[0116] Specifically, after extracting the time-frequency features of the real-time ultrasonic signals during the welding process and outputting a defect probability distribution map through a deep neural network model, the obtained defect spatial information needs to be further fused with multi-physical quantity information such as the temperature field and stress field in the welding area. The goal of this step is to construct a state vector that comprehensively reflects the welding state, providing a decision-making basis for the subsequent strategy judgment and parameter adjustment of the reinforcement learning control module. The construction of this state vector should fully reflect the defect trend, thermal field distribution characteristics, and their mutual coupling effects in the welding area.

[0117] In this embodiment, the defect probability distribution map obtained in step S4 is used first. Based on the core foundation, combined with temperature field information With stress field information Construct a multi-physical quantity state representation vector.

[0118] Generally, to simplify the dimensionality of state variables and enhance their responsiveness to control strategies, spatial reduction and statistical extraction methods can be used to convert two-dimensional or three-dimensional physical field information into a set of representative statistical indicators. Specifically, the constructed state vector... It includes the following three dimensional components:

[0119] ;

[0120] in: For at any time The state vector is a three-dimensional vector structure. This represents the defect probability distribution function, with a range of values ​​of 1. This indicates the possibility of a defect in a weld at a certain location; The area of ​​the spatial detection region corresponding to the defect probability map, in mm. 2 , used for normalization processing; This represents the maximum temperature value in the temperature field at the current time point, in degrees Celsius (°C). Represents the spatial variance of the stress field, in MPa. 2 This reflects the non-uniformity of stress distribution.

[0121] In one possible implementation, the maximum temperature The temperature distribution matrix can be obtained by scanning it pixel-by-pixel, and can be used to determine whether there is a risk of overheating or temperature unevenness in the welding area. Furthermore, the variance of the stress field can be calculated based on the following form:

[0122] ;

[0123] in: The average stress of the stress field at that moment, in MPa, is expressed as:

[0124] ;

[0125] Through the above processing method, the state vector It can express the coupled state of multiple physical quantities in a low-dimensional form, which not only preserves the defect information, but also takes into account the changing characteristics of the thermal field and the force field.

[0126] In some embodiments, to enhance the sensitivity of the state vector to sudden anomalies, a difference form or sliding window statistical feature can be introduced. For example, the difference between the state vectors at two consecutive time points can be calculated:

[0127] ;

[0128] in: The sampling time interval is used to identify rapid fluctuations in the welding status over a short period of time.

[0129] As an alternative, to improve the robustness of the state vector, a confidence interval evaluation mechanism can be introduced during the construction process to automatically remove or weight outlier data. This mechanism can be implemented using a distribution model trained on a historical dataset.

[0130] S6. The reinforcement learning control module outputs corresponding welding process parameter adjustment actions based on the input state vector.

[0131] Specifically, after constructing the multi-physical quantity state vector, to achieve closed-loop adaptive control in the welding process, a reinforcement learning control mechanism is needed to make policy judgments on the current state and output corresponding welding process parameter adjustment actions. This step, by introducing a data-driven intelligent decision-making method, enables the welding host to automatically adjust process execution parameters based on the perceived state, thereby dynamically adapting to complex changes in the welding process. This is one of the core components of the intelligent control achieved in this invention.

[0132] In this embodiment, the state vector constructed in step S5 is used. The input is fed into the reinforcement learning control module, which makes strategy decisions and outputs welding process parameter adjustment actions.

[0133] In general, the state vector The physical information indicators include three dimensions: the spatial mean of defect probability, the peak welding temperature, and the spatial variance of stress distribution. These indicators comprehensively characterize the forming stability and structural health of the current welded area.

[0134] The reinforcement learning control module is based on the policy function. The input state vector is evaluated, and the action vector is output.

[0135] ;

[0136] in: This refers to the action of adjusting welding process parameters, expressed in vector form. This indicates the adjustment amount of the stirring head speed, in rpm; This indicates the adjustment amount of the applied downward pressure, expressed in N; This indicates the adjustment amount for the welding travel speed, expressed in mm / s.

[0137] Specifically, the policy function It can be trained using value-based reinforcement learning methods (such as DQN), policy gradient-based methods (such as PPO), or model-based control methods. During training, the optimization objectives are to minimize the defect rate and maintain welding performance, and a mapping relationship from the state space to the action space is established.

[0138] In one possible implementation, the optimization objective function of the control policy is... It can be defined as:

[0139] ;

[0140] in: This represents the average value of the defect probability map; An index representing the smoothness of speed changes, used to suppress drastic fluctuations; The energy efficiency coefficient represents the control action and is used to balance energy consumption; , Weighting factors are set according to specific needs. State vector The probability distribution; State vector Follows distribution Expectations at that time.

[0141] As an alternative, if a certain index in the state vector undergoes a sudden change (such as a sharp increase in the variance of the stress field), the system can trigger a priority control strategy to adjust the response weight of a certain process parameter accordingly, ensuring that the control decision is timely and directional.

[0142] In some embodiments, to avoid lag in the response of the control system, an application window for the process adjustment action can be set so that the parameter changes are completed gradually within a finite time, expressed as:

[0143] ;

[0144] in: The current rotational speed, in rpm; To smoothly adjust the time step, the unit is the number of sampling periods; This is the maximum adjustment duration, used to balance control speed and system stability.

[0145] Through the linkage between the above strategy output and the process adjustment interface, a new process execution command is finally generated. This command is sent to the stirring head, pressure mechanism and feed mechanism through the welding host control system to realize the dynamic updating of process parameters.

[0146] S7. Combining welding process parameters and state vectors, a physical information neural network is used to predict and analyze the microstructure and macroscopic mechanical properties of the weld.

[0147] Specifically, after implementing dynamic adjustment of process parameters based on reinforcement learning, the welding process has acquired real-time perception and strategy response capabilities. However, the final quality of the weld still requires comprehensive evaluation through the microstructure and mechanical properties of the weld. Therefore, it is necessary to introduce a performance prediction mechanism oriented towards result feedback to model and analyze the weld formation quality, and combine this with a set performance threshold to trigger parameter correction strategies, forming a complete closed-loop control process. This step uses a physical information neural network to achieve structure-performance correlation modeling, providing key support for realizing prediction-driven self-correcting process control.

[0148] In this embodiment, the state vector constructed in step S5 is used. The data is input into a trained Physics-Informed Neural Network (PINN) model to predict and analyze the microstructure and macroscopic mechanical properties of the weld.

[0149] Typically, this neural network uses the defect probability, peak temperature, and stress fluctuation indices contained in the state vector as input variables, and outputs key performance indicators representing weld quality. These include average grain size. and weld tensile strength .

[0150] Specifically, the mapping relationship of the network can be represented as:

[0151] ;

[0152] in: This indicates the average grain size of the weld zone, expressed in micrometers (μm). This indicates the ultimate tensile strength of the weld, expressed in megapascals (MPa). This represents the function mapping relationship of a neural network for physical information.

[0153] In one possible implementation, the physical information neural network not only introduces supervision signals during training, but also embeds physical equations such as grain evolution laws and thermo-mechanical coupling microscopic models as priors into the loss function to improve the model's generalization ability and physical consistency to different process disturbances.

[0154] Alternatively, the loss function of this model can be designed as a composite form as follows:

[0155] ;

[0156] in: These are the model's predicted values; These are experimental or simulated values; These are weighting coefficients;

[0157] The last term is the physical constraint term, which reflects the relationship between stress changes and grain structure gradients.

[0158] After the prediction is completed, the predicted tensile strength value will be... Compared with preset performance threshold Comparison:

[0159] Trigger parameter reset mechanism;

[0160] If the judgment condition is met, i.e., the performance does not meet the design requirements, then the parameter correction stage begins. This stage no longer relies on the original adjustment actions, but instead activates compensatory process actions. And reset the current process parameters, as detailed below:

[0161] ;

[0162] in: This refers to the stirring head rotation speed; The axial pressure applied by the stirring head to the material to be welded; The speed at which the stirring head propels the workpiece; These are the initial process parameters. The compensation actions calculated by the control module based on the current state include speed correction, pressure correction, and feed rate correction, with units of rpm, N, and mm / s, respectively.

[0163] In some embodiments, to avoid false triggering, the parameter reset mechanism can be judged by combining the mean or confidence interval of the prediction results within the sliding window, so as to increase the robustness of the triggering criterion.

[0164] Furthermore, in actual industrial environments, comparative matching analysis can be performed using historical weld databases to assist in the rationality assessment of parameter reset strategies and improve decision-making quality.

[0165] The ultrasonic-based friction stir welding system described below and the ultrasonic-based friction stir welding method described above can be referred to as correspondingly.

[0166] Please see the appendix Figure 2 A friction stir welding system based on ultrasonic testing, comprising:

[0167] Friction stir welding host, used to perform the welding process and support the adjustment of welding parameters;

[0168] The ultrasonic testing module includes a phased array probe and a signal processing unit, used to acquire and process ultrasonic signals;

[0169] The thermal imaging device and stress sensor module are used to acquire temperature field and stress field information of the welding area;

[0170] The neural network analysis unit is used to extract time-frequency features from the ultrasonic signal and output a defect probability map.

[0171] The reinforcement learning control module is used to adjust welding process parameters based on the state vector of multiple physical quantities.

[0172] Physical information neural network units are used to evaluate the microstructure and macroscopic mechanical properties of welds after welding.

[0173] The control execution unit is used to modify the welding process execution instructions based on the calculated adjustment actions.

[0174] The system in this embodiment can be used to execute the above method embodiments, and its principle and technical effect are similar, so they will not be described again here.

[0175] The embodiments described in this specific implementation are preferred embodiments of this application and are not intended to limit the scope of protection of this application. Identical components are represented by the same reference numerals. Therefore, all equivalent changes made to the structure, shape, and principle of this application should be covered within the scope of protection of this application.

Claims

1. A friction stir welding method based on ultrasonic testing, characterized in that, Includes the following steps: S1. Obtain the physical parameters of the material to be welded, calculate the acoustic impedance value based on the obtained physical parameters, and match a more adaptable wavelet basis function based on the acoustic impedance value. S2. Based on the wavelet basis function, perform pre-welding ultrasonic testing on the area to be welded, acquire the original ultrasonic signal, perform filtering preprocessing, and establish a reference signal template. S3. Based on the reference signal template, real-time ultrasonic signals are collected synchronously during the welding process, and combined with the temperature field and stress field information of the welding area, a multi-source data foundation of multi-physical coupling state is formed. S4. Based on the real-time ultrasound signal, the wavelet basis function is used to extract time-frequency domain features, and the features are input into a pre-trained neural network model to output a defect probability distribution map. S5. Using the defect probability distribution map as the core input, and combining the temperature field and stress field information, construct a state vector and input it into the reinforcement learning control module for policy judgment. The steps for constructing the state vector include: Combining defect probability diagrams Temperature field information and stress field information Construct a multi-physical quantity state vector: ； in: For at any time The state vector; This represents the defect probability distribution function; This represents the area of ​​the spatial detection region corresponding to the defect probability map. This represents the maximum temperature value in the temperature field at the current time point; Represents the spatial variance of the stress field; S6. The reinforcement learning control module outputs the corresponding welding process parameter adjustment action based on the input state vector; S7. Combining the welding process parameters and state vectors, a physical information neural network is used to predict and analyze the microstructure and macroscopic mechanical properties of the weld, specifically including: If the tensile strength predicted by the physical information neural network is... Less than the set threshold This triggers a parameter reset, the parameter reset steps of which include: ； in: This refers to the stirring head rotation speed; The axial pressure applied by the stirring head to the material to be welded; The speed at which the stirring head propels the workpiece; These are the initial process parameters. This is the compensation action calculated by the control module based on the current state.

2. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The steps for obtaining the physical parameters of the material to be welded include: The physical parameter of the material is density. With the speed of sound Obtained through experiments or by looking up tables; According to the density With the speed of sound Calculate acoustic impedance ; Select an appropriate wavelet basis function based on the calculated acoustic impedance value. .

3. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The steps of the pre-welding ultrasonic testing include: The wavelet basis function is applied to the area to be welded to construct an ultrasonic wave propagation model, and the original ultrasonic signal is acquired and recorded. ; For the original signal Bandpass filtering is performed to remove noise with frequencies outside the effective detection range, resulting in the processed signal. ; A reference signal template is established by performing time-frequency analysis on the processed signal.

4. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The steps for forming the multi-source data basis of the synchronous acquisition of real-time ultrasound signals and the multi-physical coupling state include: During the welding process, ultrasonic signals within the welding area are acquired in real time and synchronously. At the same time, temperature sensors are used to collect the temperature field in the welding area. Stress sensors are used to collect stress fields in the welding area. ,in: Represents the spatial coordinates of the weld section or surface. Indicates the sampling time point; These multi-source data are merged through a data synchronization algorithm to form a comprehensive data foundation under multi-physical coupling conditions.

5. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The steps of extracting time-frequency domain features and outputting the defect probability distribution map include: The real-time ultrasound signal is analyzed using the wavelet basis function. Perform continuous wavelet transform to obtain time-frequency domain features. ,in As a scale factor, For time translation; The extracted time-frequency features are input into a pre-trained deep neural network model, which outputs a spatial distribution map reflecting the probability of defects. .

6. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The reinforcement learning control module outputs corresponding welding process parameter adjustment actions based on the input state vector, including: The control module uses the input state vector Output welding parameter adjustment action: ； in: This refers to the action of adjusting welding process parameters; This is the adjustment amount for the stirring head rotation speed; This is the adjustment amount for downforce; This is the adjustment amount for the welding travel speed; The parameter adjustment action is used to dynamically correct the welding process.

7. The friction stir welding method based on ultrasonic testing according to claim 1, characterized in that, The welding performance prediction steps of the physical information neural network include: Using multi-physical quantity state vector As input, predict the grain size of the weld. and tensile strength ; By training a physical information neural network, the prediction results are optimized to achieve accurate welding quality assessment.

8. A friction stir welding system based on ultrasonic testing, applied to the friction stir welding method based on ultrasonic testing as described in any one of claims 1-7, characterized in that, include: Friction stir welding host, used to perform the welding process and support the adjustment of welding parameters; The ultrasonic testing module includes a phased array probe and a signal processing unit, used to acquire and process ultrasonic signals; The thermal imaging device and stress sensor module are used to acquire temperature field and stress field information of the welding area; The neural network analysis unit is used to extract time-frequency features from the ultrasonic signal and output a defect probability map. The reinforcement learning control module is used to adjust welding process parameters based on the state vector of multiple physical quantities. Physical information neural network units are used to evaluate the microstructure and macroscopic mechanical properties of welds after welding. The control execution unit is used to modify the welding process execution instructions based on the calculated adjustment actions.

Images