Testing method for high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical collaboration

Abstract

The present invention relates to a method for testing the high-temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical coordination. The method: install the welding material specimen on a loading device in a test cavity and set parameters; heat the welding material specimen based on a heating device, and apply dynamic stress through the loading device to obtain a specimen in a high-temperature stress state; perform multi-source signal acquisition on the specimen in the high-temperature stress state to obtain time-synchronized acoustic, optical and electrical coordinated signal data; perform noise reduction filtering and time-frequency domain feature extraction on the acoustic, optical and electrical coordinated signal data, and then input a preset feature classification model to obtain multi-modal classification feature data; input the multi-modal classification feature data into a pre-trained sequence-to-sequence neural network model to obtain crack resistance prediction results including crack initiation time, expansion rate and fracture toughness. The implementation of the present invention realizes a comprehensive and accurate evaluation of the high-temperature crack resistance of welding materials.

Description

Technical Field

[0001] The present invention relates to the technical field of welding material performance testing, and in particular to a method for testing the high-temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical collaboration. Background Art

[0002] The high-temperature crack resistance of welding materials is crucial to ensure the safety and reliability of welded structures. Traditional testing methods for the crack resistance of welding materials often rely on a single signal source and simple data analysis technology, which makes it difficult to fully and accurately evaluate the performance of materials in complex high-temperature environments. With the continuous improvement of industrial application requirements for welding material performance, the testing method of a single signal source can no longer meet the increasingly stringent evaluation needs.

[0003] At present, the high-temperature crack resistance test of welding materials faces multiple challenges. Signal acquisition in high-temperature environments is difficult, and the stability and accuracy of traditional sensors under extreme conditions are difficult to guarantee. Secondly, the collaborative analysis and fusion processing technology of multi-source signals needs to be improved, and the correlation and complementarity between different types of signals have not been fully based on. The interpretation of test data and performance prediction still rely on empirical models, lacking intelligent and adaptive analysis methods. Summary of the invention

[0004] The main purpose of the present invention is to provide a method for testing the high-temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical collaboration, so as to achieve a comprehensive and accurate evaluation of the high-temperature crack resistance of welding materials.

[0005] To achieve the above object, the present invention provides a method for testing the high temperature crack resistance of welding materials based on intelligent acoustic-optical-electrical synergy, comprising the following steps:

[0006] The welding material specimen is mounted on a loading device in the test chamber, and parameters of a heating device, an acoustic signal acquisition device, an optical signal acquisition device, and an electrical signal detection device are set to obtain an initialized test system;

[0007] The welding material specimen is heated by the heating device and dynamic stress is applied by the loading device to obtain the specimen in a high temperature stress state;

[0008] Based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device, multi-source signal acquisition is performed on the test piece in the high temperature stress state to obtain time-synchronized acoustic, optical and electrical coordinated signal data;

[0009] The acoustic, optical and electrical coordinated signal data are subjected to noise reduction filtering and time-frequency domain feature extraction, and then input into a preset feature classification model to obtain multimodal classification feature data;

[0010] The multimodal hierarchical feature data is input into a pre-trained sequence to sequence neural network model to obtain crack resistance performance prediction results including crack initiation time, propagation rate and fracture toughness.

[0011] The present invention also provides a welding material high temperature crack resistance testing device based on intelligent acoustic and optical synergy, comprising:

[0012] An initialization module is used to install the welding material specimen on the loading device in the test chamber, and set parameters of the heating device, the sound signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain an initialized test system;

[0013] A heating module, used for heating the welding material specimen based on the heating device, and applying dynamic stress through the loading device to obtain the specimen in a high temperature stress state;

[0014] A data acquisition module, used to perform multi-source signal acquisition on the test piece in the high temperature stress state based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain time-synchronized acoustic, optical and electrical coordinated signal data;

[0015] A feature extraction module is used to perform noise reduction filtering and time-frequency domain feature extraction on the acoustic, optical and electrical coordinated signal data, and then input a preset feature classification model to obtain multimodal classification feature data;

[0016] The prediction module is used to input the multimodal hierarchical feature data into a pre-trained sequence to sequence neural network model to obtain a crack resistance performance prediction result including crack initiation time, propagation rate and fracture toughness.

[0017] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0018] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0019] In summary, the technical solution provided by the present invention realizes the all-round and multi-dimensional information collection of welding materials under high temperature stress state by integrating acoustic signal, optical signal and electrical signal acquisition devices, and improves the comprehensiveness and reliability of test data. The high-precision heating device and dynamic stress loading device are used to accurately simulate the high temperature and stress state of welding materials in the actual use environment. The multi-source signals are de-noised and optimized based on advanced algorithms such as wavelet threshold denoising and adaptive Wiener filtering, which significantly improves the signal-to-noise ratio and quality of the original data. Through short-time Fourier transform and time-frequency domain feature calculation, effective feature extraction of complex signals is realized. The fuzzy membership function and hierarchical analysis process model are used to realize the intelligent classification and fusion of different types of features, and the comprehensive expression ability of feature data is improved. Based on the sequence-to-sequence neural network model, combined with advanced technologies such as multi-head attention mechanism and residual connection, high-precision prediction of the crack resistance of welding materials is realized. Through real-time feedback and adaptive control algorithm, the test parameters can be dynamically adjusted to ensure the stability and controllability of the test process. At the same time, multiple key indicators such as crack initiation time, propagation rate and fracture toughness are predicted, providing a comprehensive crack resistance evaluation result. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the steps of a method for testing high temperature crack resistance of welding materials based on intelligent acoustic-optical-electrical coordination in one embodiment of the present invention;

[0021] Figure 2 This is a structural block diagram of a welding material high temperature crack resistance testing device based on intelligent acoustic and optical synergy in one embodiment of the present invention;

[0022] Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0023] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0025] Reference Figure 1 This embodiment provides a method for testing the high temperature crack resistance of welding materials based on intelligent acoustic-optical-electrical coordination, comprising the following steps:

[0026] S1, installing the welding material specimen on the loading device in the test chamber, and setting parameters of the heating device, the sound signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain an initialized test system;

[0027] Among them, the welding material specimen is installed on the fixture in the test cavity to ensure that the fixed position and stress state of the specimen in the fixture are optimal, so as to ensure that the specimen is evenly stressed and no additional stress concentration is generated during the subsequent test process, and to avoid test errors caused by poor assembly. The temperature curve of the heating device is programmed to enable it to heat the specimen according to the preset temperature change curve, simulating the high temperature environment and temperature gradient of the welding material in the actual welding process, so as to ensure that the experimental conditions are consistent with the actual welding environment and improve the reliability and accuracy of the test results. Based on the mechanical properties of the selected welding material, the stress parameters of the loading device are set, including parameters such as stress amplitude, loading frequency and loading mode, so as to design a dynamic stress loading scheme that meets the material properties and experimental purposes, and effectively simulate the dynamic stress of the welding material under high temperature conditions. At the same time, the microphones with different frequency response ranges in the sound signal acquisition device are calibrated to ensure that they can effectively cover and collect full-band sound wave signals during the entire test process, avoiding frequency band loss or distortion during signal acquisition. Through calibration, the full-band sound wave signal acquisition configuration is obtained. Based on the performance characteristics of the high-speed spectrometer and laser scattering device in the optical signal acquisition device, the focal length and sampling frequency are adjusted to capture the changes in the optical signal in the welding area. By adjusting the focal length of the optical system, it is ensured that the spectrometer and laser scattering device can be accurately aligned with the welding area, and by optimizing the sampling frequency, it can timely capture the slight changes in the optical signal during the welding process, including material surface temperature fluctuations, changes in optical properties during crack initiation and expansion, etc. At the same time, the sampling rate and measurement range of the electrical signal detection device are optimized to ensure that it can accurately record the fluctuations of the welding current and voltage during the welding process. According to the characteristics of the welding current and voltage, an appropriate sampling rate and measurement range are set to ensure that the electrical signal detection device can accurately record the fluctuations of the welding current and voltage during the heating and stress of the specimen, thereby reflecting the changes in the conductive properties and internal structure evolution of the material under high temperature environment.

[0028] S2, heating the welding material specimen based on the heating device, and applying dynamic stress through the loading device, so as to obtain the specimen in a high temperature stress state;

[0029] Specifically, the power of the heating device is regulated according to the heating curve to achieve continuous change control of the temperature of the welding material specimen. This process takes into account the size of the heating power, and flexibly adjusts the heating rate and the holding time according to the actual test requirements to ensure that the specimen can experience a stable and controllable high temperature state during the experiment. In this process, the infrared thermal imager is used to scan the temperature distribution of the specimen surface in real time, effectively monitor the temperature change of the specimen surface, and generate a two-dimensional or three-dimensional data matrix reflecting the temperature distribution of the specimen surface. Based on the temperature distribution data matrix, the temperature gradient and uniformity index are analyzed by calculating the temperature difference between each point. The temperature gradient can reflect the temperature difference distribution between different areas on the specimen surface, while the uniformity index can quantify the uniformity of the temperature distribution on the specimen surface. Taking these indicators into consideration, the heating uniformity evaluation result is obtained to help determine whether the current heating process meets the experimental requirements and whether there are problems of local overheating or uneven heating. If the heating uniformity is poor, the heating uniformity evaluation result is input into the PID controller to adjust the power of the heating device in a closed loop, dynamically adjust the distribution and control strategy of the heating power, and achieve uniformity and stability of the temperature of the entire specimen. According to the dynamic stress loading scheme, the control signal of the loading device is generated. The loading control signal contains key parameters such as stress amplitude, loading frequency and phase information. At the same time, the strain and force conditions of the specimen are collected through high-precision strain gauges and force sensors installed on the surface of the specimen to generate stress-strain response data, reflecting the mechanical behavior of the specimen under high-temperature dynamic stress. The stress-strain response data is subjected to Fourier transform and harmonic analysis to obtain the stress-strain response characteristics in the frequency domain, and compared with the preset theoretical loading scheme to obtain the deviation between the actual loading state and the theoretical loading scheme. The deviation includes slight deviations in stress amplitude, frequency and phase, reflecting the influence of the response lag or nonlinear factors of the loading device on the loading effect. The deviation data is input into the adaptive control algorithm, and the loading control sequence is corrected in real time in the next cycle. The difference between the actual loading state and the theoretical loading scheme is reduced by adjusting the control signal of the loading device to ensure that the specimen is always in an ideal high-temperature stress state.

[0030] S3, performing multi-source signal acquisition on the test piece in a high temperature stress state based on an acoustic signal acquisition device, an optical signal acquisition device and an electrical signal detection device to obtain time-synchronized acoustic, optical and electrical coordinated signal data;

[0031] It should be noted that, through the multiple microphones with different frequency response ranges in the acoustic signal acquisition device, the test piece is subjected to all-round acoustic wave acquisition, and the acoustic signal characteristics of the test piece in different frequency bands during the crack evolution process are captured, and the original acoustic signal data of multiple channels are obtained. The original acoustic signal data of multiple channels is subjected to time synchronization and channel fusion processing. By introducing a high-precision time reference, the acoustic signal data of each channel is accurately aligned and corrected, and the time error caused by the response time difference between different microphones is eliminated, so as to obtain a comprehensive acoustic signal under a unified time reference. Based on the high-speed spectrometer in the optical signal acquisition device, the surface of the test piece is spectrally scanned to obtain a sequence of spectral data that changes with time. The high-speed spectrometer can quickly acquire the optical signals reflected or scattered on the surface of the test piece at different wavelengths, and record the changes of the spectral characteristics caused by the microstructure changes on the surface of the welding material under high temperature and dynamic stress. In order to obtain the change information of the surface morphology of the test piece, the laser scattering device in the optical signal acquisition device is used to scan the surface of the test piece in real time. The laser scattering device irradiates the surface of the specimen with a laser beam and receives the scattered light signal to detect the changes in the microscopic morphology of the surface of the specimen and generate time series data reflecting the changes in the surface morphology. The spectral data sequence obtained by the high-speed spectrometer and the surface morphology change data obtained by the laser scattering device are time-aligned and data-fused to obtain a comprehensive optical signal feature that can comprehensively reflect the surface state of the specimen. At the same time, the electrical signal detection device performs high-frequency sampling on the current and voltage signals of the specimen during the welding process to obtain the original current and voltage waveform data. Since the current and voltage fluctuations in the welding process are often closely related to the changes in the conductivity, heat conduction process and evolution of the internal stress field of the specimen, the original current and voltage waveform data are subjected to power spectrum analysis and phase calculation to extract the electrical signal characteristic parameters reflecting the changes in the internal state of the specimen, including the spectrum distribution, phase shift and harmonic components of the current and voltage. The comprehensive acoustic signal, comprehensive optical signal characteristics and electrical signal characteristic parameters are time-stamp aligned and the data format is unified. Through the timestamp alignment technology, we can ensure that all kinds of signal data are synchronized under the same time reference, and on this basis, we can unify the data format and ensure that the characteristic parameters of all kinds of signals are effectively integrated and compared. We can obtain a time-synchronized data set of acoustic, optical, and electrical coordinated signals.

[0032] S4, performing noise reduction filtering and time-frequency domain feature extraction on the acoustic, optical and electrical coordinated signal data, and then inputting a preset feature classification model to obtain multimodal classification feature data;

[0033] Specifically, the acoustic, optical, and electrical coordinated signal data are denoised by wavelet threshold. The signal is decomposed into the wavelet domain to remove noise, and then the denoised signal data is reconstructed to retain effective information and eliminate noise interference. Based on the denoised signal data, adaptive Wiener filtering is used for optimization processing. By dynamically adjusting the filter parameters, the influence of residual noise is reduced to obtain purer signal data. The optimized coordinated signal data is subjected to short-time Fourier transform to convert the signal from the time domain to the time-frequency domain to obtain a time-frequency spectrum, which intuitively displays the changes in the frequency components of the signal at different time points. Based on the time-frequency spectrum, time-frequency domain features such as energy entropy, spectrum centroid, and band power ratio are calculated to form a feature vector. Energy entropy describes the uniformity of signal energy distribution, spectrum centroid indicates the center of signal frequency distribution, and band power ratio reflects the energy distribution of the signal in different frequency bands. Dimensionality reduction methods such as principal component analysis or linear discriminant analysis are used to convert the original high-dimensional feature vector into a low-dimensional feature set to reduce data redundancy and improve analysis efficiency. Based on the feature set after dimensionality reduction, support vector machine is used for classification to obtain preliminary classification results. The support vector machine divides the data into different categories by finding the optimal hyperplane. Based on the preliminary classification results, a fuzzy membership function is constructed to assign a fuzzy grading index to each feature. The fuzzy grading index is input into the hierarchical analysis process model, and the weights and relative importance of each feature are combined to obtain multimodal graded feature data.

[0034] In this embodiment, the preliminary classification results are statistically analyzed to analyze the distribution of each feature in different categories, including calculating the frequency, mean, variance and other statistical indicators of each feature in each category, and obtaining the typical manifestation of the feature in different categories. Based on the distribution, the fuzzy set boundary of each feature is calculated to determine the boundary area between the features in each category. The fuzzy set boundary is expressed as the upper and lower limit intervals of the feature value, that is, the fuzzy interval, which reflects the transitional nature and uncertainty of the feature between different categories. In this way, the fuzziness of each feature between the categories is preliminarily described. Based on the fuzzy interval, the Gaussian function is used to fit it to obtain the membership function of each feature. Due to its smooth curve shape, the Gaussian function can well describe the changing trend of the feature in the fuzzy interval, and map each feature value in the fuzzy interval to the membership value between [0, 1], and quantify the degree of belonging of the feature to different categories. The membership function is normalized to eliminate the dimensional differences between different features and obtain a standardized fuzzy membership function. Based on the standardized fuzzy membership function, the fuzzy classification index of each feature is calculated to form an initial fuzzy classification matrix. The matrix contains the fuzzy membership of all features in different categories, reflecting the classification of each feature in each category. The judgment matrix of the hierarchical analysis process is constructed, and each indicator in the initial fuzzy grading matrix is ​​compared pairwise. According to the importance of the feature or the degree of influence on the target, a relative score is given to obtain the target relationship matrix. The elements in the target relationship matrix represent the quantitative relationship between the relative importance of each feature. Based on the target relationship matrix, the eigenvalue method is used to calculate its maximum eigenvalue and the corresponding eigenvector to obtain the weight vector of each indicator. The maximum eigenvalue reflects the consistency of the target relationship matrix, while the eigenvector represents the relative importance weight of each feature relative to the target. The preliminary weight vector is checked for consistency to ensure that the construction of the judgment matrix is ​​reasonable and logically consistent. If the consistency test passes, the weight vector is retained; if not, the judgment matrix is ​​modified, and each indicator is compared pairwise again, and the values ​​in the matrix are adjusted. The above calculation process is repeated until the consistency test passes and a weight vector that meets the consistency requirements is obtained. The final weight vector is weighted summed with the initial fuzzy grading matrix to calculate the comprehensive score vector. Based on the comprehensive scoring vector, the features are sorted and clustered to identify the importance and similarity of the features and obtain multimodal hierarchical feature data.

[0035] S5, input the multimodal hierarchical feature data into the pre-trained sequence to sequence neural network model to obtain the crack resistance performance prediction results including crack initiation time, propagation rate and fracture toughness.

[0036] Among them, the multimodal hierarchical feature data is processed by sliding window and divided into several overlapping input sequence sets. The window size of each input sequence is set to 100 time steps, the step length is 10 time steps, and there is an overlap of 90 time steps between two adjacent input sequences. The sliding window processing method ensures that there is a high degree of continuity between the input sequences, and at the same time captures the dynamic information of the characteristics of the welding material under high temperature stress state. The input sequence set is input into an encoder containing a three-layer bidirectional long short-term memory network (Bi-LSTM). The number of units in each layer of LSTM is set to 256, and the activation function uses the tanh function. The bidirectional LSTM can consider the forward and backward information of the input sequence at the same time, that is, it not only pays attention to the feature changes before the current time step, but also pays attention to the subsequent change trends, effectively capturing the temporal relationship of the multimodal features of the welding material under high temperature dynamic stress. The encoded feature sequence is obtained through the encoding processing of the three-layer bidirectional LSTM. Based on the encoded feature sequence, the attention weight is calculated using the multi-head attention mechanism. The multi-head attention mechanism divides the feature sequence into 8 different attention heads, each with an attention dimension of 64, and performs weighted summation on the feature sequence from different angles to capture the dependencies between each time step in the feature sequence. The important influence of key time nodes such as crack initiation and expansion of welding materials on the overall prediction is highlighted. The weighted context vector calculated by the multi-head attention mechanism effectively aggregates the important information in the original feature sequence. The weighted context vector is input into the decoder containing 2 layers of LSTM, each layer has 512 LSTM units, and the activation function is ReLU. The decoder generates the initial prediction sequence by gradually parsing the context vector. Compared with the tanh activation function of the encoder, ReLU maintains the stability of the gradient in a deeper network structure, avoids the vanishing or exploding gradient, and enhances the decoder's decoding ability of context information. Residual connection is applied to the initial prediction sequence, and layer normalization is performed. The residual connection can reduce the gradient vanishing problem caused by the deep network during the training process by adding the mapping of the original input sequence to the initial prediction sequence, while the layer normalization can ensure that the output distribution of each layer of neurons is within a stable range by standardizing the output of each layer, thereby improving the convergence speed and stability of the network. The normalization parameter is set to a learnable affine transformation so that the model can adaptively adjust the normalization effect to obtain the optimized prediction sequence. The optimized prediction sequence is passed through three parallel fully connected layers, with 64, 32 and 16 neurons in each fully connected layer, and the activation function is Sigmoid. The decreasing number of neurons layer by layer can extract information at different levels and gradually aggregate to the target output. Finally, the original prediction values ​​of crack initiation time, propagation rate and fracture toughness are obtained. In order to ensure the smoothness and stability of the prediction results, the original prediction values ​​are smoothed by exponential moving average, and the smoothing coefficient is set to 0.9.Exponential moving average assigns different weights to time series data, making the most recent data more important, which can better smooth short-term fluctuations and obtain more stable prediction results. The stable prediction results are compared with the preset material performance threshold to determine whether the crack resistance of the welding material meets the requirements, help identify the critical state of the welding material under high temperature stress, determine the time point of crack initiation, the speed of expansion rate and the strength of fracture toughness, and obtain the crack resistance prediction results.

[0037] In one example, a welding material specimen is mounted on a loading device in a test chamber, and parameters of a heating device, an acoustic signal acquisition device, an optical signal acquisition device, and an electrical signal detection device are set to obtain an initialized test system, including:

[0038] The welding material specimen is mounted on the fixture in the test chamber to obtain a fixed and evenly stressed specimen assembly, and the temperature curve of the heating device is programmed to obtain a temperature rise process that simulates the actual welding environment;

[0039] The stress parameters of the loading device are set based on the mechanical properties of the welding material to obtain a dynamic stress loading scheme including stress amplitude, frequency and loading mode, and multiple microphones with different frequency response ranges in the sound signal acquisition device are calibrated to obtain a full-band sound wave signal acquisition configuration;

[0040] Based on the high-speed spectrometer and laser scattering device of the optical signal acquisition device, the focal length and sampling frequency are adjusted to obtain the acquisition settings that can accurately capture the changes in optical signals in the welding area. The sampling rate and measurement range of the electrical signal detection device are optimized to obtain detection parameters that can accurately record welding current and voltage fluctuations, thereby obtaining an initialized test system.

[0041] In this example, the welding material specimen is mounted on a fixture in the test chamber to ensure that the specimen remains fixed and evenly stressed throughout the test. The temperature curve of the heating device is programmed to simulate the temperature rise process in the actual welding environment. In order to accurately simulate the temperature changes experienced by the specimen during the welding process, a reasonable temperature curve is designed according to the requirements of the welding process. Assume that the relationship between temperature and time is:

[0042] ;

[0043] in, Indicates time The temperature of the moment, is the initial temperature, usually the ambient temperature, is the coefficient of the temperature rise rate, It is a nonlinear index of the heating process, reflecting the rate of temperature rise. According to the thermal conductivity characteristics of different materials and welding process requirements, and Can be adjusted to achieve a linear or nonlinear temperature rise curve. For example, in the simulation of welding some high-strength steels, it is necessary to set a larger Values ​​(such as 2 or 3) are selected to better simulate the thermal shock effect during the rapid rise in high temperature stage. The programming of the temperature curve uses a computer control system to accurately adjust the power of the heating device, monitor and feedback temperature changes in real time, so that the temperature distribution on the surface and inside of the specimen is consistent with the actual welding process, and more realistically simulates the welding environment. After completing the setting of the heating device, the stress parameters of the loading device are set based on the mechanical properties of the welding material. The core function of the loading device is to apply dynamic stress so that the specimen is subjected to a complex mechanical stress environment while being heated. The core of stress parameter setting is to determine the stress amplitude, frequency and loading mode. Stress amplitude It is determined according to the yield strength of the material and the actual working conditions to ensure that the applied stress does not exceed the yield limit of the material. The loading mode is uniaxial tension, compression or multi-axis alternating loading, so as to more comprehensively study the crack initiation and expansion behavior of welding materials under different stress states. At the same time, in order to ensure the accuracy and comprehensiveness of acoustic signal acquisition, multiple microphones with different frequency response ranges in the acoustic signal acquisition device are calibrated. The frequency response range of each microphone is different. During the calibration process, a standard sound source is used to emit a signal of known sound pressure level at different frequencies, and the output signal strength of each microphone is recorded. The frequency response is corrected according to the deviation between the actual output and the theoretical value to obtain the full-band acoustic signal acquisition configuration. Different microphones work together to cover the wide frequency band of acoustic signal acquisition. The focal length and sampling frequency of the optical signal acquisition device are adjusted. The high-speed spectrometer is used to monitor the spectral changes on the surface of the welding material. Its focal length adjustment must ensure that the field of view of the spectrometer completely overlaps with the welding area, so as to accurately capture the changes in its surface optical properties (such as reflectivity, luminous intensity, etc.) when the material is heated. At the same time, the laser scattering device is used to monitor the changes in the microscopic morphology of the material surface. The scattering angle of the laser beam and the position of the receiver need to be precisely calibrated to avoid measurement errors. The sampling frequency is set according to the dynamic characteristics of the material surface changes. Usually a higher frequency (for example, above 1kHz) is selected to capture rapid optical change signals. The sampling rate and measurement range of the electrical signal detection device are optimized. The current and voltage signals in the welding process usually change dramatically, and the sampling rate needs to be high enough to capture the details of the rapid fluctuations.

[0044] In one example, a welding material specimen is heated by a heating device and a dynamic stress is applied by a loading device to obtain a specimen in a high temperature stress state, including:

[0045] The power of the heating device is adjusted according to the heating curve to obtain the continuous change process of the test piece temperature, and the temperature distribution of the test piece surface is scanned in real time based on the infrared thermal imager to obtain the temperature distribution data matrix;

[0046] The temperature gradient and uniformity index are calculated based on the temperature distribution data matrix to obtain the heating uniformity evaluation result, and the heating uniformity evaluation result is input into the PID controller to perform closed-loop regulation on the power of the heating device to obtain a test piece with uniform temperature;

[0047] According to the dynamic stress loading scheme, the control signal of the loading device is generated to obtain the loading control sequence containing stress amplitude, frequency and phase information. The strain and force conditions of the specimen are collected in real time based on high-precision strain gauges and force sensors to obtain stress-strain response data.

[0048] Based on the stress-strain response data, Fourier transform and harmonic analysis are performed to obtain the deviation between the actual loading state and the theoretical loading scheme. The deviation data is input into the adaptive control algorithm to make real-time corrections to the loading control sequence to obtain the specimen in a high-temperature stress state.

[0049] In this example, the power of the heating device is controlled to ensure that the temperature of the specimen can change continuously according to the preset heating curve. The heating device is composed of a resistance heater or an induction heater, and the control system dynamically adjusts the heating power according to the real-time temperature feedback provided by the temperature sensor (such as a thermocouple or an infrared thermal imager). Assume that the heating curve is expressed as:

[0050] ;

[0051] in, It's time The target temperature at the moment, is the initial temperature, is the coefficient of temperature rise, controlling the maximum increment of the target temperature, is the time adjustment coefficient, which reflects the rate at which temperature changes over time. and The power regulation of the heating device is achieved through a PID controller (proportional-integral-differential controller), and its control goal is to make the temperature of the specimen Real-time tracking of target temperature The control formula of the PID controller is:

[0052] ;

[0053] in, For the heating device at time The output power, is the temperature error, , and They are the proportional, integral and differential coefficients, respectively, which determine the response characteristics of the PID controller. Able to respond quickly to current error, integral term Used to eliminate steady-state errors, differential terms Then it reacts to the rate of change of the error to prevent excessive oscillation. In practical applications, by adjusting these three parameters, it is ensured that the heating power accurately follows the heating curve and maintains the temperature stability of the specimen. During the heating process, in order to monitor the temperature distribution on the specimen surface in real time, an infrared thermal imager is used to continuously scan the specimen surface to generate a temperature distribution data matrix. Each element of the matrix represents the temperature value of a small area on the specimen surface. Assume that the resolution of the infrared thermal imager is , then the temperature distribution data matrix is ​​expressed as:

[0054] ;

[0055] in, It is Row, No. The temperature gradient and uniformity index of the specimen surface are calculated based on the temperature distribution data matrix. The calculation formula is:

[0056] ;

[0057] in, and Respectively indicate the temperature and The uniformity index is expressed by calculating the variance of the temperature of each area from the average temperature. Assuming the average temperature is , uniformity index It can be expressed as:

[0058] ;

[0059] Uniformity Index The smaller it is, the more uniform the temperature distribution on the specimen surface is. The temperature gradient and uniformity index are input into the PID controller to perform closed-loop regulation on the power of the heating device. If the uniformity index exceeds the set threshold, the PID controller will appropriately adjust the distribution of the heating power, optimize the temperature field of the specimen, make the surface temperature of the specimen more uniform, reduce local overheating or cooling, and obtain a specimen with uniform temperature. At the same time, according to the dynamic stress loading scheme, the control signal of the loading device is generated. The control signal of the loading device includes the stress amplitude ,frequency and phase information The basic form of the loading signal is expressed as:

[0060] ;

[0061] in, For time The instantaneous stress at is the maximum stress amplitude, is the loading frequency, is the loading phase. According to the mechanical properties of the welding material and the experimental requirements, reasonable stress amplitude and frequency are set. For example, a smaller stress amplitude and a lower frequency are selected for brittle materials to avoid premature failure, while a larger stress amplitude and a higher frequency can be selected for ductile materials to simulate the performance under high stress conditions. Phase information is used to control the synchronization of multi-axis loading, ensure the consistency of stress phases in each loading direction, and better simulate the complex stress state in actual working conditions. In order to monitor the true stress-strain response of the test piece under dynamic stress loading, high-precision strain gauges and force sensors are used to collect the strain and force conditions of the test piece in real time. The strain gauge is used to measure the tiny deformation of the test piece under stress. Its output signal is linearly related to the strain of the test piece surface, while the force sensor can accurately record the external force applied by the loading device. The stress and strain data are transmitted to the data acquisition system in real time. Generate stress-strain response data to form a time series matrix:

[0062] ;

[0063] in, and Respectively represent The stress and strain values ​​at each time step. Based on these data, Fourier transform and harmonic analysis are performed to obtain the deviation between the actual loading state and the theoretical loading scheme. Fourier transform is used to convert stress-strain data in the time domain to the frequency domain. By analyzing the amplitude and phase of each frequency component, it is identified whether there are frequency components inconsistent with the theoretical loading scheme during the loading process. Assume that the frequency domain after Fourier transform is represented as:

[0064] ;

[0065] in, is the angular frequency, is an imaginary unit. The Fourier transform can reveal whether there are harmonic components or noise interference in the stress-strain response, which will cause deviations between the actual loading state and the theoretical loading scheme. The deviation data is input into the adaptive control algorithm, such as model predictive control (MPC) or incremental PID control, to make real-time corrections to the loading control sequence to reduce the deviation, make the loading state more in line with the experimental design requirements, and obtain the specimen in the high temperature stress state.

[0066] In one example, a multi-source signal acquisition is performed on a specimen in a high temperature stress state based on an acoustic signal acquisition device, an optical signal acquisition device, and an electrical signal detection device to obtain time-synchronized acoustic, optical, and electrical coordinated signal data, including:

[0067] Based on multiple microphones with different frequency response ranges in the sound signal acquisition device, the test piece is omni-directionally sound wave collected to obtain multi-channel original sound signal data;

[0068] The multi-channel original acoustic signal data is subjected to time synchronization and channel fusion processing to obtain a comprehensive acoustic signal with a unified time reference, and the high-speed spectrometer based on the optical signal acquisition device performs spectral scanning on the surface of the specimen to obtain a spectral data sequence that changes with time;

[0069] The laser scattering device based on the optical signal acquisition device scans the surface of the specimen in real time to obtain surface morphology change data, and performs time alignment and data fusion on the spectral data sequence and the surface morphology change data to obtain the comprehensive optical signal characteristics;

[0070] Based on the electrical signal detection device, high-frequency sampling is performed on the welding current and voltage to obtain the original current and voltage waveform data, and power spectrum analysis and phase calculation are performed on the original current and voltage waveform data to obtain electrical signal characteristic parameters;

[0071] The timestamps of the integrated acoustic signal, integrated optical signal features and electrical signal feature parameters are aligned and the data format is unified to obtain a time-synchronized acoustic, optical and electrical collaborative signal data set.

[0072] In this example, multiple microphones with different frequency response ranges in the acoustic signal acquisition device are used to collect all-round acoustic waves from the specimen to obtain multi-channel raw acoustic signal data. During the welding process, crack initiation, expansion, and fracture inside the material are accompanied by the generation of acoustic emission signals. These signals have obvious frequency characteristics and cover a wide frequency range from low frequency to high frequency. Assume that the acoustic signal acquisition system consists of three microphones, covering the response ranges of low frequency (20 Hz to 2 kHz), medium frequency (1 kHz to 10 kHz), and high frequency (5 kHz to 50 kHz), respectively. For each time step , get the original sound signal data of three channels:

[0073] ;

[0074] in, , and Respectively represent the low frequency, medium frequency and high frequency microphones at time The output voltage signal of the welding material is obtained by multi-channel acoustic signal acquisition. The acoustic emission characteristics of welding materials in different frequency bands are fully recorded. The original acoustic signal data of multiple channels are synchronized and channel fused to obtain a comprehensive acoustic signal with a unified time reference. Assume that the acoustic signal sampling rate of each channel is , and , the corresponding sampling time step is , and By introducing a unified time base And data alignment algorithm, resample the acoustic signals with different sampling rates to the same time base to obtain the integrated acoustic signal with unified time base Channel fusion integrates the acoustic signal features of different channels through weighted summation. At the same time, the high-speed spectrometer based on the optical signal acquisition device performs spectral scanning on the surface of the specimen to obtain a spectral data sequence that changes over time. The high-speed spectrometer can measure the changes in spectral intensity of the specimen surface at different wavelengths with a time resolution of microseconds, reflecting the changes in the microstructure and chemical composition of the material under high temperature stress conditions. Assume that at a certain time step , the spectral data is represented as a vector:

[0075] ;

[0076] in, Indicates time Time in wavelength The spectral intensity under is the number of wavelength sampling points of the spectrometer. The spectral data of each time step are arranged in chronological order to obtain a complete spectral data sequence, which describes the dynamic evolution of the optical properties of the material surface during welding. The laser scattering device in the optical signal acquisition device is used to scan the surface of the specimen in real time to obtain the surface morphology change data. The laser scattering device emits a beam of laser to the surface of the specimen and receives the scattered light signal to detect the roughness, microcracks and other morphological changes on the surface of the specimen. Assume that the laser scattering signal is expressed as:

[0077] ;

[0078] in, Indicates time The laser scattering signal at time The intensity at each detection angle, is the angular resolution of the laser scattering detector. By comparing the changes in the intensity of the scattering signal at different time steps, the microscopic changes on the surface of the specimen during the welding process can be identified. The spectral data sequence and the surface morphology change data are time-aligned and fused. Time alignment is based on the respective data sampling timestamps, and the spectral data and laser scattering data of different sampling frequencies are mapped to a unified time reference. Data fusion uses feature fusion methods, such as weighted summation of the main components of the spectral data and scattering data, or splicing after extracting high-dimensional features through neural networks, to obtain comprehensive optical signal features that can simultaneously reflect the spectral changes and morphological changes on the surface of the specimen. At the same time, based on the electrical signal detection device, the current and voltage in the welding process are sampled at high frequency to obtain the original current and voltage waveform data. The changes in welding current and voltage are closely related to the electrical conductivity, molten pool state and thermal conductivity characteristics inside the material. Assume that the current signal and voltage signal are expressed as:

[0079] ;

[0080] ;

[0081] in, and The time The current and voltage at the moment, and are the amplitudes of current and voltage respectively, are the frequencies of the current and voltage, and is the phase angle of current and voltage. In order to analyze the characteristics of electrical signals, power spectrum analysis and phase calculation are performed on the original current and voltage waveform data. Power spectrum analysis is used to identify the main frequency components and their energy distribution in current and voltage, revealing the periodic characteristics of electrical signals during welding. Phase calculation can reflect the phase difference between current and voltage, thereby judging the conductivity characteristics and internal state changes of the material. The characteristic parameters of the electrical signals obtained by power spectrum analysis and phase calculation are time-stamp aligned and the data format is unified with the characteristics of the integrated acoustic signal and the integrated optical signal. Timestamp alignment is based on the interpolation or resampling technology of the data sampling time, mapping the multimodal signal data of different frequencies to the same time axis. Data format unification converts data of different modes into the same format through normalization and data structure transformation, such as converting all data into a standardized two-dimensional matrix or multi-dimensional tensor. A time-synchronized acoustic, optical, and electrical synergistic signal data set is obtained, which contains the multimodal response characteristics of the specimen under high temperature stress state.

[0082] In one example, the acoustic, optical, and electrical coordinated signal data are subjected to noise reduction filtering and time-frequency domain feature extraction, and then input into a preset feature classification model to obtain multimodal hierarchical feature data, including:

[0083] Perform wavelet threshold denoising on the acoustic, optical and electrical coordinated signal data to obtain the denoised signal data, and perform adaptive Wiener filtering based on the denoised signal data to obtain the coordinated signal after filtering optimization;

[0084] Performing short-time Fourier transform on the filtered and optimized cooperative signal to obtain a time-frequency spectrum, and calculating energy entropy, spectrum centroid and frequency band power ratio based on the time-frequency spectrum to obtain a feature vector in the time-frequency domain;

[0085] Perform dimension reduction processing on the time-frequency domain feature vector to obtain a feature set after dimension reduction, and perform support vector machine classification on the feature set after dimension reduction to obtain a preliminary classification result;

[0086] Based on the preliminary classification results, a fuzzy membership function is constructed to obtain the fuzzy classification index of the feature, and the fuzzy classification index is input into the hierarchical analysis process model to obtain multimodal classification feature data.

[0087] In this example, wavelet threshold denoising is performed on the acoustic, optical and electrical coordinated signal data. The multimodal signals generated by welding materials under high temperature stress conditions contain noise, such as environmental noise, electromagnetic interference and the noise of the instrument itself. Wavelet threshold denoising decomposes the signal into wavelet bases of different scales and suppresses the noise in the wavelet domain to restore the true characteristics of the signal. Select a suitable wavelet basis function (such as Daubechies wavelet) and the number of decomposition layers to convert the original signal into Decomposed into a set of wavelet coefficients of different frequencies and scales ,in represents the number of decomposition levels, Represents the index of the wavelet coefficient. For each wavelet coefficient , the threshold method is used for denoising. Common threshold methods include hard threshold and soft threshold. The hard threshold method directly sets the coefficients less than the threshold to zero, while the soft threshold method proportionally reduces the coefficients less than the threshold. After denoising, the signal is reconstructed using the retained effective wavelet coefficients to obtain the denoised signal data . Perform adaptive Wiener filtering on the denoised signal data. Wiener filtering is a filtering method based on the minimum mean square error criterion, which can effectively suppress random noise while retaining the detailed characteristics of the signal. It uses the statistical characteristics of the signal and noise to adjust the frequency response of the filter so that the filter can best retain the signal components and maximize the suppression of the noise components. Adaptive Wiener filtering dynamically adjusts the filter parameters according to the local statistical characteristics of the signal (such as the local mean and variance) to improve the denoising effect. Assume that the denoised signal The expected value of , the variance is , then the output of the Wiener filter is It is expressed as:

[0088] ;

[0089] in, is the variance of the noise, which represents the estimated value of the noise intensity. Through the above processing, the collaborative signal after filtering optimization is obtained . Perform short-time Fourier transform on the optimized cooperative signal to convert the signal from the time domain to the time-frequency domain. Short-time Fourier transform can obtain the spectrum information of the signal at different time points by performing Fourier transform on the signal within the time window. Given a length of Sliding window function , the calculation formula of short-time Fourier transform is:

[0090] ;

[0091] in, is the cooperative signal after filtering optimization, is a frequency variable, is an imaginary unit. The time-frequency spectrum of the signal is obtained by short-time Fourier transform, showing the frequency distribution of the signal at different time points. Based on the time-frequency spectrum, the time-frequency domain features such as energy entropy, spectrum centroid and band power ratio are calculated. Energy entropy is an indicator to measure the uniformity of signal energy distribution. Suppose the frequency energy distribution of the signal in a certain time window is , then the energy entropy The calculation formula is:

[0092] ;

[0093] in, Indicates frequency The energy density at the location. Energy entropy reflects the complexity of the signal frequency energy distribution. The higher the entropy value, the more uniform the energy distribution and the more complex the signal. The spectrum centroid is an indicator that reflects the center of the signal frequency distribution. Its calculation formula is:

[0094] ;

[0095] in, Indicates frequency, For frequency The spectrum centroid indicates the center position of the signal spectrum distribution. The higher the centroid position is, the more high-frequency components there are in the signal. The band power ratio is an indicator used to measure the difference in energy distribution of the signal in different frequency bands. Assume that the frequency range of the low frequency band is , the frequency range of the high frequency band is , then the calculation formula of the frequency band power ratio is:

[0096] ;

[0097] The frequency band power ratio reflects the energy ratio of the signal in different frequency bands and is an important parameter for distinguishing the frequency characteristics of the signal. After calculating the time-frequency domain features, the time-frequency domain feature vector is obtained. The time-frequency domain feature vector is subjected to dimensionality reduction processing. For example, the covariance matrix of the feature vector is calculated through principal component analysis:

[0098] ;

[0099] in, Indicates The feature vector of the samples, represents the mean of the eigenvector, is the number of samples. Then the covariance matrix Perform eigenvalue decomposition, select the eigenvector corresponding to the principal component for dimensionality reduction, and obtain the feature set after dimensionality reduction Based on the feature set after dimensionality reduction Perform support vector machine classification. Support vector machine is a binary classification model that separates data samples of different categories by finding an optimal hyperplane. The objective function of the support vector machine is expressed as:

[0100] ;

[0101] subject to , , ;

[0102] in, is the normal vector of the hyperplane, is the bias, is the slack variable, Indicates The label of each sample (1 or -1), is the penalty coefficient, which indicates the degree of penalty for misclassified samples. By optimizing the objective function, the preliminary classification results of the support vector machine are obtained. Based on the preliminary classification results, a fuzzy membership function is constructed. The fuzzy membership function reflects the fuzziness of the feature by quantifying the membership of samples belonging to different categories. Suppose the membership function of the sample is , indicating the The samples belong to The construction of membership function of each category is usually based on Gaussian distribution or trigonometric function distribution. The fuzzy classification index of each feature is calculated according to the membership function. The fuzzy classification index is input into the hierarchical analysis process model, and the multimodal classification feature data is obtained through pairwise comparison of judgment matrix and eigenvalue decomposition.

[0103] In one example, a fuzzy membership function is constructed based on the preliminary classification results to obtain a fuzzy classification index of the feature, and the fuzzy classification index is input into the hierarchical analysis process model to obtain multimodal hierarchical feature data, including:

[0104] Perform statistical analysis on the preliminary classification results to obtain the distribution of each feature in different categories, and calculate the fuzzy set boundary of each feature based on the distribution to obtain the fuzzy interval of the feature;

[0105] The fuzzy interval is fitted based on the Gaussian function to obtain the membership function of each feature, and the membership function is normalized to obtain a standardized fuzzy membership function;

[0106] The fuzzy classification index of each feature is calculated based on the standardized fuzzy membership function to obtain the initial fuzzy classification matrix, and the judgment matrix of the hierarchical analysis process is constructed. The indicators in the initial fuzzy classification matrix are compared pairwise to obtain the target relationship matrix.

[0107] Based on the eigenvalue method, the maximum eigenvalue and corresponding eigenvector of the target relationship matrix are calculated to obtain the weight vector of each indicator;

[0108] Perform consistency check on the weight vector. If it passes, it is retained. If it fails, the judgment matrix is ​​modified. Repeat the calculation until it passes the consistency check to obtain the final weight vector.

[0109] The final weight vector is weighted and summed with the initial fuzzy grading matrix to obtain a comprehensive scoring vector, and the features are sorted and clustered based on the comprehensive scoring vector to obtain multimodal graded feature data.

[0110] In this example, the distribution of each feature in different categories is extracted from the preliminary classification results. The preliminary classification results come from the prediction output of the support vector machine or other machine learning model, including the label of the category to which each feature vector belongs. Assume that samples, each with a feature vector Indicates that , is the number of features. The classification result is , indicating the sample Belongs to the category label (a total of categories). Perform statistical analysis on the classification results. , calculate the mean in each category and standard deviation ,in Indicates the category. Mean Reflects the characteristics In category The central tendency in It reflects the distribution breadth of the feature in this category. The formula is as follows:

[0111] ;

[0112] ;

[0113] in, Yes Category The number of samples in is the indicator function, when 1 when , otherwise 0. Based on the mean and standard deviation, a Gaussian distribution model is constructed for each feature in different categories. In order to quantify the fuzzy boundaries between features in different categories, the fuzzy set boundary of each feature is calculated based on the feature distribution. The fuzzy set boundary reflects the transition interval between the features in different categories. , defined in the category The fuzzy interval in is:

[0114] ;

[0115] in, and Respectively represent characteristics In category The minimum and maximum fuzzy boundaries in is an expansion factor, usually 2 or 3, to cover the distribution of most samples. In this way, a fuzzy interval is defined for each feature and each category to reflect the uncertainty and ambiguity of the feature in each category. Based on the fuzzy interval, the Gaussian function is used to fit it to obtain the membership function of each feature. Membership function It is expressed as:

[0116] ;

[0117] This function is used to quantify the eigenvalues Belongs to category The membership degree of is in the range of [0, 1]. The larger the membership value, the more likely the feature value belongs to the category. , and obtain a standardized fuzzy membership function:

[0118] ;

[0119] Through standardization, it is ensured that the sum of the membership of each feature in all categories is 1, that is, the distribution of each feature value between different categories is complete. Based on the standardized fuzzy membership function, the fuzzy classification index of each feature is calculated. Fuzzy classification index Representation characteristics In category The importance of is usually calculated by weighted average method:

[0120] ;

[0121] in, It is a feature The weight reflects the importance of the feature value. Combining all features and all categories of fuzzy grading indicators, construct the initial fuzzy grading matrix:

[0122] ;

[0123] Each element in the matrix Reflects the characteristics In category The fuzzy grading value in . Construct the judgment matrix of the hierarchical analysis process. The judgment matrix is ​​used to quantify the relative importance of each feature, which is determined by pairwise comparison. Assume that The form of the judgment matrix is:

[0124] ;

[0125] in, Representation characteristics Relative to the feature The relative importance of the , use the eigenvalue method to calculate the maximum eigenvalue of the target relationship matrix and its corresponding eigenvector ,in Indicates The importance weight of each feature. The calculation formula for the eigenvalue problem is:

[0126] ;

[0127] For the weight vector Perform consistency test. The consistency ratio of the judgment matrix The calculation formula is:

[0128] ;

[0129] in, is the consistency indicator, is a random consistency indicator. Usually, when When , the judgment matrix is ​​considered to have satisfactory consistency, otherwise the judgment matrix needs to be adjusted Until the consistency check is passed. With the initial fuzzy classification matrix Perform weighted summation to obtain the comprehensive score vector :

[0130] ;

[0131] in, Each element of represents the comprehensive score of the feature in all categories. Based on the comprehensive score vector, the features are sorted and clustered to obtain multimodal hierarchical feature data. In this way, the most important features and their distribution characteristics in different categories are identified.

[0132] In one example, multimodal hierarchical feature data is input into a pre-trained sequence-to-sequence neural network model to obtain crack resistance prediction results including crack initiation time, propagation rate, and fracture toughness, including:

[0133] The multimodal hierarchical feature data is processed with a sliding window, with a window size of 100 time steps and a step length of 10, to obtain a set of overlapping input sequences;

[0134] Input the input sequence set into an encoder consisting of 3 layers of bidirectional LSTM, with 256 LSTM units in each layer and tanh as the activation function, to obtain the encoded feature sequence;

[0135] Based on the encoded feature sequence, the attention weight is calculated based on the multi-head attention mechanism, with 8 heads and 64 attention dimensions, to obtain the weighted context vector;

[0136] The weighted context vector is input into a decoder consisting of 2 layers of LSTM, each layer has 512 LSTM units and the activation function is ReLU to obtain the initial prediction sequence;

[0137] Apply residual connection to the initial prediction sequence and perform layer normalization. The normalization parameter is a learnable affine transformation to obtain the optimized prediction sequence.

[0138] The optimized prediction sequence is passed through three parallel fully connected layers, with 64, 32, and 16 neurons in each layer, and the activation function is Sigmoid, to obtain the original prediction values ​​of crack initiation time, propagation rate, and fracture toughness;

[0139] Exponential moving average smoothing was applied to the original prediction values ​​with a smoothing coefficient of 0.9 to obtain stable prediction results, which were then compared with the preset material property thresholds to obtain the crack resistance prediction results.

[0140] In this example, sliding window processing is performed on multimodal hierarchical feature data to split the time series data into multiple subsequences in order to capture the short-term dynamic change characteristics in the data. Multimodal feature time series , where each is a vector containing multi-dimensional features. Through sliding window processing, the time series data is divided into multiple subsequences of length 100, with a step length of 10, that is, a new subsequence is generated every 10 time steps. The resulting overlapping input sequence set is expressed as:

[0141] ;

[0142] Each input subsequence contains a multimodal feature vector of 100 time steps, and there is an overlap of 90 time steps between adjacent subsequences to ensure continuity and information sharing between each subsequence. The overlapping input sequence set is input into an encoder containing 3 layers of bidirectional LSTM. Each layer of LSTM contains 256 hidden units and the activation function is tanh. Bidirectional LSTM can consider the forward and backward information of time series data at the same time, which helps to capture the dependency of multimodal features on the entire time series. For each input subsequence After processing by 3 layers of bidirectional LSTM, an encoded feature sequence is obtained ,in Represents the time step The hidden state at a certain moment contains a comprehensive representation of the information before and after that time step. The hidden state of LSTM is represented as:

[0143] ;

[0144] in, is the weight matrix, is the bias vector, is the hidden state of the previous time step, is the input feature vector of the current time step, is the hyperbolic tangent activation function. Through the processing of bidirectional LSTM, the multimodal features of each subsequence are encoded into a feature sequence containing contextual information. Based on the encoded feature sequence , the attention weights are calculated using the multi-head attention mechanism to obtain the weighted contextual question. The multi-head attention mechanism focuses on important moments or feature patterns in the feature sequence from different perspectives. , mapping it to query, key and value vector spaces, represented as , where the dimension of each vector is 64, and the attention weight is calculated by the following formula:

[0145] ;

[0146] in, is the dimension of the key vector, which is usually used for scaling to avoid the inner product being too large. The attention weight matrix Indicates the relevance of the hidden state of each time step to other time steps. The multi-head attention mechanism concatenates the results of multiple different attention heads, weights and aggregates the features from multiple perspectives, and obtains the final weighted context vector . The weighted context vector As input, it is passed to the decoder containing 2 layers of LSTM. The number of LSTM units in each layer of the decoder is 512, and the activation function is ReLU. Unlike the encoder, the decoder is mainly used to decode the context vector into a specific output prediction. For each weighted context vector , the decoder gradually decodes through two layers of LSTM units to obtain the initial prediction sequence The prediction at each decoding time step is represented as:

[0147] ;

[0148] in, is the decoded hidden state at the current time step, is the weight matrix of the decoder output layer, is the bias vector. Residual connections and layer normalization are applied to the initial prediction sequence. The residual connection is achieved by adding the weighted context vector of the decoder input With the initial prediction sequence Add together to form a new fusion feature sequence, enhance the stability of the model and avoid the gradient disappearance problem. Layer normalization is performed through the following formula:

[0149] ;

[0150] in, and The current time step The mean and standard deviation of and is a learnable affine transformation parameter. The normalized feature sequence has better distribution characteristics, which helps to improve the training effect of the model and obtain the optimized prediction sequence . The optimized prediction sequence Through three parallel fully connected layers, the number of neurons in each layer is 64, 32 and 16 respectively, and the activation function is Sigmoid. Each fully connected layer is used to predict different crack resistance performance indicators, such as crack initiation time, propagation rate and fracture toughness. The output of each fully connected layer is expressed as:

[0151] ;

[0152] in, is the Sigmoid activation function, is the original predicted value of the i-th anti-cracking performance index, and are the weight matrix and bias vector of the fully connected layer respectively. In order to obtain a stable prediction result, the original prediction value is smoothed by exponential moving average. The formula of exponential moving average is:

[0153] ;

[0154] in, is the smoothing coefficient, is the original prediction value of the current time step, is the smoothed prediction value of the previous time step. Through exponential moving average processing, the short-term fluctuation of the prediction value is reduced to obtain a smoother and more stable prediction result. The smoothed prediction result is compared with the preset material performance threshold to determine whether the crack resistance of the welding material meets the standard.

[0155] In one example, the input sequence set is input into an encoder containing 3 layers of bidirectional LSTM, each layer has 256 LSTM units, and the activation function is tanh, and the encoded feature sequence is obtained, including:

[0156] Batch the input sequence set to obtain the input tensor, and input the input tensor into the first layer of bidirectional LSTM to obtain the forward and reverse hidden state outputs;

[0157] The forward and reverse hidden state outputs are concatenated to obtain the output features of the first layer of bidirectional LSTM, and the output features of the first layer of bidirectional LSTM are input into the second layer of bidirectional LSTM to obtain the forward and reverse hidden states of the second layer;

[0158] The forward and reverse hidden states of the second layer are concatenated and nonlinearly transformed to obtain the output features of the second layer of bidirectional LSTM, and the output features of the second layer of bidirectional LSTM are input into the third layer of bidirectional LSTM to obtain the final forward and reverse hidden states;

[0159] The final forward and reverse hidden states are weightedly summed, with the weights being learnable parameters, to obtain the fused hidden state. The self-attention weights are calculated based on the fused hidden state, and the weighted summation is performed to obtain the encoded feature sequence.

[0160] In this example, the input sequence set is batched to convert it into input tensors that the model can process. The original time series data is divided into multiple batches to facilitate parallel computing and accelerate the model training process. Assume that the input sequence set contains sequences, each of which has a length of , the feature dimension of each time step is , then the shape of the input tensor after batch processing is ,in is the batch size, which indicates the number of sequences contained in each batch. The input tensor is represented as ,in , each subset Corresponding to a length of The time step feature of the sequence is , each For one dimensional feature vector. Input to the first layer of bidirectional LSTM. Bidirectional LSTM consists of two LSTM units in opposite directions, one processing sequence data from front to back and the other processing data from back to front. The forward hidden state of and the backward hidden state They represent the comprehensive information of all time steps before and after the time step respectively. The state update formula of the forward LSTM is:

[0161] ;

[0162] in, is the forward hidden state of the previous time step, is the weight matrix of the forward LSTM, is the bias vector, is the input feature of the current time step, is the activation function. Similarly, the state update formula of the backward LSTM is:

[0163] ;

[0164] in, is the backward hidden state of the next time step, is the weight matrix of the backward LSTM, is the bias vector, is the input feature of the current time step. Concatenate the forward hidden state and backward hidden state of each time step. and Connect together to get the output features of the first layer of bidirectional LSTM:

[0165] ;

[0166] in, For the first layer of bidirectional LSTM at time step The output feature vector of . The output features of the first layer of bidirectional LSTM As input, it is passed to the second layer of bidirectional LSTM. Each time step of the second layer of bidirectional LSTM outputs a forward hidden state and a backward hidden state , its state update formula is:

[0167] ;

[0168] ;

[0169] are the weight matrices of the second layer forward and backward LSTM, respectively. and is the bias vector, is the output feature of the first layer. The forward hidden state and reverse hidden state of the second layer bidirectional LSTM are concatenated to obtain the output feature of the second layer:

[0170] ;

[0171] The output features include the further processing and abstract representation of the first layer output features by the second layer bidirectional LSTM. In order to improve the expressiveness of the features, the output features of the second layer bidirectional LSTM Perform nonlinear transformation. Use a fully connected layer or activation function, such as ReLU, to enhance the nonlinear expression ability of features. Assume that the nonlinear transformation is:

[0172] ;

[0173] in, is the weight matrix of the fully connected layer, The ReLU activation function sets the negative part of the feature to 0, thereby introducing nonlinear characteristics. As input, it is passed to the third layer of bidirectional LSTM. The structure of the third layer of bidirectional LSTM is the same as the first two layers. After being processed by the third layer, the final forward hidden state is obtained. and the backward hidden state . Its state update formula is:

[0174] ;

[0175] ;

[0176] The final forward hidden state and backward hidden state are concatenated to obtain the output features of the third-layer bidirectional LSTM:

[0177] ;

[0178] In order to aggregate the feature information of the bidirectional LSTM at different time steps, the final forward and reverse hidden states Perform weighted summation. The weight of the weighted summation is a learnable parameter, which is used to dynamically adjust the weight of different time step information in the final feature. Assume that the weight of the weighted summation is , the weighted summation formula is:

[0179] ;

[0180] in, is a learnable weight parameter that satisfies Based on the fused hidden state , calculate the self-attention weight. The self-attention mechanism can capture the global dependencies between different time steps in the sequence. It dynamically adjusts the importance of features by calculating the similarity of each time step in the sequence. Assume that the hidden state after fusion is the query , the output feature of the encoder is key Sum , then the calculation formula of self-attention is:

[0181] ;

[0182] in, is the dimension of the key vector, and the softmax function is used to convert the similarity into a probability distribution. Through self-attention calculation, the weighted summed context vector is obtained. , expressed as:

[0183] ;

[0184] Finally, the encoded feature sequence is obtained , which contains a comprehensive representation of the information at different time steps in the sequence.

[0185] In one example, the weighted context vector is input into a decoder containing 2 layers of LSTM, each layer has 512 LSTM units, and the activation function is ReLU, and the initial prediction sequence is obtained, including:

[0186] The weighted context vector is concatenated with the initialized decoder hidden state to obtain the input of the first layer LSTM, and the input of the first layer LSTM is gated and updated to obtain the output of the first layer LSTM and the updated hidden state.

[0187] Apply the ReLU activation function to the output of the first layer LSTM to obtain the nonlinear feature representation of the first layer, and perform a residual connection between the nonlinear feature representation and the original weighted context vector to obtain the fused feature;

[0188] The fused features are layer-normalized to obtain the standardized first-layer output, and the standardized first-layer output is input into the second-layer LSTM to perform gating operations and state updates to obtain the output of the second-layer LSTM;

[0189] The ReLU activation function and dropout operation are applied to the output of the second layer LSTM to obtain the final decoder output, and the attention score is calculated based on the final decoder output and weighted summed with the encoder output to obtain the initial prediction sequence.

[0190] In this example, the context vector generated by the encoder is effectively combined with the initial hidden state of the decoder. Assume that the context vector is C and the initial hidden state of the decoder is . These two vectors are concatenated to form the input vector of the first layer LSTM of the decoder:

[0191] ;

[0192] in, is the concatenation of the context vector and the initial hidden state, which contains the comprehensive features of the context information and the decoder state. The input is sent to the first layer of LSTM units for processing. The core of the LSTM unit is its gating mechanism, including the input gate, forget gate, and output gate, which work together to control the flow of information between different time steps. The input gate determines the contribution of the current input information to the update of the hidden state, the forget gate controls the degree of retention of past information, and the output gate determines the contribution of the current hidden state to the output of the next layer of the network or the next time step. The hidden state at the current moment and cell state (cell state) The update formula is:

[0193] ;

[0194] ;

[0195] in, , and Respectively represent the activation values ​​of the forget gate, input gate, and output gate, which control the storage and output of information in the cell state; © represents element-wise multiplication. After completing the gating operation and state update, the output of the first layer of LSTM is obtained. and the updated hidden state . Output of the first layer LSTM Apply the ReLU activation function. The function of the ReLU activation function is to set the negative part to zero and keep the positive part unchanged, thereby enhancing the expressiveness of the model. The feature representation after nonlinear activation is:

[0196] ;

[0197] Combine the nonlinear feature representation with the original weighted context vector Perform residual connection to obtain fusion features:

[0198] ;

[0199] The role of residual connection is to introduce the input of the deep network directly into the output, effectively alleviating the gradient vanishing problem and better retaining and utilizing the original feature information. Perform layer normalization. Layer normalization eliminates the differences in feature distribution between different samples by standardizing the feature vectors of each time step, thereby improving the stability of the model and the training effect. The normalization formula is:

[0200] ;

[0201] in, and represent the mean and standard deviation of the features respectively, and is the learnable affine transformation parameter. Through this step, the standardized first layer output is obtained The standardized output is input to the second layer LSTM, which processes the input features through a similar gating mechanism. The second layer LSTM receives features that have been nonlinearly transformed and normalized, so it can capture the complex relationship between features at a higher level. After gating operations and state updates, the output of the second layer LSTM is obtained. And the updated hidden state. In order to improve the nonlinear expression ability of the model, the output of the second layer LSTM Apply the ReLU activation function to obtain nonlinear feature representation:

[0202] ;

[0203] After the activation function, the dropout operation is applied to prevent the model from overfitting. Dropout enhances the generalization ability of the model by randomly blocking the output of some neurons. Assume that the probability of dropout is , then the output after dropout is:

[0204] ;

[0205] Through the dropout operation, a certain number of neuron outputs are retained, while some neuron outputs are blocked, effectively avoiding the overfitting of the model to the training data. Based on the above processing, the final decoder output is obtained. The attention score is calculated based on this output. The attention score is used to measure the correlation between the current state of the decoder and the encoder output. The attention weight is calculated using the following formula:

[0206] ;

[0207] in, is the time step The attention weight, is the encoder at time step The output of the attention weight is normalized by the softmax function so that its sum is 1. The attention weight is weighted and summed with the output of the encoder to obtain the weighted context vector , expressed as:

[0208] ;

[0209] Weighted context vector Contains the most important information for the current decoding state in all time steps of the encoder, which helps the decoder to select context information more accurately when generating output. As input for the initial prediction sequence.

[0210] Reference Figure 2 This embodiment provides a welding material high temperature crack resistance testing device based on intelligent acoustic and optical synergy, including:

[0211] Initialization module 1, used to install the welding material specimen on the loading device in the test chamber, and set parameters of the heating device, the sound signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain an initialized test system;

[0212] The heating module 2 is used to heat the welding material specimen based on the heating device and apply dynamic stress through the loading device to obtain the specimen in a high temperature stress state;

[0213] The data acquisition module 3 is used to perform multi-source signal acquisition on the specimen in the high temperature stress state based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain time-synchronized acoustic, optical and electrical coordinated signal data;

[0214] The feature extraction module 4 is used to perform noise reduction filtering and time-frequency domain feature extraction on the acoustic, optical and electrical coordinated signal data, and then input a preset feature classification model to obtain multi-modal classification feature data;

[0215] The prediction module 5 is used to input the multimodal hierarchical feature data into the pre-trained sequence-to-sequence neural network model to obtain the crack resistance performance prediction results including crack initiation time, propagation rate and fracture toughness.

[0216] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0217] Reference Figure 3 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0218] Those skilled in the art will understand that Figure 3 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0219] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0220] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0221] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0222] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation based on the contents of the present invention description and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical synergy, characterized in that: The following steps are involved: The welding material specimen is mounted on a loading device in the test chamber, and parameters of a heating device, an acoustic signal acquisition device, an optical signal acquisition device, and an electrical signal detection device are set to obtain an initialized test system; The welding material specimen is heated by the heating device and dynamic stress is applied by the loading device to obtain the specimen in a high temperature stress state; Based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device, multi-source signal acquisition is performed on the test piece in the high temperature stress state to obtain time-synchronized acoustic, optical and electrical coordinated signal data; The acoustic, optical and electrical collaborative signal data are subjected to denoising filtering and time-frequency domain feature extraction, and then input into a preset feature grading model to obtain multimodal graded feature data; specifically including: performing wavelet threshold denoising on the acoustic, optical and electrical collaborative signal data to obtain denoised signal data, and performing adaptive Wiener filtering based on the denoised signal data to obtain a filtered optimized collaborative signal; performing short-time Fourier transform on the filtered optimized collaborative signal to obtain a time-frequency spectrum, and calculating energy entropy, spectrum centroid and frequency band power ratio based on the time-frequency spectrum to obtain a time-frequency domain feature vector; performing dimensionality reduction processing on the time-frequency domain feature vector to obtain a reduced-dimensional feature set, and performing support vector machine classification on the reduced-dimensional feature set to obtain a preliminary classification result; constructing a fuzzy membership function based on the preliminary classification result to obtain a fuzzy grading index of the feature, and inputting the fuzzy grading index into a hierarchical analysis process model to obtain multimodal graded feature data; The multimodal hierarchical feature data is input into a pre-trained sequence-to-sequence neural network model to obtain a crack resistance prediction result including crack initiation time, propagation rate and fracture toughness; specifically comprising: performing sliding window processing on the multimodal hierarchical feature data, with a window size of 100 time steps and a step length of 10, to obtain an overlapping input sequence set; inputting the input sequence set into an encoder comprising 3 layers of bidirectional LSTM, with 256 LSTM units in each layer and an activation function of tanh, to obtain an encoded feature sequence; based on the encoded feature sequence, calculating the attention weight based on a multi-head attention mechanism, with 8 heads and 64 attention dimensions, to obtain a weighted context vector; inputting the weighted context vector into an encoder comprising 2 layers of bidirectional LSTM, with 256 LSTM units in each layer and an activation function of tanh, to obtain an encoded feature sequence; based on the encoded feature sequence, calculating the attention weight based on a multi-head attention mechanism, with 8 heads and 64 attention dimensions, to obtain a weighted context vector; The decoder of LSTM has 512 LSTM units in each layer and ReLU as the activation function, so as to obtain an initial prediction sequence; residual connection is applied to the initial prediction sequence, and layer normalization is performed, and the normalization parameter is a learnable affine transformation, so as to obtain an optimized prediction sequence; the optimized prediction sequence is passed through three parallel fully connected layers, and the number of neurons in each layer is 64, 32, and 16, respectively, and the activation function is Sigmoid, so as to obtain original prediction values ​​of crack initiation time, propagation rate, and fracture toughness; exponential moving average smoothing is applied to the original prediction value, and the smoothing coefficient is 0.9, so as to obtain a stable prediction result, and the stable prediction result is compared with a preset material performance threshold to obtain a crack resistance performance prediction result.

2. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical collaboration according to claim 1 is characterized in that: The welding material test piece is installed on the loading device in the test cavity, and the parameters of the heating device, the sound signal acquisition device, the optical signal acquisition device and the electric signal detection device are set to obtain an initialized test system, including: The welding material specimen is mounted on the fixture in the test chamber to obtain a fixed and evenly stressed specimen assembly, and the temperature curve of the heating device is programmed to obtain a temperature rise process that simulates the actual welding environment; The stress parameters of the loading device are set based on the mechanical properties of the welding material to obtain a dynamic stress loading scheme including stress amplitude, frequency and loading mode, and a plurality of microphones with different frequency response ranges in the acoustic signal acquisition device are calibrated to obtain a full-band acoustic wave signal acquisition configuration; The focal length and sampling frequency are adjusted based on the high-speed spectrometer and laser scattering device of the optical signal acquisition device to obtain an acquisition setting that accurately captures the changes in the optical signal in the welding area, and the sampling rate and measurement range of the electrical signal detection device are optimized to obtain detection parameters that can accurately record the fluctuations in welding current and voltage, thereby obtaining an initialized test system.

3. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical synergy according to claim 2 is characterized in that: The step of heating the welding material specimen based on the heating device and applying dynamic stress through the loading device to obtain the specimen in a high temperature stress state comprises: The power of the heating device is adjusted according to the temperature rise curve to obtain a continuous change process of the test piece temperature, and the temperature distribution of the test piece surface is scanned in real time based on an infrared thermal imager to obtain a temperature distribution data matrix; Calculating the temperature gradient and uniformity index based on the temperature distribution data matrix to obtain a heating uniformity evaluation result, and inputting the heating uniformity evaluation result into a PID controller to perform closed-loop regulation on the power of the heating device to obtain a test piece with uniform temperature; According to the dynamic stress loading scheme, the loading device is controlled to generate a signal to obtain a loading control sequence including stress amplitude, frequency and phase information, and the strain and force conditions of the specimen are collected in real time based on a high-precision strain gauge and a force sensor to obtain stress-strain response data; Based on the stress-strain response data, Fourier transform and harmonic analysis are performed to obtain the deviation between the actual loading state and the theoretical loading scheme, and the deviation data is input into the adaptive control algorithm to perform real-time correction on the loading control sequence to obtain the test piece in a high-temperature stress state.

4. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical collaboration according to claim 3 is characterized in that: The method of performing multi-source signal acquisition on the test piece in the high temperature stress state based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain time-synchronized acoustic, optical and electrical coordinated signal data includes: Based on the multiple microphones with different frequency response ranges in the acoustic signal acquisition device, the test piece is omni-directionally acquired to obtain multi-channel original acoustic signal data; Performing time synchronization and channel fusion processing on the multi-channel original acoustic signal data to obtain a comprehensive acoustic signal with a unified time reference, and performing a spectral scan on the surface of the specimen based on a high-speed spectrometer of the optical signal acquisition device to obtain a spectral data sequence that changes with time; The laser scattering device of the optical signal acquisition device scans the surface of the specimen in real time to obtain surface morphology change data, and performs time alignment and data fusion on the spectral data sequence and the surface morphology change data to obtain a comprehensive optical signal feature; Based on the electrical signal detection device, high-frequency sampling is performed on the welding current and voltage to obtain original current and voltage waveform data, and power spectrum analysis and phase calculation are performed on the original current and voltage waveform data to obtain electrical signal characteristic parameters; The timestamps of the integrated acoustic signal, the integrated optical signal characteristics and the electrical signal characteristic parameters are aligned and the data formats are unified to obtain a time-synchronized acoustic, optical and electrical coordinated signal data set.

5. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical synergy according to claim 1 is characterized in that: The method of constructing a fuzzy membership function based on the preliminary classification result to obtain a fuzzy classification index of the feature, and inputting the fuzzy classification index into a hierarchical analysis process model to obtain multimodal hierarchical feature data includes: Performing statistical analysis on the preliminary classification results to obtain the distribution of each feature in different categories, and calculating the fuzzy set boundary of each feature based on the distribution to obtain the fuzzy interval of the feature; Fitting the fuzzy interval based on a Gaussian function to obtain a membership function of each feature, and normalizing the membership function to obtain a standardized fuzzy membership function; Calculating the fuzzy grading index of each feature based on the standardized fuzzy membership function to obtain an initial fuzzy grading matrix, and constructing a judgment matrix of the hierarchical analysis process, comparing the indicators in the initial fuzzy grading matrix pairwise to obtain a target relationship matrix; Calculate the maximum eigenvalue and the corresponding eigenvector of the target relationship matrix based on the eigenvalue method to obtain the weight vector of each indicator; Perform a consistency check on the weight vector, if it passes, it is retained, if it fails, the judgment matrix is ​​modified, and the calculation is repeated until it passes the consistency check to obtain the final weight vector; The final weight vector is weighted and summed with the initial fuzzy grading matrix to obtain a comprehensive scoring vector, and the features are sorted and clustered based on the comprehensive scoring vector to obtain multimodal graded feature data.

6. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical synergy according to claim 1 is characterized in that: The input sequence set is input into an encoder including 3 layers of bidirectional LSTM, the number of LSTM units in each layer is 256, and the activation function is tanh, to obtain an encoded feature sequence, including: Batch processing the input sequence set to obtain an input tensor, and input the input tensor into the first layer of bidirectional LSTM to obtain forward and reverse hidden state outputs; The forward and reverse hidden state outputs are concatenated to obtain output features of the first layer of bidirectional LSTM, and the output features of the first layer of bidirectional LSTM are input into the second layer of bidirectional LSTM to obtain forward and reverse hidden states of the second layer; The forward and reverse hidden states of the second layer are concatenated and nonlinearly transformed to obtain the output features of the second layer of bidirectional LSTM, and the output features of the second layer of bidirectional LSTM are input into the third layer of bidirectional LSTM to obtain the final forward and reverse hidden states; The final forward and reverse hidden states are weightedly summed, with the weights being learnable parameters, to obtain a fused hidden state, and the self-attention weights are calculated based on the fused hidden state, and the weighted summation is performed to obtain an encoded feature sequence.

7. The method for testing the high temperature crack resistance of welding materials based on intelligent acoustic, optical and electrical synergy according to claim 1 is characterized in that: The weighted context vector is input into a decoder including two layers of LSTM, each layer of LSTM has 512 units and the activation function is ReLU, to obtain an initial prediction sequence, including: The weighted context vector is concatenated with the initialized decoder hidden state to obtain the input of the first layer LSTM, and the input of the first layer LSTM is gated and updated to obtain the output of the first layer LSTM and the updated hidden state; Applying a ReLU activation function to the output of the first layer of LSTM to obtain a nonlinear feature representation of the first layer, and performing a residual connection between the nonlinear feature representation and the original weighted context vector to obtain a fused feature; Performing layer normalization processing on the fusion features to obtain a standardized first-layer output, and inputting the standardized first-layer output into the second-layer LSTM, performing gating operations and state updates, and obtaining the output of the second-layer LSTM; Apply the ReLU activation function and dropout operation to the output of the second layer LSTM to obtain the final decoder output, and calculate the attention score based on the final decoder output, and perform weighted summation with the encoder output to obtain the initial prediction sequence.

8. A welding material high temperature crack resistance testing device based on intelligent acoustic and optical synergy, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: An initialization module is used to install the welding material specimen on the loading device in the test chamber, and set parameters of the heating device, the sound signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain an initialized test system; A heating module, used for heating the welding material specimen based on the heating device, and applying dynamic stress through the loading device to obtain the specimen in a high temperature stress state; A data acquisition module, used to perform multi-source signal acquisition on the test piece in the high temperature stress state based on the acoustic signal acquisition device, the optical signal acquisition device and the electrical signal detection device to obtain time-synchronized acoustic, optical and electrical coordinated signal data; A feature extraction module is used to perform noise reduction filtering and time-frequency domain feature extraction on the acoustic, optical and electrical coordinated signal data, and then input a preset feature classification model to obtain multimodal classification feature data; The prediction module is used to input the multimodal hierarchical feature data into a pre-trained sequence to sequence neural network model to obtain a crack resistance performance prediction result including crack initiation time, propagation rate and fracture toughness.

Images