Welding defect monitoring method and system based on vibration acoustics

Abstract

The invention provides a welding defect monitoring method and system based on vibration acoustics, and relates to the technical field of welding defect monitoring, and the method comprises the steps: applying a low-frequency vibration signal to a monitoring target, and obtaining a response signal; performing spectral analysis on the response signal to obtain modal parameters; judging whether the modal parameters belong to a preset parameter range or not, if yes, applying a real-time high-frequency ultrasonic signal to the monitoring target, obtaining a real-time feedback signal, and extracting signal characteristics of the real-time feedback signal; if not, outputting an alarm signal; judging whether the signal characteristics conform to expectation or not, and if yes, applying the low-frequency vibration signals to the monitoring target again after a preset duration; and if not, outputting an alarm signal. The monitoring target is subjected to dual verification by combining the low-frequency vibration signal and the high-frequency ultrasonic signal, so that the accuracy of a welding defect monitoring result is remarkably improved.

Description

Technical Field

[0001] The present application relates to the technical field of welding defect monitoring, and in particular to a welding defect monitoring method and system based on vibroacoustics. Background Art

[0002] In industrial production, welding, as an important connection technology, is widely used in the manufacturing and assembly of various structures. However, various defects may occur during the welding process, such as cracks and incomplete penetration, which will seriously affect the strength, toughness and service life of the welded structure, and may even cause structural failure and lead to safety accidents.

[0003] Traditional welding defect monitoring methods mainly include visual monitoring, magnetic particle monitoring, penetration monitoring, etc. Although these methods can detect welding defects to a certain extent, they have many limitations. For example, visual monitoring relies on the experience and skill level of the inspectors, and it is difficult to detect small or deeply hidden defects; although magnetic particle monitoring and penetration monitoring can monitor the surface defects of welded parts after welding is completed, they cannot monitor internal defects. Summary of the invention

[0004] In order to improve the accuracy of welding defect monitoring, the present application provides a welding defect monitoring method and system based on vibroacoustics.

[0005] In the first aspect, the present application provides a welding defect monitoring method based on vibroacoustics, which adopts the following technical solution: A welding defect monitoring method based on vibroacoustics comprises the following steps: Data collection: including the first collection and the second collection; First acquisition: applying a low-frequency vibration signal to the monitored target and obtaining a response signal; Second acquisition: applying real-time high-frequency ultrasonic signals to the monitoring target to obtain real-time feedback signals; Data processing: including first processing and second processing; First processing: perform spectrum analysis on the response signal to obtain modal parameters; Second processing: extracting signal features of real-time feedback signals; Defect judgment: including the first judgment and the second judgment; First judgment: judging whether the modal parameter belongs to the preset parameter range, if yes, executing the second acquisition step; if no, executing the output step; Second judgment: judging whether the signal characteristics meet expectations, if so, executing the first acquisition step after a preset interval; if not, executing the output step; Output: Output alarm signal.

[0006] This application applies a low-frequency vibration signal to the monitoring target and obtains a corresponding response signal. The low-frequency vibration signal can excite the modal vibration of the monitoring target, and the obtained response signal contains information about the modal parameters of the monitoring target, such as modal frequency, modal damping, etc. After that, the response signal is subjected to spectrum analysis, the frequency components and energy distribution in the response signal are extracted, and then the modal parameters of the monitoring target are obtained. After that, by comparing the modal parameters with the preset parameter range, it can be preliminarily determined whether there are welding defects. If the modal parameters exceed the preset parameter range, it indicates that there are welding defects and an alarm signal needs to be issued. Otherwise, it is necessary to perform the second acquisition step, and use the high-frequency ultrasonic signal to further check whether the monitoring target has welding defects such as incomplete penetration. Real-time high-frequency ultrasonic signals can obtain the internal details of the welding structure. When the real-time high-frequency ultrasonic signal propagates in the monitoring target, reflection, scattering and other phenomena will occur when encountering welding defects. These phenomena can be captured and used to analyze the location, size and type of welding defects. After that, the feedback signal is obtained, and the signal characteristics of the feedback signal are extracted. These signal characteristics include the intensity of the reflected wave, phase change, etc. By comparing these features with the expected values, it is possible to further determine whether there are welding defects in the monitored target and the type and location of the welding defects. Then, it is determined whether the signal characteristics meet expectations. If the signal characteristics do not meet expectations, it indicates that there are welding defects in the monitored target, and then an alarm signal is issued. This application uses a dual judgment mechanism to monitor welding defects, which improves the accuracy and reliability of welding defect monitoring results.

[0007] Optionally, before executing the step of data collection, the method further includes: The third collection: collect historical images during welding and add category labels to the historical images; Fourth acquisition: collect real-time images of the monitoring target; First, modeling and training: establish a CNN model, and use historical images and added category labels to train the CNN model to obtain a trained CNN model; First prediction: input the real-time image into the trained CNN model to obtain the prediction result; The third judgment: judge whether the prediction result meets the expectation. If so, execute the first acquisition step; if not, execute the output step.

[0008] This application provides a data basis for subsequent model training by collecting historical images during welding and adding category labels (such as normal welding, cracks, incomplete penetration, etc.) to these historical images. Then, a CNN model is established, and it is trained using historical images and category labels so that the CNN model can learn the deep features in the image. Through training, the CNN model can accurately identify the types of defects in the welding process, improve the accuracy of monitoring, and then input the real-time image into the trained CNN model. The CNN model can predict the image based on the learned features and judge whether there are welding defects in the welding process, reducing the subjectivity and error of human judgment. Before executing the data collection step, the category label of the real-time image is first predicted by the CNN model to determine whether the prediction result meets expectations. If the prediction result meets expectations (that is, there are no welding defects), continue to execute the first acquisition step and continue to monitor whether there are welding defects. If the prediction result does not meet expectations (that is, there are defects in the welding process), directly execute the output step and send an alarm signal. This application realizes multi-dimensional monitoring of welding defects by combining visual monitoring with vibration acoustics and ultrasonic detection technology. Visual monitoring can intuitively reflect the surface condition of the welding process, while vibration acoustics and ultrasonic monitoring can deeply detect the internal defects of the welding structure. This comprehensive monitoring method improves the comprehensiveness and accuracy of monitoring and provides a more reliable guarantee for welding quality.

[0009] Optionally, after performing the fourth acquisition step and before performing the first modeling and training step, the method further includes: The third processing: grayscale processing is performed on the real-time image to obtain the grayscale value of each pixel in the real-time image; Fourth processing: using a region growing algorithm to segment the grayscale real-time image to obtain a target image block containing the monitored target and a remaining image block not containing the monitored target; Reset grayscale value: Set the grayscale values ​​of the corresponding pixels of the remaining image blocks in the real-time image to 0 to obtain a new real-time image.

[0010] This application first grayscales the real-time image to remove the interference of color information on image processing and retain the morphological and structural features of the image. Then, the region growing algorithm is used to segment the grayscale real-time image, and the image is divided into different regions according to the similarity of pixels, namely, the target image block containing the monitored target and the remaining image block that does not contain the monitored target. Then, in the real-time image, the grayscale value of the corresponding pixel of the remaining image block is set to 0 to remove the background information of the real-time image. By removing the image area that does not contain the monitored target, the interference of background noise on target recognition can be reduced, and the stability and accuracy of target recognition can be improved.

[0011] Optionally, after executing the third processing step and before executing the fourth processing step, the method further includes: Fifth processing: integrating the grayscale values ​​greater than the preset grayscale threshold into a first data set, and integrating the remaining grayscale values ​​into a second data set; First calculation: Calculate the difference between the first data set and the second data set. The calculation model of the difference is as follows: ; Where r represents the degree of difference; A is the number of grayscale values ​​in the first data set; B is the number of grayscale values ​​in the second data set; is the average value of all gray values ​​in the first data set; is the average value of all gray values ​​in the second data set; Fourth judgment: judging whether the difference degree is less than a preset difference threshold, if so, executing the fifth judgment step; if not, executing the fourth processing step; Fifth judgment: judging whether A is greater than B, if so, executing the first adjustment step; if not, executing the second adjustment step; Adjustment threshold: including first adjustment and second adjustment; First adjustment: respectively calculating the difference between each gray value in the first data set and the preset gray threshold, recording it as the first difference, selecting the gray value corresponding to the minimum first difference as the new preset gray threshold, and executing the fifth processing step; Second adjustment: Calculate the difference between the grayscale value in the second data set and the preset grayscale threshold, record it as the second difference, select the grayscale value corresponding to the minimum second difference as the new preset grayscale threshold, and execute the fifth processing step.

[0012] The present application first divides the grayscale values ​​into two groups by a preset grayscale threshold, and constructs a first data set and a second data set respectively accordingly, then quantifies the difference between the two data sets through a difference degree calculation model, thereby quantifying the difference between the two groups of grayscale values, and then compares the difference degree with the preset difference threshold. When the difference degree is less than the preset difference threshold, it indicates that the distinction between the highlight area and the dark area in the real-time image is not obvious enough, and the preset difference threshold needs to be further adjusted. Therefore, the present application selects a new preset grayscale threshold according to the difference between the grayscale values ​​in the first data set and the second data set and the preset grayscale threshold, so that the difference degree can be greater than the preset difference threshold. The present application can make the new grayscale threshold more in line with the actual situation of the image data, and improve the accuracy and efficiency of subsequent processing.

[0013] Optionally, after executing the fourth judgment step and before executing the fourth processing step, the method further includes: Fifth collection: recording the difference degree calculated in the first calculation step as first data, and integrating all the first data into a third data set; Extracting the maximum value: extracting the largest first data in the third data set, recording it as the second data, and obtaining the preset grayscale threshold corresponding to the second data, recording it as the third data; First search: search the gray value with the smallest difference with the third data in the first data set and the second data set, and record it as the fourth data; Determine the seed point: use the pixel point corresponding to the fourth data as the seed point in the fourth processing step.

[0014] The present application integrates the degree of difference obtained after each execution of the first calculation step into a third data set, then finds the maximum degree of difference (i.e., the second data) in the third data set, and obtains the corresponding preset grayscale threshold (i.e., the third data) to obtain the preset grayscale threshold corresponding to the maximum degree of difference. The maximum degree of difference reflects the most significant or important feature changes in the image data. Then, the grayscale value closest to the third data (i.e., the fourth data) is found in the first data set and the second data set, and then the pixel point corresponding to the fourth data is used as the seed point in the fourth processing step, thereby improving the accuracy and efficiency of image segmentation. The present application not only improves the accuracy and efficiency of image processing by deeply analyzing the degree of difference data, extracting key feature points, and determining appropriate seed points, but also helps to discover potential laws and features in image data, providing more critical information for subsequent image segmentation.

[0015] Optionally, after performing the first search step and before performing the seed point determination step, the method further includes: Sixth judgment: judging whether the quantity of the fourth data is equal to one, if so, executing the step of determining the seed point; if not, executing the step of integrating; Integration: Integrate all fourth data into a fourth data set; First acquisition: acquiring a pixel point corresponding to the k-th fourth data, recorded as a first pixel point, and acquiring eight neighboring pixel points of the first pixel point, recorded as second pixel points; Second acquisition: obtaining the gray value of the second pixel, recorded as the fifth data; Seventh judgment: using the kth fourth data and the fifth data to judge whether the first pixel is a noise point, if so, executing the first deletion step; if not, executing the iteration step; First deletion: delete the kth fourth data in the fourth data set; Iteration: take the k+1th fourth data as the new kth fourth data, and execute the sixth judgment step.

[0016] The present application determines the execution order of subsequent steps according to the number of fourth data. When the number of fourth data is equal to one, it means that there is only one most likely seed point candidate, and the step of determining the seed point is directly executed. When the number of fourth data is greater than one, it means that there may be noise points. After that, all the fourth data are integrated into a fourth data set, and the pixels are screened to delete the noise points. After that, the present application obtains the pixel point (i.e., the first pixel point) and its eight neighboring pixel points (i.e., the second pixel point) corresponding to the kth fourth data, and then compares the kth fourth data with the fifth data to determine whether the first pixel point is a noise point. If the first pixel point is a noise point, the kth fourth data is deleted from the fourth data set to remove possible noise points and improve the accuracy of subsequent image segmentation. After that, the k+1th fourth data is used as the new kth fourth data, and the sixth judgment step is re-executed to analyze and process all the fourth data one by one until all possible noise points are removed. This application effectively removes possible noise points by carefully analyzing the fourth data and the fifth data, making the final determined seed points more accurate and reliable. This application not only improves the accuracy and efficiency of image segmentation, but also helps to reduce the interference of noise on real-time image segmentation results.

[0017] Optionally, after executing the iterative step and before executing the seed point determination step, the method further includes: Third acquisition: acquiring the remaining fourth data in the fourth data set, recorded as sixth data; Fourth acquisition: acquiring sixth data capable of forming a closed figure, recorded as seventh data; Second deletion: retain the seventh data in the fourth data set and delete the remaining fourth data; In the step of determining the seed point, a pixel point corresponding to any one of the seventh data in the second deleting step is used as the seed point.

[0018] The present application obtains the remaining fourth data in the fourth data set and records it as the sixth data, then obtains the sixth data that can form a closed figure, recorded as the seventh data, then retains the fourth data corresponding to the seventh data in the fourth data set, deletes the remaining fourth data, and then uses the pixel point corresponding to any seventh data in the fourth data set (i.e., the fourth data after being filtered by the second deletion step) as the seed point, thereby selecting a more suitable starting point for subsequent processing, which helps to improve the efficiency and accuracy of image processing. By adopting the above scheme, the present application can determine whether the pixel point has connectivity with the surrounding pixel points based on the position distribution of the pixel point, thereby determining whether it is suitable as a seed point.

[0019] Optionally, after executing the second judgment step and before executing the output step, the method further includes: Sixth collection: collecting historical feedback signals of the monitoring target; Add label: add a type label to the historical feedback signal, and record the historical feedback signal after adding the type label as the first signal; Second modeling and training: establishing an RNN model, using the first signal to train the RNN model, and obtaining a trained RNN model; Recognition: Input the real-time feedback signal into the trained RNN model to obtain the predicted type label; The eighth judgment: judge whether the predicted type label is an abnormal label, if so, execute the output step; if not, execute the first acquisition step.

[0020] This application collects historical feedback signals of the monitoring target, then adds type labels (such as normal, abnormal, etc.) to the historical feedback signals, then establishes an RNN (recurrent neural network) model, and uses the first signal (i.e., the historical feedback signal after adding the type label) to train the RNN model, so that the trained RNN model can classify the real-time feedback signal, and then by inputting the real-time feedback signal into the trained RNN model, the type prediction result of the signal can be obtained, that is, whether it is an abnormal signal. If the prediction result is abnormal, the output step is executed; if the prediction result is normal, the first acquisition step is returned to continue monitoring or perform other related operations. This application can learn and identify abnormal behaviors or states of the monitoring target based on historical data, so as to take corresponding measures in time.

[0021] Optionally, after performing the second modeling and training step and before performing the recognition step, the method further includes: Seventh acquisition: acquiring the historical high-frequency ultrasonic signal corresponding to the historical feedback signal, and the difference between the time of acquiring the historical feedback signal and the time of acquiring the historical high-frequency ultrasonic signal; Model expansion: In the trained RNN model, multiple hidden layers are added to obtain an expanded RNN model. The added hidden layer in the expanded RNN model is recorded as the first hidden layer, and the original hidden layer in the expanded RNN model is recorded as the second hidden layer. The third training: fixing the second hidden layer, using the historical high-frequency ultrasonic signal and the difference to train the first hidden layer, and obtaining a retrained RNN model; Model update: Use the retrained RNN model as the new trained RNN model and perform the recognition steps; In the recognition step, the difference between the time of collecting the real-time feedback signal and the time of collecting the real-time high-frequency ultrasonic signal is also input into the new trained RNN model.

[0022] The present application collects the historical high-frequency ultrasonic signals corresponding to the historical feedback signals, and the difference in the acquisition time of these signals. If there is a welding defect, the time difference between sending the high-frequency ultrasonic signal and receiving the feedback signal is different. Therefore, the present application also collects the time difference between the two and retrains the trained RNN model, so that the retrained RNN model can further judge whether there is a welding defect based on the time difference between the feedback signal and the high-frequency ultrasonic signal. The present application also adds multiple hidden layers to the trained RNN model, and fixes the second hidden layer (i.e., the original hidden layer) in the RNN model, and only uses the historical high-frequency ultrasonic signal and the time difference to train the first hidden layer (i.e., the newly added hidden layer), thereby introducing new signal types and features while maintaining the original RNN model structure and learning results. By fixing the second hidden layer, the useful information that the original RNN model has learned can be retained during the retraining process without destroying it, and the training of the first hidden layer of the RNN model can be focused on, so that it can focus on learning new signal features and patterns. This application introduces high-frequency ultrasonic signals and time differences as new input features, and expands and trains the RNN model accordingly, so that the RNN model can learn richer information, thereby improving the accuracy and reliability of welding defect monitoring results.

[0023] In the second aspect, the present application provides a welding defect monitoring system based on vibroacoustics, which adopts the following technical solution: A welding defect monitoring system based on vibroacoustics, comprising: The data acquisition module includes a first acquisition unit and a second acquisition unit; The first acquisition unit is used to apply a low-frequency vibration signal to the monitoring target and obtain a response signal; The second acquisition unit is used to apply a real-time high-frequency ultrasonic signal to the monitoring target to obtain a real-time feedback signal; A data processing module, comprising a first processing unit and a second processing unit; A first processing unit is used to perform spectrum analysis on the response signal to obtain modal parameters; A second processing unit, used for extracting signal features of the real-time feedback signal; A defect judgment module, comprising a first judgment unit and a second judgment unit; A first judging unit is used to judge whether the modal parameter belongs to a preset parameter range, and if so, trigger the second acquisition unit; otherwise, trigger the output module; The second judgment unit is used to judge whether the signal characteristics meet expectations. If so, the first acquisition unit is triggered after a preset time interval; otherwise, the output module is triggered; Output module, used to output alarm signals.

[0024] In this application, a low-frequency vibration signal is applied by the first acquisition unit to obtain a response signal for preliminary evaluation, and a real-time high-frequency ultrasonic signal is applied by the second acquisition unit to obtain detailed feedback. The data processing module performs spectrum analysis and feature extraction on the signal. The defect judgment module judges the monitoring target state according to the preset parameter range and signal characteristics, and triggers the output module to output an alarm signal when an abnormality is found. This application combines multiple signals for continuous monitoring of welding defects, thereby improving the accuracy of welding defect monitoring.

[0025] In summary, the present application includes at least one of the following beneficial technical effects: 1. This application adopts a dual judgment mechanism to monitor welding defects, which improves the accuracy and reliability of welding defect monitoring.

[0026] 2. This application realizes multi-dimensional monitoring of welding defects by combining visual monitoring with vibration acoustics and ultrasonic monitoring technology. Visual monitoring can intuitively reflect the surface condition of the welding process, while vibration acoustics and ultrasonic monitoring technology can deeply detect the internal defects of the welding structure. This comprehensive monitoring method improves the comprehensiveness and accuracy of monitoring and provides a more reliable guarantee for welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a flow chart of Example 1 of the present application; Figure 2 This is a flowchart from S01 third collection to S08 third judgment of Example 2 of the present application; Figure 3 This is a flowchart of the fifth process from S51 to S66 of adjusting the threshold value in Embodiment 2 of the present application; Figure 4 This is the flow chart from S71 third acquisition to S73 second deletion in Example 2 of the present application; Figure 5 This is a flow chart of Example 3 of the present application. DETAILED DESCRIPTION

[0028] The following combination Figures 1 to 5 This application is described in further detail.

[0029] Embodiment 1: This embodiment discloses a welding defect monitoring method based on vibroacoustics, referring to Figure 1The monitoring method includes: S1 data acquisition, S2 data processing, S3 defect judgment and S4 output. S1 data acquisition includes: two stages: first acquisition (applying low-frequency vibration signal to obtain response signal) and second acquisition (applying real-time high-frequency ultrasonic signal to obtain feedback signal). Then, S2 data processing (including spectrum analysis and feature extraction) is performed, and then S3 defect judgment (judging whether modal parameters and signal characteristics meet preset standards) is used to decide whether to continue monitoring or output an alarm signal. The process of this embodiment is as follows: S1 data collection, including S11 first collection and S12 second collection.

[0030] S11 is the first acquisition, applying a low-frequency vibration signal to the monitored target and obtaining a response signal.

[0031] Low-frequency vibration signals are used to stimulate the modal response of the monitored target. The mode is the vibration mode of the welded structure at a specific frequency and is closely related to the physical properties of the welded structure (such as mass, stiffness and damping). By applying a low-frequency vibration signal and measuring the resulting response signal, the modal behavior of the monitored target can be captured.

[0032] The response signal contains information about the target modal parameters to be monitored, such as modal frequency, modal damping, and mode shape.

[0033] S12 is the second acquisition, applying a real-time high-frequency ultrasonic signal to the monitoring target to obtain a real-time feedback signal.

[0034] Real-time high-frequency ultrasonic signals are used to detect the monitoring target in more detail. High-frequency ultrasonic signals have the characteristics of strong penetration and high resolution, and can penetrate deep into the material and reflect its microstructure and defects.

[0035] By applying real-time high-frequency ultrasonic signals to the monitoring target and receiving the real-time feedback signals generated thereby, it is possible to capture detailed information about the internal state of the monitoring target, including the density, thickness, cracks, slag inclusions, etc. of the welding material, thereby determining whether there are welding defects in the monitoring target.

[0036] S2 data processing, including S21 first processing and S22 second processing.

[0037] S21 is the first processing, performing spectrum analysis on the response signal to obtain modal parameters.

[0038] Spectral analysis technology converts the response signal from the time domain to the frequency domain, thereby revealing the strength of different frequency components in the signal. Through spectrum analysis, modal parameters in the response signal, such as modal frequency and modal damping, can be extracted.

[0039] S22 is a second processing step of extracting signal features of the real-time feedback signal, wherein the signal features include: amplitude, frequency, phase, etc.

[0040] S3 defect judgment, including S31 first judgment and S32 second judgment.

[0041] S31 is the first judgment, which determines whether various modal parameters belong to the preset parameter range. If so, it indicates that the monitored target may be in a normal state, and then S12 is executed for the second acquisition to further determine whether the monitored target has welding defects; if not, it indicates that the monitored target has defects or abnormalities, and S4 is executed for output.

[0042] S32 is the second judgment, judging whether the signal characteristics meet expectations. If so, it indicates that the monitored target is in a normal state. After a preset interval, S11 is executed for the first acquisition. If not, it indicates that the monitored target has defects or abnormalities, and S4 is executed for output.

[0043] The signal characteristics meet the expected standards when the signal characteristics such as amplitude, frequency, and phase are within the preset range.

[0044] S4 output, output alarm signal.

[0045] This embodiment includes S1 data acquisition (including S11 first acquisition of applying a low-frequency vibration signal to the monitoring target to obtain a response signal, and S12 second acquisition of applying a real-time high-frequency ultrasonic signal to obtain a real-time feedback signal), S2 data processing (including S21 first processing of performing spectral analysis on the response signal to obtain modal parameters, and S22 second processing of extracting signal characteristics of the real-time feedback signal), S3 defect judgment (including S31 first judgment based on whether the modal parameters are within the preset parameter range and S32 second judgment based on whether the signal characteristics meet expectations) and S4 output, so as to accurately determine whether the monitoring target has welding defects.

[0046] Embodiment 2: This embodiment differs from Embodiment 1 in that, referring to Figure 2 Before executing S1 data collection, the method further includes: S01 The third acquisition is to collect historical images in different welding processes and add category labels to the historical images. These historical images include images of different welding materials under different welding parameters. The category labels added to these historical images include normal welding labels and abnormal welding labels. The abnormal welding labels also include: cracks, pores, slag inclusions, etc.

[0047] S02 fourth acquisition, acquiring real-time images of the monitored target.

[0048] S03: The third processing is to perform grayscale processing on the real-time image to obtain the grayscale value of each pixel in the real-time image.

[0049] S04 is a fourth process, using a region growing algorithm to segment the grayscale real-time image to obtain a plurality of image blocks, wherein the image blocks include: a target image block containing the monitored target and a remaining image block not containing the monitored target.

[0050] S05 resets the grayscale value, and sets the grayscale values ​​of the pixels corresponding to the remaining image blocks in the real-time image to 0, thereby more clearly highlighting the target image blocks containing the monitored targets, while reducing the interference of background information, and uses the image with the set grayscale value as a new real-time image.

[0051] S06 First modeling and training, build a CNN model, and use historical images and added category labels to train the CNN model to obtain a trained CNN model. Through training, a CNN model that can accurately identify the category of the monitoring target (such as whether the welding state is normal, the type of welding defects, etc.) can be obtained.

[0052] S07 is a first prediction, in which the new real-time image in the grayscale value reset in S05 is input into the trained CNN model to obtain a prediction result.

[0053] S08 is the third judgment, judging whether the prediction result is in line with expectations. If so, S11 is executed to collect and continue to monitor whether the welding structure has defects; if not, S4 is executed to output.

[0054] This embodiment first collects welding history images and marks the categories, then obtains the real-time image of the monitoring target, then grays the real-time image, then uses the region growing algorithm to segment the image to distinguish the target from the background, sets the gray value of the background to 0 to highlight the target area, then builds and trains a CNN model to identify the state of the monitoring target, inputs the real-time image into the trained model to obtain the prediction result, and then decides whether to continue collecting or outputting the alarm signal based on whether the prediction result meets the expectation. This embodiment, on the basis of applying low-frequency vibration signals and high-frequency ultrasonic signals to the monitoring target, combines historical graphics and real-time images to further improve the monitoring capability of welding defects, thereby realizing the function of accurately monitoring welding defects.

[0055] Reference Figure 3 In other embodiments, after executing the third process S03 and before executing the fourth process S04, the method further includes: S51 fifth processing, based on the gray value of each pixel in the real-time image in the third processing S03, the gray values ​​greater than the preset gray threshold are integrated into the first data set, and the remaining gray values ​​are integrated into the second data set.

[0056] S52: First calculation, calculating the difference between the first data set and the second data set. The calculation model of the difference is as follows: ; Where r represents the degree of difference; A is the number of grayscale values ​​in the first data set; B is the number of grayscale values ​​in the second data set; is the average value of all gray values ​​in the first data set; is the average value of all gray values ​​in the second data set.

[0057] S53 is the fourth judgment, judging whether the difference degree calculated in the first calculation of S52 is less than the preset difference threshold. If so, it indicates that the currently set preset grayscale threshold is not enough to distinguish the target area and background of the real-time image, and it is necessary to execute S65 fifth judgment to adjust the preset grayscale threshold; if not, execute S54 fifth acquisition.

[0058] S54 is a fifth collection, recording the difference degree calculated in the first calculation of S52 as the first data, and obtaining the first data calculated after each execution of the first calculation of S52, and integrating all the first data into a third data set.

[0059] S55 extracts the maximum value, extracts the largest first data in the third data set, records it as the second data, and obtains the preset gray threshold corresponding to the second data, records it as the third data.

[0060] S56 first search, respectively calculate the difference between the grayscale value in the first data set and the third data, respectively calculate the difference between the grayscale value in the second data set and the third data, obtain the minimum value among all the differences, and determine the grayscale value corresponding to the minimum value, and record the grayscale value corresponding to the minimum value as the fourth data.

[0061] S57 is a sixth judgment, determining whether the quantity of the fourth data is equal to one, if so, executing S64 to determine the seed point; if not, executing S58 to integrate.

[0062] S58 integrates all the fourth data into a fourth data set.

[0063] S59 first obtains, obtains the pixel point corresponding to the k-th fourth data, records it as the first pixel point, and determines the eight neighboring pixel points of the first pixel point, records them as the second pixel points.

[0064] S60: second acquisition, obtaining the grayscale value of the second pixel, recorded as fifth data.

[0065] S61 is the seventh judgment, using the kth fourth data and the fifth data to judge whether the first pixel is a noise point, if so, executing S62 for the first deletion; if not, executing S63 for iteration.

[0066] The method of using the kth fourth data and the fifth data to determine whether the first pixel is a noise point is: determine whether the brightness of the first pixel is significantly different from the brightness of the second pixel. If so, the first pixel is considered to be a noise point; otherwise, the first pixel is considered not to be a noise point.

[0067] S62 first deletes, deleting the kth fourth data in the fourth data set to remove noise points.

[0068] In S63, it is iterated to take the k+1th fourth data as the new kth fourth data, and to execute the sixth judgment in S57 until all the pixels are compared with the grayscale values ​​of the eight neighboring pixels.

[0069] S64 determines the seed point, and uses the pixel point corresponding to the fourth data that has not been deleted in the fourth data set as the seed point in the fourth processing S04.

[0070] S65 is the fifth judgment, judging whether A is greater than B. If so, it indicates that the number of samples in the first data set is greater than the number of sample data in the second data set, and then S661 is executed for the first adjustment; if not, S662 is executed for the second adjustment.

[0071] S66 adjusts the threshold, including S661 a first adjustment and S662 a second adjustment.

[0072] S661: First adjustment, respectively calculating the difference between each gray value in the first data set and the preset gray threshold, recording it as the first difference, selecting the gray value corresponding to the minimum first difference as the new preset gray threshold, and executing S51: Fifth processing.

[0073] S662: second adjustment, calculating the difference between the grayscale value in the second data set and the preset grayscale threshold, recording it as the second difference, selecting the grayscale value corresponding to the minimum second difference as the new preset grayscale threshold, and executing S51: fifth processing.

[0074] This embodiment performs the fifth processing based on the grayscale values ​​of the pixels in the real-time image, and divides the grayscale values ​​into a first data set and a second data set; then, the difference between the two data sets is calculated, and based on the comparison result of the difference between the two data sets and the preset threshold, it is decided whether to adjust the preset grayscale threshold; through a series of calculation, judgment and iteration processes, noise points are identified and removed, and seed points are determined; finally, based on the number of samples of the first data set and the second data set, different adjustment strategies are selected to optimize the preset grayscale threshold, including respectively calculating the difference between the grayscale values ​​in the two data sets and the preset grayscale threshold, and selecting the grayscale value that minimizes the difference as the new preset grayscale threshold, and then re-executing the steps of the fifth processing to improve the accuracy and efficiency of image processing.

[0075] Reference Figure 4In other embodiments, after executing S63 iteration and before executing S64 to determine the seed point, the method further includes: S71: third acquisition, obtaining the remaining fourth data in the fourth data set, recorded as sixth data.

[0076] S72 fourth acquisition, marking the pixel points corresponding to the sixth data in the real-time image, acquiring the pixel points that can form a closed figure, and recording the corresponding sixth data as the seventh data.

[0077] S73: second deletion, retaining the seventh data in the fourth data set and deleting the remaining fourth data.

[0078] In determining the seed point in S64, a pixel point corresponding to any seventh data in the fourth data set in the second deletion in S73 is used as the seed point.

[0079] This embodiment first obtains the remaining data in the fourth data set as the sixth data, and then selects the partial data corresponding to the pixel points that can form a closed figure as the seventh data, and then only retains the seventh data in the fourth data set and deletes the remaining data. Finally, in the step of determining the seed point, any pixel point corresponding to the retained seventh data is selected as the seed point for subsequent image processing or data analysis tasks. This embodiment further selects the fourth data, and further improves the efficiency of image processing.

[0080] Example 3: Reference Figure 5 The difference between this embodiment and embodiment 2 is that after executing the second judgment in S32 and before executing the output in S4, it also includes: S81 Sixth collection, collecting historical feedback signals of the monitoring target.

[0081] S82 adds a label, adds a type label to the historical feedback signal, and records the historical feedback signal after the type label is added as the first signal.

[0082] S83 second modeling and training, establishing an RNN model, using the first signal to train the RNN model, and obtaining a trained RNN model.

[0083] S84 seventh acquisition, collecting the historical high-frequency ultrasonic signals corresponding to the historical feedback signals, recording their timestamps, and calculating the difference between the two timestamps.

[0084] S85 model expansion: in the trained RNN model, multiple hidden layers are added to obtain the expanded RNN model. The added hidden layer in the expanded RNN model is recorded as the first hidden layer, and the original hidden layer in the expanded RNN model is recorded as the second hidden layer.

[0085] S86 The third training is to fix the second hidden layer, and use the historical high-frequency ultrasonic signal and the difference to train the first hidden layer to obtain the retrained RNN model.

[0086] S87 model update, the retrained RNN model is used as the new trained RNN model.

[0087] S88 identifies and inputs the difference between the real-time feedback signal, the time for collecting the real-time feedback signal, and the time for collecting the real-time high-frequency ultrasonic signal into the trained RNN model to obtain a predicted type label.

[0088] S89 is the eighth judgment, judging whether the predicted type label is an abnormal label, if so, executing S4 output, if not, executing S11 first collection.

[0089] This embodiment first collects the historical feedback signals of the monitoring target and adds type labels to them, then establishes an RNN model and uses these labeled signals for training. Then, the historical high-frequency ultrasonic signals and time differences associated with the feedback signals are collected, the RNN model is expanded, and some hidden layers are fixed, and only the newly added hidden layers are trained for high-frequency ultrasonic signals and time differences. After the model is updated, the real-time feedback signal and time difference are input into the model for type identification, and it is determined whether to perform abnormal processing or continue monitoring operations based on the prediction results. This embodiment determines whether the feedback signal is an abnormal signal by judging the label of the real-time feedback signal, and then outputs an alarm signal, otherwise the first acquisition step is executed.

[0090] Embodiment 4: This embodiment discloses a welding defect monitoring system based on vibroacoustics, the monitoring system comprising: The data acquisition module includes a first acquisition unit and a second acquisition unit.

[0091] The first acquisition unit is used to apply a known low-frequency vibration signal to the monitored target and then obtain the target's response to the signal (ie, obtain the corresponding signal).

[0092] The second acquisition unit is used to apply a real-time high-frequency ultrasonic signal to the monitoring target and obtain a real-time feedback signal.

[0093] The data processing module includes a first processing unit and a second processing unit.

[0094] The first processing unit is connected to the first acquisition unit for performing spectrum analysis on the response signal to obtain modal parameters (such as modal frequency, modal damping, etc.).

[0095] The second processing unit is communicatively connected to the second acquisition unit and is used to extract signal features of the real-time feedback signal.

[0096] The defect judgment module includes a first judgment unit and a second judgment unit.

[0097] The first judgment unit is connected to the first processing unit and the second acquisition unit for judging whether the modal parameter belongs to the preset parameter range. If it does, the second acquisition unit is triggered; otherwise, the output module is triggered. By comparing the degree of deviation between the modal parameter and the preset range, the first judgment unit can evaluate whether there are abnormalities or defects in the welding structure of the monitoring target. If the modal parameter is within the preset range, the monitoring target is considered to be in a normal state; otherwise, it is considered that there is a potential problem.

[0098] The second judgment unit is connected to the second processing unit and the first acquisition unit for judging whether the signal characteristics meet expectations. If they meet expectations, the first acquisition unit is triggered after a preset time interval; otherwise, the output module is triggered. By comparing and analyzing the signal characteristics with the expected values, the second judgment unit can detect whether abnormal changes or potential defects have occurred in the monitoring target. If the signal characteristics match the expected values ​​or the deviation is within an acceptable range, the monitoring target is considered to be in a normal state; otherwise, it is considered to be abnormal.

[0099] The output module is connected to the defect judgment module for outputting an alarm signal.

[0100] This embodiment integrates a data acquisition module (including a first acquisition unit that applies a low-frequency vibration signal to obtain a response signal, and a second acquisition unit that applies a real-time high-frequency ultrasonic signal to obtain a feedback signal), a data processing module (including a first processing unit that performs spectrum analysis to obtain modal parameters, and a second processing unit that extracts real-time signal characteristics), a defect judgment module (including a first judgment unit that judges whether the modal parameters are abnormal to trigger corresponding operations, and a second judgment unit that evaluates whether the signal characteristics meet expectations to determine subsequent steps) and an output module to achieve accurate monitoring of welding defects.

[0101] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A welding defect monitoring method based on vibroacoustics, characterized in that: include: Data collection: including the first collection and the second collection; First acquisition: applying a low-frequency vibration signal to the monitored target and obtaining a response signal; Second acquisition: applying real-time high-frequency ultrasonic signals to the monitoring target to obtain real-time feedback signals; Data processing: including first processing and second processing; First processing: perform spectrum analysis on the response signal to obtain modal parameters; Second processing: extracting signal features of real-time feedback signals; Defect judgment: including the first judgment and the second judgment; First judgment: judging whether the modal parameter belongs to the preset parameter range, if yes, executing the second acquisition step; if no, executing the output step; Second judgment: judging whether the signal characteristics meet expectations, if so, executing the first acquisition step after a preset interval; if not, executing the output step; Output: Output alarm signal.

2. The welding defect monitoring method based on vibroacoustics according to claim 1, characterized in that: Before executing the step of data collection, the method further comprises: The third collection: collect historical images during welding and add category labels to the historical images; Fourth acquisition: collect real-time images of the monitoring target; First, modeling and training: establish a CNN model, and use historical images and added category labels to train the CNN model to obtain a trained CNN model; First prediction: input the real-time image into the trained CNN model to obtain the prediction result; The third judgment: judge whether the prediction result meets the expectation. If so, execute the first acquisition step; if not, execute the output step.

3. The welding defect monitoring method based on vibroacoustics according to claim 2 is characterized in that: After performing the fourth acquisition step and before performing the first modeling and training step, the method further includes: The third processing: grayscale processing is performed on the real-time image to obtain the grayscale value of each pixel in the real-time image; Fourth processing: using a region growing algorithm to segment the grayscale real-time image to obtain a target image block containing the monitored target and a remaining image block not containing the monitored target; Reset grayscale value: Set the grayscale values ​​of the corresponding pixels of the remaining image blocks in the real-time image to 0 to obtain a new real-time image.

4. The welding defect monitoring method based on vibroacoustics according to claim 3 is characterized in that: After executing the step of the third process and before executing the step of the fourth process, the method further includes: Fifth processing: integrating the grayscale values ​​greater than the preset grayscale threshold into a first data set, and integrating the remaining grayscale values ​​into a second data set; First calculation: Calculate the difference between the first data set and the second data set. The calculation model of the difference is as follows: ； Where r represents the degree of difference; A is the number of grayscale values ​​in the first data set; B is the number of grayscale values ​​in the second data set; is the average value of all gray values ​​in the first data set; is the average value of all gray values ​​in the second data set; Fourth judgment: judging whether the difference degree is less than a preset difference threshold, if so, executing the fifth judgment step; if not, executing the fourth processing step; Fifth judgment: judging whether A is greater than B, if so, executing the first adjustment step; if not, executing the second adjustment step; Adjustment threshold: including first adjustment and second adjustment; First adjustment: respectively calculating the difference between each gray value in the first data set and the preset gray threshold, recording it as the first difference, selecting the gray value corresponding to the minimum first difference as the new preset gray threshold, and executing the fifth processing step; Second adjustment: Calculate the difference between the grayscale value in the second data set and the preset grayscale threshold, record it as the second difference, select the grayscale value corresponding to the minimum second difference as the new preset grayscale threshold, and execute the fifth processing step.

5. The welding defect monitoring method based on vibroacoustics according to claim 4 is characterized in that: After executing the fourth judgment step and before executing the fourth processing step, the method further includes: Fifth collection: recording the difference degree calculated in the first calculation step as first data, and integrating all the first data into a third data set; Extracting the maximum value: extracting the largest first data in the third data set, recording it as the second data, and obtaining the preset grayscale threshold corresponding to the second data, recording it as the third data; First search: search the gray value with the smallest difference with the third data in the first data set and the second data set, and record it as the fourth data; Determine the seed point: use the pixel point corresponding to the fourth data as the seed point in the fourth processing step.

6. The welding defect monitoring method based on vibroacoustics according to claim 5, characterized in that: After performing the first search step and before performing the seed point determination step, the method further includes: Sixth judgment: judging whether the quantity of the fourth data is equal to one, if so, executing the step of determining the seed point; if not, executing the step of integrating; Integration: Integrate all fourth data into a fourth data set; First acquisition: acquiring a pixel point corresponding to the k-th fourth data, recorded as a first pixel point, and acquiring eight neighboring pixel points of the first pixel point, recorded as second pixel points; Second acquisition: obtaining the gray value of the second pixel, recorded as the fifth data; Seventh judgment: using the kth fourth data and the fifth data to judge whether the first pixel is a noise point, if so, executing the first deletion step; if not, executing the iteration step; First deletion: delete the kth fourth data in the fourth data set; Iteration: take the k+1th fourth data as the new kth fourth data, and execute the sixth judgment step.

7. The welding defect monitoring method based on vibroacoustics according to claim 6, after the step of performing iteration and before the step of determining a seed point, further comprising: Third acquisition: acquiring the remaining fourth data in the fourth data set, recorded as sixth data; Fourth acquisition: acquiring sixth data capable of forming a closed figure, recorded as seventh data; Second deletion: retain the seventh data in the fourth data set and delete the remaining fourth data; In the step of determining the seed point, a pixel point corresponding to any one of the seventh data in the second deleting step is used as the seed point.

8. The welding defect monitoring method based on vibroacoustics according to any one of claims 1 to 7, characterized in that: After executing the second judgment step and before executing the output step, the method further includes: Sixth collection: collecting historical feedback signals of the monitoring target; Add label: add a type label to the historical feedback signal, and record the historical feedback signal after adding the type label as the first signal; Second modeling and training: establishing an RNN model, using the first signal to train the RNN model, and obtaining a trained RNN model; Recognition: Input the real-time feedback signal into the trained RNN model to obtain the predicted type label; The eighth judgment: judge whether the predicted type label is an abnormal label, if so, execute the output step; if not, execute the first acquisition step.

9. The welding defect monitoring method based on vibroacoustics according to claim 8, characterized in that: After performing the second modeling and training step and before performing the recognition step, the method further includes: Seventh acquisition: acquiring the historical high-frequency ultrasonic signal corresponding to the historical feedback signal, and the difference between the time of acquiring the historical feedback signal and the time of acquiring the historical high-frequency ultrasonic signal; Model expansion: In the trained RNN model, multiple hidden layers are added to obtain an expanded RNN model. The added hidden layer in the expanded RNN model is recorded as the first hidden layer, and the original hidden layer in the expanded RNN model is recorded as the second hidden layer. The third training: fixing the second hidden layer, using the historical high-frequency ultrasonic signal and the difference to train the first hidden layer, and obtaining a retrained RNN model; Model update: Use the retrained RNN model as the new trained RNN model and perform the recognition steps; In the recognition step, the difference between the time of collecting the real-time feedback signal and the time of collecting the real-time high-frequency ultrasonic signal is also input into the new trained RNN model.

10. A welding defect monitoring system based on vibroacoustics, the system being applicable to the method according to any one of claims 1 to 9, characterized in that: include: The data acquisition module includes a first acquisition unit and a second acquisition unit; The first acquisition unit is used to apply a low-frequency vibration signal to the monitoring target and obtain a response signal; The second acquisition unit is used to apply a real-time high-frequency ultrasonic signal to the monitoring target to obtain a real-time feedback signal; A data processing module, comprising a first processing unit and a second processing unit; A first processing unit is used to perform spectrum analysis on the response signal to obtain modal parameters; A second processing unit, used for extracting signal features of the real-time feedback signal; A defect judgment module, comprising a first judgment unit and a second judgment unit; A first judging unit is used to judge whether the modal parameter belongs to a preset parameter range, and if so, trigger the second collecting unit; Otherwise, the output module is triggered; The second judgment unit is used to judge whether the signal characteristics meet expectations. If they meet expectations, the first collection unit is triggered after a preset time interval; Otherwise, the output module is triggered; Output module, used to output alarm signals.

Images