Multi-modal fusion metal micro-crack ultrasonic detection system and method

Abstract

The present invention relates to the field of nondestructive testing technology, in particular to a multi-modal fusion metal microcrack ultrasonic detection system and method thereof, comprising: ultrasonic signal acquisition, signal preprocessing, feature alignment network, image segmentation, time-frequency analysis and segmentation result fusion module. The system uses three modes of electromagnetic ultrasound, laser ultrasound and phased array ultrasound to detect metal samples, realizing real-time nondestructive signal data acquisition and preprocessing. The feature alignment network aligns and fuses the signal features of the three modes, and the U‑Net segmentation neural network performs sub-millimeter crack segmentation on the fused features. The time-frequency analysis module extracts signal frequency and amplitude information, and combines it with crack labels to evaluate the fracture strength. Finally, the segmentation result fusion module integrates crack labels and time series features, and classifies the presence or absence of cracks through a weighted convolutional neural network. The system has significantly improved detection performance, practicality and adaptability, providing a new solution for accurate detection of metal microcracks.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, in particular to a multi-modal fusion metal micro-crack ultrasonic testing system and method thereof. Background Art

[0002] The detection of metal microcracks is of vital importance in high-end manufacturing industries such as aerospace, nuclear power, and petrochemicals. With the continuous advancement of industrial technology, the safety and reliability requirements for metal components are increasing, which requires more accurate and efficient microcrack detection technology. Traditional ultrasonic detection methods, such as phased array ultrasound, electromagnetic ultrasound, and laser ultrasound, each have their own advantages and limitations.

[0003] Phased array ultrasonic technology is popular for its flexible beam control and focusing capabilities. However, it has blind spots when detecting surface or near-surface microcracks and requires a coupling agent, which may cause inconvenience in some application scenarios. Electromagnetic ultrasonic technology overcomes the problem of requiring a coupling agent and is particularly suitable for detecting high temperature or rough surfaces, but its signal-to-noise ratio is low and its ability to detect tiny cracks is limited. Laser ultrasonic technology is known for its non-contact and high-resolution characteristics, but it has not yet been widely used in industrial applications due to factors such as its low excitation efficiency and expensive equipment.

[0004] In recent years, researchers have tried to combine different ultrasonic detection methods to make up for the shortcomings of a single technology. For example, a study proposed a dual-modal fusion method of phased array ultrasound and electromagnetic ultrasound, which improved the comprehensiveness of detection to a certain extent. However, this simple dual-modal fusion still cannot give full play to the advantages of various ultrasonic technologies, especially when dealing with complex detection scenarios, its performance improvement is still insufficient.

[0005] In addition, existing fusion methods often use simple data superposition or feature cascade, lacking in-depth understanding and intelligent processing of the characteristics of different modal signals. This may lead to redundancy or inconsistency in the fused information, affecting the accuracy and reliability of detection. At the same time, existing methods are inefficient in processing large amounts of multimodal data and are difficult to meet the needs of real-time detection.

[0006] In the face of these challenges, there is an urgent need for an advanced detection system that can fully utilize the advantages of multiple ultrasonic detection technologies and intelligently process and fuse multimodal signals. This system should be able to effectively improve the detection rate and positioning accuracy of microcracks while maintaining high detection efficiency and anti-interference capabilities. Summary of the invention

[0007] The present invention aims at the above technical problems and proposes a multi-modal fusion metal micro-crack ultrasonic detection system and method. The invention realizes the efficient fusion of three ultrasonic diagnostic modes: electromagnetic ultrasound, laser ultrasound and phased array ultrasound through innovative system architecture and algorithm design.

[0008] The present invention proposes a multi-modal fusion metal micro-crack ultrasonic detection system, comprising:

[0009] Ultrasonic signal acquisition module, used for:

[0010] Metal samples are tested based on three ultrasonic diagnostic modalities: electromagnetic ultrasound, laser ultrasound and phased array ultrasound;

[0011] Obtain lossless signal data in real time;

[0012] The signal preprocessing module is connected to the ultrasonic signal acquisition module and is used to:

[0013] Acquiring lossless signal data collected by the ultrasonic signal acquisition module;

[0014] Preprocessing the lossless signal data and outputting a lossless signal;

[0015] A feature alignment network, connected to the signal preprocessing module, is used to:

[0016] Built on a cross-modal feature extraction module;

[0017] Align input lossless signal characteristics of three ultrasonic diagnostic modalities;

[0018] Perform feature-level fusion on the input lossless signal and output fusion signal features;

[0019] An image segmentation module, connected to the feature alignment network, is used to:

[0020] Build a segmentation neural network based on U-Net;

[0021] Perform sub-millimeter metal micro-crack segmentation on the input fusion signal features;

[0022] Output sub-millimeter scale crack binary labels;

[0023] The time-frequency analysis module is connected to the ultrasound signal acquisition module and the image segmentation module and is used to:

[0024] Perform time-frequency analysis on the three ultrasonic signals to obtain frequency and amplitude information;

[0025] Aligning with the lossless signal to obtain frequency and amplitude information of each lossless signal;

[0026] Output spectrum and amplitude diagrams of three ultrasonic signals;

[0027] Based on the frequency, amplitude information and crack binary label, the fracture strength of the ultrasonic signal is obtained to determine the crack;

[0028] The segmentation result fusion module is connected to the image segmentation module and the time-frequency analysis module and is used to:

[0029] Aligning the crack binary label and the spectrum graph and amplitude graph of the three ultrasonic signals to obtain time series features;

[0030] Inputting the crack binary label and the time series feature into a weighted convolutional neural network;

[0031] Perform classification and output the classification results of whether there are cracks or not.

[0032] Preferably, the feature alignment network comprises:

[0033] Three groups of feature extractors, each of which is composed of the following units in series:

[0034] A convolution operation unit, used to extract multi-scale features from input signals;

[0035] Multiple residual dense blocks are used to extract and fuse different scale features of multimodal lossless signals;

[0036] The global attention module is used to weight the features extracted by each residual dense block according to their global importance.

[0037] The channel connection module is used to use the weighted features as the scale-channel relationship matrix, perform matrix multiplication on the input features and the scale-channel relationship matrix, and map the feature channel relationship to the global feature relationship through scale transformation to achieve feature alignment.

[0038] Preferably, the U-Net network in the image segmentation module includes:

[0039] Encoder, including:

[0040] Input layer;

[0041] The first convolutional layer is used to extract the fusion signal features and obtain a coding feature map;

[0042] A deconvolution layer, used to deconvolve the encoded feature map to obtain a decoded feature map;

[0043] Decoder, including:

[0044] The second convolutional layer;

[0045] Output layer;

[0046] The encoder further includes two atrous convolution connection modules, which respectively convolve the encoding feature map output by the encoder, and use maximum pooling and deconvolution to perform upsampling operations to obtain the encoding feature map.

[0047] Preferably, the weighted convolutional neural network in the segmentation result fusion module includes:

[0048] The third convolutional layer is used to perform convolution operations on the input features;

[0049] Pooling layer, used to reduce the dimension of convolutional features;

[0050] The third convolution layer adopts a multi-scale convolution kernel to adapt to crack features of different scales.

[0051] Preferably, the feature alignment network further comprises:

[0052] Waveform enhancement module for:

[0053] The laser signal and the electromagnetic ultrasound signal are dynamically fused in the channel dimension through the gating module;

[0054] Output enhanced waveform characteristics;

[0055] Cross-modal attention module for:

[0056] The phased array ultrasonic signal is processed through linear transformation and self-attention mechanism on the feature map channel;

[0057] Improve the signal-to-noise ratio of the image.

[0058] Preferably, the waveform enhancement module comprises:

[0059] A feature enhancement unit, used for performing preliminary enhancement on the input signal;

[0060] Feature dimension reduction unit, used to reduce feature dimension;

[0061] Door control unit, comprising:

[0062] Characterize enhancer units;

[0063] A transposed matrix subunit, used for multiplying the electromagnetic ultrasonic signal to obtain a signal one;

[0064] A feature fusion unit, used for performing fusion processing on signal one;

[0065] A feature dimension increasing unit is used to increase the feature dimension and obtain signal 2;

[0066] The convolution unit is used to perform convolution operation on signal 2 and output the final gating module output.

[0067] Preferably, the cross-modal attention module comprises:

[0068] A linear transformation unit, used for performing a linear transformation on the phased array ultrasonic characteristic matrix to obtain a characteristic matrix 1;

[0069] Self-attention unit, used to:

[0070] Combining electromagnetic ultrasonic feature matrix and laser ultrasonic feature matrix;

[0071] Based on feature matrix 1, the weight values ​​of different modal features are obtained through the self-attention mechanism;

[0072] A weight product unit, used for multiplying the weight value by the feature matrix 1 to obtain the feature matrix 2;

[0073] Feature fusion unit, used for:

[0074] Adding the laser ultrasonic feature matrix and the electromagnetic ultrasonic feature matrix into feature matrix 2;

[0075] Obtain feature matrix three through fusion operation;

[0076] The normalization unit is used to perform linear transformation and Sigmoid normalization on feature matrix 3 to obtain the final fusion feature matrix.

[0077] As a preference, it also includes:

[0078] The parameter optimization module is connected to the ultrasonic signal acquisition module and is used to:

[0079] According to the detection requirements of metal microcracks, the frequency range, the aperture size of the phased array ultrasound, and the distance between the phased array ultrasound transducer and the metal workpiece to be tested are determined by integrating three sets of ultrasonic data;

[0080] The scanning path is determined by integrating the maximum irradiation range of the phased array ultrasonic transducer and the geometric features of the inspected workpiece;

[0081] Determine the coil spacing and oscillation frequency of the electromagnetic ultrasonic transducer according to the conductivity, thickness and frequency range of the metal workpiece to be tested;

[0082] The laser power, frequency and polarization direction are determined by combining the geometric characteristics and scanning path of the laser ultrasonic transducer.

[0083] Preferably, the parameter optimization module is further used for:

[0084] According to the formula , calculate the distance between the electromagnetic ultrasonic transducer and the metal workpiece to be tested ,in is the amplitude of the electromagnetic field sound pressure generated by the electromagnetic ultrasonic transducer coil, is the performance parameter of electromagnetic ultrasonic transducer;

[0085] According to the formula , calculate the contact resistance between the electromagnetic ultrasonic transducer and the metal workpiece to be tested, where is the conductivity of the metal workpiece to be tested, is the projected area of ​​the electromagnetic ultrasonic transducer surface, is the contact resistance, is the resistivity of the metal workpiece to be tested;

[0086] According to the relationship between the oscillation frequency of the electromagnetic ultrasonic transducer and the contact resistance, the oscillation frequency of the electromagnetic ultrasonic transducer coil is calculated;

[0087] According to the thickness of the metal workpiece to be measured , electromagnetic ultrasonic transducer coil oscillation frequency , dielectric loss factor and distance , determine the optimal parameters, the formula is as follows:

[0088] Optimal parameters = ,

[0089] in, is the correlation coefficient, which is determined by experimental fitting.

[0090] The multi-modal fusion metal micro-crack ultrasonic detection method comprises the following steps:

[0091] S1: Through the ultrasonic signal acquisition module, the metal sample is tested based on three ultrasonic diagnostic modes: electromagnetic ultrasound, laser ultrasound and phased array ultrasound, and non-destructive signal data is obtained in real time;

[0092] S2: Preprocessing the lossless signal data through a signal preprocessing module to output a lossless signal;

[0093] S3: Through the feature alignment network, based on the cross-modal feature extraction module, the lossless signal features of the three input ultrasound diagnostic modalities are aligned, the input lossless signals are fused at the feature level, and the fused signal features are output;

[0094] S4: Through the image segmentation module, a segmentation neural network is constructed based on U-Net to perform sub-millimeter metal micro-crack segmentation on the input fusion signal features and output sub-millimeter scale crack binary labels;

[0095] S5: Perform time-frequency analysis on the three ultrasonic signals through the time-frequency analysis module to obtain frequency and amplitude information, align them with the lossless signal, obtain frequency and amplitude information of each lossless signal, and output spectrum diagrams and amplitude diagrams of the three ultrasonic signals; based on the frequency, amplitude information and crack binary labels, obtain the fracture strength of the ultrasonic signal and determine the crack;

[0096] S6: Through the segmentation result fusion module, the crack binary label and the spectrum diagram and amplitude diagram of the three ultrasonic signals are aligned to obtain the time series features, and the crack binary label and the time series features are input into the weighted convolutional neural network for classification, and the classification result of whether there is a crack is output.

[0097] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0098] First, from a macroscopic perspective, the present invention realizes the complementary advantages and synergistic effect of multimodal ultrasonic detection technology. By integrating the non-contact advantages of electromagnetic ultrasound, the high-resolution characteristics of laser ultrasound, and the flexible focusing capabilities of phased array ultrasound, the system can adapt to various complex detection environments and significantly improve the detection rate and detection reliability of microcracks. For example, in our experiments, for complex components such as aircraft engine turbine blades, the detection rate of this system reached 98.5%, far exceeding the single mode and simple dual-mode fusion methods.

[0099] Secondly, the present invention designs a feature alignment network, which effectively solves the scale and feature differences between different modal signals. This feature alignment method based on deep learning can adaptively learn the important features of different modal signals and realize intelligent feature fusion. This not only improves the fusion effect, but also greatly enhances the generalization ability of the system, enabling it to adapt to the detection needs of different materials and structures.

[0100] Furthermore, the image segmentation module of the present invention adopts an improved U-Net structure, combined with multi-scale feature extraction and dilated convolution technology, which greatly improves the positioning accuracy and size measurement accuracy of microcracks. In our experiments, the positioning accuracy of this system reached ±0.05mm, and the size measurement error was only 5.2%, which is of great significance for the accurate characterization of micron-level cracks.

[0101] In addition, the collaborative work of the time-frequency analysis module and the segmentation result fusion module of the present invention effectively improves the system's anti-noise ability. By comprehensively utilizing time domain and frequency domain information and combining the powerful feature extraction capabilities of deep learning, the system can still maintain excellent detection performance in a low signal-to-noise ratio environment. This feature makes the present invention particularly suitable for application in industrial environments with complex noise.

[0102] Finally, although the present invention integrates multiple ultrasonic detection technologies, it achieves fast detection speed through optimized parallel processing architecture and efficient algorithm design. In our experiment, it only takes 2.5 minutes to complete a comprehensive detection of a complex sample, which greatly improves the detection efficiency and lays the foundation for large-scale, real-time industrial detection.

[0103] In summary, the present invention has not only achieved a qualitative leap in detection performance, but also demonstrated great potential in practicality and adaptability. It provides a new solution for the precise detection of metal microcracks, and is expected to play an important role in key fields such as aerospace, nuclear power, and petrochemicals, and make important contributions to improving the safety and reliability of industrial production. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 It is the main system flow chart of the present invention;

[0105] Figure 2 is an internal structure diagram of the feature alignment network of the present invention;

[0106] Figure 3 is an internal structure diagram of the image segmentation module of the present invention;

[0107] Figure 4 is an internal structure diagram of the waveform enhancement module of the present invention;

[0108] Figure 5 : is the internal structure diagram of the cross-modal attention module of the present invention;

[0109] Figure 6 It is the internal structure diagram of the segmentation result fusion module of the present invention. DETAILED DESCRIPTION

[0110] like Figure 1 - Figure 6 As shown, the present invention provides a multi-modal fusion metal microcrack ultrasonic detection system and method thereof. The system achieves high-precision detection of metal microcracks by integrating three ultrasonic diagnostic modes: electromagnetic ultrasound, laser ultrasound, and phased array ultrasound.

[0111] First, please refer to Figure 1 The system of the present invention includes an ultrasonic signal acquisition module 1, which is used to detect metal samples based on three ultrasonic diagnostic modes and obtain non-destructive signal data in real time. Preferably, the ultrasonic signal acquisition module 1 adopts high-frequency sampling technology, and the sampling frequency can reach 50MHz to ensure that the high-frequency signal generated by the tiny crack is captured.

[0112] Secondly, the signal preprocessing module 2 is connected to the ultrasonic signal acquisition module 1 to obtain and preprocess the lossless signal data. In one embodiment of the present invention, the preprocessing includes the steps of denoising, filtering and signal enhancement. For example, wavelet transform can be used for denoising, a bandpass filter with a cutoff frequency of 10 MHz can be used for filtering, and adaptive gain control can be used for signal enhancement.

[0113] Next, the feature alignment network 3 is connected to the signal preprocessing module 2 and constructed based on the cross-modal feature extraction module to align the lossless signal features of the three ultrasonic diagnostic modalities and perform feature-level fusion. The feature alignment network 3 of the present invention adopts an innovative multi-scale convolution structure, which can effectively process the scale differences of different modal signals.

[0114] Please refer to Figure 2 The feature alignment network 3 includes three groups of feature extractors 31, each of which is composed of a convolution operation unit 311, a plurality of residual dense blocks 312, a global attention module 313 and a channel connection module 314 in series. The convolution operation unit 311 uses a multi-scale convolution kernel with kernel sizes of 3x3, 5x5 and 7x7 to extract features of different scales. The residual dense block 312 uses an improved DenseNet structure, each block contains 4 convolution layers, and the growth rate is set to 32, which effectively improves the feature extraction capability.

[0115] The global attention module 313 uses the self-attention mechanism to weight the features. Specifically, the module calculates the attention weight w as follows:

[0116] softmax ,

[0117] in, , and is the linear transformation function, is the input feature. In this way, the model can adaptively focus on important features, and T is the matrix transpose operation.

[0118] The channel connection module 314 maps the feature channel relationship to the global feature relationship through matrix multiplication and scale transformation. This operation can be expressed as:

[0119] ,

[0120] in, and is a learnable parameter, is the sigmoid activation function, is the input feature.

[0121] Please refer to Figure 3 The image segmentation module 4 of the present invention is connected to the feature alignment network 3, and a segmentation neural network is constructed based on U-Net to perform sub-millimeter-level metal microcrack segmentation on the fusion signal features. The U-Net network includes an encoder 41 and a decoder 42.

[0122] The encoder 41 includes an input layer 411, a first convolutional layer 412 and a deconvolutional layer 413. Preferably, the first convolutional layer 412 uses a dilated convolution, and the dilation rate increases layer by layer from 1 to 8 to increase the receptive field. The deconvolutional layer 413 uses a transposed convolution with a step size of 2, which effectively improves the resolution of the feature map.

[0123] In addition, the encoder 41 also includes two dilated convolutional connection modules 414. These modules extract multi-scale context information through parallel multi-scale dilated convolutions (with dilation rates of 1, 2, 4, and 8, respectively), and then upsample through maximum pooling and deconvolution to obtain rich encoded feature maps.

[0124] The decoder 42 includes a second convolutional layer 421 and an output layer 422. The decoder 42 adopts a skip connection structure to fuse the feature map of the encoder 41 with the feature map in the decoding process, effectively retaining high-resolution detail information.

[0125] The network structure design of the present invention is particularly suitable for detecting tiny cracks on metal surfaces. For example, for cracks with a width of only 10 μm, experiments show that the detection rate of the system can reach more than 95%, which is much higher than traditional methods.

[0126] Through the above innovative network structure and algorithm design, the system of the present invention can effectively fuse multi-modal ultrasonic signals, extract rich feature information, and achieve accurate positioning and segmentation of metal microcracks. This not only improves the accuracy and reliability of detection, but also can adapt to various complex detection environments, providing a powerful tool for the safety assessment of metal components.

[0127] Next, the system of the present invention further includes a time-frequency analysis module 5, which is connected to the ultrasonic signal acquisition module 1 and the image segmentation module 4. The main function of the time-frequency analysis module 5 is to perform time-frequency analysis on the three ultrasonic signals, obtain frequency and amplitude information, and align with the lossless signal. Preferably, the time-frequency analysis module 5 adopts the short-time Fourier transform (STFT) method, the window length is set to 256 sampling points, and the overlap rate is 50% to obtain good time-frequency resolution.

[0128] In a preferred embodiment of the present invention, the time-frequency analysis module 5 also uses the wavelet packet transform (WPT) technology to perform multi-scale decomposition on the signal. Using a 5-layer decomposition and using the db4 wavelet basis, the transient characteristics caused by the crack can be effectively captured. By comparing the energy distribution at different scales, the time-frequency analysis module 5 can identify the abnormal frequency components caused by the crack.

[0129] The time-frequency analysis module 5 outputs the spectrum diagram and amplitude diagram of the three ultrasonic signals, and obtains the fracture strength of the ultrasonic signal based on the frequency, amplitude information and crack binary label to determine the crack. Preferably, the calculation formula of the fracture strength S is as follows:

[0130] ,

[0131] in, is the amplitude of the crack position, is the background amplitude, is the main frequency of the crack location, is the background main frequency, and are weight coefficients, determined through experiments, with typical values ​​of 0.7 and 0.3 respectively.

[0132] The system of the present invention further includes a segmentation result fusion module 6, which is connected to the image segmentation module 4 and the time-frequency analysis module 5. The function of the segmentation result fusion module 6 is to align the crack binary label and the spectrum diagram and amplitude diagram of the three ultrasonic signals, obtain the time series features, and input these features into the weighted convolutional neural network for classification, and finally output the classification result of whether there is a crack or not.

[0133] Please refer to Figure 6 In one embodiment of the present invention, the weighted convolutional neural network adopts an innovative multi-scale structure. Specifically, the network includes a third convolutional layer 61 and a pooling layer 62. The third convolutional layer 61 adopts parallel multi-scale convolution, and the convolution kernel sizes are 3x3, 5x5 and 7x7, respectively, and the number of convolution kernels of each size is 64. This design can capture crack features of different scales at the same time and improve the robustness of detection.

[0134] The pooling layer 62 adopts a hybrid pooling strategy, combining the advantages of maximum pooling and average pooling. The pooling operation can be expressed as:

[0135] ,

[0136] in, is the input feature, is the weight coefficient, the preferred value is 0.6, For input features The maximum pooling operation performed, For input features The average pooling operation performed. This hybrid pooling strategy not only retains significant features but also takes into account the overall information, which is beneficial to improving classification performance.

[0137] The system of the present invention further includes a waveform enhancement module 7 and a cross-modal attention module 8 in the feature alignment network 3, which further improves the effect of feature extraction and fusion. The waveform enhancement module 7 is mainly used to realize dynamic feature fusion of laser signals and electromagnetic ultrasound signals in the channel dimension through the gating module. Preferably, the waveform enhancement module 7 adopts a gated linear unit (GLU) structure, which can be expressed as:

[0138] ,

[0139] and is the input feature, represents element-wise multiplication, is the sigmoid function. This structure allows the network to dynamically select important feature channels and improve the effect of feature fusion.

[0140] The cross-modal attention module 8 is mainly used to process phased array ultrasonic signals, and improve the signal-to-noise ratio of the image through linear transformation and self-attention mechanism on the feature map channel. In a preferred embodiment of the present invention, the cross-modal attention module 8 adopts an improved attention mechanism, and the attention weight w is calculated as follows:

[0141] softmax ,

[0142] in, , and are query, key, and value matrices respectively, is the dimension of the key vector, T is the matrix transpose operation, is a learnable position encoding matrix. This improved attention mechanism can better capture the long-range dependencies between different modal signals.

[0143] Please refer to Figure 4 In a preferred embodiment of the present invention, the waveform enhancement module 7 includes a feature enhancement unit 71, a feature dimension reduction unit 72, a gating unit 73, a first feature fusion unit 74, a feature dimension enhancement unit 75 and a convolution unit 76. This multi-level structural design can more finely process and fuse ultrasonic signals of different modes.

[0144] The feature enhancement unit 71 is used to perform preliminary enhancement on the input signal. In one embodiment of the present invention, the feature enhancement unit 71 adopts a residual structure, which includes two The convolution layer is connected to a short circuit. This design can improve the feature extraction capability without increasing the network depth. Preferably, the output of the feature enhancement unit 71 can be expressed as:

[0145] ,

[0146] in, represents the residual mapping, is the input feature.

[0147] The feature dimension reduction unit 72 is used to reduce the feature dimension. In the present invention, the feature dimension reduction unit 72 uses Convolution performs channel dimension reduction, which can effectively reduce the amount of calculation and extract the correlation between channels. Preferably, the dimension reduction ratio can be set to 2:1, that is, the number of input channels is halved. The gate control unit 73 is the core part of the waveform enhancement module 7, including a feature enhancement subunit and a transposed matrix subunit. The feature enhancement subunit is used to further enhance the feature. The transposed matrix subunit is used to multiply the electromagnetic ultrasound signal to obtain signal one. The synergy of these two subunits can realize intelligent feature selection and fusion.

[0148] The first feature fusion unit 74 is used to perform fusion processing on the signal 1. In one embodiment of the present invention, the first feature fusion unit 74 fuses the output of the gating unit 73 and the original input in a weighted summation manner. Preferably, the fusion process can be expressed as:

[0149] ,

[0150] in, is the input feature, is the gate control unit output, 1 is the weight coefficient obtained through learning.

[0151] The feature dimension increasing unit 75 is used to increase the feature dimension to obtain signal 2. In the present invention, the feature dimension increasing unit 75 uses a transposed convolution operation to restore the feature dimension to its original size. This step ensures that the output feature maintains the same spatial resolution as the input. Preferably, the step size of the transposed convolution can be set to 2, and the kernel size is , to obtain a good upsampling effect.

[0152] Finally, the convolution unit 76 is used to perform a convolution operation on the signal 2 and output the final gate module output. In one embodiment of the present invention, the convolution unit 76 uses Convolution performs the final adjustment and fusion of features. This step can be seen as a synthesis of all previous processing, ensuring that the number of channels and distribution of the output features are suitable for subsequent processing.

[0153] The waveform enhancement module 7 of the present invention can effectively process and fuse signals from different ultrasonic detection modes, and improve the system's ability to detect microcracks. For example, in practical applications, for complex welded joints, the module can enhance weak crack signals while suppressing background noise, thereby significantly improving the system's detection sensitivity and reliability.

[0154] Please refer to Figure 5 In a preferred embodiment of the present invention, the structure of the cross-modal attention module 8 is carefully designed, including a linear transformation unit 81, a self-attention unit 82, a weight product unit 83, a second feature fusion unit 84 and a normalization unit 85. This structural design is intended to make full use of the complementary information between different modal signals.

[0155] The linear transformation unit 81 is used to perform a linear transformation on the phased array ultrasound characteristic matrix to obtain a characteristic matrix 1. In one embodiment of the present invention, the linear transformation can be expressed as:

[0156] ,

[0157] in, is the input phased array ultrasonic characteristic matrix, is the learnable weight matrix, is the bias term, is the obtained feature matrix 1.

[0158] The functions of the self-attention unit 82 include two aspects: first, combining the electromagnetic ultrasound feature matrix and the laser ultrasound feature matrix; second, based on the feature matrix 1, obtaining the weight values ​​of different modal features through the self-attention mechanism. In the present invention, the calculation of the self-attention mechanism can be expressed as:

[0159] ,

[0160] in, is the learnable weight matrix, 1 is the feature dimension, is the obtained attention weight matrix, and T is the matrix transpose operation.

[0161] The weight multiplication unit 83 is used to multiply the weight value with the feature matrix 1 to obtain the feature matrix 2. This step can be expressed as:

[0162] ,

[0163] in, is the weight matrix obtained from the self-attention unit, is the feature matrix 1, is the obtained feature matrix 2.

[0164] The function of the second feature fusion unit 84 includes two steps: first, adding the laser ultrasound feature matrix and the electromagnetic ultrasound feature matrix to the feature matrix 2; then, obtaining the feature matrix 3 through a fusion operation. In one embodiment of the present invention, the fusion operation can be performed in a weighted summation manner:

[0165] ,

[0166] in, is the feature matrix 2, is the electromagnetic ultrasonic characteristic matrix, is the laser ultrasound characteristic matrix, , , is the learnable weight coefficient, The obtained feature matrix three.

[0167] Finally, the normalization unit 85 is used to perform linear transformation and Sigmoid normalization on the feature matrix 3 to obtain the final fusion feature matrix. This step can be expressed as:

[0168] ,

[0169] in, is the feature matrix three, is the learnable weight matrix, is the bias term, is the Sigmoid function, is the final fusion feature matrix.

[0170] Through the cross-modal attention mechanism, the present invention can effectively integrate information from different ultrasonic detection modalities and make full use of the advantages of various modalities. For example, when detecting microcracks in complex material structures, the module can adaptively adjust the weights of different modal signals to improve the accuracy and reliability of detection. For some difficult-to-detect cracks, such as microcracks perpendicular to the surface, the module can rely more on the information of phased array ultrasound; while for surface cracks, it may make more use of the high-resolution characteristics of laser ultrasound. This flexible feature fusion strategy enables the present invention to maintain excellent performance in various complex detection scenarios.

[0171] Through the above detailed module design and algorithm optimization, the multi-modal fusion metal micro-crack ultrasonic detection system of the present invention can effectively integrate the advantages of different ultrasonic detection methods and greatly improve the accuracy and reliability of detection. For example, in practical applications, for a steel plate with a thickness of 10 mm, the system can detect micro-cracks with a length of only 0.5 mm and a width of 20 μm, with a detection rate of 98%, far exceeding the traditional single-modal detection method. This is of great significance for ensuring the safety of metal components and extending their service life.

[0172] In a preferred embodiment of the present invention, the system further comprises a parameter optimization module connected to the ultrasonic signal acquisition module 1. The main function of the parameter optimization module is to automatically optimize various ultrasonic detection parameters according to specific detection requirements and characteristics of the detected metal material to obtain the best detection effect.

[0173] The parameter optimization module first comprehensively analyzes the three sets of ultrasonic data according to the detection requirements of metal microcracks to determine the most suitable frequency range. Preferably, for typical metal microcrack detection, the frequency range is usually set between 2MHz and 10MHz. This range can provide good resolution while ensuring sufficient penetration depth.

[0174] In addition, the parameter optimization module is also responsible for determining the aperture size of the phased array ultrasound and the distance between the phased array ultrasound transducer and the metal workpiece to be tested. In one embodiment of the present invention, the selection of the aperture size takes into account the balance between the beam focusing effect and the spatial resolution. For example, for a steel plate with a thickness of 20 mm, the aperture size can be set to 32 array elements, with a spacing of 0.5 mm between each array element, so that a lateral resolution of about 1 mm can be obtained at a depth of 5 mm.

[0175] The parameter optimization module is also responsible for determining the optimal scanning path by integrating the maximum irradiation range of the phased array ultrasonic transducer and the geometric features of the inspected workpiece. Preferably, an adaptive scanning strategy is used to automatically generate a scanning path based on the shape of the workpiece. For example, for workpieces with complex shapes, a curved surface following technology can be used to ensure that the ultrasonic beam is always perpendicular to the inspected surface.

[0176] For electromagnetic ultrasonic testing, the parameter optimization module determines the electromagnetic ultrasonic transducer coil spacing and oscillation frequency according to the conductivity, thickness and frequency range of the metal workpiece to be tested. In one embodiment of the present invention, for an aluminum alloy with a conductivity of 5.8×107 S / m and a thickness of 10 mm, the coil spacing can be selected to be 1 mm and the oscillation frequency to be 2 MHz, so that good spatial resolution can be obtained while ensuring sufficient excitation energy.

[0177] Finally, the parameter optimization module is also responsible for integrating the geometric characteristics and scanning path of the laser ultrasonic transducer to determine the laser power, frequency and polarization direction. Preferably, for a metal surface with a surface roughness Ra of 1.6μm, a Nd:YAG laser with a power of 100mJ, a wavelength of 1064nm, and a pulse width of 10ns can be selected, so that ultrasonic waves of sufficient intensity can be generated without damaging the surface being inspected.

[0178] The parameter optimization module of the present invention also includes a series of innovative algorithms for accurately calculating various detection parameters. For example, the optimal distance d between the electromagnetic ultrasonic transducer and the metal workpiece to be tested can be calculated by the following formula:

[0179] ,

[0180] in is the amplitude of the electromagnetic field sound pressure generated by the electromagnetic ultrasonic transducer coil, The performance parameters of electromagnetic ultrasonic transducer are determined through experiments. For common metal materials, The value of is usually between 0.1 and 0.5.

[0181] In addition, the contact resistance R between the electromagnetic ultrasonic transducer and the metal workpiece to be tested can be calculated by the following formula:

[0182] ,

[0183] in, is the conductivity of the metal workpiece to be tested, is the projected area of ​​the electromagnetic ultrasonic transducer surface, is the resistivity of the metal workpiece to be measured. This formula takes into account material properties and geometric factors to accurately estimate the contact resistance and thus optimize the efficiency of the transducer.

[0184] When determining the oscillation frequency of the electromagnetic ultrasonic transducer coil, the present invention adopts an innovative optimization algorithm. The algorithm takes into account the thickness of the metal workpiece to be measured. , electromagnetic ultrasonic transducer coil oscillation frequency and dielectric loss factor , the optimal parameters are determined by the following formula:

[0185] Optimal parameters = ,

[0186] in, is the correlation coefficient, which is obtained by fitting a large amount of experimental data. For example, for common carbon steel materials, the typical value may be 0.05, , , This parameter optimization method can adapt to different materials and detection conditions, greatly improving the versatility and adaptability of the system.

[0187] The thickness of the metal workpiece to be tested directly affects the propagation path and attenuation characteristics of the ultrasonic wave in the material. Thicker workpieces will cause increased attenuation of the ultrasonic signal, thus affecting the detection accuracy. It determines the distance that the ultrasound needs to penetrate, which in turn affects the signal strength and resolution. Therefore, when optimizing the detection parameters, the influence of the workpiece thickness must be considered. The oscillation frequency determines the wavelength and detection depth of the ultrasound. Higher frequencies can provide higher resolution, but weaker penetration; lower frequencies have the opposite effect, with stronger penetration but lower resolution. Oscillation frequency It directly affects the penetration depth and resolution of ultrasound. Therefore, when optimizing the detection parameters, it is necessary to select the appropriate frequency according to the specific conditions of the workpiece. The dielectric loss factor indicates the degree of absorption and dissipation of ultrasonic energy by the material. A larger dielectric loss factor means more energy loss, which will affect the signal quality and detection accuracy. Dielectric loss factor Reflects the internal loss characteristics of the material, which affects the propagation and attenuation of the ultrasonic signal. This factor must be considered when optimizing the detection parameters to ensure signal quality. The distance between the transducer and the workpiece affects the propagation path and signal strength of the ultrasonic wave. Too large a distance will cause signal attenuation, while too small a distance may cause poor coupling or interference. Distance It determines the propagation path length of the ultrasonic wave and the degree of signal attenuation. Therefore, when optimizing the detection parameters, the appropriate distance must be selected to ensure the best signal quality and detection effect.

[0188] Finally, the present invention also provides a multi-modal fusion metal micro-crack ultrasonic detection method corresponding to the above system. The method comprises the following steps:

[0189] First, the ultrasonic signal acquisition module 1 is used to detect the metal sample based on three ultrasonic diagnostic modes: electromagnetic ultrasound, laser ultrasound, and phased array ultrasound, to obtain non-destructive signal data in real time. In this step, the three ultrasonic detection modes are performed synchronously to ensure the time consistency of the data.

[0190] Secondly, the acquired lossless signal data is preprocessed by the signal preprocessing module 2, and the preprocessed lossless signal is output. The preprocessing process includes operations such as denoising, filtering and signal enhancement to improve the accuracy of subsequent analysis.

[0191] Then, through the feature alignment network 3, based on the cross-modal feature extraction module, the lossless signal features of the three input ultrasound diagnostic modalities are aligned, the input lossless signals are fused at the feature level, and the fused signal features are output. This step is the key to achieving multimodal fusion. Through the innovative network structure design, the scale and feature differences between different modal signals are effectively solved.

[0192] Next, through the image segmentation module 4, a segmentation neural network is constructed based on U-Net to perform sub-millimeter metal micro-crack segmentation on the input fusion signal features and output sub-millimeter scale crack binary labels. This step utilizes the powerful image segmentation capabilities of deep learning to accurately locate and characterize micro-cracks.

[0193] Subsequently, the three ultrasonic signals are subjected to time-frequency analysis by the time-frequency analysis module 5 to obtain frequency and amplitude information, which are aligned with the lossless signal to obtain the frequency and amplitude information of each lossless signal, and the spectrum and amplitude diagram of the three ultrasonic signals are output. Based on the obtained frequency and amplitude information and the binary crack label, the fracture strength of the ultrasonic signal is obtained to determine the crack. This step makes comprehensive use of time domain and frequency domain information to improve the reliability of crack detection.

[0194] Finally, the crack binary label and the spectrum and amplitude diagrams of the three ultrasonic signals are aligned through the segmentation result fusion module 6 to obtain the time series features, and the crack binary label and time series features are input into the weighted convolutional neural network for classification, and the crack classification results are output. This step finally gives a reliable crack detection conclusion by fusing multiple features and classification results.

[0195] Through this multi-step, multi-modal fusion method, the present invention can fully utilize the advantages of different ultrasonic detection technologies and significantly improve the detection accuracy and reliability of metal microcracks. For example, in practical applications, for complex aviation aluminum alloy structures, this method can simultaneously utilize the non-contact advantages of electromagnetic ultrasound, the high resolution of laser ultrasound, and the directionality of phased array ultrasound to achieve accurate detection of tiny fatigue cracks, with a detection rate of up to 99.5%, which greatly exceeds the performance of traditional single-modal detection methods. This is of great significance for improving the product quality and safety of high-end manufacturing industries such as aerospace, nuclear power, etc.

[0196] In order to verify the superiority of the multi-modal fusion metal micro-crack ultrasonic detection system and method of the present invention, we selected a typical application scenario for simulation experiments. This scenario simulates the micro-crack detection process of aircraft engine turbine blades, which is a very challenging detection task because turbine blades are usually made of complex high-temperature alloys and are prone to tiny fatigue cracks in harsh working environments.

[0197] The experimental setup is as follows:

[0198] We selected 10 simulated turbine blade samples, each with a size of 100mm×50mm×5mm and made of nickel-based high-temperature alloy Inconel 718. On each sample, we artificially created 10 microcracks of different sizes, with a crack length ranging from 0.1mm-2mm and a width of 10μm-50μm. These microcracks are distributed at different locations and depths of the samples to simulate various crack conditions that may occur under actual working conditions.

[0199] The following three methods were compared:

[0200] Embodiment: Multi-modal fusion metal micro-crack ultrasonic detection system of the present invention

[0201] Comparative Example 1: Traditional single-mode phased array ultrasonic testing method

[0202] Comparative Example 2: Dual-mode fusion (phased array ultrasound + electromagnetic ultrasound) detection method

[0203] The detection indicators and methods are as follows:

[0204] 1. Detection rate: the number of correctly detected cracks divided by the total number of actual cracks.

[0205] 2. Positioning accuracy: the average deviation between the detected crack center and the actual crack center.

[0206] 3. Size measurement error: the average relative error between the detected crack size and the actual crack size.

[0207] 4. Minimum detectable crack: The smallest crack size that can be reliably detected.

[0208] 5. Testing speed: the average time required to complete a comprehensive test of a sample.

[0209] 6. Noise immunity: changes in detection rate under different signal-to-noise ratio (SNR) conditions.

[0210] The test results are shown in the following table:

[0211] Detection indicators Example Comparative Example 1 Comparative Example 2 Detection rate 98.5% 85% 92% Positioning accuracy ±0.05mm ±0.2mm ±0.1mm Dimensional measurement error 5.2% 15.8% 9.6% Minimum detectable crack 0.15mm 0.5mm 0.3mm Detection speed 2.5min / sample 5min / sample 3.5min / sample Noise immunity (detection rate at SNR=5dB) 95% 70% 85%

[0212] From the above results, it can be seen that the multimodal fusion detection system of the present invention is significantly superior to the single-modal and dual-modal fusion methods in all indicators. It is particularly noteworthy that the present invention is particularly outstanding in terms of detection rate, positioning accuracy and minimum detectable crack. This is mainly due to the innovative multimodal fusion architecture and advanced signal processing algorithm of the present invention.

[0213] The specific analysis is as follows:

[0214] 1. Detection rate: The detection rate of the present invention reaches 98.5%, which is much higher than the other two methods. This shows that multimodal fusion can effectively utilize the complementary advantages of different ultrasonic detection methods and greatly improve the comprehensiveness and reliability of detection.

[0215] 2. Positioning accuracy: The positioning accuracy of the present invention reaches ±0.05mm, which is an excellent result for microcrack detection. Such high accuracy is mainly due to the innovative feature alignment network and image segmentation module in the present invention, which can accurately capture and locate tiny crack features.

[0216] 3. Dimensional measurement error: The measurement error of the present invention is only 5.2%, which means that it can more accurately assess the severity of the crack and provide a reliable basis for subsequent maintenance decisions.

[0217] 4. Minimum detectable crack: The present invention can reliably detect tiny cracks as small as 0.15 mm, which is crucial for early fault detection. In critical equipment such as aircraft engines, early detection of tiny cracks can significantly improve safety and reduce maintenance costs.

[0218] 5. Detection speed: The detection speed of the present invention is the fastest, and each sample only takes 2.5 minutes. This high efficiency is due to the parallel processing architecture and optimized algorithm design of the present invention, which enables the collection and processing of multimodal data to be highly parallelized.

[0219] 6. Anti-noise capability: Under low signal-to-noise ratio (SNR=5dB), the present invention still maintains a detection rate of 95%, which demonstrates its superior anti-noise performance. This is mainly due to the time-frequency analysis module and segmentation result fusion module in the present invention, which can effectively suppress noise interference and extract useful signal features.

[0220] Based on the above experimental results, it is found that in the turbine blade detection scenario, the best implementation scheme of the present invention is as follows:

[0221] 1. The electromagnetic ultrasonic transducer adopts meandering coil design with an operating frequency of 5MHz, which can achieve good lateral resolution while ensuring sufficient penetration depth.

[0222] 2. Laser ultrasound uses Nd:YAG laser with a wavelength of 1064nm, pulse energy of 80mJ, and pulse width of 10ns. This configuration can generate ultrasound waves of sufficient intensity without damaging the sample surface.

[0223] 3. The phased array ultrasonic transducer uses a 64-element linear array with a center frequency of 7.5MHz, which can achieve precise focusing and scanning.

[0224] 4. The feature alignment network adopts an 8-head attention mechanism and the residual dense block depth is set to 6 layers. This configuration performs best in our experiments and can effectively capture the complementary information of different modal signals.

[0225] 5. The U-Net network depth in the image segmentation module is set to 5 layers, and each layer uses a double convolution structure, which can maintain a high spatial resolution while ensuring sufficient receptive field.

[0226] 6. The time-frequency analysis module uses wavelet packet transform, with 5 decomposition layers and db4 wavelet basis. This setting can effectively capture the transient characteristics caused by cracks.

[0227] 7. The weighted convolutional neural network in the segmentation result fusion module uses ResNet-50 as the backbone network, which can alleviate the gradient vanishing problem while ensuring sufficient model capacity.

[0228] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. Multi-modal fusion metal micro-crack ultrasonic detection system, characterized in that: include: Ultrasonic signal acquisition module, used for: Metal samples are tested based on three ultrasonic diagnostic modalities: electromagnetic ultrasound, laser ultrasound and phased array ultrasound; Obtain lossless signal data in real time; The signal preprocessing module is connected to the ultrasonic signal acquisition module and is used to: Acquiring lossless signal data collected by the ultrasonic signal acquisition module; Preprocessing the lossless signal data and outputting a lossless signal; A feature alignment network, connected to the signal preprocessing module, is used to: Built on a cross-modal feature extraction module; Align input lossless signal characteristics of three ultrasonic diagnostic modalities; Perform feature-level fusion on the input lossless signal and output fusion signal features; An image segmentation module, connected to the feature alignment network, is used to: Build a segmentation neural network based on U-Net; Perform sub-millimeter metal micro-crack segmentation on the input fusion signal features; Output sub-millimeter scale crack binary labels; The time-frequency analysis module is connected to the ultrasound signal acquisition module and the image segmentation module and is used to: Perform time-frequency analysis on the three ultrasonic signals to obtain frequency and amplitude information; Aligning with the lossless signal to obtain frequency and amplitude information of each lossless signal; Output spectrum and amplitude diagrams of three ultrasonic signals; Based on the frequency, amplitude information and crack binary label, the fracture strength of the ultrasonic signal is obtained to determine the crack; The segmentation result fusion module is connected to the image segmentation module and the time-frequency analysis module and is used to: Aligning the crack binary label with the spectrum graph and amplitude graph of the three ultrasonic signals to obtain time series features; Inputting the crack binary label and the time series feature into a weighted convolutional neural network; Perform classification and output the classification results of whether there are cracks or not.

2. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 1 is characterized in that: The feature alignment network includes: Three groups of feature extractors, each of which is composed of the following units in series: A convolution operation unit, used to extract multi-scale features from input signals; Multiple residual dense blocks are used to extract and fuse different scale features of multimodal lossless signals; The global attention module is used to weight the features extracted by each residual dense block according to their global importance. The channel connection module is used to use the weighted features as the scale-channel relationship matrix, perform matrix multiplication on the input features and the scale-channel relationship matrix, and map the feature channel relationship to the global feature relationship through scale transformation to achieve feature alignment.

3. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 1 is characterized in that: The U-Net network in the image segmentation module includes: Encoder, including: Input layer; The first convolutional layer is used to extract the fusion signal features and obtain a coding feature map; A deconvolution layer, used to deconvolve the encoded feature map to obtain a decoded feature map; Decoder, including: The second convolutional layer; Output layer; The encoder further includes two atrous convolution connection modules, which respectively convolve the encoding feature map output by the encoder, and use maximum pooling and deconvolution to perform upsampling operations to obtain the encoding feature map.

4. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 1, characterized in that: The weighted convolutional neural network in the segmentation result fusion module includes: The third convolutional layer is used to perform convolution operations on the input features; Pooling layer, used to reduce the dimension of convolutional features; The third convolution layer adopts a multi-scale convolution kernel to adapt to crack features of different scales.

5. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 1, characterized in that: The feature alignment network also includes: Waveform enhancement module for: The laser signal and the electromagnetic ultrasound signal are dynamically fused in the channel dimension through the gating module; Output enhanced waveform characteristics; Cross-modal attention module for: The phased array ultrasonic signal is processed through linear transformation and self-attention mechanism on the feature map channel; Improve the signal-to-noise ratio of the image.

6. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 5, characterized in that: The waveform enhancement module comprises: A feature enhancement unit, used for performing preliminary enhancement on the input signal; Feature dimension reduction unit, used to reduce feature dimension; Door control unit, including: Characterize enhancer units; A transposed matrix subunit, used for multiplying the electromagnetic ultrasonic signal to obtain a signal one; A first feature fusion unit, used for performing fusion processing on signal one; A feature dimension increasing unit is used to increase the feature dimension and obtain signal 2; The convolution unit is used to perform convolution operation on signal 2 and output the final gating module output.

7. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 5, characterized in that: The cross-modal attention module includes: A linear transformation unit, used for performing a linear transformation on the phased array ultrasonic characteristic matrix to obtain a characteristic matrix 1; Self-attention unit, used to: Combining electromagnetic ultrasonic feature matrix and laser ultrasonic feature matrix; Based on feature matrix 1, the weight values ​​of different modal features are obtained through the self-attention mechanism; A weight product unit, used for multiplying the weight value by the feature matrix 1 to obtain the feature matrix 2; The second feature fusion unit is used to: Adding the laser ultrasonic feature matrix and the electromagnetic ultrasonic feature matrix into feature matrix 2; Obtain feature matrix three through fusion operation; The normalization unit is used to perform linear transformation and Sigmoid normalization on feature matrix 3 to obtain the final fusion feature matrix.

8. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 1, characterized in that: Also includes: The parameter optimization module is connected to the ultrasonic signal acquisition module and is used to: According to the detection requirements of metal microcracks, the frequency range, the aperture size of the phased array ultrasound, and the distance between the phased array ultrasound transducer and the metal workpiece to be tested are determined by integrating three sets of ultrasonic data; The scanning path is determined by integrating the maximum irradiation range of the phased array ultrasonic transducer and the geometric features of the inspected workpiece; Determine the coil spacing and oscillation frequency of the electromagnetic ultrasonic transducer according to the conductivity, thickness and frequency range of the metal workpiece to be tested; The laser power, frequency and polarization direction are determined by combining the geometric characteristics and scanning path of the laser ultrasonic transducer.

9. The multi-modal fusion metal micro-crack ultrasonic detection system according to claim 8, characterized in that: The parameter optimization module is also used for: According to the public , calculate the distance between the electromagnetic ultrasonic transducer and the metal workpiece to be tested ,in is the amplitude of the electromagnetic field sound pressure generated by the electromagnetic ultrasonic transducer coil, is the performance parameter of electromagnetic ultrasonic transducer; According to the formula , calculate the contact resistance between the electromagnetic ultrasonic transducer and the metal workpiece to be tested, where is the conductivity of the metal workpiece to be tested, is the projected area of ​​the electromagnetic ultrasonic transducer surface, is the contact resistance, is the resistivity of the metal workpiece to be tested; According to the relationship between the oscillation frequency of the electromagnetic ultrasonic transducer and the contact resistance, the oscillation frequency of the electromagnetic ultrasonic transducer coil is calculated; According to the thickness of the metal workpiece to be measured , electromagnetic ultrasonic transducer coil oscillation frequency , dielectric loss factor and distance , determine the optimal parameters, the formula is as follows: Optimal parameters = , in, is the correlation coefficient, which is determined by experimental fitting.

10. A multi-modal fusion metal microcrack ultrasonic detection method, characterized in that: The following steps are involved: S1: Through the ultrasonic signal acquisition module, the metal sample is tested based on three ultrasonic diagnostic modes: electromagnetic ultrasound, laser ultrasound and phased array ultrasound, and non-destructive signal data is obtained in real time; S2: Preprocessing the lossless signal data through a signal preprocessing module to output a lossless signal; S3: Through the feature alignment network, based on the cross-modal feature extraction module, the lossless signal features of the three input ultrasound diagnostic modalities are aligned, the input lossless signals are fused at the feature level, and the fused signal features are output; S4: Through the image segmentation module, a segmentation neural network is constructed based on U-Net to perform sub-millimeter metal micro-crack segmentation on the input fusion signal features and output sub-millimeter scale crack binary labels; S5: Perform time-frequency analysis on the three ultrasonic signals through the time-frequency analysis module to obtain frequency and amplitude information, align them with the lossless signal, obtain frequency and amplitude information of each lossless signal, and output spectrum diagrams and amplitude diagrams of the three ultrasonic signals; based on the frequency, amplitude information and crack binary labels, obtain the fracture strength of the ultrasonic signal and determine the crack; S6: Through the segmentation result fusion module, the crack binary label and the spectrum diagram and amplitude diagram of the three ultrasonic signals are aligned to obtain the time series features, and the crack binary label and the time series features are input into the weighted convolutional neural network for classification, and the classification result of whether there is a crack is output.

Images