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

Abstract

The application discloses a metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion. The method synchronously excites and collects full-waveband spectrum and ultrasonic signals on the surface of a sample by laser. After the signals are preprocessed, they are input into a multimodal deep learning model integrated with a LANet network for feature extraction and fusion. The fused features are used for multidimensional detection. The plasma spectrum is corrected for analyzing element composition and chemical defects. A double-flow time sequence network is constructed to analyze the time domain features of the ultrasonic signals in parallel to evaluate physical defects and analyze the frequency domain sound velocity information to evaluate residual stress. The attenuation features of the ultrasonic signals are modeled by using the deep learning model, the ultrasonic attenuation coefficient is calculated, and the probability distribution of the grain size is obtained. By fusing the two signals and using the powerful feature mining and nonlinear mapping capability of deep learning, the accuracy of detection is improved, and multidimensional comprehensive evaluation of metal 3D printing components is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of laser material detection, in particular to a metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion. BACKGROUND

[0002] Metal 3D printing, also known as additive manufacturing, is a revolutionary manufacturing technology that builds three-dimensional entities by layering materials. It has great potential in high-precision fields such as aerospace and medical devices. However, during the printing process of metal components, various defects are easily introduced due to complex physical and chemical changes, posing a serious threat to the service performance and safety of the components. Therefore, efficient and accurate quality detection of 3D printed high-performance alloy components is a pressing scientific and technical problem.

[0003] Currently, the detection technology for metal 3D printed parts mainly falls into two categories:

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

[0005] Physical defect detection technology: represented by laser ultrasonic, it analyzes the propagation characteristics of ultrasonic waves excited by laser in materials to detect internal pores, cracks and other physical defects.

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

[0007] (1) Detection information is one-sided, unable to form a comprehensive evaluation of quality: existing detection systems usually only detect a single indicator, either analyzing chemical composition or detecting physical defects. However, the final performance of a component is a combination of chemical and physical properties. A component with uniform chemical composition may have fatal internal micro-cracks; conversely, a component with a dense physical structure may not meet the standards for corrosion resistance or high-temperature performance due to the loss or segregation of key elements. This "blind man touching an elephant" type of single-dimensional detection cannot provide a complete and reliable judgment of the quality of the component;

[0008] (2) Monitoring mode lags behind, making it difficult to achieve real process control: Most of the detection, especially ultrasonic detection, is offline analysis after the complete printing of the component. The disadvantage of this mode is fatal: for a complex component that needs to be printed for tens of hours, if an irreversible defect is generated at an early stage, the defect can only be discovered after a large amount of expensive metal powder, energy and valuable machine time are wasted, causing huge economic losses. The prior art lacks the ability to perform real-time online monitoring during printing, and cannot achieve "early detection, early warning and early intervention" of defects;

[0009] (3) Data island effect, unable to perform root cause analysis of defects: Even if chemical and physical detection are performed on the same component, the two sets of data obtained are independent "islands" of each other. Due to the lack of precise synchronization in time and space, it is difficult to accurately correlate a physical defect (if a pore) at a certain location with a chemical anomaly (such as the local absence of a certain element) at the same location. This hinders the in-depth understanding of the formation mechanism of the defect, for example, we cannot determine whether the burning of a certain element caused the instability of the melt pool at that location, resulting in a pore.

[0010] Therefore, there is an urgent need for a new detection method that can overcome the above limitations, which not only synchronously acquires and fuses information in both chemical and physical dimensions, but also integrates into the printing process to achieve in-situ and real-time monitoring, thereby providing a comprehensive, deep and forward-looking assessment and control of the quality of metal 3D printing. SUMMARY

[0011] In order to solve the technical problems in the prior art, the application provides a metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion.

[0012] The metal 3D printing detection method and device based on emission spectrum and ultrasonic fusion provided by the application adopts the following technical scheme:

[0013] A metal 3D printing detection method based on emission spectrum and ultrasonic fusion, comprising:

[0014] S1, synchronously acquiring full-band spectrum signals and ultrasonic signals generated when a laser excites a sample surface;

[0015] S2, preprocessing the acquired full-band spectrum signals and ultrasonic signals;

[0016] S3, input the pre-processed full-band spectral signal and the ultrasonic signal into a multi-modal deep learning correction model, and perform feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using an attention mechanism and a deep convolution operation in the LANet network module;

[0017] S4, establish a spectral correction model based on the extracted fused features, correct the original plasma spectral calibration curve of the laser using the spectral correction model, and analyze the element composition and chemical defects of the sample using the corrected spectral calibration curve;

[0018] S5, construct a dual-flow time sequence network, process the pre-processed ultrasonic signal through the dual-flow time sequence network, the dual-flow time sequence network analyzes the time domain waveform features of the ultrasonic signal in parallel to evaluate the physical defects of the sample, and analyzes the frequency domain sound velocity information to obtain the stress of the sample, and then fuses the dual-flow analysis results to output the physical defect probability and stress value of the sample;

[0019] S6, apply a deep learning model to analyze the pre-processed ultrasonic signal, model the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtain the probability distribution of the grain size of the sample according to the ultrasonic attenuation coefficient.

[0020] In some embodiments, the full-band spectral signal and the ultrasonic signal generated when the laser excites the surface of the sample are synchronously collected, specifically including:

[0021] A laser pulse emitted by a laser ablates the surface of the sample, and a plasma for spectral analysis and an ultrasonic wave excited by the back pressure generated by the expansion of the plasma are synchronously generated at the same point;

[0022] A displacement platform is used to translate the sample to scan the surface of the sample point by point to define a plurality of sample points to be measured;

[0023] At each of the sample points, a spectrometer with a wavelength range of 200-1000 nm is used to collect the full-band spectral signal, and a non-contact interferometer is used to collect the ultrasonic signal.

[0024] In some embodiments, the method for pre-processing the collected full-band spectral signal includes:

[0025] A plurality of full-band spectral signal data of each sample point to be measured are averaged into one to improve the signal-to-noise ratio of the full-band spectral signal;

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

[0027] Adaptive Iteratively Reweighted Penalized Least Squares is used to deduct the background baseline from the smoothed full-band spectrum signal.

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

[0029] In some embodiments, the method of preprocessing the collected ultrasound signal comprises:

[0030] The collected ultrasound signal is denoised by a multi-step process. First, a wavelet filtering algorithm is used, and its inner product formula is:

[0031] ,

[0032] wherein, is the result of wavelet transform, called wavelet coefficient, which is a two-dimensional function, and its value represents the degree of similarity between the original input signal x(t) and the wavelet function at a specific scale factor and a specific position τ, is the scale factor, which is used to stretch the basic wavelet , is the displacement factor, which controls the translation position of the wavelet function on the time axis, x(t) is the original input signal, which is a function of time t, is the basic wavelet, which is a function with a specific waveform and a finite length, and is the prototype of all analysis wavelets, the superscript asterisk * represents the complex conjugate, if the basic wavelet is a real function, then is itself; if the basic wavelet is a complex function, the conjugate needs to be taken;

[0033] The selected wavelet base is , which is used to remove noise, and the threshold selection method is a heuristic threshold, and its calculation formula is:

[0034] ,

[0035] wherein λ is the threshold value calculated finally, in the wavelet denoising process, if the absolute value of the wavelet coefficient is less than this threshold value, it will be considered as noise and set to zero, if it is greater than this threshold value, it will be considered as an effective signal and be retained or processed accordingly, is the signal value, which represents all data points of the entire original ultrasound signal, is the number of signal values, i.e. the total length or total number of sampling points of the signal, ​The k value corresponding to the minimum value of the risk function Risk(k) is taken as the risk function Risk(k) defined as:

[0036] ,

[0037] wherein k is a test variable used to calculate the risk function, which takes an integer value ranging from 1 to N, and f(j) represents the jth value in the sequence formed by arranging the square values of the wavelet transform coefficients of all sampling points of the signal in ascending order;

[0038] Then the denoised time domain signal is converted to the frequency domain, and the denoised ultrasonic signal is subjected to fast Fourier transform to obtain a frequency domain representation of the signal; then an adaptive filtering algorithm is used in the frequency domain to subtract background noise from the ultrasonic signal.

[0039] The denoised ultrasonic signal is converted from the time domain signal to a two-dimensional time-frequency graph by a continuous wavelet transform method, and the imresize function in the image processing library is used to adjust the two-dimensional time-frequency graph to a specified resolution required by the deep learning model, so as to serve as an input for subsequent analysis.

[0040] In some embodiments, the multi-modal deep learning correction model comprises two parallel feature extraction modules as follows:

[0041] The Stage1 module adopts a double-flow one-dimensional residual network, and two branches of the network process laser-induced breakdown spectroscopy and laser ultrasonic time domain signals respectively; each branch contains four cascaded residual modules composed of one-dimensional convolution layers, batch normalization layers and ReLU activation functions, which are used to extract spectral full width at half maximum, element interference and ultrasonic amplitude information by step-2 downsampling after hierarchical spectral feature extraction.

[0042] The Stage2 module adopts an improved ResNet50 network to extract time-frequency features from the converted two-dimensional time-frequency graph. After four-level residual block processing, a 2048-dimensional feature vector is output. The output 2048-dimensional feature vector is further reduced to 1024-dimensional by two-level fully connected layers to extract signal area information.

[0043] In some embodiments, in step S4, the training process of the spectral correction model adopts the following compound loss function optimization strategy:

[0044] The spectral reconstruction loss adopts the mean square error function LMSE to measure the difference between the model output and the real spectrum, and its calculation formula is:

[0045] ,

[0046] Wherein, E[] represents the expectation operation, that is, the theoretical average value, S represents the data of the real spectrum, and S^ represents the spectrum data reconstructed by the model. By minimizing the error, the model can accurately restore the spectrum characteristics;

[0047] For element quantification tasks, the element quantification loss adopts a weighted mean absolute error L_MAE, and its calculation formula is:

[0048] ,

[0049] Wherein, is the element content predicted by the model, is the actual content, and the weight coefficient is the importance and measurement difficulty of the element, so that the model focuses on the prediction accuracy of key elements during the training process;

[0050] The learning rate adjustment adopts a cosine annealing algorithm, and its formula is:

[0051] ,

[0052] Wherein, represents the learning rate of the t-th training cycle, and are the minimum and maximum values of the learning rate, respectively, and T is the total number of training cycles.

[0053] In some embodiments, the dual-flow time sequence network comprises:

[0054] a) a time domain processing flow that receives a pre-processed 1024-point ultrasonic signal and processes it using three cascaded one-dimensional residual convolution blocks and a LANet network module to strengthen and extract the ultrasonic signal amplitude attenuation features based on the physical principle that defects will cause ultrasonic signal amplitude attenuation, and finally output the physical defect probability after time sequence modeling;

[0055] b) a frequency domain processing flow that processes a two-dimensional time-frequency map generated by converting the ultrasonic signal in parallel through a lightweight ResNet18 network to extract and output a 512-dimensional frequency domain feature vector that captures the sound velocity change information from the two-dimensional time-frequency map based on the acoustic elasticity theory related to the stress and sound velocity change rate in the material;

[0056] c) a physical constraint fusion and calculation layer for interacting the time domain and frequency domain features, and mapping the 512-dimensional frequency domain feature vector to a stress value through an acoustic elasticity calculation layer, wherein the acoustic elasticity coefficient K in the calculation layer is set as a trainable parameter, and its initial value is the material value calibrated by the tensile specimen test; and by using an amplitude-sound velocity joint loss function for model training, the stress value is output synchronously with the physical defect probability.

[0057] In some embodiments, the deep learning model is a physically constrained attenuation feature decoupling network, and the construction and application of the attenuation feature decoupling network comprises:

[0058] a) Automatic attenuation coefficient calculation: the attenuation feature decoupling network comprises a trainable peak positioning module, which automatically captures the amplitudes A1 and A2 of consecutive echoes by analyzing the ultrasonic echo sequence, and calculates the ultrasonic attenuation coefficient based thereon;

[0059] b) Feature fusion and size prediction: the attenuation coefficient and the frequency domain features after dimension reduction by a fully connected layer are fused in a physically constrained layer, and are jointly input into a grain size prediction sub-network to output the final probability distribution of the grain size;

[0060] c) Dual supervision mechanism training: the training of the network adopts a dual supervision mechanism, wherein a main loss function adopts Huber loss to constrain the predicted value of the grain size, and a auxiliary physical loss function ensures that the calculation of the attenuation coefficient satisfies the Rayleigh scattering theorem.

[0061] The application also provides a metal 3D printing detection device based on emission spectrum and ultrasonic fusion, comprising:

[0062] An acquisition module is configured to synchronously acquire a full-band spectrum signal and an ultrasonic signal generated when a laser excites a sample surface;

[0063] A preprocessing module is configured to preprocess the acquired full-band spectrum signal and ultrasonic signal;

[0064] A signal fusion module is configured to input the preprocessed full-band spectrum signal and ultrasonic signal into a multi-modal deep learning correction model, and perform feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using an attention mechanism and a deep convolution operation in the LANet network module;

[0065] A correction module is configured to establish a spectrum correction model based on the extracted fused features, correct a laser original plasma spectrum calibration curve using the spectrum correction model, and analyze element composition and chemical defects of a sample using the corrected spectrum calibration curve;

[0066] The detection module is configured to construct a double-flow time sequence network, process the preprocessed ultrasonic signal through the double-flow time sequence network, analyze time domain waveform features of the ultrasonic signal in parallel through the double-flow time sequence network to evaluate physical defects of the sample, analyze frequency domain sound velocity information of the ultrasonic signal to obtain stress of the sample, and fuse double-flow analysis results to output a physical defect probability and a stress value of the sample; meanwhile, the detection module is also configured to apply a deep learning model to analyze the preprocessed ultrasonic signal, model attenuation features of the ultrasonic signal to calculate an ultrasonic attenuation coefficient, and obtain a probability distribution of a grain size of the sample according to the ultrasonic attenuation coefficient.

[0067] The application further provides a computer readable storage medium storing one or more programs, which can be executed by one or more processors to implement steps in the metal 3D printing detection method based on emission spectrum and ultrasonic fusion.

[0068] In summary, the application includes at least one of the following beneficial technical effects:

[0069] 1. Multi-dimensional signal fusion detection, improving accuracy and stability: the application realizes multi-dimensional detection of element composition, chemical defects, physical defects, residual stress and grain size of the metal 3D printing component by simultaneously collecting laser-induced breakdown spectroscopy and ultrasonic signals, uses the powerful feature extraction capability of the deep learning model to automatically mine deep features in the spectroscopy signal and the ultrasonic signal, and realizes spectrum correction, uses the nonlinear mapping capability of the deep learning model to establish a complex relationship between the spectroscopy signal and the ultrasonic signal, significantly improves the accuracy of detection, and the robustness of the deep learning model enhances the stability of the detection, which can effectively resist noise and complex environment interference, through processing and analysis of the ultrasonic signal, defect, residual stress and grain size distribution map of the printing component in the manufacturing process can be obtained, and finally the measured sample can be intuitively and multifunctionally represented, realizing visualization of material detection and analysis, using the deep learning method, combining the laser plasma spectroscopy-ultrasonic fusion probe technology, realizing one excitation, two signals and multiple defects simultaneous detection, achieving the effect of "1+1>2".

[0070] 2. Real-time online detection, improving the monitoring capability of the 3D printing process of metal components: Traditional 3D printing quality detection methods are often performed after printing is completed, and real-time monitoring cannot be achieved. However, the present application realizes real-time online detection of each layer of powder production process by synchronously performing laser excitation and signal acquisition during the manufacturing process of 3D printed components. The deep learning model can quickly process the collected signals and output the detection results, so that the detection process is synchronized with the printing process. This real-time detection capability is crucial for timely discovering and correcting defects in the printing process, and helps to improve the quality and reliability of 3D printed components, while reducing material waste and production costs.

[0071] 3. Non-contact non-destructive testing, no need to prepare samples, simple operation: The present application collects ultrasonic signals using a non-contact interferometer, avoiding physical contact and damage to 3D printed components, and facilitating integration in industrial applications, especially suitable for detecting components that are difficult to contact or may be damaged by contact. In addition, since the technology does not require sample preparation, it simplifies the detection process, reduces the complexity of operation, and improves the detection efficiency. It can be widely used in aerospace, defense industry, nuclear industry and other fields, and has important significance for improving the performance and service life of alloy components. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 The flowchart of the metal 3D printing detection method based on emission spectrum and ultrasonic fusion provided by an embodiment of the present application;

[0073] Figure 2 The structural diagram of the related algorithm model of the present application;

[0074] Figure 3 The application scenario diagram of the metal 3D printing detection method based on emission spectrum and ultrasonic fusion in Figure 1

[0075] Figure 4 The typical laser-induced breakdown spectroscopy diagram of the present application embodiment;

[0076] Figure 5 The typical laser ultrasonic signal of the present application embodiment;

[0077] Figure 6 The laser ultrasonic signal processed by the filtering and denoising algorithm of the present application embodiment;

[0078] Figure 7 The time-frequency diagram of the laser ultrasonic signal after continuous wavelet transform of the present application embodiment;

[0079] Figure 8 The contrast diagram of the laser ultrasonic signal without defects of the present application embodiment. DETAILED DESCRIPTION​

[0080] The embodiments of the technical scheme of the present application will be described in detail below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical scheme of the present application, and therefore only serve as examples, but cannot limit the protection scope of the present application.

[0081] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in the present application should be understood as the usual meaning understood by the skilled person in the field to which the present application belongs.

[0082] As shown in Figure 1 and Figure 2 The present application provides a metal 3D printing detection method based on emission spectrum and ultrasound fusion, comprising:

[0083] S1, synchronously collecting full-band spectrum signals and ultrasound signals generated when a laser excites a sample surface;

[0084] S2, preprocessing the collected full-band spectrum signals and ultrasound signals;

[0085] S3, inputting the preprocessed full-band spectrum signals and ultrasound signals into a multi-modal deep learning correction model, and performing feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using its internal attention mechanism and deep convolution operation;

[0086] S4, establishing a spectrum correction model based on the extracted fusion features, correcting the original plasma spectrum calibration curve of the laser using the spectrum correction model, and analyzing the element composition and chemical defects of the sample using the corrected spectrum calibration curve;

[0087] S5, constructing a dual-flow time sequence network, processing the preprocessed ultrasound signals through the dual-flow time sequence network, the dual-flow time sequence network analyzing the time domain waveform features of the ultrasound signals in parallel to evaluate the physical defects of the sample, and analyzing the frequency domain sound velocity information to obtain the stress of the sample, and then fusing the dual-flow analysis results to output the physical defect probability and stress value of the sample;

[0088] S6, applying a deep learning model to analyze the preprocessed ultrasound signals, modeling the attenuation characteristics of the ultrasound signals to calculate the ultrasound attenuation coefficient, and obtaining the probability distribution of the grain size of the sample according to the ultrasound attenuation coefficient.

[0089] The specific steps are as follows:

[0090] Step S1: synchronously collecting full-band spectrum signals and ultrasound signals generated when a laser excites a sample surface;

[0091] In a specific embodiment, the data acquisition process is performed with reference to the apparatus as shown in Figure 3 The specific data acquisition process is as follows:

[0092] Under the unified control of the computer 1, the entire data acquisition process is precisely time-synchronized by the synchronization controller 5, as follows:

[0093] (1) Sample placement and scanning: The sample 16 to be measured is placed on the displacement platform 4, which is used to move the sample 16 so that the laser 8 can perform precise point-by-point scanning of the surface of the sample 16.

[0094] (2) Synchronous excitation and signal generation: The powder laying machine lays powder layer by layer, and prints five samples of each of the three substrates of aluminum alloy, stainless steel, and high-temperature alloy with known element types and concentrations. The laser 8 follows the printing member to perform ablation excitation, and the synchronization controller 5 sends a trigger instruction to the laser 8 to make it emit a laser pulse with a wavelength of 1064 nm and an energy of 80 mJ. The laser beam is guided and focused to the surface of the sample 16 (the defocusing amount is set to -3 mm) via the mirror 9 and the focusing lens 10, thereby ablating the sample and synchronously forming the plasma 15 and the ultrasonic signal 14.

[0095] (3) Signal synchronous acquisition: Under the precise delay (2 μs) trigger of the synchronization controller 5, the spectral acquisition branch and the ultrasonic acquisition branch start working at the same time.

[0096] Spectral acquisition: The light emitted by the plasma 15 is collected by the collection lens 13 and introduced into the optical fiber probe 11, and then introduced into the ME5000 echelle spectrometer 2 for detection, with a gate width of 9 μs.

[0097] Ultrasonic acquisition: At the same time, the detection head 12 of the wavelength 1064 nm double-wave mixing interferometer 7 non-contact detects the ultrasonic signal 14 on the surface of the sample.

[0098] (4) Data processing and transmission: The two acquired signals are sent to the data acquisition and control unit including the synchronization controller 5 and the acquisition card 6, and are digitized by the acquisition card 6 in the unit at a sampling rate of 400 kHz, and finally transmitted to the computer 1 for storage and subsequent analysis. At each scanning point, the above process is repeated 400 times to acquire 400 pairs of full-waveband LIBS spectra and laser ultrasonic signals. Typical spectral and ultrasonic waveforms are shown in Figure 4 and Figure 5 .

[0099] Step 2: Preprocessing of the acquired full-waveband spectral signals and selection of characteristic spectral lines, filtering and denoising processing of the acquired ultrasonic signals, and frequency domain conversion.

[0100] 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.

[0101] 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.

[0102] The Savitzky-Golay algorithm was used to smooth the LIBS spectra of the sample under test.

[0103] 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.

[0104] 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:

[0105] ,

[0106] 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.

[0107] The selected wavelet basis is This is used for noise removal, and the threshold selection method is heuristic thresholding, calculated using the following formula:

[0108] ,

[0109] where λ is the threshold value calculated finally, in the wavelet denoising process, if the absolute value of the wavelet coefficient is less than the threshold value, it will be considered as noise and be set to zero, if greater than the threshold value, it will be considered as an effective signal and be retained or processed accordingly, S j is the signal value, which represents all data points of the entire original ultrasonic signal, j is a natural number greater than 0, is the number of signal values, that is, the total length or total number of sampling points of the signal, is the value of k corresponding to the minimum value of the risk function Risk(k), and the risk function Risk(k) is defined as:

[0110] ,

[0111] where k is a test variable used to calculate the risk function, its value range is an integer between 1 and N, and f(j) represents the jth value in the sequence formed by arranging the square values of the wavelet transform coefficients of all sampling points of the signal in ascending order.

[0112] In order to better utilize the frequency characteristics of the ultrasonic signal, it is necessary to convert the denoised time domain signal to the frequency domain, and perform fast Fourier transform on the denoised ultrasonic signal to obtain the frequency domain representation of the signal.

[0113] In the frequency domain, an adaptive filtering algorithm can be used to subtract the background noise of the ultrasonic signal of the sample to be tested, and the background noise subtraction can further improve the signal-to-noise ratio of the signal, making the ultrasonic signal clearer, which is helpful for subsequent feature extraction and analysis. The typical ultrasonic signal after filtering and denoising is shown in Figure 6 .

[0114] The denoised ultrasonic signal is converted from the time domain to the time-frequency domain using continuous wavelet transform, and the typical time-frequency diagram generated is shown in Figure 7 . The imresize function in the image processing library is used to adjust the wavelet transform result to the specified resolution (256x256) required by the deep learning model, so as to meet the input requirements of the subsequent deep learning model.

[0115] Step S3, input the preprocessed full-band spectral signal and ultrasonic signal into a multi-modal deep learning correction model, and perform feature extraction and fusion through the LANet network module integrated in the multi-modal deep learning correction model. The LANet network module uses its internal attention mechanism and deep convolution operation to extract and fuse features.

[0116] wherein, please refer to Figure 2 , the multi-modal deep learning correction model includes the following two parallel feature extraction modules:

[0117] Stage1 module adopts a double-flow one-dimensional residual network, two branches of which respectively process laser-induced breakdown spectroscopy and laser ultrasonic time-domain signals; wherein each branch contains four cascaded residual modules composed of one-dimensional convolution layer, batch normalization layer and ReLU activation function, to extract spectral full width at half maximum, element interference and ultrasonic amplitude information through step 2 downsampling after hierarchical spectral feature extraction;

[0118] Stage2 module: an improved ResNet50 network is adopted to extract time-frequency features from the converted two-dimensional time-frequency graph. The extracted time-frequency features are processed by four residual blocks (channel number 64→128→256→512) to output a 2048-dimensional feature vector. The output 2048-dimensional feature vector is further reduced to 1024-dimensional by two fully connected layers to extract signal area information.

[0119] Further, 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 suitable for multi-modal fusion data processing. Its unique structure and processing flow significantly improves the efficiency and effectiveness of feature extraction and fusion.

[0120] Specifically, the main part contains channel attention module, spatial attention module, deep convolution module and wavelet pooling module. These modules are connected in turn to form an efficient data processing flow. First, the feature parameters of stage1 and stage2 are input as input features into the channel attention module. In this module, the input features are subjected to average pooling and maximum pooling operations to extract global average features and maximum features, respectively. These pooled features are then input into a shared multi-layer perceptron to generate channel attention weights. These weights can dynamically adjust the importance of each channel feature, highlight key features and suppress unimportant features, thereby enhancing the LANet's ability to capture key information.

[0121] Further, the features processed by the channel attention module are input into the spatial attention module. The full width at half maximum, area and other features are first processed by convolution layers to capture the feature correlation in the spatial dimension. Then, by combining maximum pooling and average pooling operations, spatial attention weights are generated. These weights are used to further enhance the features of important spatial positions in the feature map while suppressing unimportant regions, so that the model can more accurately locate and analyze the key regions in the signal.

[0122] Further, the features enter a deep convolution module, which processes the features using deep convolution operations. Each input channel is individually convolved, and the results are then spliced. This operation not only effectively reduces the amount of calculation, but also extracts more discriminative deep features, enabling the model to achieve stronger feature expression capabilities under limited computing resources. Finally, the features processed by the deep convolution module are sent to a wavelet pooling module. The wavelet pooling module takes 3-channel input and gradually increases the number of channels of the feature map through a series of filters and feature map mapping, achieving dimensionality reduction of the features. In the dimensionality reduction process, wavelet pooling can retain more high-frequency information, which is crucial for capturing subtle changes in signals.

[0123] Further, in step S4, the training process of the spectral correction model adopts the following composite loss function optimization strategy:

[0124] The spectral reconstruction loss uses the mean square error function LMSE to measure the difference between the model output and the real spectrum. The calculation formula is:

[0125] ,

[0126] Where E[] represents the expectation operation, i.e. the theoretical average value, S represents the data of the real spectrum, and S^ represents the spectrum data reconstructed by the model. By minimizing this error, the model can accurately restore the spectral characteristics;

[0127] For element quantification tasks, the element quantification loss uses the weighted average absolute error L_MAE. The calculation formula is:

[0128] ,

[0129] Where, is the predicted element content of the model, is the actual content, and the weight coefficient is set according to the importance and difficulty of measurement of the element, so that the model focuses on the prediction accuracy of key elements during the training process;

[0130] The learning rate adjustment uses the cosine annealing algorithm, whose formula is:

[0131] ,

[0132] Where, is the learning rate of the t-th training cycle, and are the minimum and maximum values of the learning rate, respectively, and T is the total number of training cycles.

[0133] 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.

[0134] 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:

[0135] (1)

[0136] (2)

[0137] (3)

[0138] (4)

[0139] (5)

[0140] 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 The predicted value for each sample, This represents the average of all sample observations.

[0141] 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.

[0142] The amplitude of the ultrasonic signal collected in the defect-free area is significantly greater than that in the defective area, because a part of the ultrasonic signal is reflected by the defect surface during the propagation of the surface laser ultrasonic wave, thereby causing energy loss, resulting in a smaller amplitude of the finally collected signal as shown in FIG. 3B, wherein the laser ultrasonic signals of the defective and defect-free areas are compared. At the same time, the ultrasonic propagation signal collected by the non-contact interferometer, the actual propagation time of the ultrasonic wave is calculated by the delay time, and the sound velocity can be obtained by the propagation distance. Based on the acoustic elasticity theory, it is known that: Figure 8

[0143] The stress is related to the change rate, by analyzing the ultrasonic time domain signal of the receiving point, the ratio of the residual stress sound velocity and the sound velocity in the stress-free state can be obtained, and the change rate of the acoustic elasticity coefficient can be obtained. The acoustic elasticity coefficient can be measured by tensile specimen test.

[0144] Further, the present application constructs a double-flow time sequence network in the laser ultrasonic signal processing branch to realize synchronous detection of defects and stress. The time domain processing receives 1024 points of pre-processed ultrasonic signals, and extracts waveform features through three cascaded one-dimensional residual convolution blocks (channel number 32→64→128, kernel size 5-3-3). Each residual block is input into a LANet network module, and the amplitude attenuation features caused by defects are strengthened through learnable weights.

[0145] Further, the parallel frequency domain processing flow inputs the time-frequency diagram generated by the continuous wavelet transform into the lightweight ResNet18 network, and outputs a 512-dimensional frequency domain feature vector to capture the sound velocity change information.

[0146] The double-flow features interact in the physical constraint fusion layer: the time domain features output the defect probability after time sequence modeling by LSTM:

[0147] ,

[0148] wherein is the time sequence hidden state, is the frequency domain feature;

[0149] The frequency domain feature realizes stress mapping through the acoustic elasticity calculation layer

[0150] ,

[0151] wherein the sound velocity is automatically decoupled from the time-frequency feature by the network, and K is initialized as the material calibration value as a trainable parameter;

[0152] ​The amplitude-sound velocity joint loss function is used in the training;

[0153] ,

[0154] Step S6: applying a deep learning model to analyze the preprocessed ultrasonic signal, modeling the attenuation characteristics of the ultrasonic signal to calculate the ultrasonic attenuation coefficient, and obtaining the probability distribution of the grain size of the sample according to the ultrasonic attenuation coefficient.

[0155] The propagation of ultrasonic waves in metal materials will produce echoes due to the reflection and scattering of grains. The loss of acoustic energy in this process is manifested as attenuation, which is related to the grain size of the material.

[0156] Specifically, the grain size of the metal material affects the scattering and absorption of the acoustic wave, resulting in the attenuation of the ultrasonic energy. This attenuation can be obtained by measuring the change in the amplitude of the continuous echo, and the ultrasonic attenuation coefficient is a linear ultrasonic attribute that quantifies the loss of ultrasonic amplitude due to scattering, absorption or friction when the ultrasonic wave propagates through the material. When the Rayleigh scattering theorem becomes the main attenuation mechanism, it can be obtained by the following formula:

[0157] ,

[0158] When calculating , the peak amplitudes of the first peak and the second peak echo are used as the values of and . After obtaining the ultrasonic attenuation coefficient, the correlation between the ultrasonic attenuation coefficient and the grain size can be established, and the change of the grain size can be obtained through the ultrasonic attenuation coefficient.

[0159] Further, the deep learning model is an attenuation feature decoupling network based on physical constraints. When constructing the attenuation feature decoupling network, the processed ultrasonic echo sequence is first subjected to peak value detection, and the attenuation coefficient is calculated by a differentiable algorithm:

[0160] ,

[0161] where A1, A2 are automatically captured by a trainable peak positioning module.

[0162] The coefficient is input into the grain size prediction sub-network together with the frequency domain feature: after the frequency domain feature is reduced by 5 layers of full connection (512→256→128→64→32), it is fused with the attenuation coefficient in the physical constraint layer:

[0163] ,

[0164] In the formula, C is set as a learnable parameter as a material constant, and f is the center frequency of 5MHz.

[0165] The network output end is designed with a double supervision mechanism: the main loss function adopts Huber loss to constrain the grain size prediction value , and the auxiliary physical loss ensures that the attenuation coefficient satisfies the Rayleigh scattering theorem

[0166] .

[0167] The application also provides a metal 3D printing detection device based on emission spectrum and ultrasonic fusion, which comprises an acquisition module, a preprocessing module, a signal fusion module, a correction module and a detection module.

[0168] The acquisition module is used for synchronously collecting full-band spectrum signals and ultrasonic signals generated when a sample surface is excited by laser;

[0169] The preprocessing module is used for preprocessing the collected full-band spectrum signals and ultrasonic signals;

[0170] The signal fusion module is used for inputting the preprocessed full-band spectrum signals and ultrasonic signals into a multi-modal deep learning correction model, and performing feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using an attention mechanism and a deep convolution operation in the LANet network module;

[0171] The correction module is used for establishing a spectrum correction model based on the extracted fusion features, correcting a laser original plasma spectrum calibration curve by using the spectrum correction model, and analyzing element composition and chemical defects of a sample by using the corrected spectrum calibration curve;

[0172] The detection module is used for constructing a double-flow time sequence network, processing the preprocessed ultrasonic signals through the double-flow time sequence network, and analyzing time domain waveform features of the ultrasonic signals in parallel through the double-flow time sequence network to evaluate physical defects of the sample and analyze frequency domain sound velocity information of the sample to obtain stress of the sample, and then fusing double-flow analysis results to output a physical defect probability and a stress value of the sample; meanwhile, the detection module is also used for applying a deep learning model to analyze the preprocessed ultrasonic signals, modeling attenuation characteristics of the ultrasonic signals to calculate an ultrasonic attenuation coefficient, and obtaining a probability distribution of grain size of the sample according to the ultrasonic attenuation coefficient.

[0173] The application also provides a computer readable storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to implement steps in the metal 3D printing detection method based on emission spectrum and ultrasonic fusion.

[0174] Those skilled in the art will appreciate that embodiments of the application can be readily used as a method, a system or a computer program product. Accordingly, the application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer-readable program code thereon for use by or in connection with an instruction execution system. Program code embodied on one or more computer-usable storage media can be downloaded over a network

[0175] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0176] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0177] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can still be modified, or some or all of the technical features can be replaced by equivalent replacements; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and the specification of the present application.

Claims

1. A metal 3D printing detection method based on emission spectrum and ultrasound fusion, characterized in that, The method comprises the following steps: S1, synchronously collecting full-band spectral signals and ultrasonic signals generated when a laser excites a sample surface; S2, preprocessing the collected full-band spectral signals and ultrasonic signals; S3, inputting the preprocessed full-band spectral signals and ultrasonic signals into a multi-modal deep learning correction model, and performing feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using an internal attention mechanism and a deep convolution operation; S4, establishing a spectral correction model based on the extracted fused features, correcting a laser original plasma spectral calibration curve using the spectral correction model, and analyzing the element composition and chemical defects of the sample using the corrected spectral calibration curve; S5, constructing a double-flow time sequence network, processing the preprocessed ultrasonic signals through the double-flow time sequence network, the double-flow time sequence network analyzes the time domain waveform features of the ultrasonic signals in parallel to evaluate the physical defects of the sample, and analyzes the frequency domain acoustic velocity information to obtain the stress of the sample, and then fuses the double-flow analysis results to output the physical defect probability and stress value of the sample; S6, applying a deep learning model to analyze the preprocessed ultrasonic signals, modeling the attenuation characteristics of the ultrasonic signals to calculate the ultrasonic attenuation coefficient, and obtaining the probability distribution of the grain size of the sample according to the ultrasonic attenuation coefficient; The double-flow time sequence network comprises: a) a time domain processing flow that receives 1024-point preprocessed ultrasonic signals and processes them using three cascaded one-dimensional residual convolution blocks and a LANet network module to strengthen and extract the ultrasonic signal amplitude attenuation features based on the physical principle that defects can cause ultrasonic signal amplitude attenuation, and finally outputs the physical defect probability after time sequence modeling; b) a frequency domain processing flow that processes a two-dimensional time-frequency map generated by converting the ultrasonic signals through a lightweight ResNet18 network in parallel to extract and output a 512-dimensional frequency domain feature vector that captures acoustic velocity change information from the two-dimensional time-frequency map based on the acoustic elasticity theory that relates stress and acoustic velocity change rate in materials; c) a physical constraint fusion and calculation layer that interacts with the time domain and frequency domain features and maps the 512-dimensional frequency domain feature vector to a stress value through an acoustic elasticity calculation layer, wherein the acoustic elasticity coefficient K in the calculation layer is set as a trainable parameter with an initial value of the material value calibrated by a tensile specimen test, and a magnitude-acoustic velocity joint loss function is used for model training to realize synchronous output of the stress value with the physical defect probability.

2. The method for metal 3D printing detection based on emission spectrum and ultrasound fusion according to claim 1, characterized in that: Synchronously collecting full-band spectral signals and ultrasonic signals generated when a laser excites a sample surface, specifically comprising: Ablating the sample surface with laser pulses emitted by a laser to synchronously generate plasma for spectral analysis at the same point and ultrasonic waves excited by the back pressure generated by the expansion of the plasma; Translating the sample using a displacement platform to perform point-by-point scanning on the surface of the sample to define a plurality of sample points to be measured; At each of the sample points, a full-band spectral signal is collected using a spectrometer with a wavelength range of 200-1000 nm, and an ultrasonic signal is collected using a non-contact interferometer.

3. The method for metal 3D printing detection based on emission spectrum and ultrasound fusion according to claim 2, characterized in that: The method for pre-processing the collected full-band spectral signal includes: The full-band spectral signal data of each sample point is averaged to improve the signal-to-noise ratio of the full-band spectral signal; The averaged full-band spectral signal is smoothed using a Savitzky-Golay algorithm; The smoothed full-band spectral signal is background-subtracted using an adaptive iteratively reweighted penalized least squares method; After the above pre-processing, the entire full-band spectral signal data is used for subsequent analysis in a full-spectrum selection manner.

4. The method for metal 3D printing detection based on emission spectrum and ultrasound fusion according to claim 1, characterized in that: The method for pre-processing the collected ultrasonic signal includes: The collected ultrasonic signal is denoised using a multi-step processing, first using a wavelet filtering algorithm with an inner product formula: ， where is the result of the wavelet transform, called wavelet coefficient, which is a two-dimensional function, the value of which indicates how similar the original input signal x(t) is to the wavelet function at a certain scale factor and a certain position τ, is the scale factor, which acts as a scaling factor for the basic wavelet , is the shift factor, which controls the shift position of the wavelet function on the time axis, x(t) is the original input signal, which is a function of time t, is the basic wavelet, which is a function with a specific waveform and a finite length, and is the prototype of all analysis wavelets, the superscript asterisk * represents the complex conjugate, if the basic wavelet is a real function, then is itself; if the basic wavelet is a complex function, the conjugate thereof needs to be taken; The selected wavelet base is , for removing noise, the threshold selection method is a heuristic threshold, and the calculation formula is: ， where λ is the threshold value calculated finally, in the process of wavelet denoising, if the absolute value of wavelet coefficient is less than this threshold value, it will be considered as noise and be set to zero, if it is greater than this threshold value, it will be considered as effective signal and be reserved or processed accordingly, is the signal value, which represents all data points of the entire original ultrasonic signal, is the number of signal values, i.e. the total length or total number of sampling points of the signal, is the value of k corresponding to the minimum value of the risk function Risk(k), and the risk function Risk(k) is defined as: ， where k is a test variable used to calculate the risk function, taking an integer value between 1 and N, and f(j) represents the square value of the wavelet transform coefficient of all sample points of the signal arranged in ascending order, and the jth value in the sequence formed; The denoised time-domain signal is then converted to the frequency domain, and the denoised ultrasonic signal is subjected to a fast Fourier transform to obtain a frequency-domain representation of the signal; then, in the frequency domain, an adaptive filtering algorithm is used to subtract background noise from the ultrasonic signal; The denoised ultrasonic signal is converted from the time-domain signal to a two-dimensional time-frequency graph using a continuous wavelet transform method, and the imresize function in the image processing library is used to adjust the two-dimensional time-frequency graph to the specified resolution required by the deep learning model, serving as the input for subsequent analysis.

5. The method for metal 3D printing detection based on emission spectrum and ultrasound fusion according to claim 4, characterized in that: The multi-modal deep learning correction model includes the following two parallel feature extraction modules: The Stage1 module uses a dual-flow one-dimensional residual network, with two branches processing laser-induced breakdown spectroscopy and laser ultrasonic time-domain signals, respectively; each branch includes four cascaded residual modules composed of one-dimensional convolution layers, batch normalization layers, and ReLU activation functions, to extract spectral full-width at half-maximum, element interference, and ultrasonic amplitude information through step-2 downsampling to realize hierarchical spectral features; The Stage2 module uses an improved ResNet50 network to extract time-frequency features from the converted two-dimensional time-frequency graph, and the extracted time-frequency features are processed through four residual blocks to output a 2048-dimensional feature vector, which is then reduced to 1024 dimensions through two fully connected layers to extract signal area information.

6. The method for metal 3D printing detection based on emission spectrum and ultrasound fusion according to claim 5, characterized in that: In step S4, the training process of the spectral correction model uses the following compound loss function optimization strategy: The spectral reconstruction loss uses the mean square error function LMSE to measure the difference between the model output and the true spectrum, and its calculation formula is: ， where E[] represents the expectation operation, i.e., the theoretical average value, S represents the true spectrum data, and S^ represents the reconstructed spectrum data of the model; by minimizing the error, the model can accurately restore the spectral features; For element quantification tasks, the element quantification loss uses the weighted average absolute error L_MAE, and its calculation formula is: ， wherein, is the predicted content of the element by the model, is the actual content, the weight coefficient is set for the importance of the element and the difficulty of measurement, so that the model focuses on the prediction accuracy of key elements during the training process; The learning rate adjustment adopts a cosine annealing algorithm, and the formula is: ， wherein, denotes the learning rate for the t-th training cycle, and are the minimum and maximum values of the learning rate, respectively, and T is the total number of training cycles.

7. The method of metal 3D printing detection based on fusion of emission spectrum and ultrasound according to claim 1, characterized in that: 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 comprises: a) Automatic attenuation coefficient calculation: the decay feature decoupling network comprises a trainable peak positioning module, which automatically captures the amplitudes A1 and A2 of consecutive echoes by analyzing the ultrasonic echo sequence, and calculates the ultrasonic attenuation coefficient based thereon; b) Feature fusion and size prediction: the attenuation coefficient and the frequency domain features after dimension reduction by the full connection layer are fused in a physical constraint layer, and are jointly input into a grain size prediction sub-network to output the final probability distribution of the grain size; c) Dual supervision mechanism training: the training of the network adopts a dual supervision mechanism, wherein the main loss function adopts Huber loss to constrain the predicted value of the grain size, and a auxiliary physical loss function ensures that the calculation of the attenuation coefficient satisfies the Rayleigh scattering theorem.

8. A metal 3D printing detection device based on emission spectrum and ultrasound fusion, characterized in that: Comprise: An acquisition module is configured to synchronously acquire full-band spectral signals and ultrasonic signals generated when a laser excites a sample surface; A preprocessing module is configured to preprocess the acquired full-band spectral signals and ultrasonic signals; A signal fusion module is configured to input the preprocessed full-band spectral signals and ultrasonic signals into a multi-modal deep learning correction model, and perform feature extraction and fusion through a LANet network module integrated in the multi-modal deep learning correction model; the LANet network module extracts and fuses features by using an attention mechanism and a deep convolution operation therein; A correction module is configured to establish a spectral correction model based on the extracted fused features, correct a laser original plasma spectral calibration curve using the spectral correction model, and analyze element composition and chemical defects of a sample using the corrected spectral calibration curve; A detection module is configured to construct a dual-flow time sequence network, process the preprocessed ultrasonic signals through the dual-flow time sequence network, analyze time domain waveform features of the ultrasonic signals in parallel through the dual-flow time sequence network to evaluate physical defects of the sample, analyze frequency domain sound velocity information of the sample to obtain stress of the sample, fuse results of dual-flow analysis to output a physical defect probability and a stress value of the sample, and analyze the preprocessed ultrasonic signals using a deep learning model to model attenuation features of the ultrasonic signals to calculate an ultrasonic attenuation coefficient, and obtain a probability distribution of grain size of the sample according to the ultrasonic attenuation coefficient; The dual-flow time sequence network comprises: a) a time domain processing flow that receives 1024-point preprocessed ultrasonic signals, and processes the signals using three cascaded one-dimensional residual convolution blocks and a LANet network module to strengthen and extract amplitude attenuation features of the ultrasonic signals based on a physical principle that defects cause amplitude attenuation of the ultrasonic signals, and finally outputs the physical defect probability after time sequence modeling; b) a frequency domain processing stream that processes a two-dimensional time-frequency map generated by the ultrasonic signal conversion through a lightweight ResNet18 network in parallel to extract and output a 512-dimensional frequency domain feature vector capturing the sound speed change information from the two-dimensional time-frequency map for the acoustic elasticity theory related to the stress and sound speed change rate in the material; c) a physical constraint fusion and calculation layer for interacting the time domain and frequency domain features and mapping the 512-dimensional frequency domain feature vector to a stress value through an acoustic elasticity calculation layer, wherein the acoustic elasticity coefficient K in the calculation layer is set as a trainable parameter with an initial value of the material value calibrated by the tensile specimen test; and the model training is performed by using an amplitude-sound speed joint loss function to achieve the synchronous output of the stress value with the probability of the physical defects.

9. A computer-readable storage medium, characterized in that, The computer readable storage medium stores one or more programs that can be executed by one or more processors to implement the steps of the metal 3D printing detection method based on emission spectrum and ultrasonic fusion as claimed in any one of claims 1-7.

Images