A welding quality online monitoring method and system based on acoustic signals and deep learning

Abstract

The application discloses a kind of based on acoustic signal and deep learning's welding quality online monitoring method and system, comprising the following steps: S1, the acoustic signal of welding is collected, and A / D conversion is carried out, and original acoustic signal is obtained;S2, the frequency spectrum of original acoustic signal is filtered and denoising analysis processing using wavelet packet decomposition method, and two-dimensional form's spectral data is obtained;S3, welding quality state identification classification model is constructed using deep learning technology and cross attention fusion neural network;S4, two-dimensional form's spectral data is identified and classified using welding quality state identification classification model, and the four states of welding are obtained, realize the online monitoring of welding quality state.This method compared with traditional technology, acoustic signal is processed using wavelet packet decomposition method, with higher frequency resolution, cross attention fusion neural network in deep learning technology is used for identification and classification, overcome the problem of poor generalization ability of traditional model.

Description

Technical Field

[0001] The present invention relates to the technical field of welding quality monitoring, and more specifically, to an online welding quality monitoring method and system based on acoustic signals and deep learning. Background Art

[0002] Laser-arc hybrid welding is a new joining method in the welding field. Lasers offer high power density and a large aspect ratio, stabilizing arc and droplet transfer. Laser-arc hybrid welding leverages the unique characteristics of both lasers and arcs to compensate for the shortcomings of laser and arc welding. The arc improves the workpiece's absorption of the laser, increasing the efficiency of laser energy transmission and weld penetration. Laser-arc hybrid welding combines the high precision, high efficiency, and low heat input of laser welding with the excellent bridging properties of arc welding, offering a potentially valuable welding method.

[0003] The welding process involves complex physical and chemical changes, making weld quality monitoring particularly important. Currently, weld quality is difficult to measure directly. Many existing testing methods rely primarily on post-process destructive testing, which is inefficient and relatively costly. Practical and reliable testing methods are lacking in practice, particularly in the field of laser-arc hybrid welding. Laser-arc hybrid welding is often accompanied by acoustic, optical, and electrical signals. In recent years, many researchers have indirectly measured weld quality based on changes in current, voltage, or optical signals caused by physical, chemical, or energy changes during the welding process. These methods typically employ various sensors based on photoelectric dynamic signals to collect photoelectric data, which are then analyzed and tested using traditional feature extraction algorithms. However, adverse factors such as electromagnetic interference and smoke during the welding process can affect the accuracy of photoelectric dynamic signal acquisition. Therefore, using these signals to assess weld quality remains inaccurate and inconvenient.

[0004] The acoustic signal generated during laser-arc hybrid welding is the result of the laser beam energy radiating onto the base material during the welding process, forming dynamic features such as keyholes, molten pools, and spatter. The addition of the arc makes these dynamic features even more complex and variable. These changes cause oscillations in the surrounding air, generating non-stationary random signals. These random signals contain information about the welding state and are closely related to the quality of the weld. Currently, welding quality detection based on acoustic signals is most commonly seen in arc welding. Traditional methods, such as mapping between training set data and welding quality, establish a mapping relationship between the quality of the weld and the quality of the weld by performing feature analysis on the acoustic signal in the time domain, frequency domain, or time-frequency domain. However, these models have poor generalization capabilities and are unable to effectively utilize the big data from the welding process for self-learning and self-correction.

[0005] The prior art discloses a method for online monitoring and evaluation of GMAW welding quality based on integrated deep learning. The method includes: collecting welding information during the welding process, including welding pool images, arc sound signals, welding voltage signals, and welding current signals; preprocessing the welding information and performing feature extraction; concatenating and fusing the features obtained from different signal sources to construct a fused feature vector; constructing a quality-based detection model, inputting the fused feature vector into the quality-based detection model, and outputting the welding quality in the current state; using different quality-based detection models to obtain the welding quality in the current state, and obtaining a final decision based on the results of the different quality-based detection models. A drawback of this solution is that it cannot solve the problem of high-frequency signal loss during sound signal processing.

[0006] To this end, in combination with the above requirements and the defects of the existing technology, this application proposes an online monitoring method and system for welding quality based on acoustic signals and deep learning. Summary of the Invention

[0007] The present invention provides an online monitoring method for welding quality based on acoustic signals and deep learning, which can effectively reduce the interference of adverse factors such as electromagnetic interference and smoke on dynamic signal acquisition during laser-arc hybrid welding. It has higher frequency resolution, can overcome the problem of high-frequency signal loss during analysis and processing, can overcome the problem of poor generalization ability of traditional models, and can effectively utilize big data in the welding process for self-learning and self-correction.

[0008] The primary purpose of the present invention is to solve the above technical problems, and the technical solutions of the present invention are as follows:

[0009] A first aspect of the present invention provides a method for online monitoring of welding quality based on acoustic signals and deep learning, the method comprising the following steps:

[0010] S1. Collect the welding sound signal and perform A / D conversion to obtain the original sound signal.

[0011] S2. Use the wavelet packet decomposition method WPD to filter and reduce noise on the spectrum of the original sound signal to obtain two-dimensional spectrum data.

[0012] S3. Use deep learning technology and cross-attention fusion neural network CAFNET to build a welding quality status recognition and classification model.

[0013] S4. Use the welding quality status recognition and classification model to identify and classify the two-dimensional spectrum data to obtain four welding states, thereby realizing online monitoring of the welding quality status.

[0014] Further, the four states of the welding include: non-penetration, partial penetration, just penetration and over penetration; the A / D conversion is specifically: converting the collected acoustic signals into electric signals, and then into digital signals.

[0015] Further, the step S2 is specifically: defining a set of orthogonal wavelet packet functions, obtaining the wavelet packet functions under the two scales through conversion, decomposing both the high frequency and low frequency signals, obtaining the wavelet packet decomposition functions and the complete signal decomposition tree, and obtaining the two-dimensional form of the frequency spectrum data.

[0016] The wavelet packet decomposition denoising is based on the wavelet decomposition, further refines and decomposes the high frequency components which are not processed in the wavelet transformation, has higher frequency resolution, decomposes both the high frequency components and the low frequency components of the signal, is more accurate and comprehensive than the wavelet decomposition, can self-adaptively select the frequency band of the feature vector, and corresponds the feature to the frequency spectrum, thereby improving the time-frequency resolution of the signal.

[0017] Further, the wavelet packet function is specifically:

[0018]

[0019] wherein, φ(t) represents the low frequency wavelet packet function; Ψ(t) represents the high frequency wavelet packet function; {h n} n∈Z , {g n} n∈Z , g n = (-1) n h 1-n are all real coefficient filters; k represents the position coordinate; t represents the time.

[0020] Further, the wavelet packet function under the two scales is specifically:

[0021]

[0022]

[0023] wherein, μ represents the wavelet packet, the subscript of μ represents the position of the wavelet packet in the level, when n = 0, μ0(t) = (t) is the scale function, μ1(t) = (t) is the wavelet function, and μ n is a recursively defined function, wherein μ is the wavelet packet, which is determined by the orthogonal scale function μ0(t) = (t).

[0024] Further, the wavelet packet decomposition function is specifically:

[0025]

[0026] wherein, represents the low frequency part of the signal; Represents the high-frequency part of the signal;

[0027] express.

[0028] Furthermore, the step S3 is specifically as follows: according to the four welding states, the two

[0029] Four data sets were established based on dimensional spectrum data, and the cross attention fusion neural network CAFNET was used to learn the data sets to obtain a welding model that can capture acoustic features and identify and classify welding states.

[0030] Classification model for quality status identification.

[0031] Furthermore, the cross-attention fusion neural network CAFNET consists of two branches of two-dimensional convolutional neural networks 2D-CNNs and a cross-attention module CA; wherein the two-dimensional spectral data serves as the signal input of the first branch and the second branch of the two-dimensional convolutional neural network 2D-CNNs.

[0032] The two-dimensional spectrum data is input to the first branch and passes through the first convolution module and the second convolution module in sequence.

[0033] After the module, it is input into the cross attention CA module; the two-dimensional spectrum data is input into the second branch, passes through the third convolution module and the fourth convolution module in sequence, and then input into the cross attention CA module.

[0034] The output of the cross attention module and the output of the second convolution module are added element by element.

[0035] After passing through the fifth and sixth convolution modules and performing global average pooling, the output of the cross attention module and the output of the fourth convolution module are added element by element and then passed through the seventh and sixth convolution modules in sequence.

[0036] The convolution module and the eighth convolution module are then input to the fully connected layer after global average pooling.

[0037] The output of the fully connected layer is identified and classified through the normalized exponential function softmax, and the welding type recognition result is output.

[0038] A second aspect of the present invention provides an online monitoring system for welding quality based on acoustic signals and deep learning, comprising: a movable platform, an arc welding gun, a welding machine, a laser welding gun, a fiber laser and a sound collection system installed on a work platform; wherein the parts to be welded are arranged on the movable platform, the laser welding gun is fixed directly above the movable platform, the arc welding gun is fixed on one side of the movable platform, and the sound collection system is fixed on the other side of the movable platform opposite to the arc welding gun; the arc welding gun is connected to the welding machine, the laser welding gun is connected to the fiber laser, and the fiber laser and the welding machine are both connected to the sound collection system.

[0039] Furthermore, the sound collection system includes a microphone, a data acquisition card and a computer;

[0040] The microphone is connected to a data acquisition card for collecting sound signals, converting them into electrical signals and storing them; the data acquisition card is connected to the computer; the computer is used to control the welder and the fiber laser, and the computer stores an online monitoring method for welding quality based on sound signals and deep learning.

[0041] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0042] The present invention provides a method and system for online monitoring of welding quality based on acoustic signals and deep learning. The method uses a wavelet packet decomposition method to process acoustic signals, decomposing both the high-frequency and low-frequency components of the signal, thereby having higher frequency resolution and overcoming the problem of high-frequency signal loss during analysis and processing. The method adopts the cross-attention fusion neural network in deep learning technology for recognition and classification, thereby overcoming the problem of poor generalization ability of traditional models and effectively utilizing big data in the welding process for self-learning and self-correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a flow chart of an online welding quality monitoring method based on acoustic signals and deep learning in the present invention.

[0044] Figure 2 Schematic diagram of a cubic decomposition tree of wavelet packet transform in one embodiment of the present invention.

[0045] Figure 3 Schematic diagram of the structure of the cross-attention fusion neural network in the present invention.

[0046] Figure 4 This is a schematic diagram of an online welding quality monitoring system based on acoustic signals and deep learning in the present invention.

[0047] Figure 5 Schematic diagram of the overall process in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0050] Example 1

[0051] like Figure 1 As shown, the present invention provides a method for online monitoring of welding quality based on acoustic signals and deep learning, which includes the following steps:

[0052] S1. Collect the welding sound signal and perform A / D conversion to obtain the original sound signal.

[0053] S2. Use the wavelet packet decomposition method WPD to filter and reduce noise on the spectrum of the original sound signal to obtain two-dimensional spectrum data.

[0054] S3. Use deep learning technology and cross-attention fusion neural network CAFNET to build a welding quality status recognition and classification model.

[0055] S4. Use the welding quality status recognition and classification model to identify and classify the two-dimensional spectrum data to obtain four welding states, thereby realizing online monitoring of the welding quality status.

[0056] Furthermore, the four welding states include: incomplete penetration, partial penetration, just enough penetration and excessive penetration; the A / D conversion specifically includes: converting the collected acoustic signal into an electrical signal, and then converting it into a digital signal.

[0057] Furthermore, step S2 is specifically as follows: defining a set of orthogonal wavelet packet functions, obtaining wavelet packet functions at two scales through transformation, decomposing both high-frequency and low-frequency signals, obtaining wavelet packet decomposition functions and obtaining a complete signal decomposition tree, and obtaining spectrum data in two-dimensional form.

[0058] Among them, wavelet packet decomposition denoising is based on wavelet decomposition, and further refines the unprocessed high-frequency components in the wavelet transform, with higher frequency resolution. It decomposes both the high-frequency and low-frequency components of the signal, which is more refined and comprehensive than the wavelet decomposition. The eigenvector can adaptively select the frequency band, and the features correspond to the spectrum, thereby improving the time-frequency resolution of the signal.

[0059] In a specific embodiment, a signal decomposition tree obtained by decomposing a group of high-frequency and low-frequency signals through a wavelet packet function is as follows: Figure 2 As shown, S represents the original signal, A represents the low-frequency signal, and D represents the high-frequency signal.

[0060] Furthermore, the wavelet packet function is specifically:

[0061]

[0062] Among them, φ(t) represents the low-frequency wavelet packet function; Ψ(t) represents the high-frequency wavelet packet function; {h n} n∈Z 、{g n} n∈Z 、g n =(-1) n h 1-n All are real coefficient filters; k represents the position coordinate; t represents time.

[0063] Furthermore, the wavelet packet function under the two scales is specifically:

[0064]

[0065]

[0066] Where μ represents a wavelet packet (the subscript of μ represents the position of the wavelet packet in the classification); when n = 0, μ0(t) = (t) is the scaling function, μ1(t) = (t) is the wavelet function, and μ n is a recursively defined function, where is a wavelet packet determined by the orthogonal scaling function μ0(t)=(t).

[0067] Furthermore, the wavelet packet decomposition function is specifically:

[0068]

[0069] in, Represents the low-frequency part of the signal; Represents the high-frequency part of the signal; express.

[0070] Furthermore, step S3 is specifically as follows: four data sets are established according to the two-dimensional spectrum data obtained in step S2 according to the four welding states, and the cross-attention fusion neural network CAFNET is used to learn the data sets to obtain a welding quality state recognition and classification model that can capture acoustic features and identify and classify welding states.

[0071] Further, such as Figure 3As shown, the cross-attention fusion neural network CAFNET consists of two branches of two-dimensional convolutional neural networks 2D-CNNs and a cross-attention module CA; wherein, two-dimensional spectrum data is used as the signal input of the first branch and the second branch of the two-dimensional convolutional neural network 2D-CNNs.

[0072] The two-dimensional spectral data is input into the first branch, passes through the first convolution module and the second convolution module in sequence, and then is input into the cross-attention CA module; the two-dimensional spectral data is input into the second branch, passes through the third convolution module and the fourth convolution module in sequence, and then is input into the cross-attention CA module.

[0073] The output of the cross-attention module and the output of the second convolution module are added element by element, and then passed through the fifth convolution module and the sixth convolution module in sequence, and input into the fully connected layer after global average pooling; the output of the cross-attention module and the output of the fourth convolution module are added element by element, and then passed through the seventh convolution module and the eighth convolution module in sequence, and input into the fully connected layer after global average pooling.

[0074] The output of the fully connected layer is identified and classified through the normalized exponential function softmax, and the welding type recognition result is output.

[0075] Example 2

[0076] Based on the above embodiment 1, combined Figure 4 This embodiment describes in detail the second aspect of the present invention, an online welding quality monitoring system based on acoustic signals and deep learning.

[0077] like Figure 4 As shown, the second aspect of the present invention provides an online monitoring system for welding quality based on acoustic signals and deep learning, comprising: a movable platform, an arc welding gun, a welding machine, a laser welding gun, a fiber laser and a sound collection system installed on a work platform; wherein the parts to be welded are arranged on the movable platform, the laser welding gun is fixed directly above the movable platform, the arc welding gun is fixed on one side of the movable platform, and the sound collection system is fixed on the other side of the movable platform opposite to the arc welding gun; the arc welding gun is connected to the welding machine, the laser welding gun is connected to the fiber laser, and the fiber laser and the welding machine are both connected to the sound collection system.

[0078] Furthermore, the sound acquisition system includes a microphone, a data acquisition card and a computer; the microphone is connected to the data acquisition card for collecting sound signals, converting them into electrical signals and storing them; the data acquisition card is connected to the computer; the computer is used to control the welder and the fiber laser, and the computer stores an online monitoring method for welding quality based on sound signals and deep learning.

[0079] The method for online monitoring of welding quality based on acoustic signals and deep learning includes the following steps:

[0080] S1. Collect the welding sound signal and perform A / D conversion to obtain the original sound signal.

[0081] S2. Use the wavelet packet decomposition method WPD to filter and reduce noise on the spectrum of the original sound signal to obtain two-dimensional spectrum data.

[0082] S3. Use deep learning technology and cross-attention fusion neural network CAFNET to build a welding quality status recognition and classification model.

[0083] S4. Use the welding quality status recognition and classification model to identify and classify the two-dimensional spectrum data to obtain four welding states, thereby realizing online monitoring of the welding quality status.

[0084] Furthermore, the four welding states include: incomplete penetration, partial penetration, just enough penetration and excessive penetration; the A / D conversion specifically includes: converting the collected acoustic signal into an electrical signal, and then converting it into a digital signal.

[0085] Furthermore, step S2 is specifically as follows: defining a set of orthogonal wavelet packet functions, obtaining wavelet packet functions at two scales through transformation, decomposing both high-frequency and low-frequency signals, obtaining wavelet packet decomposition functions and obtaining a complete signal decomposition tree, and obtaining spectrum data in two-dimensional form.

[0086] Furthermore, the wavelet packet function is specifically:

[0087]

[0088] Among them, φ(t) represents the low-frequency wavelet packet function; Ψ(t) represents the high-frequency wavelet packet function; {h n} n∈Z 、{g n} n∈Z 、g n =(-1) n h 1-n All are real coefficient filters; k represents; t represents.

[0089] Furthermore, the wavelet packet function under the two scales is specifically:

[0090]

[0091]

[0092] Where μ represents a wavelet packet (the subscript of μ represents the position of the wavelet packet in the classification); when n = 0, μ0(t) = (t) is the scaling function, μ1(t) = (t) is the wavelet function, and μ n is a recursively defined function, where is a wavelet packet determined by the orthogonal scaling function μ0(t)=(t).

[0093] Furthermore, the wavelet packet decomposition function is specifically:

[0094]

[0095] in, Represents the low-frequency part of the signal; Represents the high-frequency part of the signal; express.

[0096] Furthermore, step S3 is specifically as follows: four data sets are established according to the two-dimensional spectrum data obtained in step S2 according to the four welding states, and the cross-attention fusion neural network CAFNET is used to learn the data sets to obtain a welding quality state recognition and classification model that can capture acoustic features and identify and classify welding states.

[0097] Furthermore, the cross-attention fusion neural network CAFNET consists of two branches of two-dimensional convolutional neural networks 2D-CNNs and a cross-attention module CA; wherein the two-dimensional spectral data serves as the signal input of the first branch and the second branch of the two-dimensional convolutional neural network 2D-CNNs.

[0098] The two-dimensional spectral data is input into the first branch, passes through the first convolution module and the second convolution module in sequence, and then is input into the cross-attention CA module; the two-dimensional spectral data is input into the second branch, passes through the third convolution module and the fourth convolution module in sequence, and then is input into the cross-attention CA module.

[0099] The output of the cross-attention module and the output of the second convolution module are added element by element, and then passed through the fifth convolution module and the sixth convolution module in sequence, and input into the fully connected layer after global average pooling; the output of the cross-attention module and the output of the fourth convolution module are added element by element, and then passed through the seventh convolution module and the eighth convolution module in sequence, and input into the fully connected layer after global average pooling.

[0100] The output of the fully connected layer is identified and classified through the normalized exponential function softmax, and the welding type recognition result is output.

[0101] Example 3

[0102] Based on the above embodiment 1 and embodiment 2, combined Figure 5, this embodiment describes in detail the specific process of the present invention.

[0103] In a specific embodiment, Figure 5 As shown, laser-arc hybrid welding is performed on a welding table using a shielding gas composed of 99% argon and 1% carbon dioxide. Arc-assisted laser welding is used, with the laser in the front and the arc in the back. The laser and arc welding guns are fixed in position, and the workpiece is clamped on a motion platform, which drives the workpiece. A microphone is fixed to the work platform after setting the optimal acquisition angle. It collects acoustic signals from the side of the work platform during the welding process. The sound acquisition system primarily includes a microphone, a data acquisition card, and a computer, which continuously collects and stores information during the welding process.

[0104] The acoustic signal during welding uses air as the medium and applies different pressures to the vibrating membrane of the microphone in the form of waves. The internal magnet will generate different currents, converting the acoustic signal into an electrical signal, which is then transmitted to the digital acquisition card for A / D conversion, converting the electrical signal into a digital signal, saving it and uploading it to the computer.

[0105] In addition to the acoustic signals that need to be collected during welding, there are also many irrelevant noises and ambient noise. Therefore, before extracting the characteristic signals, the data needs to be filtered and denoised. Unlike most scholars, this study uses wavelet packet decomposition and reconstruction (WPD) to perform noise reduction analysis.

[0106] Wavelet packet decomposition (WPD) decomposes the collected acoustic signal spectrum into 256 spectral segments. The frequency response range of a typical free-field microphone is typically 3.5-20,000 Hz, and according to Shannon's sampling theorem, the effective sampling range is between 1.75 and 10,000 Hz. In this embodiment, the acoustic signal is sampled 20 times per millisecond. Each decomposed spectral segment has 400 nodes, forming a two-dimensional spectral data format of 256 times 400. Based on this, four data sets are established based on the weld penetration status: incomplete penetration, partial penetration, just-penetrated penetration, and excessive penetration. This allows the Cross-Attention Fusion Neural Network (CAFNET) to capture acoustic features for effective recognition and classification.

[0107] The Cross Attention Fusion Neural Network (CAFNet) consists of two branches of two-dimensional convolutional neural networks (2D-CNNs) and a cross attention (CA) module. The 2D convolutional neural network takes a two-dimensional signal as input and extracts features through four convolution (CONV) modules. The cross attention module fuses the two features through the cross attention mechanism. The fused features are recognized and classified using the normalized exponential function softmax.

[0108] The icons in the accompanying drawings that describe the structural positional relationships are for illustrative purposes only and should not be construed as limitations on this patent.

[0109] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A method for online monitoring of welding quality based on acoustic signals and deep learning, wherein the welding is laser-arc hybrid welding, characterized in that: The following steps are involved: S1, collect welding sound signals and perform A / D conversion to obtain original sound signals; S2. Using wavelet packet decomposition to perform filtering and noise reduction analysis on the spectrum of the original sound signal, the wavelet packet decomposition decomposes the spectrum of the original sound signal into 256 spectrum fragments, each of which has 400 nodes, forming a two-dimensional spectrum data in the form of 256 times 400; S3. Using a cross-attention fusion neural network on the two-dimensional spectrum data obtained in S2 to construct a welding quality status recognition and classification model; the cross-attention fusion neural network is composed of a two-branch two-dimensional convolutional neural network and a cross-attention module; wherein the two-dimensional spectrum data is used as a signal input for the first branch and the second branch of the two-dimensional convolutional neural network; The two-dimensional spectrum data is input into the first branch, passes through the first convolution module and the second convolution module in sequence, and then is input into the cross attention module; the two-dimensional spectrum data is input into the second branch, passes through the third convolution module and the fourth convolution module in sequence, and then is input into the cross attention module; The output of the cross attention module and the output of the second convolution module are added element by element, and then passed through the fifth convolution module and the sixth convolution module in sequence, and are input to the fully connected layer after global average pooling; the output of the cross attention module and the output of the fourth convolution module are added element by element, and then passed through the seventh convolution module and the eighth convolution module in sequence, and are input to the fully connected layer after global average pooling; the output of the fully connected layer is identified and classified by a normalized exponential function, and a welding type recognition result is output; S4. Use the welding quality status recognition and classification model to identify and classify the two-dimensional spectrum data to obtain four welding states, thereby realizing online monitoring of the welding quality status.

2. The method for online monitoring of welding quality based on acoustic signals and deep learning according to claim 1, characterized in that: The four welding states include: incomplete penetration, partial penetration, just enough penetration and excessive penetration; the A / D conversion specifically includes: converting the collected acoustic signal into an electrical signal, and then converting it into a digital signal.

3. The method for online monitoring of welding quality based on acoustic signals and deep learning according to claim 1, characterized in that: The step S2 specifically includes: defining a set of orthogonal wavelet packet functions, obtaining wavelet packet functions at two scales through transformation, decomposing both high-frequency and low-frequency signals, obtaining wavelet packet decomposition functions and a complete signal decomposition tree, and obtaining two-dimensional spectrum data.

4. The method for online monitoring of welding quality based on acoustic signals and deep learning according to claim 3, characterized in that: The step S3 specifically comprises: establishing four data sets according to the four welding states using the two-dimensional spectrum data obtained in step S2, and using a cross-attention fusion neural network to learn the data sets to obtain a welding quality state recognition and classification model that can capture acoustic features and identify and classify welding states.

5. A welding quality online monitoring system based on acoustic signals and deep learning, the system is used to implement the welding quality online monitoring method based on acoustic signals and deep learning according to any one of claims 1 to 4, characterized in that: The system comprises: a movable platform installed on a work platform, an arc welding gun, a welding machine, a laser welding gun, a fiber laser, and a sound collection system; wherein the parts to be welded are arranged on the movable platform, the laser welding gun is fixed directly above the movable platform, the arc welding gun is fixed to one side of the movable platform, and the sound collection system is fixed to the other side of the movable platform opposite to the arc welding gun; the arc welding gun is connected to the welding machine, the laser welding gun is connected to the fiber laser, and the fiber laser and the welding machine are both connected to the sound collection system; The sound acquisition system includes a microphone, a data acquisition card and a computer; the microphone is connected to the data acquisition card, and is used to collect sound signals, convert them into electrical signals and store them; the data acquisition card is connected to the computer; the computer is used to control the welder and the fiber laser, and the computer is used to execute the online monitoring method for welding quality based on sound signals and deep learning.

Images