A Pretreatment and Feature Extraction Method for Deep Penetration K-TIG Welding Multi-Source Information

Abstract

The present invention relates to a method for preprocessing and feature extraction of multi-source information in keyhole plasma arc welding (K-TIG welding), which includes signal processing and feature extraction of voltage signals, arc sound signals, arc-keyhole-molten pool images, and arc pressure signals in K-TIG welding. Different preprocessing algorithms are developed for different characteristic information in the welding process, which is beneficial to suppressing noise components and retaining useful detailed features, thereby greatly improving the signal-to-noise ratio of welding signals. The multi-source signals in the welding process contain rich information on the welding dynamic process. The present invention specifically extracts effective features of the multi-source welding information, which is beneficial to establishing the mapping relationship between signal features, the welding dynamic process, and welding quality. The present invention preprocesses different characteristic information in the welding process, improves the signal-to-noise ratio of welding signals, specifically extracts effective features of the multi-source welding information, is beneficial to establishing the mapping relationship between signal features, the welding dynamic process, and welding quality, and is applicable to the field of welding information processing.

Description

Technical Field

[0001] The present invention belongs to the technical field of welding signal processing, and particularly relates to a method for preprocessing and feature extraction of multi-source information in keyhole plasma tungsten inert gas (K-TIG) welding. Background Art

[0002] To achieve the automation and intelligence of welding technology, it is necessary to monitor the dynamic welding process and provide decision-making information for regulating welding parameters to achieve high-quality welding. The multi-source information in the dynamic welding process refers to various physical information with different physical sources during the welding process, reflecting the real-time welding state, and needs to be collected by various sensors. Through the established multi-source information sensing system for the K-TIG welding process, electrical signals, arc sound signals, and visual signals during the welding process are synchronously collected. Through the designed sandwich experiment and arc pressure measurement experiment, arc-pool-keyhole profile images and arc pressure distribution information can be obtained.

[0003] However, the obtained original electrical signals and arc sound signals are all noisy signals. Especially for the arc sound signal, it also contains the sound of the cooling fan of the robot control cabinet and the sound of the welder cooling collected. It is urgent to develop corresponding signal preprocessing methods for noise reduction of electrical signals and obtaining pure arc sound signals. Then, by extracting multi-angle features from the preprocessed signals, the real physical welding process can be better mapped. At the same time, due to the characteristics of strong arc light in the K-TIG welding process, the arc light received by the arc-pool-keyhole profile image is stronger than that received by the traditional front pool image and the back keyhole exit image. It is urgent to develop an algorithm that can process strong arc light images and can robustly extract geometric features representing the dynamic welding process from the arc-pool-keyhole profile images. The original pressure signal obtained from the arc pressure measurement is similar to the original electrical signal and is interfered by the acquisition loop noise. Accurate arc pressure data can be obtained only after preprocessing.

[0004] The existing technologies for signal processing and feature extraction algorithms for the essential characteristics of different physical signals in the welding process are relatively lacking, making it difficult to further study the arc physical characteristics and penetration recognition of K-TIG welding. Summary of the Invention

[0005] Aiming at the technical problems existing in the prior art, the purpose of the present invention is to provide a method for preprocessing and feature extraction of multi-source information in keyhole plasma tungsten inert gas (K-TIG) welding, which can perform different preprocessing on different characteristic information in the welding process, greatly improve the signal-to-noise ratio of welding signals, and specifically extract effective features of welding multi-source information, which is beneficial to establishing the mapping relationship between signal features, the dynamic welding process, and welding quality.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A method for processing multi-source heterogeneous information and extracting features in keyhole plasma tungsten inert gas (K-TIG) welding, which performs signal processing and feature extraction on the voltage signal, arc sound signal, arc-keyhole-pool image, and arc pressure signal of K-TIG welding, including the following steps:

[0008] Combining the wavelet coefficient distribution characteristics of the voltage signal and the general threshold method to perform wavelet denoising on the voltage signal, retaining the detailed characteristics of the voltage signal, performing time-domain statistical analysis on the denoised voltage signal, and extracting time-domain characteristic parameters including variance D1, root mean square R1, and kurtosis coefficient K1; performing frequency-domain power spectrum processing on the denoised voltage signal, and extracting the characteristic frequencies f2, f4, and f5 of the 2nd, 4th, and 5th frequency bands as frequency-domain characteristic parameters;

[0009] Introducing an improved spectral subtraction method to perform denoising on the arc sound signal, and extracting the time-domain statistical characteristics of the arc sound signal including root mean square R2 and kurtosis coefficient K2; re-dividing the frequency of the arc sound signal into 10 groups of Mel frequencies and designing a group of band-pass filter banks to perform band-pass filtering on the frequency domain, and then combining the Mel frequency with cepstrum analysis to obtain Mel frequency cepstral coefficients MFCC1 to MFCC10, and also extracting the band energy E of the signal spectrum;

[0010] Selecting three regions of interest for the arc-pool-keyhole image in the penetration state, namely the torch region ROI1, the pool compression region ROI2, and the keyhole channel region ROI3, extracting the edge lines of the three ROIs based on the weldment edge detection algorithm combining Hough transform and slope selection, and further extracting the geometric characteristics of the arc-keyhole-pool including the pool compression depth, arc diameter, arc endpoints, and the lengths of the front and rear parts of the keyhole channel;

[0011] Using moving average filtering to preprocess the arc pressure signal and perform a time-to-space conversion on the region of interest of the arc pressure signal, and extracting arc pressure characteristic parameters including the arc pressure peak value P max and the effective arc pressure region [L1, L2], where L1 is the length of the left half of the effective region and L2 is the length of the right half of the effective region.

[0012] Furthermore, using the db2 wavelet function to perform wavelet transform on the voltage signal, decomposing the voltage signal into 10 layers, selecting the first 5 decomposition layers to suppress the wavelet coefficients of the noise. In each layer, the wavelet coefficients of the voltage signal have corresponding distributions at the start and end of the arc. The voltage signal before the start of the arc is all noise signal, and its corresponding wavelet coefficients are the wavelet coefficients of the noise component. After effectively estimating the noise component, an appropriate threshold is selected, and the threshold selection is determined by the following formula:

[0013] T 1i = max(CDi ) i = 1, 2, …, 5;

[0014]

[0015] T i = max(T 1i , T 2i ) i = 1, 2, …, 5;

[0016] Wherein, T 1i is the threshold value of the i-th layer obtained by analyzing the distribution characteristics of the wavelet coefficients of the voltage signal, CD′ i is the wavelet coefficient of the pre-arc noise signal of the i-th layer, T 2i is the corresponding threshold value obtained by the general threshold method, N i is the length of the wavelet coefficients of the i-th layer, T i is the threshold value after noise reduction.

[0017] Furthermore, after obtaining the noise reduction threshold, the original wavelet transform coefficients are processed according to certain rules. The wavelet coefficients of each layer are shrunk towards 0 by a distance of one threshold value to suppress the noise components, and variants such as penalty factors are also added to meet the noise reduction requirements of the voltage signal. The soft threshold function rule is shown in the following formula.

[0018]

[0019] Wherein, CD i is the original wavelet transform coefficient of the voltage signal, CD * i is the wavelet transform coefficient of the processed voltage signal.

[0020] Furthermore, the extraction of the variance D1, root mean square R1, and kurtosis coefficient K1 of the time-domain characteristic parameters is carried out according to the following formula.

[0021]

[0022]

[0023]

[0024] Wherein, s is the voltage signal to be analyzed, n is the length of the sample data, s i is the i-th sample data of the voltage signal, is the average value of the sample data, is the fourth-order central moment of the signal, σ s is the standard deviation of the signal.

[0025] Furthermore, the power spectral energy of the voltage signal is mainly concentrated in five frequency bands, namely 0 - 1.5 kHz, 3.9 - 4.8 kHz, 8.2 - 9.1 kHz, 14.1 - 14.9 kHz, and 18.4 - 19.3 kHz; for frequency band 1, except for the significant DC component, the characteristic frequency is 300 Hz, and the remaining 600 Hz, 900 Hz, 1200 Hz, and 1500 Hz are the harmonics of 300 Hz; the characteristic frequency of frequency band 2 is 4312 Hz, and the characteristic frequency of frequency band 3 is 8624 Hz, which is the harmonic of the characteristic frequency of frequency band 2; the characteristic frequency of frequency band 4 is 14540 Hz; the characteristic frequency of frequency band 5 is 18847 Hz.

[0026] Furthermore, when using the improved spectral subtraction method to denoise the arc sound signal, the arc sound signal is windowed and framed with a rectangular window, and then the spectral subtraction is performed on each frame using the over-subtraction technique. The improved spectral subtraction method is processed according to the following formula:

[0027]

[0028] In the formula, Y(ω) is the original sound signal collected, S(ω) is the pure arc sound signal, N(ω) is the environmental noise signal, α is the over-subtraction factor, which mainly controls the peak value of the noise, and β is the gain compensation factor, which mainly fills the trough of the noise.

[0029] Furthermore, the cepstrum analysis and cepstral coefficient analysis include the following steps:

[0030] Separate the vocal tract spectrum H(ω) using logarithmic operations to highlight the lower-energy frequency components in the spectrum, smooth the spectrogram, obtain the envelope determined by the formants in the spectrum, and then perform the inverse Fourier transform F -1 (·), and the calculation formula is:

[0031] S(ω) = H(ω)E(ω);

[0032] lgS(ω) = lgH(ω) + lgE(ω);

[0033] s′(t) = h′(t) + e′(t) = F -1 (lgH(ω)) + F -1 (lgE(ω));

[0034] In the formula, S(ω), H(ω), and E(ω) are the frequency-domain signals of the arc sound signal, the vocal tract signal, and the arc energy respectively, and s′(t), h′(t), and e′(t) are the cepstrum signals of the arc sound signal, the vocal tract signal, and the arc energy respectively.

[0035] When performing cepstral coefficient analysis, first convert the spectrum of the signal into a Mel spectrum, and then perform cepstral analysis on the Mel spectrum of the signal. The low-frequency part h′(t) of the cepstrum is the cepstral coefficient.

[0036] Furthermore, to extract the geometric features of the ROI, the reference lines of the ROI need to be determined. Select the geometric center of the welding torch as the reference line of ROI1, select the parallel line 300 pixel points away from the position of the welding torch as the reference line of ROI2, and select the center line of the welded part as the reference line of ROI3. The pixel position P of the welding torch Torch For

[0037]

[0038] In the formula, m is the number of pictures to be processed, and n1 is the number of horizontal lines of the pictures for extracting the left and right edge points of the welding torch; l ij is the x coordinate of the left edge point of the welding torch on the j-th horizontal line in the i-th picture, and r ij is the x coordinate of the right edge point of the welding torch on the j-th horizontal line in the i-th picture.

[0039] Furthermore, the welding part edge detection algorithm based on the Hough transform and slope selection includes the following process.

[0040] First, use the Canny operator to perform edge detection on the ROI, convert the grayscale image with a size of m×n and single-pixel values in the range [0, 255] into a binary image with only edge pixel information. Then, traverse from n / 2 along the reference line of each ROI towards the upper half of the image to obtain the upper edge position of the molten pool. Then, use the Hough transform to convert the remaining edge pixel information into the Hough space, directionally select the elements with θ = π / 2, where θ is the angle between the edge line and the y-axis, and take the top two of its voting numbers as ρ to obtain the upper and lower edge lines of each ROI. Furthermore, calculate the center line of the welding part based on the upper and lower edge lines, and calculate the compression depth of the molten pool based on the upper edge line and the upper boundary line of the molten pool.

[0041] Furthermore, when using sliding mean filtering to process the signal as a whole, set the sliding window to 100 sampling points, with a corresponding time window of 10 ms. Then, perform coordinate transformation on the processed arc pressure data, convert it from the reverse time coordinate to the spatial coordinate that conforms to the coordinate system rules, set the threshold of the arc pressure to 100 Pa, and further obtain the arc pressure distribution area.

[0042] Generally speaking, the present invention has the following beneficial effects:

[0043] (1) Different preprocessing algorithms are developed for different characteristic information in the welding process. For example, wavelet denoising of voltage signals is performed by combining the wavelet coefficient distribution characteristics of voltage signals and the general threshold method, and an improved spectral subtraction method is introduced to denoise the arc sound signals, which helps to suppress noise components and retain useful detailed features, thereby greatly improving the signal-to-noise ratio of welding signals.

[0044] (2) The multi-source signals in the welding process contain rich information on the welding dynamic process. The present invention specifically extracts the effective features of the welding multi-source information, which is conducive to establishing the mapping relationship between signal features, the welding dynamic process, and welding quality. Description of the Drawings

[0045] Figure 1 is the flowchart of multi-source heterogeneous information processing and feature extraction for keyhole plasma arc welding (K-TIG);

[0046] Figure 2 is the voltage signal collected during the welding process;

[0047] Figure 3 is the voltage signal and its wavelet coefficients at different scales;

[0048] Figure 4 is the wavelet transform denoised voltage signal and its signal-to-noise ratio;

[0049] Figure 5 is the schematic diagram of the division of the region of interest of the voltage signal;

[0050] Figure 6 is the power spectrum of the welding voltage signal;

[0051] Figure 7 are the characteristic frequencies of each frequency band of the voltage signal power spectrum;

[0052] Figure 8 are the arc sound signal, environmental noise, and their power spectra;

[0053] Figure 9 is the flowchart of the improved spectral subtraction method based on the over-subtraction technique;

[0054] Figure 10 are the arc sound signal after denoising and its power spectrum;

[0055] Figure 11 is the conversion relationship diagram between frequency and Mel frequency;

[0056] Figure 12 are the Mel frequency cepstral coefficients of the arc sound signal;

[0057] Figure 13 is the schematic diagram of the selection of the region of interest (ROI) for the arc-keyhole-pool image;

[0058] Figure 14 It is the flow chart of arc-keyhole-molten pool geometric feature extraction algorithm;

[0059] Figure 15 This is a flow chart of weldment edge detection algorithm based on Hough transform and slope selection;

[0060] Figure 16 It is the arc pressure signal preprocessing process;

[0061] Figure 17 It is a schematic diagram of characteristic parameters of arc pressure signal. DETAILED DESCRIPTION

[0062] The present invention is described in further detail below.

[0063] Figure 1 This is a flowchart for multi-source heterogeneous information processing and feature extraction of deep penetration K-TIG welding. Signal processing and feature extraction are performed on the voltage signal, arc sound signal, arc-keyhole-molten pool image and arc pressure signal of K-TIG welding. It includes:

[0064] (1) Combining the wavelet coefficient distribution characteristics of the voltage signal with the universal threshold method, the voltage signal is subjected to wavelet denoising, retaining the detailed characteristics of the voltage signal. The voltage signal after denoising is subjected to time domain statistical analysis, and the time domain characteristic parameters including variance D1, root mean square R1 and kurtosis coefficient K1 are extracted. The voltage signal after denoising is subjected to frequency domain power spectrum processing, and the characteristic frequencies f2, f4 and f5 of the 2nd, 4th and 5th frequency bands are taken as frequency domain characteristic parameters.

[0065] (2) An improved spectral subtraction method is introduced to reduce the noise of the arc sound signal and extract the time domain statistical characteristics of the arc sound signal, including the root mean square R2 and the kurtosis coefficient K2; the frequency of the arc sound signal is re-divided into 10 groups of Mel frequencies and a group of bandpass filters are designed to perform bandpass filtering in the frequency domain. The Mel frequency is then combined with cepstral analysis to obtain the Mel frequency cepstral coefficients MFCC1-MFCC10, and the frequency band energy E of the signal spectrum is also extracted;

[0066] (3) Three regions of interest (ROIs) are selected from the arc-molten pool-keyhole image in the fully melted state, namely the welding gun area ROI1, the molten pool compression area ROI2, and the keyhole channel area ROI3. The edge lines of the three ROIs are extracted based on the weld edge detection algorithm based on Hough transform and slope selection, and then the geometric features of the arc-keyhole-molten pool are extracted, including the molten pool compression depth, arc diameter, arc endpoint, and the front and rear lengths of the keyhole channel.

[0067] (4) Use sliding mean filtering to preprocess the arc pressure signal and perform time domain to space conversion on the region of interest of the arc pressure signal to extract arc pressure characteristic parameters including arc pressure peak value Pmax And the effective arc pressure area [L1, L2], L1 is the length of the left half of the effective area, and L2 is the length of the right half of the effective area.

[0068] The present invention conducted welding experiments on 10.8 mm thick DSS. The specific welding parameters are shown in Table 1. The welding voltage signal and arc sound signal were synchronously collected in all experiments.

[0069] Table 1 Duplex stainless steel welding experimental parameters

[0070]

[0071] Figure 2 The image fully demonstrates the voltage signal changes throughout the welding process. Region A, before arc initiation, represents noise introduced by the circuit. The DC offset in the signal is clearly visible, necessitating the removal of both noise and DC offset introduced by the acquisition circuit. Region B, during the welding process, reveals rich details in the voltage signal. Due to the DC welding method, the voltage fluctuates around the mean. Therefore, preprocessing the electrical signal for noise reduction requires removing the interference of DC offset and noise while preserving as much detailed information as possible about the voltage signal during the welding process.

[0072] First, we need to remove the zero drift of the signal. We intercept the voltage signal in the 10s period before arcing, calculate the DC component of the zero drift by averaging, and remove the drift from the overall signal. The calculation formula is as follows:

[0073] V′(t)=V(t)-mean(V 10s (t));

[0074] Where V'(t) is the signal after zero drift is removed, and V(t) is the original collected voltage signal. After removing the noise after zero drift, the signal is shifted downward by the DC offset. The noise segment signal is symmetrical about the time axis.

[0075] The wavelet transform, known as a "mathematical microscope," has a fixed resolution cell area. That is, the product of the time window width and the frequency window height is a constant defined by the Heisenberg uncertainty principle. To meet the varying demands of high- and low-frequency measurement, its time window width and frequency window height can be adaptively adjusted. Therefore, when using the wavelet transform to decompose a signal into different levels, the frequency components of the signal will be decomposed into different levels as the time window resolution changes. Generally, noise has high-frequency characteristics and poor correlation between different levels, primarily distributed at high-resolution levels. However, the wavelet transform coefficients of a signal are highly correlated across different levels, especially at edges or sudden changes. Therefore, the basic idea of denoising based on the wavelet transform is to perform a wavelet transform on the noisy signal. Based on the different characteristics of the noise and the useful signal at different levels, the wavelet coefficients of the noise are suppressed to reconstruct the de-noised signal.

[0076] The voltage signal is subjected to wavelet transform using the db2 wavelet function, and the voltage signal is decomposed into 10 layers. The voltage signal, the low-frequency component, and the high-frequency components of each layer are as Figure 3 shown. The low-frequency component shows the DC component of the voltage signal during the DC K-TIG welding process, which is similar to the low-pass filtering operation on the original voltage signal; as the level deepens, the resolution decreases, and the high-frequency detail components are mainly retained in the first 3 levels. At each level, at the mutation points of the signal, that is, the arc-starting information and the arc-ending information are both retained. Considering that deepening the level will significantly increase the computational resources consumed by signal processing, Figure 3 and no more detailed information can be shown after the 5th layer in , it is more appropriate to set the decomposition level to 5 layers.

[0077] To suppress the wavelet coefficients of the noise, it is necessary to effectively estimate the noise components and select an appropriate threshold. The present invention proposes a method for threshold selection by combining the distribution characteristics of the wavelet coefficients of the voltage signal and the general threshold method. As Figure 3 shown, at each level, the wavelet coefficients of the voltage signal have corresponding distributions at the arc-starting and arc-ending points. The voltage signals before arc-starting are all noise signals, and their corresponding wavelet coefficients are the wavelet coefficients of the noise components, which should be suppressed separately at each level. As shown in the following formula,

[0078] T 1i = max(CD i ′) i = 1, 2, …, 5;

[0079]

[0080] T i = max(T 1i , T 2i ) i = 1, 2, …, 5;

[0081] T 1i is the threshold of the i-th layer obtained by analyzing the distribution characteristics of the wavelet coefficients of the voltage signal, and CD' i is the wavelet coefficient of the noise signal before arc-starting in the i-th layer. T 2i is the corresponding threshold obtained by the general threshold method, where N i is the length of the wavelet coefficients in the i-th layer. Finally, by combining the thresholds obtained by the two methods, the optimal threshold can be obtained on the premise of specific analysis of the distribution of the wavelet coefficients of the signal, and the maximum degree of noise suppression can be carried out. [[ID=4l]]

[0082] After obtaining the noise reduction threshold, it is necessary to process the original wavelet transform coefficients according to a certain rule, and this rule is generally called the threshold function. The commonly used threshold functions are the soft threshold function and the hard threshold function, and the general threshold method is generally combined with the soft threshold function. The rule of the soft threshold function is shown in the following formula,

[0083]

[0084] The wavelet coefficients of each layer are shrunk towards 0 by a distance of a threshold to suppress the noise components. Usually, there are also variants such as adding a penalty factor to the soft threshold function. After experiments, the rules shown in the above formula have met the noise reduction requirements of the voltage signal. The voltage signal after wavelet transform noise reduction is as Figure 4 shown. Generally, the signal-to-noise ratio (SNR) can be used to evaluate the signal quality. The calculation of the signal-to-noise ratio is shown in the following formula. After wavelet transform noise reduction, the noise components of the voltage signal are greatly suppressed, and the signal-to-noise ratio rises from 43 to 81.

[0085]

[0086] The multi-source signals in the welding process contain rich information on the welding dynamic process. However, the redundant information in the original signals cannot be ignored. The purpose of processing the multi-source signals is to extract the effective features in the signals and establish the mapping relationship between the signal features and the welding dynamic process and welding quality.

[0087] As an electro-signal is a typical time-domain signal, considering the DC welding characteristics of K-TIG welding, it is appropriate to extract and analyze the statistical features of the electro-signal. As Figure 5 shown, drawing on the ROI idea in image processing, an ROI operation is performed on the electro-signal to extract the voltage signal during the stable welding process, that is, the signal segment from 4 s after high-frequency arc starting to 6 s before arc extinguishing, for feature extraction. Generally, the commonly used statistical features are variance D, root mean square R, and kurtosis coefficient K, and the calculation formulas are shown as follows,

[0088]

[0089]

[0090]

[0091] In the formula, s is the voltage signal to be analyzed, n is the length of the sample data. In the present invention, a rectangular window with a length of 3000 is used to window and frame the signal, s i is the i-th sample data of the voltage signal, is the average value of the sample data, is the fourth-order central moment of the signal, and σ s is the standard deviation of the signal.

[0092] After performing time-domain statistical analysis on the DC characteristics of the K-TIG welding voltage signal, considering the relationship between the voltage signal and the arc length and even arc stability, it is still necessary to analyze the frequency-domain detailed features of the electro-signal. Figure 6The power spectrum of the electrical signal after removing the zero drift is shown. The power spectra of the voltage signals are very similar in shape, with only slight differences in the frequencies corresponding to the peak powers. Taking 40 dB as the dividing line, it can be seen that the energy of the voltage signal in K-TIG welding is mainly concentrated in five frequency bands, namely 0 - 1.5 kHz, 3.9 - 4.8 kHz, 8.2 - 9.1 kHz, 14.1 - 14.9 kHz, and 18.4 - 19.3 kHz.

[0093] According to the physical meaning of the power spectrum of the signal, there must be signal components corresponding to the five frequency bands in the K-TIG welding voltage signal, as Figure 7 shown. In frequency band 1, in addition to the significant DC component, with 300 Hz as the characteristic frequency, the remaining 600 Hz, 900 Hz, 1200 Hz, and 1500 Hz are the harmonics of 300 Hz. In the time domain of the signal, it corresponds to the time domain component with a period of 1 / 300 Hz; in frequency band 2, the characteristic frequency is 4312 Hz, corresponding to a very small time domain component with a period of 1 / 4312 Hz in the time domain of the signal. At this time, it is difficult to see the clear physical meaning after further subdivision of this time domain component. Therefore, the corresponding time domain components of the subsequent frequency bands will not be analyzed; in frequency band 3, the characteristic frequency is 8624 Hz, which is the harmonic of the characteristic frequency of frequency band 2; the characteristic frequencies of frequency bands 4 and 5 are 14540 Hz and 18847 Hz respectively. In summary, taking the characteristic frequencies of the voltage signal in each frequency band as the frequency domain characteristics, since Figure 6 the characteristic frequencies of the two signals in frequency band 1 overlap, and the characteristic frequency of frequency band 3 is the harmonic of the characteristic frequency of frequency band 2, only the characteristic frequencies of frequency bands 2, 4, and 5 are used as the frequency domain characteristics.

[0094] Based on the established multi-source information sensing system for the K-TIG welding process, while the voltage signal is being collected, the arc sound signal is also collected synchronously. Due to the interference of DC components in the acquisition circuit, there is also a certain DC bias in the original arc sound signal. Its processing method is the same as that of removing the zero drift of the voltage signal. All the subsequent arc sound signals mentioned are signals after removing the zero drift. The collected arc sound signal can be divided into four parts, namely the initial ambient noise part, the ambient noise part after the gas starting time before arc starting, the arc sound part during welding, and the ambient noise part during the gas prolonging time after arc extinguishing. In the initial ambient noise part, the sound signal is mainly composed of the sound of the radiator fan of the robot control cabinet and the sound of the radiator fan of the welding power source. Compared with the initial ambient noise part, the sound of the shielding gas injection is also collected in the sound signals during the gas starting time before arc starting and the gas prolonging time after arc extinguishing. Therefore, the arc sound signal during welding can be regarded as the coupling of the composite ambient noise and the pure arc sound.

[0095] Compared with the arc sound signal, the energy of the ambient noise is relatively low, and it is difficult to intuitively distinguish the frequency domain differences between the two using the traditional FFT. The Welch method can obtain a relatively smooth distribution of the signal power spectrum or power spectral density. By converting the signal power to the logarithmic domain, it is possible to more clearly compare the performance of the two in the entire frequency domain. Select Figure 8 the ambient noise from 8 - 12 s in (a) and the arc sound signal from 30 - 34 s, and obtain the power spectra of the arc sound signal and the ambient noise as shown in Figure 8 (b). It can be seen that the power spectra of the two are highly coupled in the low - frequency band of 0 - 1800 Hz and the mid - frequency band of 5000 - 13000 Hz. The arc sound signal has two relatively obvious characteristic frequency bands of 3000 - 5000 Hz and 18000 - 19600 Hz. The key to processing the arc sound signal is to remove the influence of the ambient noise from the original arc sound signal.

[0096] When a skilled welder is performing welding operations, the sound signal received by the ear is the same as the sound signal collected by the sound sensor. However, a skilled welder can distinguish the pure arc sound from the original signal mixed with the sound of the robot control cabinet radiator fan, the sound of the welding power source radiator fan, the sound of the shielding gas injection, and the pure arc sound, and make decisions to achieve high - quality welding. The logarithmic response characteristic of the human ear to sound frequency has a significant impact on distinguishing sounds with different frequency characteristics. Therefore, by distinguishing the ambient noise and the arc sound signal in the frequency domain and removing the frequency - domain components of the ambient noise, a relatively pure arc sound can be obtained. From the analysis of the sound source, the arc sound signal and the ambient noise are independent of each other, and the ambient noise is additive noise. A natural idea is to "subtract" the ambient noise from the arc sound signal, and the algorithm implementation of this idea is called spectral subtraction. Spectral subtraction, a speech enhancement algorithm, is based on the idea of subtracting the noise power spectrum from the power spectrum of the noisy signal to obtain a relatively pure speech signal.

[0097] The collected sound signal, i.e., the noisy signal y(t), can be expressed as the superposition of the ambient noise n(t) and the pure arc sound s(t):

[0098] y(t) = s(t) + n(t);

[0099] Converting the noisy signal to the frequency domain, that is, taking the Fourier transform of the above formula, we can get:

[0100] Y(ω) = S(ω) + N(ω);

[0101] According to the above formula, the power spectrum of the noisy signal can be expressed as:

[0102]

[0103] Wherein, Re(S(ω)N(ω)) is the cross-correlation coefficient between the ambient noise and the pure arc sound. Since the ambient noise is additive noise and is independent of the pure arc sound, this term is 0. Therefore, the amplitude of the pure arc sound can be expressed as:

[0104]

[0105] In the process of using spectral subtraction to denoise the arc sound signal, it is necessary to window and frame the arc sound signal and perform spectral subtraction on each frame. The present invention uses a rectangular window for framing, the window size is 60 ms, and the overlapping rate of two adjacent frames is 10%. In the process of spectral subtraction of each frame of the original arc sound signal, when the noise is overestimated, |Y(ω)| 2 –|N(ω)| 2 will appear negative. A natural processing method is to set the negative value to zero, but this method will cause independent small peaks to appear in the spectrum of the denoised signal. When converted to the time domain and listened to, it will be like a screeching sound, which is called "musical noise". In order to cope with the problem that the appearance of negative values inside the radical and simple processing will introduce "musical noise", an over-subtraction technique is often used to perform spectral subtraction on each frame.

[0106]

[0107] The above formula is the operation process of each frame using the improved spectral subtraction with the over-subtraction technique. α is the over-subtraction factor, which mainly controls the peak of "musical noise"; β is the gain compensation factor, which mainly fills the valley of "musical noise"; after multiple experiments, in the present invention, α = 5 and β = 0.002. The main idea of the improved spectral subtraction based on the over-subtraction technique is to cut the peak and fill the valley of "musical noise" in the frequency domain. The algorithm flow is as Figure 9 shown.

[0108] The arc sound signal after denoising by the improved spectral subtraction is as Figure 10 shown. As can be seen from Figure 10 (a), the ambient noise of the arc sound signal after denoising is greatly suppressed, and the arc sound part is retained to the greatest extent. Due to the principle of subtracting the frequency-domain energy by spectral subtraction, the amplitude of the arc sound part decreases slightly. In terms of the frequency domain, as Figure 10As shown in (b), the frequency-domain components where the arc sound signal overlaps with the ambient noise are suppressed, and the characteristic frequency bands of the arc sound signal itself are retained. To evaluate the noise reduction effect of the improved spectral subtraction method used in this paper on the arc sound signal, the arc sound signal is denoised using a moving average filter and a wavelet denoising algorithm respectively. The signal-to-noise ratios obtained by different denoising algorithms are shown in Table 2. The wavelet transform-based denoising algorithm used is an algorithm for denoising the voltage signal, which has the ability to retain the DC component and the detailed features of the signal, but is not suitable for denoising the sound signal, resulting in a decrease in the signal-to-noise ratio. The moving average filter is equivalent to performing a low-pass filtering operation on the signal, and the signal-to-noise ratio is improved but a large amount of detailed information is lost. After denoising the arc sound signal based on the improved spectral subtraction method, the signal-to-noise ratio is improved, as shown in (a), and the detailed information of the signal is retained. Figure 10 As shown in (a), the detailed information of the signal is retained.

[0109] Table 2 Signal-to-noise ratios of different denoising algorithms

[0110]

[0111] As a time-series signal, the time-domain characteristics of the arc sound signal can be analyzed and extracted from a statistical perspective. Similar to the extraction of the time-domain statistical characteristics of the voltage signal, the root mean square R, kurtosis coefficient K, and variance D of the arc sound signal in the time domain are extracted. It is difficult to distinguish the two penetration states from the perspective of time-domain statistical characteristics. It should be noted that the root mean square R and variance D are very similar because the denoised arc sound signal is symmetric about the time axis, s≈0, which makes D≈R. 2 Therefore, considering that the root mean square R can better reflect the energy of the signal, in the subsequent analysis, only the root mean square R and kurtosis coefficient K of the arc sound signal in the time domain are retained as statistical characteristics.

[0112] Skilled welders can perform perfect welding operations based on the arc sound they hear. Since the human ear is more sensitive to frequency information, it is very necessary to analyze the arc sound signal in the frequency domain. From the power spectrum of the denoised sound signal, it can be clearly seen that there are two frequency bands with higher energy. These two frequency bands correspond to frequency band 2 and frequency band 5 of the voltage signal respectively, and the frequency values corresponding to the frequency components with the largest energy in each characteristic frequency band of the arc sound signal are the same as those of the voltage signal, indicating that the arc sound signal is generated by the excitation of arc energy. However, although the characteristic frequencies are the same, there are significant differences in the energies corresponding to the frequency components. Taking f1 = 3000 Hz and f2 = 6000 Hz, the energies of the characteristic frequency bands of the signal spectrum are calculated using the following formula.

[0113]

[0114] Considering the relationship between the human ear and the arc sound, the Mel Frequency Cepstral Coefficients (MFCCs), which are widely used in the field of speech recognition, can be considered for introduction. The human ear has the characteristics of linear response to low frequencies and logarithmic response to high frequencies. Therefore, the new frequency obtained by re - dividing the frequency according to the response characteristics of the human ear is called the Mel frequency. The division method is to design a set of band - pass filters as shown in Figure 11 to perform band - pass filtering on the frequency domain. When the divided Mel frequency is 10 groups, the segmentation of the frequency domain shows the characteristics of large low - frequency resolution and low high - frequency resolution.

[0115]

[0116] Cepstrum analysis is a commonly used method for decomposing and analyzing convolutional acoustic signals. The so - called Cepstrum is the reverse writing of the first four letters of Spectrum, and its processing method also has this meaning. Generally, the arc sound signal s(t) is considered to be excited by the arc energy e(t) in the linear time - invariant sound channel h(t) composed of the arc - molten pool - keyhole channel, and its frequency - domain relationship is shown by the following formula

[0117] s(t) = h(t)*e(t);

[0118] S(ω) = H(ω)E(ω);

[0119] The sound channel h(t) is generally considered to show the low - frequency information of the spectrum in the frequency domain, that is, the envelope determined by the resonance peaks in the spectrum. To obtain this envelope, H(ω) is separated, and a logarithmic operation is performed on the above formula. The multiplicative signal becomes additive after the logarithmic operation. At the same time, the logarithmic operation will highlight the frequency components with lower energy in the spectrum, smooth the spectrogram, and facilitate obtaining the envelope. As shown in the following formula, the inverse Fourier transform F –1 (·) is performed on the logarithmized spectrum to obtain the cepstrum s'(t).

[0120] s′(t) = h′(t)+e′(t) = F -1 (lgH(ω))+F -1 (lgE(ω));

[0121] The analysis of Mel - frequency cepstral coefficients combines the Mel frequency with cepstrum analysis. First, the spectrum of the signal is converted into a Mel spectrum, and then cepstrum analysis is performed on the Mel spectrum of the signal. The low - frequency part h'(t) of the cepstrum is the cepstral coefficient. The MFCCs corresponding to the collected arc sound signal are as shown in Figure 12 shown.

[0122] The arc-molten pool-keyhole visual information sensing system based on the sandwich welded part is used to capture the arc-molten pool-keyhole images in the typical critical penetration state. During the welding process, the molten metal in the keyhole is extruded outward by the plasma arc and flows towards the rear of the molten pool. The molten metal near the keyhole channel part of the molten pool presents a slope under the extrusion of the arc. There is a gap between the upper surface of the welded part and the upper surface of the molten pool, which is called the compression depth. Due to the thermal deformation of the sandwich welded part and the initial assembly error during the welding process, there is a small black area on the upper surface of the sandwich welded part in the image. It is caused by the small-angle deformation of the welded part in the direction perpendicular to the camera axis due to thermal deformation, which is equivalent to capturing the upper surface of the welded part. The upper and lower surfaces of the quartz glass are opaque in the Z-axis direction. During the K-TIG welding process, the arc light is very strong. Due to the special structure of the sandwich welded part, the arc light on the duplex stainless steel side is reflected to the quartz glass side, and the workbench surface is relatively smooth and flat, resulting in the reflection of strong arc light, that is, the bright area at the bottom of the image. These image defects occur outside the arc-molten pool-keyhole area and will not affect the extraction of the geometric features of the arc-molten pool-keyhole, but only increase the redundant information of the whole image. Therefore, the region of interest ROI can be selected according to Figure 13 as shown, and the geometric features are extracted for the ROI. The feature extraction algorithm flow is as shown in Figure 14 as shown.

[0123] The camera and the welding torch are relatively stationary, that is, the position of the welding torch in the picture remains unchanged, which will be used as the reference line for geometric feature extraction. The position of the welding torch (P Torch ) is extracted in ROI1 as shown in the following formula:

[0124]

[0125] In the formula, m is the number of pictures used for processing; n is the number of horizontal lines used to extract the left and right edge points of the welding torch in each picture; l ij is the x coordinate of the left edge point of the welding torch on the jth horizontal line in the ith picture; similarly, r ij is the x coordinate of the right edge point of the welding torch.

[0126] The molten pool is compressed by the arc and is overall in a slope shape. The molten metal in the front of the molten pool continuously flows backward and forms a weld after cooling and solidification. Therefore, macroscopically, the upper surface of the molten pool oscillates like a wave from the keyhole side to the tail. The extraction of the compression depth of the molten pool also requires a reference line for definition. After analyzing the image, a parallel line 300 pixel points away from the position of the welding torch P Torch is selected as the reference line for extracting the compression depth of the molten pool. This reference line is in Figure 14It is marked green in ROI2. ROI2 is not only used for extracting the compression depth of the molten pool, but also can find the center line of the sandwich welded part, which is used as the reference line for extracting the geometric features of the keyhole channel in ROI3. After edge detection, there are many pseudo-edges around the upper and lower edges of the welded part. This is because the upper and lower edges of the welded part are the upper and lower demarcation lines of the molten pool, with a large temperature gradient, resulting in a drastic change in gray level, which reduces the robustness of the Canny operator at this location. According to prior knowledge, the upper and lower edges of the welded part should be horizontal lines in the image. Therefore, an edge detection algorithm for the welded part based on Hough transform and slope selection is proposed, and its process is as Figure 15 shown.

[0127] First, use the Canny operator to perform edge detection on ROI2, converting a grayscale image with a single-pixel value of [0, 255] of size m×n into a binary image with only edge pixel information. Since the upper edge of the molten pool is generally distributed in the upper half of the image and the pixels inside the molten pool are uniform without pseudo-edges detected, traversing along P Torch –300 from n / 2 to the upper half of the image can obtain the position of the upper edge of the molten pool. At this time, it is necessary to extract the upper and lower edges of the welded part from the binary image with only edge pixel information. From prior knowledge, the upper and lower edges of the welded part are horizontal in the image, that is, the slope of the upper and lower edges of the welded part extracted should be 0, and the included angle with the y-axis in the image coordinate system (with the upper left corner of the image as the origin, the horizontal direction as the x-axis, and the vertical direction as the y-axis) is π / 2. Considering the analysis of the edge slope, the Hough Transform has the ability to transform elements in the polar coordinate system to the Hough space, realizing the one-to-one correspondence between points in the polar coordinate system and lines in the Hough space, and lines in the polar coordinate system and points in the Hough space. Therefore, use the Hough transform to convert the remaining edge pixel information to the Hough space, directionally select elements with θ = π / 2, and take the top two of their voting numbers as ρ to obtain the upper and lower edge lines of the welded part. Therefore, the center line of the welded part can be calculated based on the upper and lower edge lines, and the compression depth of the molten pool can also be calculated based on the upper edge line and the upper boundary line of the molten pool.

[0128] Since the keyhole area is completely covered by the arc passing through it, it is the brightest area in the picture. However, due to the non-wetting characteristics of the molten metal and the high-temperature resistant quartz glass, part of the arc is distributed at the interface between the molten metal and the high-temperature resistant quartz glass. At the same time, the specular reflection on the surface of the workbench also expands the high-brightness area in the image, which all increases the difficulty of extracting the geometric features of the keyhole. Therefore, similar to the method of extracting the compression depth of the molten pool in ROI2, the center line of the welded part extracted from ROI2 is used as the reference line to calculate the geometric features of the keyhole channel. The geometric features of the keyhole channel mainly include the front length of the keyhole channel and the rear length of the keyhole channel. The front length of the keyhole channel is the distance from the front wall of the keyhole channel to the center line of the welding torch P TorchDistance; the length of the rear part of the keyhole channel is the distance from the rear wall of the keyhole channel to the center line P of the welding torch along the center line of the welded part. Torch The distance. The extraction method is as Figure 14 shown. In the keyhole region ROI3, the front and rear walls of the keyhole channel are the boundaries of the keyhole channel with respect to the front and rear molten metal, clearly separating the extremely high-brightness region and the sub-high-brightness region in the image. After using the Canny operator to extract the features of ROI3, it can be clearly seen that the edge information is distributed on both sides of ROI3, and it is very convenient to obtain the edge information of the front and rear walls of the keyhole channel.

[0129] During the welding process, since the height of the tip of the tungsten electrode from the surface of the welded part is fixed, the most important geometric parameter of the arc shape is the change in the shape of the bell-shaped arc above the upper surface of the welded part. In this paper, the arc diameter is defined as the farthest distance of the arc distribution, which is the distance between the front end point and the rear end point of the arc. The front end point of the arc is the farthest point reached by the area above the upper surface of the welded part in the welding direction (the positive x-axis direction). Similarly, the rear end point of the arc is the farthest point reached by the area above the upper surface of the welded part in the negative welding direction. Since the arc shape changes significantly during the actual welding process, the front end of the arc warps up due to the accumulation of molten metal and the blockage of the workpiece to be welded, and the rear end point of the arc shows a warping and a state of converging into the keyhole channel in the non-penetration and penetration states respectively. Therefore, it is difficult to select a fixed ROI. Considering the gray distribution characteristics of the area above the upper surface of the welded part, threshold segmentation is directly performed on the area above the upper surface. The bright arc area is segmented out because most of the pixel values are 255, and the front and rear end points of the arc can be easily obtained.

[0130] Using the arc pressure measurement device and method, the arc pressure at a welding current of 350 A is as Figure 16 shown. In the measured original data, the zero drift and noise of the signal are significantly present. The zero drift is processed according to Equation (3-1), where the mean value is calculated using the part of the arc pressure signal after 15 s. After removing the zero drift, the mean value of the noise part of the signal should be 0. The sliding mean filter is used to process the whole signal. Considering that the peak part of the arc pressure is relatively narrow, in order to avoid large errors brought to the signal by the filtering operation, the sliding window is set to 100 sampling points, and the corresponding time window is 10 ms. The filtered signal is as Figure 16 shown in the right half part. The signal is relatively smooth, and the arc pressure shows a normal distribution.

[0131] The welding arc travels in a fixed direction, and the measured arc pressure data reflects the distribution of the arc pressure in this direction. What the present invention focuses on is the spatial distribution of the arc pressure. Therefore, taking the 5s point (the peak of the arc pressure) as the center of the arc pressure, ±8mm is intercepted, that is, the arc pressure data from 3 - 7s is the region of interest. Since the pinhole measurement point is fixed and the welding torch passes through the keyhole measurement point at a fixed welding speed, then the right endpoint B of the arc first passes through the pinhole measurement point. In the original time - series data and image with time as the abscissa ( Figure 16 the left - hand part), the arc pressure data of the right endpoint B of the arc is measured first. As the arc passes through the pinhole measurement point, the arc pressure data from the right endpoint B to the arc center and finally to the left endpoint A of the arc is gradually measured. To conveniently represent the distribution of the arc pressure in the anode region, the origin of the coordinate system is set at the pinhole measurement point, that is, the peak position of the arc pressure center. Therefore, it is necessary to perform coordinate transformation on the processed arc pressure data, converting it from the reverse time coordinate to the spatial coordinate that conforms to the coordinate system rules. After coordinate transformation, the arc pressure is as Figure 16 shown in the right - hand part. <>

[0132] For the arc pressure signal, not only the peak value P of the arc pressure max needs to be considered, but also the shape of the arc pressure distribution. Taking 100 Pa as the threshold, as Figure 17 shown, the part above the threshold ( Figure 17 the horizontal line) is regarded as the effective arc pressure region. Taking the position of the peak value P of the arc pressure max as the demarcation line, the length of the left - hand part of the effective region is L1, and the length of the right - hand part is L2. L1 and L2 reflect the shape of the spatial distribution of the arc pressure during the welding process and indirectly reflect the shape of the arc. For example, at the welding current of 350 A shown in Figure 17 , the peak value P of the arc pressure max is 3258 kPa, L1 is 5.32 mm, and L2 is 4.14 mm. That is, the length of the front part of the arc in the welding direction is less than that of the tail part, which is consistent with the experimental phenomenon under the condition of a small welding current.

[0133] The above - mentioned embodiments are the preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above - mentioned embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for preprocessing and feature extraction of multi-source information in keyhole K-TIG welding, characterized in that Signal processing and feature extraction are performed on the voltage signal, arc sound signal, arc-keyhole-molten pool image, and arc pressure signal of K-TIG welding, including the following steps: Combined with the wavelet coefficient distribution characteristics of the voltage signal and the general threshold method, wavelet denoising processing is performed on the voltage signal to retain the detailed characteristics of the voltage signal. Time-domain statistical analysis is performed on the denoised voltage signal to extract time-domain characteristic parameters, including variance D1, root mean square R1, and kurtosis coefficient K1. Frequency-domain power spectrum processing is performed on the denoised voltage signal to extract the characteristic frequencies f2, f4, and f5 of the second, fourth, and fifth frequency bands as frequency-domain characteristic parameters. The improved spectral subtraction method is introduced to perform denoising processing on the arc sound signal, and the time-domain statistical characteristics of the arc sound signal, including root mean square R2 and kurtosis coefficient K2, are extracted. The frequencies of the arc sound signal are re-divided into 10 groups of Mel frequencies, and a set of band-pass filter banks is designed for band-pass filtering in the frequency domain. Then, the Mel frequency and cepstrum analysis are combined to obtain the Mel frequency cepstrum coefficients MFCC1 to MFCC10, and the band energy E of the signal spectrum is also extracted. For the arc-keyhole-molten pool image in the penetration state, three regions of interest are selected, namely the welding torch region ROI1, the molten pool compression region ROI2, and the keyhole channel region ROI3. The edge lines of the three ROIs are extracted based on the welding part edge detection algorithm combining Hough transform and slope selection, and then the geometric characteristics of the arc-keyhole-molten pool image are extracted, including the molten pool compression depth, arc diameter, arc endpoints, and the lengths of the front and rear parts of the keyhole channel. Preprocess the arc pressure signal using moving average filtering and perform a time-domain to space conversion on the region of interest of the arc pressure signal, and extract arc pressure characteristic parameters including the arc pressure peak value P max and the effective arc pressure region [L1, L2], where L1 is the length of the left half of the effective region and L2 is the length of the right half of the effective region.

2. The multi-source information preprocessing and feature extraction method for keyhole K-TIG welding according to claim 1, wherein The db2 wavelet function is used to perform wavelet transform on the voltage signal, and the voltage signal is decomposed into 10 layers. The first 5 decomposition layers are selected to suppress the wavelet coefficients of the noise. In each layer, the wavelet coefficients of the voltage signal have corresponding distributions at the start and end of the arc. The voltage signal before the start of the arc is all noise signal, and its corresponding wavelet coefficients are the wavelet coefficients of the noise component. After effectively estimating the noise component, a suitable threshold is selected, and the threshold selection is determined by the following formula. T 1i = max(CD i ′) for i = 1, 2, …, 5; T i = max(T 1i , T 2i ) for i = 1, 2, …, 5; Wherein, T 1i is the threshold value of the i-th layer obtained by analyzing the distribution characteristics of the wavelet coefficients of the voltage signal, and CD′ i is the wavelet coefficient of the pre-arc noise signal of the i-th layer, T 2i is the corresponding threshold value obtained by the general threshold method, N i is the length of the wavelet coefficients of the i-th layer, and T i is the threshold value after noise reduction.

3. A method for preprocessing and feature extraction of multi-source information in keyhole K-TIG welding according to claim 2, characterized in that, Obtain the noise reduction threshold T i After that, the original wavelet transform coefficients are processed according to certain rules. The wavelet coefficients of each layer are shrunk towards 0 by a distance of one threshold to suppress the noise components, and a penalty factor is also added to meet the noise reduction requirements of the voltage signal. The soft threshold function rule is shown in the following formula: Wherein, CD i is the original wavelet transform coefficient of the voltage signal, and CD * i is the wavelet transform coefficient of the processed voltage signal.

4. A multi-source information preprocessing and feature extraction method for keyhole K-TIG welding according to claim 1, characterized in that The extraction of the time-domain characteristic parameters variance D1, root mean square R1, and kurtosis coefficient K1 is performed according to the following formula. Where s is the voltage signal to be analyzed, n is the length of the sample data, and s i is the i-th sample data of the voltage signal, is the average value of the sample data, is the fourth-order central moment of the signal, and σ s is the standard deviation of the signal.

5. A method for preprocessing and feature extraction of multi-source information in keyhole K-TIG welding according to claim 1, characterized in that, The power spectrum energy of the voltage signal is concentrated in five frequency bands, namely 0 - 1.5 kHz, 3.9 - 4.8 kHz, 8.2 - 9.1 kHz, 14.1 - 14.9 kHz, and 18.4 - 19.3 kHz. In addition to the significant DC component in frequency band 1, the characteristic frequency is 300 Hz, and the remaining 600 Hz, 900 Hz, 1200 Hz, and 1500 Hz are the multiples of 300 Hz. The characteristic frequency of frequency band 2 is 4312 Hz, and the characteristic frequency of frequency band 3 is 8624 Hz, which is twice the characteristic frequency of frequency band 2. The characteristic frequency of frequency band 4 is 14540 Hz, and the characteristic frequency of frequency band 5 is 18847 Hz.

6. The preprocessing and feature extraction method for multi-source information of keyhole K-TIG welding according to claim 1, wherein When using the improved spectral subtraction method to perform denoising processing on the arc sound signal, rectangular windowing and framing are performed on the arc sound signal, and then spectral subtraction is performed on each frame using the over-subtraction technique. The improved spectral subtraction method is processed according to the following formula. Wherein, Y(ω) is the collected original acoustic signal, S(ω) is the arc acoustic signal, N(ω) is the environmental noise signal, α is the over-subtraction factor to control the peak value of the noise, and β is the gain compensation factor to fill the valley of the noise.

7. A method for preprocessing and feature extraction of multi-source information in keyhole K-TIG welding according to claim 1, characterized in that, The cepstrum analysis and obtaining cepstrum coefficients include the following steps: The channel spectrum H(ω) is separated by using logarithmic operations, so that the frequency components with lower energy in the spectrum are highlighted to smooth the spectrogram, the envelope determined by the formants in the spectrum is obtained, and then the inverse Fourier transform F -1 (·) is performed, and the calculation formula is S(ω) = H(ω)E(ω); lgS(ω) = lgH(ω) + lgE(ω); s′(t) = h′(t) + e′(t) = F -1 (lg|H(ω)|) + F -1 (lg|E(ω)|); Wherein, S(ω), H(ω), and E(ω) are respectively the frequency-domain signals of the arc acoustic signal, vocal tract signal, and arc energy, and s′(t), h′(t), and e′(t) are respectively the cepstrum signals of the arc acoustic signal, vocal tract signal, and arc energy; When performing cepstrum coefficient analysis, first convert the spectrum of the signal into a Mel spectrum, and then perform cepstrum analysis on the Mel spectrum of the signal. The low-frequency part h′(t) of the cepstrum is the cepstrum coefficient.

8. A method for preprocessing and feature extraction of multi-source information in keyhole K-TIG welding according to claim 1, characterized in that, To extract the geometric features of the ROI, it is necessary to determine the reference line of the ROI. The geometric center of the welding torch is selected as the reference line of ROI1, the parallel line 300 pixel points away from the position of the welding torch is selected as the reference line of ROI2, and the center line of the welded part is selected as the reference line of ROI3. The pixel position P of the welding torch Torch is Where m is the number of images to be processed, and n1 is the number of horizontal lines of the images for extracting the left and right edge points of the welding torch; l ij is the x coordinate of the left edge point of the welding torch on the j-th horizontal line in the i-th image, and r ij is the x coordinate of the right edge point of the welding torch on the j-th horizontal line in the i-th image.

9. A multi-source information preprocessing and feature extraction method for keyhole K-TIG welding according to claim 1, characterized in that The welding part edge detection algorithm based on Hough transform and slope selection includes the following steps: First, use the Canny operator to perform edge detection on the ROI, convert the grayscale image with a single-pixel value of [0, 255] and a size of m×n into a binary image with only edge pixel information. Therefore, traverse from n / 2 along the baseline of each ROI to the upper half of the image to obtain the upper edge position of the molten pool. Then, use the Hough transform to convert the remaining edge pixel information into the Hough space, directionally select the elements with θ = π / 2, where θ is the angle between the edge line and the y-axis, and take the top two votes as ρ to obtain the upper and lower edge lines of each ROI. Furthermore, calculate the center line of the welding part according to the upper and lower edge lines, and calculate the compression depth of the molten pool according to the upper edge line and the upper boundary line of the molten pool.

10. A deep penetration K-TIG welding multi-source information preprocessing and feature extraction method according to claim 1, characterized in that, When using moving average filtering to process the signal as a whole, set the moving window to 100 sampling points, and the corresponding time window is 10 ms. Then, perform coordinate transformation on the processed arc pressure data, convert it from the reverse time coordinate to the spatial coordinate that conforms to the coordinate system rules, set the threshold of the arc pressure to 100 Pa, and further obtain the arc pressure distribution area.

Images