Online detection method and device for keyhole type hole in laser welding seam and storage medium
Through machine learning model combined with multi-source information, the internal hole detection of welds is solved, and the problems of single signal and weak data analysis methods are realized, high-precision online detection of internal holes in welds is suitable for long-range or batch laser welding, and the detection is improved System stability and applicability.
Patent Information
- Application Number
- CN202411174097.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-29
AI Technical Summary
The prior art has problems such as single signal and weak data analysis methods in the online detection of internal holes in the weld, which leads to insufficient detection accuracy. Especially in long-range welding or batch welding, process-type air holes formed by keyhole collapse due to external interference. Larger size affects joint strength, airtightness and fatigue performance.
By obtaining the joint assembly information before laser welding, the optical radiation characteristic information during welding, and the actual measured values of the weld morphological characteristics after welding, it is divided into multiple windows along the time axis, and a machine learning model is established, and multi-source information is used for Training, calculate the deviation between the predicted value of the weld morphological characteristics and the measured value, and realize the online detection of keyhole-shaped holes inside the weld.
It realizes full integration of multi-source data, improves detection accuracy and stability, reduces dependence on expert experience, expands applicable scenarios, can efficiently and accurately identify large-size holes inside the weld, and improves the detailed perception of the detection system. ability.
Smart Images

Figure CN119046734A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of welding detection, and in particular to an online detection method, equipment and storage medium for keyhole-shaped holes inside a laser weld. Background Art
[0002] Laser welding has the advantages of good flexibility, strong adaptability, and high degree of automation, and is widely used in the manufacturing industry. However, the interaction between the welded material and the laser is easily disturbed and unstable, and the collapse of the keyhole can easily form large holes inside the weld, which will cause the strength, sealing and fatigue performance of the joint to decrease significantly. Online identification of holes inside the weld is critical to ensuring product quality and safety. Due to the low level of operator participation in the laser welding process, it is difficult to detect internal defects in the weld in time, so it is urgent to develop an intelligent detection method for weld holes.
[0003] The search found that the representative achievements in this field include: the paper "Weld X-ray Image Defect Detection Based on Edge AI" proposed a weld X-ray image analysis based on a deep learning model with a lightweight inverse residual structure, and the paper "Research on Weld Defect Detection Based on Adaptive Fusion of Multi-domain Multi-scale Deep Features" used deep learning to mine the anomalies of weld ultrasonic signals, thereby realizing internal weld defect detection. These methods are based on data analysis of X-ray or ultrasonic signals. Although deep learning technology helps to improve detection accuracy, dedicated X-ray imaging and ultrasonic detection equipment limit the application of related methods. Chinese Patent Announcement No. CN104655564B discloses a double-shoulder stir friction weld hole defect detection method, which is based on the fluctuation of welding torque signals to identify internal weld defects. No X-ray imaging equipment is required, but the basis for signal identification is difficult to determine, and the accuracy of identification mainly depends on expert experience, which makes it difficult to promote such methods. Chinese patent publication number CN104959719B discloses a rapid nondestructive detection method for hole defects in fusion welding seams. The weld morphology is detected and compared with the plane of the parent material to calculate the excess height area, thereby inferring the internal holes of the weld. Due to insufficient relevant information and a single means of information extraction, this type of method is difficult to guarantee detection accuracy in practical applications.
[0004] In summary, there is currently a lack of effective means for online detection of holes inside welds. Methods based on X-ray imaging or ultrasonic echo signals are limited in terms of cost, efficiency, and site. Detection methods based on welding process signals or weld morphology have shortcomings such as single signals and weak data analysis methods. With the promotion of laser welding, process-type pores formed by keyhole collapse due to external interference in long-distance welding or batch welding are usually large in size, which has a significant adverse effect on joint strength, air tightness, and fatigue performance. An online detection method for holes inside welds is urgently needed. Summary of the invention
[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide an online detection method, equipment and storage medium for keyhole-type holes inside laser welds, so as to solve or partially solve the problems of single signal and weak data analysis means in the detection means.
[0006] The purpose of the present invention can be achieved by the following technical solutions:
[0007] One aspect of the present invention provides an online detection method for keyhole-shaped holes inside a laser weld, comprising the following steps:
[0008] Step S1, obtaining the seam assembly information before laser welding, the light radiation characteristic information during laser welding, and the measured value of the weld morphology characteristic after welding, and obtaining multiple samples by dividing the information into multiple windows along the time axis, taking each window as a sample;
[0009] Step S2, using the multiple samples, taking the seam assembly information and the light radiation characteristic information as inputs of the machine learning model, taking the measured values of the weld morphology characteristics as outputs of the machine learning model, and training the machine learning model;
[0010] Step S3, using the trained machine learning model and the plurality of samples as input to calculate the predicted value of the weld morphology feature;
[0011] Step S4, calculating the deviation between the predicted value of the weld morphology feature and the measured value of the weld morphology feature, and obtaining the online detection result of the keyhole-shaped hole inside the weld.
[0012] As a preferred technical solution, step S3 includes the following sub-steps:
[0013] Step S301, calculating the deviation between the predicted value of the weld morphology feature and the measured value of the weld morphology feature to obtain a long-range weld marking curve, in which the window is divided into multiple levels according to the deviation;
[0014] Step S302, based on a preset decision rule, an abnormal interval of the long-range weld marking curve is obtained as the position of a large-size hole, that is, an online detection result of a keyhole-type hole inside the weld.
[0015] As a preferred technical solution, the decision rule is:
[0016] Deviation windows occur continuously or frequently, and / or windows with high deviation levels occur.
[0017] As a preferred technical solution, in step S1, the window width of the divided window satisfies the following conditions:
[0018] The number of line laser scans and optical radiation signal data points within a single window exceeds a preset value.
[0019] As a preferred technical solution, the weld morphology characteristics include weld width, excess height, undercut depth and expansion area.
[0020] As an optimal technical solution, the expansion area is calculated based on the morphology difference before and after welding. In the process of calculating the morphology difference, the contour inflection points on both sides of the seam and the seam center point are selected for the missing contour data at the assembly seam, and filling is achieved through quadratic polynomial fitting.
[0021] As a preferred technical solution, the light radiation characteristic information is obtained through two visible light band sensors and two infrared light band sensors, and the light radiation characteristic information includes the mean of each sensor signal, the ratio of the two visible light band sensors, the ratio of the two infrared light band sensors, and the intensity ratio between the three-level wavelet transforms of each sensor signal.
[0022] As a preferred technical solution, the seam assembly information is obtained by line laser profiling and laser triangulation, and the seam assembly information includes seam width, undercut depth and excess height area.
[0023] Another aspect of the present invention provides an online detection device for keyhole-shaped holes inside a laser weld, comprising a data processing terminal and a photoelectric sensor array and a line laser scanning device respectively connected to the data processing terminal, wherein the photoelectric sensor array comprises:
[0024] A first visible light sensor is aimed at a laser entry point on the surface of the welding workpiece;
[0025] The second visible light sensor is aligned 8-10 mm above the laser entry point;
[0026] A first infrared light sensor is aimed at the laser incident point;
[0027] The second infrared light sensor is aimed at the rear end of the molten pool.
[0028] Another aspect of the present invention provides a computer-readable storage medium comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the aforementioned online detection method for keyhole-type holes inside a laser weld.
[0029] Compared with the prior art, the present invention has at least one of the following beneficial effects:
[0030] (1) Fully integrate multi-source data to achieve online monitoring: The present invention integrates multi-source information including the seam assembly information before laser welding, the light radiation characteristic information during laser welding, and the measured values of the weld morphology characteristics after welding, extracts characteristic parameters and imports them into the corresponding window, divides the long-range weld into several samples, establishes a mapping relationship through a machine learning model, and obtains the common laws of the entire weld; then, the deviation between the model calculated value and the measured value is analyzed to find the local windows that do not conform to the common laws, and based on the frequency of window appearance and the degree of deviation, the large-sized holes inside the weld are determined.
[0031] (2) High detection accuracy and stability: The present invention combines knowledge and data dual drive, and focuses on keyhole collapse behavior based on expert experience in signal selection, feature design, and model construction. In terms of model training and deviation calculation, it fully utilizes the consistency of the welding process itself and the model learning ability. The dual drive architecture reduces the reliance on expert experience, improves detection accuracy and stability, and expands the applicable scenarios.
[0032] (3) Strong ability to perceive details in the welding process: Visible light sensors are arranged up and down to extract information on the metal vapor ejection behavior, and infrared sensors are arranged front and back to extract information on the oscillation of the molten pool. The characteristic parameters are designed to characterize the fluctuations in the relationship between the signal strengths of multiple sensors, thereby improving the detection system's ability to perceive details of the welding process. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 Flow chart of the online detection method of keyhole-shaped holes inside a laser weld in an embodiment;
[0034] Figure 2 Schematic diagram of an online detection device for keyhole-shaped holes inside a laser weld in an embodiment;
[0035] Figure 3 It is the deviation distribution diagram in the embodiment;
[0036] Figure 4 The cross section of the weld is cut to determine the area where the weld has a large-sized hole in the embodiment.
[0037] Among them, 1. a first visible light sensor, 2. a second visible light sensor, 3. a first infrared light sensor, 4. a second infrared light sensor, and 5. a data processing terminal. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0039] Example 1
[0040] In view of the problems existing in the aforementioned prior art, the present embodiment provides an online detection method for keyhole-type holes inside laser welds for long-distance laser welding or batch laser welding. It integrates the information of the entire welding process and is based on the formation mechanism of keyhole collapse-type holes. For large-sized holes inside welds, an online hole detection method is implemented without the need for X-ray imaging or ultrasonic echo detection. At the same time, it reduces the dependence on expert experience, thereby providing an efficient and high-precision welding defect detection technology for long-distance laser welding or batch laser.
[0041] See also Figure 1 , the method comprises the following steps:
[0042] Step S1, setting the topological structure and hyperparameters of the machine learning model. The data set of the model is the assembly data such as the seam width and height difference measured before the welding operation, as well as the light radiation characteristic parameters during the welding process. The label set of the model is the weld morphology characteristic parameters. The welding process and the corresponding weld are divided into several windows, each window is a sample.
[0043] The weld is divided into several windows along the time axis (or length). Each window contains pre-welding assembly information, light radiation waveform information of the welding process, and weld morphology information. The window width must match the scanning interval and signal acquisition frequency. Each window has at least two line laser scans and at least 20 light radiation signal data points. Specifically, the window time length range is usually 5ms to 30ms. Each window is a sample, containing at least one seam scan data, 20 light signal acquisition data, and one weld morphology scan data. If this cannot be met for some reason, data filling can be achieved by cubic spline interpolation of the entire weld data.
[0044] The sample data set including multiple samples includes the seam assembly characteristic values and the welding process characteristic values, wherein the seam assembly characteristic values include the seam width, undercut depth (i.e., the concave depth of the fusion line area in the weld morphology) and excess height area (i.e., the area of the weld protruding above the workpiece surface). Based on the line laser profile and laser triangulation method, the seam assembly characteristic values can be calculated by image processing methods; usually, each sample includes at least one line laser scan. If this cannot be met, cubic spline interpolation can be used for data processing.
[0045] Step S2, set the window time width according to the interaction between the welded material and the incident laser and the welding process parameters. The window width should not be too narrow, and there is usually at least one laser scan within the weld length corresponding to the window; the window width should not be too long, and the number of windows should meet the training requirements of the machine learning model. Usually, the window time length is generally 1 to 2 times the duration of the common keyhole instability phenomenon. The window adopts a sliding mode, starting from the moment of light welding, and the window number increases along the time axis.
[0046] In this embodiment, the window length is associated with the duration of the keyhole collapse phenomenon, and the keyhole behavior is related to the welded material, laser wavelength, and welding process parameters. For high-strength steel and stainless steel, the window length is set to 5-10ms; for aluminum and aluminum alloys, the window length is set to 10-20ms; for pure copper, the window length is set to 20-30ms.
[0047] Step S3, after the workpiece to be welded is clamped and fixed, a line laser scan is carried out to obtain the three-dimensional morphology of the assembly seam according to the laser triangulation method, calculate the width of the seam and the height difference on both sides of the seam, and import the calculation results into the data set of the corresponding number window.
[0048] After the workpiece is assembled and fixed, the assembly seam is scanned by line laser. The seam width, the height difference on both sides of the seam (i.e., the bite edge depth), and the chamfer area of the seam edge (i.e., the residual height area) are determined based on the laser triangulation method, where the chamfer area is the area difference between the edge arc segment and the virtual straight line segment. This process can be carried out separately, that is, welding can be started after the entire seam measurement is completed; it can also be carried out simultaneously with the welding process, that is, the scanning device is placed a certain distance in front of the laser and a light shield is set, and laser welding is carried out while scanning and measuring.
[0049] Step S4, start the laser to start welding, and during the welding process, the photoelectric sensor collects visible light radiation and infrared light radiation in the welding area; in each window, the characteristic parameters representing the stability of the keyhole are extracted from the light radiation waveform, and the calculation results are imported into the data set of the corresponding numbered window.
[0050] Specifically, the optical radiation signal of the welding process is converted into an electrical signal by a photodiode, and then enters the computer through AD conversion after filtering and amplification. In order to obtain rich dynamic information, a sampling frequency of more than 20kHz is usually used. The characteristic value of the welding process is extracted from the optical radiation signal of the welding area. The photoelectric sensor is divided into two categories according to the wavelength, where the visible light band is 300nm~900nm and the infrared light band is 1100nm~2000nm; each section includes two sensors, where the visible light band sensor is arranged up and down, the lower sensor is aimed at the laser incident point on the surface of the workpiece, and the upper sensor is aimed at 8-10mm above. The infrared light band sensor is arranged front and back along the welding direction, the front sensor is aimed at the laser incident point on the surface of the workpiece, and the rear sensor is aimed at the rear end of the molten pool. The offset distance needs to be set according to the actual size of the molten pool, usually 1 / 2 or 2 / 3 of the length of the molten pool. Among them, the offset distance is the distance from the rear infrared sensor to the laser incident point. Under different welding conditions, the length of the molten pool can vary in the range of 3-15mm.
[0051] The characteristic values of the welding process include the mean of the four photoelectric sensor signals, the ratio of the upper and lower visible light sensor signals, the ratio of the front and rear infrared light sensor signals, and the intensity ratio between the three-level wavelet transforms of the four photoelectric sensor signals, that is, the ratio of the second-level wavelet transform intensity to the first-level wavelet transform intensity, and the ratio of the third-level wavelet transform intensity to the first-level wavelet transform intensity.
[0052] Step S5, for the completed weld or weld segment, line laser scanning is used to obtain the weld surface morphology based on the laser triangulation method, and the weld width, excess height (i.e., the height difference between the weld height and the workpiece surface), undercut depth (i.e., the depth of the concave near the fusion line), expansion area (i.e., the area surrounded by the weld contour line and the straight line on the workpiece surface) and other features are calculated, and the calculation results are imported into the label set of the corresponding window.
[0053] The weld morphology characteristic values include weld width, excess height, undercut depth and expansion area. The expansion area is calculated by calculating the morphology difference before and after welding. For the missing contour data at the assembly seam, the contour inflection points on both sides of the seam and the seam center point (the depth of this point can be set according to the expected penetration depth) are selected and fitted by a quadratic polynomial.
[0054] Step S6, after the welding operation is completed, all samples (windows) are used for model training. When the accuracy of the training set is not less than 90%, the model training is considered to be completed.
[0055] During training, each window is used as a sample. The data input of the sample includes the seam assembly information and the characteristic parameters of the light radiation during the welding process. The label of the sample is the surface morphology of the weld. Each weld consists of several samples. After welding is completed, these samples are used to train the model. The model stops when the accuracy reaches 90% or the number of iterations reaches the set value. The trained model is used to calculate the weld morphology characteristics of each window, and the deviation between the calculated morphology characteristic value and the measured value is calculated. The window with a deviation exceeding the threshold is considered to be inconsistent with the main law of the weld and is marked as different levels according to the degree of deviation. Usually, the window can be marked into 5 levels. The deviation is 0 if it is less than the threshold. The larger the deviation, the higher the marking level.
[0056] The model in this embodiment selects a support vector machine or an artificial neural network model; for long welds, the number of samples for a single welding is not less than 50, and the model training can be performed after each welding is completed; for batches of short welds, the model training can be performed after a certain number of samples are accumulated. The samples used in one training should come from the same material and welding process parameters, and different materials or different welding process parameters should be trained separately.
[0057] After welding is completed, a line laser is used to scan the weld surface morphology to calculate the weld width, excess height, undercut depth and expansion area. The expansion area is calculated based on the weld contour and the pre-weld seam contour. When the assembly seam width is less than 0.2 mm, the seam edge points are directly used for straight line fitting to supplement the missing segments. When the seam width is greater than 0.2 mm, the edge points on both sides of the seam and the seam center point (the depth of this point can be set according to the expected penetration depth) are selected and fitted using a quadratic polynomial.
[0058] Furthermore, the seam feature parameters, welding light radiation feature parameters and weld morphology feature parameters are imported into the data set and label set by window, and the machine learning model is trained using samples of the entire weld. The training stops when the model accuracy exceeds 90% or reaches the set number of iterations. If the model accuracy does not reach 85% when the training stops, it means that the model capability is insufficient or the welding process is poorly stable. In this case, the model structure or hyperparameters should be adjusted. If the requirements are still not met, the weld needs to be X-rayed.
[0059] Step S7, using the trained model to predict the weld surface morphology in all windows, calculate the deviations of weld width, excess height, undercut depth and expansion area respectively, mark the windows whose deviations exceed the threshold according to the degree of deviation, and obtain the marking curve of the long-range weld.
[0060] After the model is successfully trained, all samples are calculated, and the weld morphology calculated by the model is compared with the measured value to calculate the deviation. Samples with deviations less than the set threshold are marked as 0, and samples with deviations greater than the set threshold are marked as 1, or marked as different levels according to the deviation. The weld morphology indicators include seam width, excess height depth, undercut and expansion area. The total deviation can be calculated by weighted average method, that is, total deviation = 0.2 × seam width deviation + 0.2 × excess height deviation + 0.1 × undercut deviation + 0.5 × area deviation. The total deviation does not exceed 40% and is marked as 0, the total deviation is 40% to 60% and is marked as 1, the total deviation is 60% to 80% and is marked as 2, the total deviation is 80% to 100% and is marked as 3, and the total deviation exceeds 100% and is marked as 4.
[0061] It should be noted that this embodiment does not utilize the generalization ability of the model, but utilizes the model to find abnormal samples, and the data used in the process of model training and model prediction are the same.
[0062] Step S8, according to the decision rule, determine the abnormal interval of the weld mark curve and output it as the location of the large-sized hole.
[0063] Using the trained model, the weld morphology feature values of all samples are calculated and compared with the measured values. If the deviation is within the threshold range, it is marked as 0; if the deviation exceeds the threshold, it is marked as 1, or marked as different levels according to the deviation size. After the entire weld is marked, the internal hole position is determined according to the judgment rule. The judgment rule setting usually considers: (1) continuous or high-frequency deviation windows, and (2) the appearance of high deviation level windows.
[0064] The following is an example of a specific embodiment. Specifically, the object to be detected is a pure copper laser weld, the material thickness is 3mm, the laser power is 6kW, the laser defocus is -1mm, the welding speed is 6m / min, nitrogen is used as the side blowing shielding gas, and the flow rate is 15l / min. The weld length is 30cm. The window time length is set to 5ms. The machine learning model uses an artificial neural network, the number of hidden layers is 2, the ReLU function is used as the activation function, the loss function is the mean square error, and the learning rate is set to 0.0005.
[0065] After the workpiece is assembled, it is spot welded and fixed. First, a line scanning laser is used to obtain the 3D point cloud data of the seam area. The seam width, height difference on both sides of the seam and the residual height area are extracted from the 3D point cloud data, and the data set is imported according to the window. Then, light welding is performed. During the welding process, two silicon photodiodes are used to collect visible light signals, and two indium gallium arsenide photodiodes are used to collect infrared signals. The schematic diagram of the acquisition system structure is shown in the figure. Figure 2As shown. The characteristic parameters are calculated from the four light intensity signals, including the ratio of the upper and lower visible light sensor signals, the ratio of the front and rear infrared light sensor signals, and the intensity ratio between the three-level wavelet transforms of the four photoelectric sensor signals, that is, the ratio of the second-level wavelet transform intensity to the first-level wavelet transform intensity, and the ratio of the third-level wavelet transform intensity to the first-level wavelet transform intensity. The characteristic parameters are imported into the data set by window.
[0066] After welding is completed, the surface morphology of the weld is collected by line laser scanning, and the weld width, excess height, undercut and expansion area are calculated from the weld point cloud, and the label set is imported by window. All windows are used as sample sets for artificial neural network model training. On the PyTorch platform, the Scikit-learn toolbox and PyTorch library are used for model construction and training. The model training takes less than 3 minutes and the accuracy rate reaches 93.6%.
[0067] The trained model is used to estimate the weld formation of all windows and calculate the deviation value. The total deviation calculation formula is: deviation = 0.2×seam width deviation + 0.2×reinforcement deviation + 0.1×bite deviation + 0.5×area deviation. Figure 3 In the deviation distribution diagram shown, there is a local segment where the deviation continuously exceeds the threshold value, and it is determined that a large-sized hole exists at this location.
[0068] In order to verify the effectiveness of the method provided in this embodiment, a cross section of the weld is cut in the area where a large-sized hole is determined to exist in the weld, such as Figure 4 (a); take a cross section at other locations of the weld, such as Figure 4 (b) as shown.
[0069] Compared with the existing technology, this method has the following beneficial effects: This method integrates multi-source information before, during and after welding, extracts characteristic parameters and imports them into corresponding windows, divides long-range welds into several samples, establishes mapping relationships through machine learning models, and obtains the common laws of the entire weld; then, the deviation between the model calculation value and the measured value is analyzed, and local windows that do not conform to the common laws are found. Based on the frequency of window appearance and the degree of deviation, large-sized holes inside the weld are determined. This method is driven by both knowledge and data. In terms of signal selection, feature design and model construction, it is based on expert experience and focuses on keyhole collapse behavior. In terms of model training and deviation calculation, it fully utilizes the consistency of the welding process itself and the model learning ability. The dual-drive architecture reduces the degree of dependence on expert experience, improves detection accuracy and stability, and expands the applicable scenarios of this method.
[0070] Example 2
[0071] On the basis of Example 1, this embodiment provides an online detection device for keyhole-shaped holes inside a laser weld, including a data processing terminal 5 and a photoelectric sensor array and a line laser scanning device respectively connected to the data processing terminal 5, wherein the photoelectric sensor array includes:
[0072] The first visible light sensor 1 is aimed at the laser incident point on the surface of the welding workpiece;
[0073] The second visible light sensor 2 is aligned 8-10 mm above the laser incident point;
[0074] A first infrared light sensor 3 is aimed at the laser incident point;
[0075] The second infrared light sensor 4 is aimed at the rear end of the molten pool.
[0076] The data processing terminal 5 is respectively connected to the first visible light sensor 1 , the second visible light sensor 2 , the first infrared light sensor 3 , and the second infrared light sensor 4 , so as to implement the online detection method of the first embodiment.
[0077] The process of obtaining laser welding process information belongs to non-contact detection, does not affect the normal movement of the laser work head, and does not require a complex synchronous trigger device. By adopting the above equipment, multi-source information for detection can be accurately collected.
[0078] Example 3
[0079] This embodiment provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the online detection method for keyhole-type holes inside a laser weld as described in Example 1.
[0080] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. An online detection method for keyhole-shaped holes inside a laser weld, characterized in that: The steps include: Step S1, obtaining the seam assembly information before laser welding, the light radiation characteristic information during laser welding, and the measured value of the weld morphology characteristic after welding, and obtaining multiple samples by dividing the information into multiple windows along the time axis, taking each window as a sample; Step S2, using the multiple samples, taking the seam assembly information and the light radiation characteristic information as inputs of the machine learning model, taking the measured values of the weld morphology characteristics as outputs of the machine learning model, and training the machine learning model; Step S3, using the trained machine learning model and the plurality of samples as input to calculate the predicted value of the weld morphology feature; Step S4, calculating the deviation between the predicted value of the weld morphology feature and the measured value of the weld morphology feature, and obtaining the online detection result of the keyhole-shaped hole inside the weld.
2. The method for online detection of keyhole-shaped holes in laser welds according to claim 1, characterized in that: The step S4 includes the following sub-steps: Step S401, calculating the deviation between the predicted value of the weld morphology feature and the measured value of the weld morphology feature to obtain a long-range weld marking curve, in which the window is divided into multiple levels according to the deviation; Step S402, based on a preset decision rule, an abnormal interval of the long-range weld marking curve is obtained as a large-size hole position, that is, an online detection result of a keyhole-type hole inside the weld.
3. The method for online detection of keyhole-shaped holes inside a laser weld according to claim 2, characterized in that: The decision rule is: Deviation windows occur continuously or frequently, and / or windows with high deviation levels occur.
4. The method for online detection of keyhole-shaped holes inside a laser weld according to claim 1, characterized in that: In step S1, the window width of the divided window satisfies the following conditions: The number of line laser scans and optical radiation signal data points within a single window exceeds a preset value.
5. The method for online detection of keyhole-shaped holes inside a laser weld according to claim 1, characterized in that: The weld morphology features include weld width, excess height, undercut depth and expansion area.
6. The method for online detection of keyhole-shaped holes inside a laser weld according to claim 5, characterized in that: The expansion area is calculated based on the morphology difference before and after welding. In the process of calculating the morphology difference, the contour inflection points on both sides of the seam and the seam center point are selected for the missing contour data at the assembly seam, and the filling is achieved through quadratic polynomial fitting.
7. The method for online detection of keyhole-shaped holes in laser welds according to claim 1, characterized in that: The light radiation characteristic information is obtained through two visible light band sensors and two infrared light band sensors, and the light radiation characteristic information includes the mean value of each sensor signal, the ratio of the two visible light band sensors, the ratio of the two infrared light band sensors, and the intensity ratio between the three-level wavelet transforms of each sensor signal.
8. The method for online detection of keyhole-shaped holes inside a laser weld according to claim 1, characterized in that: The seam assembly information is obtained by line laser profiling and laser triangulation, and the seam assembly information includes seam width, undercut depth and excess height area.
9. An online detection device for keyhole-shaped holes inside laser welds, characterized in that: The method for online detection of keyhole-shaped holes in a laser weld according to any one of claims 1 to 8 is used, wherein the online detection device comprises a data processing terminal (5) and a photoelectric sensor array and a line laser scanning device respectively connected to the data processing terminal (5), wherein the photoelectric sensor array comprises: A first visible light sensor (1) is aimed at a laser incident point on the surface of a welding workpiece; A second visible light sensor (2) is aligned 8-10 mm above the laser incident point; A first infrared light sensor (3) is aimed at the laser incident point; A second infrared light sensor (4) is aimed at the rear end of the molten pool; The data processing terminal (5) is connected to the first visible light sensor (1), the second visible light sensor (2), the first infrared light sensor (3) and the second infrared light sensor (4) respectively.
10. A computer-readable storage medium, characterized in that: It comprises one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs comprise instructions for executing the online detection method for keyhole-type holes inside a laser weld as claimed in any one of claims 1-8.
Citation Information
Cited By
Low-coherence imaging quality evaluation method and device and electronic equipment
CN119887754A
Low coherence imaging quality evaluation method, device and electronic equipment
CN119887754B
