Metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion

Abstract

The invention discloses a metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion. The method comprises the following steps: synchronously exciting and collecting full-wave band spectrum and ultrasonic signals on the surface of a sample through laser, preprocessing the signals, inputting the preprocessed signals into a multi-modal deep learning model integrated with an LANet network for feature extraction and fusion, and realizing multi-dimensional detection by using the fused features; the analysis module is used for analyzing element components and chemical defects; a double-flow time sequence network is constructed, time domain characteristics of the ultrasonic signals are analyzed in parallel to evaluate physical defects, and frequency domain sound velocity information is analyzed to evaluate residual stress; and modeling the attenuation characteristics of the ultrasonic signal by using a deep learning model, and calculating an ultrasonic attenuation coefficient to obtain probability distribution of the grain size. According to the method, the two signals are fused, the detection accuracy is improved by utilizing the strong feature mining and nonlinear mapping capability of deep learning, and multi-dimensional comprehensive evaluation of the metal 3D printing component is realized.

Description

Technical Field

[0001] This application relates to the field of laser material testing technology, specifically to a metal 3D printing testing method and device based on emission spectroscopy and ultrasonic fusion. Background Technology

[0002] Metal 3D printing, also known as additive manufacturing, is a disruptive manufacturing technology that constructs three-dimensional entities by adding materials layer by layer, demonstrating enormous potential in high-precision fields such as aerospace and medical devices. However, the printing process of metal components is prone to introducing various defects due to complex physicochemical changes, posing a serious threat to the service performance and safety of the components. Therefore, efficient and accurate quality inspection of high-performance alloy components printed by 3D printing has become an urgent scientific and technological problem to be solved.

[0003] Currently, the inspection technologies for metal 3D printed parts are mainly divided into two categories:

[0004] Chemical composition analysis technology: represented by laser-induced breakdown spectroscopy (LIBS), which obtains the elemental composition information of materials by analyzing the plasma spectrum generated by laser ablation of materials;

[0005] Physical defect detection technology: represented by laser ultrasound, it detects physical defects such as pores and cracks inside materials by analyzing the propagation characteristics of ultrasonic waves excited by lasers in materials.

[0006] Despite the progress these technologies have made in their respective fields, they also have profound and interconnected limitations in practical applications, leading to key technical problems in current quality control systems:

[0007] (1) The detection information is one-sided and cannot form a comprehensive quality assessment: Existing detection systems can usually only detect a single indicator, that is, either analyze chemical composition or detect physical defects. However, the final performance of a component is a combination of chemical and physical properties. A component with a uniform chemical composition may have fatal internal microcracks; conversely, a component with a dense physical structure may fail to meet the standards for corrosion resistance or high-temperature performance due to the burning or segregation of key elements. This "blind men and the elephant" style of one-dimensional detection cannot provide a complete and reliable judgment on the quality of the component;

[0008] (2) The monitoring method is lagging behind, making it difficult to achieve true process control: Most inspections, especially ultrasonic inspections, are conducted offline after the component has been fully printed. The drawback of this model is fatal: for a complex component that requires tens of hours of printing, if irreversible defects occur in the early stages, these defects can only be discovered after a large amount of expensive metal powder, energy, and valuable machine time have been consumed, resulting in huge economic losses. Existing technologies lack the ability to conduct real-time online monitoring during the printing process, making it impossible to achieve "early detection, early warning, and early intervention" of defects;

[0009] (3) Data silo effect, making it impossible to conduct root cause analysis of defects: Even if chemical and physical tests are performed on the same component, the two sets of data obtained are independent "islands". Due to the lack of precise synchronization in time and space, it is difficult to accurately correlate a physical defect (such as a pore) at a certain location with a chemical anomaly at that location (such as a local deficiency of a certain element). This hinders a deeper understanding of the defect formation mechanism. For example, we cannot determine whether the instability of the molten pool at that location is caused by the burning off of a certain element, thus forming a pore.

[0010] Therefore, there is an urgent need for a new detection method that can overcome the above limitations. It should not only be able to acquire and integrate information from both chemical and physical dimensions simultaneously, but also be able to integrate into the printing process to achieve in-situ, real-time monitoring, so as to conduct a truly comprehensive, in-depth and forward-looking assessment and control of the quality of metal 3D printing. Summary of the Invention

[0011] To address the technical problems in the prior art, this application provides a method and apparatus for metal 3D printing inspection based on emission spectroscopy and ultrasonic fusion.

[0012] The metal 3D printing inspection method and device based on emission spectroscopy and ultrasonic fusion provided in this application adopts the following technical solution:

[0013] A metal 3D printing inspection method based on emission spectroscopy and ultrasound fusion, comprising:

[0014] S1. Simultaneously acquire full-band spectral signals and ultrasonic signals generated when the laser excites the sample surface;

[0015] S2. Preprocess the acquired full-band spectral signals and ultrasonic signals;

[0016] S3. Input the preprocessed full-band spectral signal and ultrasonic signal into a multimodal deep learning correction model, and perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features;

[0017] S4. Based on the extracted fusion features, establish a spectral correction model, use the spectral correction model to correct the original plasma spectrum calibration curve of the laser, and use the corrected spectral calibration curve to analyze the elemental composition and chemical defects of the sample.

[0018] S5. Construct a dual-stream time-series network to process the preprocessed ultrasonic signal. The dual-stream time-series network analyzes the time-domain waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency-domain sound velocity information to obtain the stress of the sample. Then, the dual-stream analysis results are fused to output the probability of physical defects and stress value of the sample.

[0019] S6. Apply a deep learning model to analyze the preprocessed ultrasonic signal, model the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtain the probability distribution of the sample's grain size based on the ultrasonic attenuation coefficient.

[0020] In some embodiments, the simultaneous acquisition of full-band spectral signals and ultrasonic signals generated when the sample surface is excited by laser specifically includes:

[0021] The laser pulse emitted by the laser ablates the sample surface, simultaneously generating plasma for spectral analysis at the same point, as well as ultrasonic waves excited by the recoil force generated by the expansion of the plasma.

[0022] The sample is translated using a displacement platform to scan the surface of the sample point by point, thereby defining several sample points to be tested;

[0023] At each sample point, a full-band spectral signal was acquired using a spectrometer with a wavelength range of 200-1000 nm, and an ultrasonic signal was acquired using a non-contact interferometer.

[0024] In some embodiments, the method for preprocessing the acquired full-band spectral signal includes:

[0025] The signal-to-noise ratio of the full-band spectral signal is improved by averaging several full-band spectral signal data of each sample point to be tested into one image.

[0026] The Savitzky-Golay algorithm is used to smooth the averaged full-band spectral signal;

[0027] An adaptive iterative reweighted penalized least squares method is used to subtract the background baseline from the smoothed full-band spectral signal.

[0028] After the above preprocessing, the entire full-band spectral signal data is used for subsequent analysis by using the full-spectrum selection method.

[0029] In some embodiments, the method for preprocessing the acquired ultrasound signals includes:

[0030] The acquired ultrasound signal is denoised using a multi-step processing method. First, a wavelet filtering algorithm is used, with the inner product formula as follows: , in, The result of wavelet transform is called wavelet coefficients, which are two-dimensional functions whose values ​​represent the changes in the original input signal x(t) at a specific scale factor. The degree of similarity between the wavelet function and the wavelet function at a specific position τ. The scaling factor acts as a scaling factor for the basic wavelet. To stretch or contract, The shift factor controls the translation of the wavelet function on the time axis, and x(t) is the original input signal, which is a function of time t. It is the basic wavelet, a function with a specific waveform and finite length. It is the prototype of all analytical wavelets. The superscript asterisk * represents complex conjugate. If the basic wavelet... If it is a real number function, then that is It itself; if the basic wavelet If it is a complex function, then its conjugate needs to be taken;

[0031] The selected wavelet basis is This is used for noise removal, and the threshold selection method is heuristic thresholding, calculated using the following formula: , Here, λ is the final calculated threshold. In the wavelet denoising process, if the absolute value of the wavelet coefficients is less than this threshold, they will be considered noise and set to zero; if they are greater than this threshold, they will be considered valid signals and preserved or processed accordingly. The signal value represents all data points of the entire original ultrasound signal. The number of signal values, i.e., the total length of the signal or the total number of sampling points. The risk function Risk(k) is defined as the value of k that minimizes the risk function. , Where k is a test variable used to calculate the risk function, and its value ranges from 1 to N. f(j) represents the j-th value in the sequence formed by arranging the squares of the wavelet transform coefficients of all sampling points of the signal in ascending order.

[0032] Then, the denoised time-domain signal is converted to the frequency domain, and a fast Fourier transform is performed on the denoised ultrasonic signal to obtain the frequency domain representation of the signal; then, an adaptive filtering algorithm is used in the frequency domain to remove background noise from the ultrasonic signal.

[0033] The denoised ultrasound signal is transformed from a time-domain signal to a two-dimensional time-frequency graph using a continuous wavelet transform method. The imresize function in the image processing library is then used to adjust the two-dimensional time-frequency graph to the specified resolution required by the deep learning model, so that it can be used as input for subsequent analysis.

[0034] In some embodiments, the multimodal deep learning correction model includes the following two parallel feature extraction modules:

[0035] The Stage1 module employs a dual-stream one-dimensional residual network, with two branches processing laser-induced breakdown spectroscopy and laser-ultrasound time-domain signals, respectively. Each branch contains four cascaded residual modules consisting of one-dimensional convolutional layers, batch normalization layers, and ReLU activation functions. These modules are used to extract spectral full width at half maximum (FWHM), elemental interference, and ultrasonic amplitude information after hierarchical spectral feature processing via step size 2 downsampling.

[0036] Stage 2 module: It uses an improved ResNet50 network to extract time-frequency features from the transformed two-dimensional time-frequency map. The extracted time-frequency features are processed by four levels of residual blocks to output a 2048-dimensional feature vector. The output 2048-dimensional feature vector is then reduced to 1024 dimensions by two levels of fully connected layers to extract signal area information.

[0037] In some embodiments, in step S4, the training process of the spectral correction model employs the following composite loss function optimization strategy:

[0038] The spectral reconstruction loss is measured using the mean squared error function (LMSE) to determine the difference between the model output and the true spectrum. The calculation formula is as follows: , Where E[] represents the expected value, i.e. the theoretical average value, S represents the actual spectral data, and S^ represents the spectral data reconstructed by the model. By minimizing this error, the model can accurately restore the spectral features.

[0039] For element quantification tasks, the element quantification loss is calculated using the weighted average absolute error L_MAE, which is given by the following formula: , in, The element content predicted by the model. For actual content, weighting coefficient The importance and measurement difficulty of the elements are set so that the model focuses on the prediction accuracy of key elements during training.

[0040] The learning rate adjustment uses the cosine annealing algorithm, and its formula is: , in, This represents the learning rate in the t-th training cycle. and These are the minimum and maximum values ​​of the learning rate, respectively, and T is the total number of training cycles.

[0041] In some embodiments, the dual-stream timing network includes:

[0042] a) The time-domain processing stream receives 1024 points of preprocessed ultrasonic signals and processes them using three cascaded one-dimensional residual convolutional blocks and a LANet network module. Based on the physical principle that defects cause ultrasonic signal amplitude attenuation, the stream enhances and extracts the ultrasonic signal amplitude attenuation features, and finally outputs the probability of the physical defect after time-series modeling.

[0043] b) Frequency domain processing stream, which processes the two-dimensional time-frequency graph generated by the ultrasonic signal conversion in parallel through a lightweight ResNet18 network, in order to extract and output a 512-dimensional frequency domain feature vector that captures the sound velocity change information from the two-dimensional time-frequency graph based on the acoustoelastic theory of stress and sound velocity change rate in the material.

[0044] c) A physical constraint fusion and computation layer is used to interact with the time-domain and frequency-domain features, and maps the 512-dimensional frequency-domain feature vector to stress values ​​through an acoustoelastic computation layer. The acoustoelastic coefficient K in the computation layer is set as a trainable parameter with an initial value of the material value calibrated by tensile test specimens. The model is trained using an amplitude-velocity joint loss function to achieve synchronous output of the stress values ​​with the probability of physical defects.

[0045] In some embodiments, the deep learning model is a decay feature decoupling network based on physical constraints, and the construction and application of the decay feature decoupling network includes:

[0046] a) Automated attenuation coefficient calculation: The attenuation feature decoupling network includes a trainable peak localization module, which automatically captures the amplitudes A1 and A2 of continuous echoes by analyzing the ultrasonic echo sequence, and calculates the ultrasonic attenuation coefficient based on this.

[0047] b) Feature fusion and size prediction: The attenuation coefficient and the frequency domain features after dimensionality reduction by the fully connected layer are fused in a physical constraint layer and input together into a grain size prediction subnetwork to output the probability distribution of the final grain size.

[0048] c) Dual-supervised training: The network is trained using a dual-supervised mechanism, where the main loss function uses Huber loss to constrain the predicted value of the grain size, while an auxiliary physical loss function ensures that the calculation of the attenuation coefficient satisfies the Rayleigh scattering theorem.

[0049] This application also provides a metal 3D printing inspection device based on emission spectroscopy and ultrasonic fusion, comprising:

[0050] The acquisition module is used to simultaneously acquire the full-band spectral signals and ultrasonic signals generated when the sample surface is excited by laser;

[0051] The preprocessing module is used to preprocess the acquired full-band spectral signals and ultrasonic signals;

[0052] The signal fusion module is used to input the preprocessed full-band spectral signal and ultrasound signal into a multimodal deep learning correction model, and to perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features.

[0053] The calibration module is used to establish a spectral calibration model based on the extracted fusion features, use the spectral calibration model to calibrate the original plasma spectral calibration curve of the laser, and use the calibrated spectral calibration curve to analyze the elemental composition and chemical defects of the sample.

[0054] The detection module is used to construct a two-stream temporal network to process the pre-processed ultrasonic signal. The two-stream temporal network analyzes the temporal waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency domain sound velocity information to obtain the stress of the sample. Then, it fuses the two-stream analysis results to output the probability of physical defects and the stress value of the sample. At the same time, the detection module is also used to apply a deep learning model to analyze the pre-processed ultrasonic signal, calculate the ultrasonic attenuation coefficient by modeling the attenuation characteristics of the ultrasonic signal, and obtain the probability distribution of the grain size of the sample based on the ultrasonic attenuation coefficient.

[0055] This application also provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion.

[0056] In summary, this application includes at least one of the following beneficial technical effects:

[0057] 1. Multi-dimensional signal fusion detection improves accuracy and stability: This application achieves multi-dimensional detection of elemental composition, chemical defects, physical defects, residual stress, and grain size of metal 3D printed components by simultaneously acquiring laser-induced breakdown spectra and ultrasonic signals. Utilizing the powerful feature extraction capabilities of deep learning models, deep-level features in spectral and ultrasonic signals are automatically mined and spectral correction is achieved. Through the nonlinear mapping capabilities of deep learning models, a complex relationship between spectral and ultrasonic signals is established, significantly improving detection accuracy. At the same time, the robustness of deep learning models enhances detection stability, effectively resisting interference from noise and complex environments. By processing and analyzing ultrasonic signals, the distribution maps of defects, residual stress, and grain size during the manufacturing process of printed components can be obtained. Ultimately, multi-functional composite characterization of the tested sample can be performed intuitively, realizing visualization of material detection and analysis. Using deep learning methods combined with laser-plasma spectroscopy-ultrasonic fusion probe technology, a single excitation, two signals, and simultaneous detection of multiple defects are achieved, achieving a "1+1>2" effect.

[0058] 2. Real-time online inspection enhances the monitoring capabilities of the 3D printing process for metal components: Traditional 3D printing quality inspection methods are often performed after printing is completed, making real-time monitoring impossible. This application, however, achieves real-time online inspection of each layer of powder spreading during the 3D printing process by simultaneously performing laser excitation and signal acquisition. The deep learning model can quickly process the acquired signals and output inspection results, allowing the inspection process to proceed synchronously with the printing process. This real-time inspection capability is crucial for timely detection and correction of defects during printing, helping to improve the quality and reliability of 3D printed components while reducing material waste and production costs.

[0059] 3. Non-contact non-destructive testing, no sample preparation required, simple operation: The technology in this application uses a non-contact interferometer to acquire ultrasonic signals, avoiding physical contact and damage to 3D printed components. It is also easy to integrate into industrial applications, and is particularly suitable for testing components that are difficult to access or may be damaged by contact. In addition, since this technology does not require sample preparation, it simplifies the testing process, reduces operational complexity, and improves testing efficiency. It can be widely used in aerospace, defense industry, nuclear industry and other fields, and is of great significance for improving the performance and service life of alloy components. Attached Figure Description

[0060] Figure 1 A schematic flowchart of a metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion provided for an embodiment of this application;

[0061] Figure 2 This is a schematic diagram of the structure of the relevant algorithm model in this application;

[0062] Figure 3 for Figure 1 A schematic diagram illustrating the application scenario of the metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion;

[0063] Figure 4 This is a typical laser-induced breakdown spectrum of an embodiment of this application;

[0064] Figure 5 This is a typical laser ultrasound signal from an embodiment of this application;

[0065] Figure 6 The laser ultrasonic signal processed by the filtering and denoising algorithm of this application embodiment;

[0066] Figure 7 This is a time-frequency diagram of the laser ultrasound signal after continuous wavelet transform according to an embodiment of this application;

[0067] Figure 8 This is a comparison diagram of defect-free laser ultrasound signals from an embodiment of this application. Detailed Implementation

[0068] The embodiments of the technical solution of this application will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of this application and are therefore merely examples, and should not be used to limit the scope of protection of this application.

[0069] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0070] like Figure 1 and Figure 2 As shown, this application provides a metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion, including:

[0071] S1. Simultaneously acquire full-band spectral signals and ultrasonic signals generated when the laser excites the sample surface;

[0072] S2. Preprocess the acquired full-band spectral signals and ultrasonic signals;

[0073] S3. Input the preprocessed full-band spectral signal and ultrasonic signal into a multimodal deep learning correction model, and perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features;

[0074] S4. Based on the extracted fusion features, establish a spectral correction model, use the spectral correction model to correct the original plasma spectrum calibration curve of the laser, and use the corrected spectral calibration curve to analyze the elemental composition and chemical defects of the sample.

[0075] S5. Construct a dual-stream time-series network to process the preprocessed ultrasonic signal. The dual-stream time-series network analyzes the time-domain waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency-domain sound velocity information to obtain the stress of the sample. Then, the dual-stream analysis results are fused to output the probability of physical defects and stress value of the sample.

[0076] S6. Apply a deep learning model to analyze the preprocessed ultrasonic signal, model the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtain the probability distribution of the sample's grain size based on the ultrasonic attenuation coefficient.

[0077] The specific steps are as follows:

[0078] Step S1: Simultaneously acquire the full-band spectral signal and ultrasonic signal generated when the laser excites the sample surface;

[0079] In one specific embodiment, the data acquisition process refers to as follows: Figure 3 The device shown is used for data acquisition. The specific data acquisition process is as follows:

[0080] Under the unified control of computer 1, the entire data acquisition process is precisely synchronized in time by synchronous controller 5, as follows:

[0081] (1) Sample placement and scanning: The sample 16 to be tested is placed on the displacement platform 4. The displacement platform 4 is used to move the sample 16 so that the laser 8 can perform precise point-by-point scanning on the surface of the sample 16.

[0082] (2) Synchronous excitation and signal generation: The powder spreader lays powder layer by layer, printing samples of five known element types and concentrations in each of the three substrates: aluminum alloy, stainless steel, and high-temperature alloy. The laser 8 follows the printed component for ablation excitation. The synchronous controller 5 sends a trigger command to the laser 8, causing it to emit a laser pulse with a wavelength of 1064 nm and an energy of 80 mJ. The laser beam is guided and focused onto the surface of the sample 16 (defocusing amount set to -3 mm) by the reflector 9 and the focusing lens 10, thereby ablating the sample and simultaneously generating plasma 15 and ultrasonic signal 14.

[0083] (3) Synchronous signal acquisition: Under the precise delay (2μs) trigger of the synchronous controller 5, the spectral acquisition branch and the ultrasonic acquisition branch start working simultaneously.

[0084] Spectral acquisition: The light emitted by the plasma 15 is collected by the acquisition lens 13 and guided to the fiber optic probe 11, and then to the ME5000 echelle grating spectrometer 2 for detection. The acquisition gate width is 9μs.

[0085] Ultrasonic acquisition: Simultaneously, the ultrasonic signal 14 on the sample surface is detected non-contactly by the probe 12 of the dual-wavelength hybrid interferometer 7 with a wavelength of 1064 nm.

[0086] (4) Data Processing and Transmission: The two acquired signals are sent to a data acquisition and control unit, including a synchronization controller 5 and an acquisition card 6. The acquisition card 6 within this unit performs digitization at a sampling rate of 400 kHz and finally transmits the data to a computer 1 for storage and subsequent analysis. At each scanning point, the above process is repeated 400 times to acquire 400 pairs of full-band LIBS spectra and laser ultrasound signals. Typical spectral and ultrasonic waveforms are as follows: Figure 4 and Figure 5 As shown.

[0087] Step 2: Preprocess the acquired full-band spectral signals and select characteristic spectral lines, and perform filtering, noise reduction, and frequency domain transformation on the acquired ultrasonic signals.

[0088] Specifically, the preprocessing of the acquired full-band spectral signal refers to spectral averaging, spectral line smoothing, and spectral background baseline subtraction of the LIBS spectrum.

[0089] Furthermore, the 400 spectra of each sample were averaged into one 10-spectrum average to perform spectral averaging on the LIBS spectrum of the sample under test.

[0090] The Savitzky-Golay algorithm was used to smooth the LIBS spectra of the sample under test.

[0091] An adaptive iterative reweighted penalized least squares method was used to subtract the spectral background baseline from the LIBS spectrum of the sample under test. The characteristic spectral lines were selected using the full spectrum selection method for subsequent analysis.

[0092] Furthermore, the ultrasonic signal preprocessing includes denoising and frequency domain transformation. The filtering and denoising algorithm for laser ultrasonic signals uses wavelet filtering, and its inner product formula is: , in, The result of wavelet transform is called wavelet coefficients, which are two-dimensional functions whose values ​​represent the changes in the original input signal x(t) at a specific scale factor. The degree of similarity between the wavelet function and the wavelet function at a specific position τ. The scaling factor acts as a scaling factor for the basic wavelet. To stretch or contract, The shift factor controls the translation of the wavelet function on the time axis, and x(t) is the original input signal, which is a function of time t. It is the basic wavelet, a function with a specific waveform and finite length. It is the prototype of all analytical wavelets. The superscript asterisk * represents complex conjugate. If the basic wavelet... If it is a real number function, then that is It itself; if the basic wavelet If it is a complex function, then its conjugate needs to be taken.

[0093] The selected wavelet basis is This is used for noise removal, and the threshold selection method is heuristic thresholding, calculated using the following formula: , Where λ is the final calculated threshold. In wavelet denoising, if the absolute value of the wavelet coefficients is less than this threshold, they are considered noise and set to zero; if they are greater than this threshold, they are considered valid signals and are retained or processed accordingly. j Let be the signal value, representing all data points of the entire original ultrasound signal, where j is a natural number greater than 0. The number of signal values, i.e., the total length of the signal or the total number of sampling points. The risk function Risk(k) is defined as the value of k that minimizes the risk function. , Where k is a test variable used to calculate the risk function, and its value ranges from 1 to N. f(j) represents the j-th value in the sequence formed by arranging the squares of the wavelet transform coefficients of all sampling points of the signal in ascending order.

[0094] To better utilize the frequency characteristics of ultrasonic signals, it is necessary to convert the denoised time-domain signal to the frequency domain and perform a fast Fourier transform on the denoised ultrasonic signal to obtain the frequency domain representation of the signal.

[0095] In the frequency domain, an adaptive filtering algorithm can be used to subtract background noise from the ultrasonic signal of the sample under test. Background noise subtraction can further improve the signal-to-noise ratio, making the ultrasonic signal clearer and facilitating subsequent feature extraction and analysis. A typical ultrasonic signal after filtering and denoising is shown below. Figure 6 As shown.

[0096] Continuous wavelet transform is used to convert the denoised ultrasonic signal from the time domain to the time-frequency domain. The typical time-frequency plot generated is shown below. Figure 7As shown, the `imresize` function from the image processing library is used to adjust the wavelet transform result to the specified resolution (256×256) required by the deep learning model, thus meeting the input requirements of the subsequent deep learning model.

[0097] Step S3: The preprocessed full-band spectral signal and ultrasonic signal are input into a multimodal deep learning correction model, and feature extraction and fusion are performed through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features.

[0098] Please refer to Figure 2 The multimodal deep learning correction model includes the following two parallel feature extraction modules:

[0099] The Stage1 module employs a dual-stream one-dimensional residual network, with two branches processing laser-induced breakdown spectroscopy and laser-ultrasound time-domain signals, respectively. Each branch contains four cascaded residual modules consisting of one-dimensional convolutional layers, batch normalization layers, and ReLU activation functions. These modules are used to extract spectral full width at half maximum (FWHM), elemental interference, and ultrasonic amplitude information after hierarchical spectral feature processing via step size 2 downsampling.

[0100] Stage 2 module: It uses an improved ResNet50 network to extract time-frequency features from the transformed two-dimensional time-frequency map. The extracted time-frequency features are processed by four levels of residual blocks (channel number 64→128→256→512) to output a 2048-dimensional feature vector. The output 2048-dimensional feature vector is then reduced to 1024 dimensions by two levels of fully connected layers to extract signal area information.

[0101] Furthermore, the processed stage1 and stage2 data are input into the LANet network model structure. The LANet network model structure is an innovative deep learning architecture that is suitable for multimodal fusion data processing. Its unique structure and processing flow significantly improve the efficiency and effectiveness of feature extraction and fusion.

[0102] Specifically, the main components include a channel attention module, a spatial attention module, a depthwise convolution module, and a wavelet pooling module. These modules are connected sequentially to form an efficient data processing flow. First, the feature parameters of stage1 and stage2 are fed into the channel attention module as input features. In this module, the input features undergo average pooling and max pooling operations to extract the global average feature and the maximum feature, respectively. These pooled features are then fed into a shared multilayer perceptron to generate channel attention weights. These weights can dynamically adjust the importance of each channel feature, highlighting key features and suppressing unimportant features, thereby enhancing LANet's ability to capture key information.

[0103] Furthermore, the features processed by the channel attention module are input into the spatial attention module. Features such as full width at half maximum (FWHM), full width at half maximum (FWHM), and area are first processed by convolutional layers to capture feature correlations in the spatial dimension. Then, spatial attention weights are generated by combining max pooling and average pooling operations. These weights are used to further enhance features at important spatial locations in the feature map while suppressing unimportant regions, enabling the model to more accurately locate and analyze key regions in the signal.

[0104] Furthermore, the features are fed into a deep convolution module, which processes them using deep convolution operations. Each input channel is convolved individually, and the results are then concatenated. This approach not only effectively reduces computation but also extracts more discriminative deep features, enabling the model to achieve stronger feature representation capabilities with limited computational resources. Finally, the features processed by the deep convolution module are fed into a wavelet pooling module. The wavelet pooling module takes a 3-channel input and gradually increases the number of channels in the feature map through a series of filters and feature map mappings, achieving feature dimensionality reduction. During dimensionality reduction, wavelet pooling retains more high-frequency information, which is crucial for capturing subtle changes in the signal.

[0105] Furthermore, in step S4, the training process of the spectral correction model employs the following composite loss function optimization strategy:

[0106] The spectral reconstruction loss is measured using the mean squared error function (LMSE) to determine the difference between the model output and the true spectrum. The calculation formula is as follows: , Where E[] represents the expected value, i.e. the theoretical average value, S represents the actual spectral data, and S^ represents the spectral data reconstructed by the model. By minimizing this error, the model can accurately restore the spectral features.

[0107] For element quantification tasks, the element quantification loss is calculated using the weighted average absolute error L_MAE, which is given by the following formula: , in, The element content predicted by the model. For actual content, weighting coefficient The importance and measurement difficulty of the elements are set so that the model focuses on the prediction accuracy of key elements during training.

[0108] The learning rate adjustment uses the cosine annealing algorithm, and its formula is: , in, This represents the learning rate in the t-th training cycle. and These are the minimum and maximum values ​​of the learning rate, respectively, and T is the total number of training cycles.

[0109] This strategy dynamically adjusts the learning rate by simulating the periodic changes of the cosine function, enabling the model to converge quickly in the early stages of training, while finely adjusting the parameters in the later stages of training to avoid getting trapped in local optima, thereby improving the model's generalization ability.

[0110] Before applying the model, it is necessary to comprehensively evaluate the performance of each module to ensure that it can complete the preprocessing, qualitative and quantitative tasks. For the prediction effect of the spectral quantitative correction module, the relative standard deviation (RSD) is selected as the evaluation index for the spectral volatility after the spectrum is processed by the LANet network of this application. The mathematical definition of RSD is shown in formula (1). For the qualitative task (the prediction effect of the category prediction module), the accuracy and confusion matrix are selected as evaluation indexes as shown in formula (2). For the quantitative task (the prediction effect of the content prediction module), absolute and relative indicators are used respectively, including the mean absolute error (MAE) shown in formula (3), the root mean square error (RMSE) shown in formula (4), and the coefficient of determination (R²) shown in formula (5). 2 The specific formulas are as follows: (1) (2) (3) (4) (5) in, The standard deviation of spectral intensity Represents the average value of the spectral intensity; The value is 1 if the sample is correctly classified, and 0 otherwise. Indicates the number of samples; Indicates the first Observations (labels) of each sample Mark the Predicted values ​​for each sample This represents the average of all sample observations.

[0111] Step S5: Construct a dual-stream time-series network to process the preprocessed ultrasonic signal. The dual-stream time-series network analyzes the time-domain waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency-domain sound velocity information to obtain the stress of the sample. Then, the dual-stream analysis results are fused to output the probability of physical defects and stress value of the sample.

[0112] The amplitude of the ultrasonic signal acquired in the defect-free area is significantly greater than that in the defective area. This is because during the propagation of the surface laser ultrasonic wave, a portion of the ultrasonic signal is reflected by the defective surface, resulting in energy loss and a smaller amplitude in the final acquired signal. Figure 8 As shown, the laser-ultrasonic signals from defective and defect-free regions are compared. Simultaneously, the ultrasonic propagation signal acquired by a non-contact interferometer is used to calculate the actual propagation time of the ultrasonic wave from the delay time, and then the speed of sound can be obtained from the propagation distance. Based on the theory of acoustoelasticity, it can be known that: ,stress and The rate of change is related; by analyzing the ultrasonic time-domain signal at the receiving point, the velocity of sound with residual stress is obtained. and the speed of sound under stress-free conditions The ratio of these two values ​​can be used to obtain the rate of change and the acoustic elastic coefficient. It can be determined through tensile testing of a specimen.

[0113] Furthermore, this application constructs a dual-stream temporal network in the laser ultrasonic signal processing branch to achieve synchronous detection of defects and stress. The temporal processing receives 1024 points of preprocessed ultrasonic signals, extracts waveform features through three cascaded one-dimensional residual convolutional blocks (channel number 32→64→128, kernel size 5-3-3), and each residual block is input into the LANet network module, which strengthens the amplitude attenuation features caused by defects through learnable weights.

[0114] Furthermore, the parallel frequency domain processing stream inputs the time-frequency map generated by the continuous wavelet transform into a lightweight ResNet18 network, and outputs a 512-dimensional frequency domain feature vector to capture the information on sound speed variation.

[0115] The two-stream features interact in the physical constraint fusion layer: the temporal features are modeled using LSTM time series to output the defect probability. , in This is a time-hidden state. Frequency domain characteristics;

[0116] Frequency domain features are achieved through stress mapping via an acoustoelastic computation layer. , Among them, the speed of sound The network automatically decouples the time-frequency features, and K is initialized as a trainable parameter to the material calibration value;

[0117] The amplitude-velocity of sound joint loss function is used during training; ,

[0118] Step S6: Apply a deep learning model to analyze the preprocessed ultrasonic signal, model the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtain the probability distribution of the sample's grain size based on the ultrasonic attenuation coefficient.

[0119] When ultrasound propagates in a metallic material, it produces echoes due to reflection and scattering by the grains. The loss of sound wave energy during this process is manifested as attenuation, which is related to the grain size of the material.

[0120] Specifically, the grain size of metallic materials affects the scattering and absorption of sound waves, thus leading to the attenuation of ultrasonic energy. This attenuation can be obtained by measuring the change in the amplitude of continuous echoes, known as the ultrasonic attenuation coefficient. , is a linear ultrasonic property that quantifies the loss of ultrasonic amplitude due to scattering, absorption, or friction when ultrasonic waves propagate through a material. Rayleigh scattering theorem is the primary attenuation mechanism and can be obtained through the following formula: ,

[0121] In calculation At that time, the peak amplitudes of the first and second peak echoes were used as... and The value of is obtained. After obtaining the ultrasonic attenuation coefficient, a correlation can be established with the grain size, and the change in grain size can be obtained through the ultrasonic attenuation coefficient.

[0122] Furthermore, the deep learning model is a physical constraint-based attenuation feature decoupling network. When constructing the attenuation feature decoupling network in this application, peak detection is first performed on the processed ultrasonic echo sequence, and the attenuation coefficient is calculated using a differentiable algorithm. , A1 and A2 are automatically captured by the trainable peak localization module.

[0123] This coefficient, along with the frequency domain features, is input into the grain size prediction subnetwork: After dimensionality reduction through 5 fully connected layers (512→256→128→64→32), the frequency domain features are fused with the attenuation coefficient in the physical constraint layer. ,

[0124] In the formula, C is set as a learnable parameter as a material constant, and f is the 5MHz center frequency.

[0125] The network output is designed with a dual-supervision mechanism: the main loss function uses Huber loss to constrain the predicted grain size. Auxiliary physical loss ensures that the attenuation coefficient satisfies Rayleigh scattering theorem. .

[0126] This application also provides a metal 3D printing inspection device based on emission spectrum and ultrasonic fusion, including an acquisition module, a preprocessing module, a signal fusion module, a calibration module and a detection module.

[0127] The acquisition module is used to simultaneously acquire full-band spectral signals and ultrasonic signals generated when the sample surface is excited by laser;

[0128] The preprocessing module is used to preprocess the acquired full-band spectral signals and ultrasonic signals;

[0129] The signal fusion module is used to input the preprocessed full-band spectral signal and ultrasonic signal into a multimodal deep learning correction model, and to perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features;

[0130] The calibration module is used to establish a spectral calibration model based on the extracted fusion features, use the spectral calibration model to calibrate the original plasma spectral calibration curve of the laser, and use the calibrated spectral calibration curve to analyze the elemental composition and chemical defects of the sample.

[0131] The detection module is used to construct a two-stream time-series network to process the pre-processed ultrasonic signal. The two-stream time-series network analyzes the time-domain waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency-domain sound velocity information to obtain the stress of the sample. Then, the two-stream analysis results are fused to output the probability of physical defects and the stress value of the sample. At the same time, the detection module is also used to apply a deep learning model to analyze the pre-processed ultrasonic signal, to calculate the ultrasonic attenuation coefficient by modeling the attenuation characteristics of the ultrasonic signal, and to obtain the probability distribution of the grain size of the sample based on the ultrasonic attenuation coefficient.

[0132] This application also provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion.

[0133] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0134] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0135] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0136] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.

Claims

1. A method for inspecting metal 3D printing based on emission spectroscopy and ultrasonic fusion, characterized in that, Includes the following steps: S1. Simultaneously acquire full-band spectral signals and ultrasonic signals generated when the laser excites the sample surface; S2. Preprocess the acquired full-band spectral signals and ultrasonic signals; S3. Input the preprocessed full-band spectral signal and ultrasonic signal into a multimodal deep learning correction model, and perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features; S4. Based on the extracted fusion features, establish a spectral correction model, use the spectral correction model to correct the original plasma spectrum calibration curve of the laser, and use the corrected spectral calibration curve to analyze the elemental composition and chemical defects of the sample. S5. Construct a dual-stream time-series network to process the preprocessed ultrasonic signal. The dual-stream time-series network analyzes the time-domain waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency-domain sound velocity information to obtain the stress of the sample. Then, the dual-stream analysis results are fused to output the probability of physical defects and stress value of the sample. S6. Apply a deep learning model to analyze the preprocessed ultrasonic signal, model the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtain the probability distribution of the sample's grain size based on the ultrasonic attenuation coefficient.

2. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 1, characterized in that: Simultaneous acquisition of full-band spectral signals and ultrasonic signals generated when the sample surface is excited by laser, specifically including: The laser pulse emitted by the laser ablates the sample surface, simultaneously generating plasma for spectral analysis at the same point, as well as ultrasonic waves excited by the recoil force generated by the expansion of the plasma. The sample is translated using a displacement platform to scan the surface of the sample point by point, thereby defining several sample points to be tested; At each sample point, a full-band spectral signal was acquired using a spectrometer with a wavelength range of 200-1000 nm, and an ultrasonic signal was acquired using a non-contact interferometer.

3. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 2, characterized in that: Methods for preprocessing the acquired full-band spectral signals include: The signal-to-noise ratio of the full-band spectral signal is improved by averaging several full-band spectral signal data of each sample point to be tested into one image. The Savitzky-Golay algorithm is used to smooth the averaged full-band spectral signal; An adaptive iterative reweighted penalized least squares method is used to subtract the background baseline from the smoothed full-band spectral signal. After the above preprocessing, the entire full-band spectral signal data is used for subsequent analysis by using the full-spectrum selection method.

4. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 1, characterized in that: Methods for preprocessing acquired ultrasound signals include: The acquired ultrasound signal is denoised using a multi-step processing method. First, a wavelet filtering algorithm is used, with the inner product formula as follows: ， in, The result of wavelet transform is called wavelet coefficients, which are two-dimensional functions whose values ​​represent the changes in the original input signal x(t) at a specific scale factor. The degree of similarity between the wavelet function and the wavelet function at a specific position τ. The scaling factor acts as a scaling factor for the basic wavelet. To stretch or contract, The shift factor controls the translation of the wavelet function on the time axis, and x(t) is the original input signal, which is a function of time t. It is the basic wavelet, a function with a specific waveform and finite length. It is the prototype of all analytical wavelets. The superscript asterisk * represents complex conjugate. If the basic wavelet... If it is a real number function, then that is It itself; if the basic wavelet If it is a complex function, then its conjugate needs to be taken; The selected wavelet basis is This is used for noise removal, and the threshold selection method is heuristic thresholding, calculated using the following formula: ， Here, λ is the final calculated threshold. In the wavelet denoising process, if the absolute value of the wavelet coefficients is less than this threshold, they will be considered noise and set to zero; if they are greater than this threshold, they will be considered valid signals and preserved or processed accordingly. The signal value represents all data points of the entire original ultrasound signal. The number of signal values, i.e., the total length of the signal or the total number of sampling points. The risk function Risk(k) is defined as the value of k that minimizes the risk function. ， Where k is a test variable used to calculate the risk function, and its value ranges from 1 to N. f(j) represents the j-th value in the sequence formed by arranging the squares of the wavelet transform coefficients of all sampling points of the signal in ascending order. Then, the denoised time-domain signal is converted to the frequency domain, and a fast Fourier transform is performed on the denoised ultrasonic signal to obtain the frequency domain representation of the signal; then, an adaptive filtering algorithm is used in the frequency domain to remove background noise from the ultrasonic signal. The denoised ultrasound signal is transformed from a time-domain signal to a two-dimensional time-frequency graph using a continuous wavelet transform method. The imresize function in the image processing library is then used to adjust the two-dimensional time-frequency graph to the specified resolution required by the deep learning model, so that it can be used as input for subsequent analysis.

5. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 4, characterized in that: The multimodal deep learning correction model includes the following two parallel feature extraction modules: The Stage1 module employs a dual-stream one-dimensional residual network, with two branches processing laser-induced breakdown spectroscopy and laser-ultrasound time-domain signals, respectively. Each branch contains four cascaded residual modules consisting of one-dimensional convolutional layers, batch normalization layers, and ReLU activation functions. These modules are used to extract spectral full width at half maximum (FWHM), elemental interference, and ultrasonic amplitude information after hierarchical spectral feature processing via step size 2 downsampling. Stage 2 module: It uses an improved ResNet50 network to extract time-frequency features from the transformed two-dimensional time-frequency map. The extracted time-frequency features are processed by four levels of residual blocks to output a 2048-dimensional feature vector. The output 2048-dimensional feature vector is then reduced to 1024 dimensions by two levels of fully connected layers to extract signal area information.

6. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 5, characterized in that: In step S4, the training process of the spectral correction model adopts the following composite loss function optimization strategy: The spectral reconstruction loss is measured using the mean squared error function (LMSE) to determine the difference between the model output and the true spectrum. The calculation formula is as follows: ， Where E[] represents the expected value, i.e. the theoretical average value, S represents the actual spectral data, and S^ represents the spectral data reconstructed by the model. By minimizing this error, the model can accurately restore the spectral features. For element quantification tasks, the element quantification loss is calculated using the weighted average absolute error L_MAE, which is given by the following formula: ， in, The element content predicted by the model. For actual content, weighting coefficient The importance and measurement difficulty of the elements are set so that the model focuses on the prediction accuracy of key elements during training. The learning rate adjustment uses the cosine annealing algorithm, and its formula is: ， in, This represents the learning rate in the t-th training cycle. and These are the minimum and maximum values ​​of the learning rate, respectively, and T is the total number of training cycles.

7. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 1, characterized in that: The dual-stream time-series network includes: a) The time-domain processing stream receives 1024 points of preprocessed ultrasonic signals and processes them using three cascaded one-dimensional residual convolutional blocks and a LANet network module. Based on the physical principle that defects cause ultrasonic signal amplitude attenuation, the stream enhances and extracts the ultrasonic signal amplitude attenuation features, and finally outputs the probability of the physical defect after time-series modeling. b) Frequency domain processing stream, which processes the two-dimensional time-frequency graph generated by the ultrasonic signal conversion in parallel through a lightweight ResNet18 network, in order to extract and output a 512-dimensional frequency domain feature vector that captures the sound velocity change information from the two-dimensional time-frequency graph based on the acoustoelastic theory of stress and sound velocity change rate in the material. c) A physical constraint fusion and computation layer is used to interact with the time-domain and frequency-domain features, and maps the 512-dimensional frequency-domain feature vector to stress values ​​through an acoustoelastic computation layer. The acoustoelastic coefficient K in the computation layer is set as a trainable parameter with an initial value of the material value calibrated by tensile test specimens. The model is trained using an amplitude-velocity joint loss function to achieve synchronous output of the stress values ​​with the probability of physical defects.

8. The metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion according to claim 7, characterized in that: The deep learning model is a decay feature decoupling network based on physical constraints. The construction and application of the decay feature decoupling network include: a) Automated attenuation coefficient calculation: The attenuation feature decoupling network includes a trainable peak localization module, which automatically captures the amplitudes A1 and A2 of continuous echoes by analyzing the ultrasonic echo sequence, and calculates the ultrasonic attenuation coefficient based on this. b) Feature fusion and size prediction: The attenuation coefficient and the frequency domain features after dimensionality reduction by the fully connected layer are fused in a physical constraint layer and input together into a grain size prediction subnetwork to output the probability distribution of the final grain size. c) Dual-supervised training: The network is trained using a dual-supervised mechanism, where the main loss function uses Huber loss to constrain the predicted value of the grain size, while an auxiliary physical loss function ensures that the calculation of the attenuation coefficient satisfies the Rayleigh scattering theorem.

9. A metal 3D printing inspection device based on emission spectroscopy and ultrasonic fusion, characterized in that: include: The acquisition module is used to simultaneously acquire the full-band spectral signals and ultrasonic signals generated when the sample surface is excited by laser; The preprocessing module is used to preprocess the acquired full-band spectral signals and ultrasonic signals; The signal fusion module is used to input the preprocessed full-band spectral signal and ultrasound signal into a multimodal deep learning correction model, and to perform feature extraction and fusion through the LANet network module integrated in the multimodal deep learning correction model; the LANet network module uses its internal attention mechanism and deep convolution operation to extract fused features. The calibration module is used to establish a spectral calibration model based on the extracted fusion features, use the spectral calibration model to calibrate the original plasma spectral calibration curve of the laser, and use the calibrated spectral calibration curve to analyze the elemental composition and chemical defects of the sample. The detection module is used to construct a two-stream temporal network to process the pre-processed ultrasonic signal. The two-stream temporal network analyzes the temporal waveform characteristics of the ultrasonic signal in parallel to evaluate the physical defects of the sample and analyzes its frequency domain sound velocity information to obtain the stress of the sample. Then, it fuses the two-stream analysis results to output the probability of physical defects and the stress value of the sample. At the same time, the detection module is also used to apply a deep learning model to analyze the pre-processed ultrasonic signal, calculate the ultrasonic attenuation coefficient by modeling the attenuation characteristics of the ultrasonic signal, and obtain the probability distribution of the grain size of the sample based on the ultrasonic attenuation coefficient.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the metal 3D printing inspection method based on emission spectroscopy and ultrasonic fusion as described in any one of claims 1-8.

Images