A non-destructive testing method for the depth of a laser non-penetration welding seam

Abstract

The application relates to the technical field of welding detection, and discloses a laser non-penetration welding seam depth nondestructive detection method; laser visual sensing and ultrasonic phased array detection are coaxially integrated in a welded joint; high-quality collection of the three-dimensional appearance of a weld seam surface and internal acoustic information is synchronously completed in an optimal time window after welding; multi-source heterogeneous sensing information is deeply fused by using a characteristic layer fusion technology; a depth prediction model is constructed based on a machine learning algorithm; the industry problem that a single sensing technology cannot establish a reliable mapping relationship between a surface feature and internal depth is effectively solved; online, rapid, nondestructive, high-precision quantitative detection and real-time quality judgment of the laser non-penetration welding seam depth are realized; the detection efficiency, precision and reliability are remarkably improved; and an effective quality control solution is provided for intelligent manufacturing.

Description

Technical Field

[0001] This invention relates to the field of welding inspection technology, specifically to a non-destructive testing method for the penetration depth of laser-assisted non-penetration welds. Background Technology

[0002] Laser welding is widely used in precision manufacturing due to its advantages such as high energy density, low heat input, and small deformation. Among them, non-penetration welding has extremely high requirements for the control of the weld penetration. If the penetration is too shallow, the connection strength will be insufficient, and if it is too deep, it may burn through the workpiece. Therefore, accurate and reliable detection of the weld penetration is the key to ensuring product quality.

[0003] Existing technologies primarily employ destructive testing (such as dissection measurement) or single non-destructive testing methods (such as traditional ultrasonic testing and laser vision inspection) to assess weld penetration depth. Destructive testing requires destroying the sample, making it unsuitable for real-time online inspection in production environments. Traditional single non-destructive testing methods are limited by their testing principles; for example, laser vision can only acquire weld surface morphology information, and traditional ultrasonic testing lacks sensitivity to minute changes in the penetration direction. Furthermore, the two types of data are typically only post-processed and simply fused, failing to achieve deep correlation at the feature layer, resulting in low penetration depth detection accuracy and poor reliability, making it difficult to meet the real-time and accuracy requirements for quality control in laser non-penetration welding processes. The main problems with existing technologies include: the destructive nature of destructive testing limits its online application capability; single testing methods provide limited information dimensions, failing to comprehensively characterize the penetration state; and traditional data fusion methods are simple (such as result comparison), failing to explore the inherent correlation of multimodal features, leading to large penetration depth prediction errors and making it difficult to meet the requirements of high-precision non-destructive testing. Therefore, a non-destructive testing method for weld penetration depth in laser non-penetration welding is proposed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a non-destructive testing method for the penetration depth of laser-guided non-penetration welds, thereby resolving the problems in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for non-destructive testing of the penetration depth of laser-welded non-penetration welds, comprising the following steps:

[0006] S1. After the laser welding process is completed, a laser vision sensor coaxially integrated with the welding laser head is used to immediately scan the weld surface that has not yet cooled, and obtain three-dimensional point cloud data containing the three-dimensional shape and grayscale information of the weld surface.

[0007] S2. Ultrasonic phased array detection is used to acquire full matrix capture (FMC) data above the weld area by encoder positioning triggering, thereby obtaining ultrasonic full matrix data in the direction of the weld cross section.

[0008] S3. Preprocess the laser point cloud data obtained in step S1 to extract key geometric feature parameters of the weld surface. The feature parameters include at least the weld centerline, weld width, weld height and surface depression features identified based on grayscale information.

[0009] S4. Perform synthetic aperture focusing (SAFT) imaging processing on the ultrasonic full matrix data obtained in step S2 to reconstruct a high-resolution acoustic image of the weld cross-section and extract acoustic feature parameters related to the weld depth from it.

[0010] S5. The geometric feature parameters of the weld surface extracted in step S3 and the acoustic feature parameters extracted in step S4 are fused in the feature layer to form a multimodal feature vector.

[0011] S6. Input the multimodal feature vector into the pre-trained weld depth prediction model, which outputs the weld depth prediction value corresponding to the current weld measurement point;

[0012] S7. Repeat steps S1 to S6 along the weld length to achieve continuous, non-destructive testing of the entire weld penetration depth.

[0013] Preferably, in step S1, the laser vision sensor is coaxially integrated with the welding laser head, and its scanning path coincides with the center line of the weld. The scanning time is within 0.5-3 seconds after welding, taking advantage of the weld metal being at a high temperature and having good thermal radiation characteristics for measurement.

[0014] Preferably, in step S2, the ultrasonic phased array detection uses a high-frequency linear array probe with a center frequency of not less than 10MHz, and the probe moves synchronously with the welding head or vision sensor through an encoder to ensure that the ultrasonic detection position corresponds to the laser vision scanning position.

[0015] Preferably, the acoustic feature parameters in step S4 include: the echo energy distribution at the fusion interface, the sound beam propagation time difference in the weld area, and the feature gradient value in the direction of weld depth based on the grayscale distribution of the SAFT image.

[0016] Preferably, the feature layer fusion in step S5 is a feature-weighted splicing fusion method, wherein the weight coefficient of acoustic features is higher than the weight coefficient of surface geometric features.

[0017] Preferably, the melting depth prediction model mentioned in step S6 is a regression model based on machine learning algorithms, which is pre-trained through the following steps:

[0018] A. Prepare a series of standard specimens with different welding parameters and known penetration depth values, wherein the known penetration depth values ​​are obtained by destructive dissection measurement;

[0019] B. For each standard sample, perform steps S1 to S5 to obtain multiple sets of multimodal feature vectors and their corresponding true melt depth values ​​to form a training dataset.

[0020] C. Using the training dataset, train and validate the selected machine learning algorithm to obtain the final melt depth prediction model.

[0021] Preferably, the machine learning algorithm is a gradient boosting decision tree (GBDT), support vector regression (SVR), or artificial neural network (ANN).

[0022] Preferably, the method further includes step S8: comparing the predicted continuous penetration depth obtained in step S7 with the threshold range required by its design, determining the quality of the weld to be qualified, and locating and marking the out-of-tolerance parts.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] This invention coaxially integrates laser vision sensing and ultrasonic phased array detection into the welding head, simultaneously acquiring high-quality three-dimensional morphology of the weld surface and internal acoustic information within the optimal post-weld time window. It utilizes feature layer fusion technology to deeply fuse multi-source heterogeneous sensor information and constructs a weld penetration prediction model based on machine learning algorithms. This effectively solves the industry challenge of establishing a reliable mapping relationship between surface features and internal weld penetration using a single sensing technology. It achieves online, rapid, non-destructive, and high-precision quantitative detection and real-time quality assessment of laser-insulated weld penetration depth, significantly improving detection efficiency, accuracy, and reliability, and providing an effective quality control solution for intelligent manufacturing.

[0025] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the coaxial integrated structure of the system of the present invention;

[0027] Figure 2 This is a flowchart of the detection method of the present invention;

[0028] Figure 3 This is a flowchart of the surface feature extraction process of the present invention;

[0029] Figure 4 This is a flowchart of the acoustic feature extraction process of the present invention;

[0030] Figure 5 This is a schematic diagram of feature fusion and model prediction in this invention;

[0031] Figure 6 This is a schematic diagram illustrating the continuous detection and quality assessment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figures 1-6 This invention discloses a non-destructive testing method for the weld penetration depth of laser-assisted non-penetrating welds. This method coaxially integrates laser vision sensing and ultrasonic phased array detection technology into the welding head, and synchronously collects information on the surface morphology and internal structure of the weld within a specific time window after welding. The acquired multi-source heterogeneous sensing data is preprocessed, feature extracted and fused, and finally, a pre-trained machine learning model is used to achieve online, rapid, non-destructive and high-precision quantitative prediction and quality judgment of the weld penetration depth.

[0034] I. Laser Vision Scanning and Data Acquisition

[0035] Equipment configuration and parameters:

[0036] The laser vision sensor can be the Keyence LJ-V7060 series, coaxially integrated with the welding laser head via an adapter bracket to ensure the scanning path coincides with the weld centerline (deviation ≤ 0.05mm). Sensor parameter settings are as follows:

[0037] Laser wavelength: 650nm (red laser, balancing safety and detection accuracy);

[0038] Scanning frequency: 200Hz (acquiring 200 lines of scan data per second);

[0039] Resolution: 0.02mm (X / Y axis direction);

[0040] Gray level quantization: 8 bits (0-255 gray levels).

[0041] Example of scanning timing control:

[0042] In the laser welding test of 304 stainless steel, the scan was started 1 second after welding. At this time, the surface temperature of the weld was about 500℃, and the metal was in the austenitic state with high thermal radiation stability. The temperature field was monitored by an infrared thermal imager, and the scan start time was dynamically adjusted (e.g., delayed to 1.5 seconds for thin plate welding to avoid excessive surface oxidation).

[0043] Data acquisition process extension:

[0044] The sensor moves along the weld length at a speed of 50 mm / s, with a single scan covering a width of 10 mm and an overlap rate of 20%. For example, when welding a 100 mm long weld, 10 scans are required, and the data from each scan is transmitted to the industrial control computer in real time via Ethernet. The original point cloud data format is:

[0045] ;

[0046] in , For planar coordinates, For height value, The value is a grayscale value (0-255).

[0047] The scanning path provides a positional synchronization reference for ultrasonic testing. The encoder triggers ultrasonic data acquisition through the moving guide rail of the laser scan, ensuring that the two are spatially aligned (deviation ≤0.1mm).

[0048] Function: High-temperature scanning can capture the morphological features of the weld in the early stage of solidification. The grayscale information reflects the degree of surface oxidation, providing a basis for subsequent depression identification.

[0049] II. Ultrasonic Phased Array Detection and Full Matrix Data Acquisition

[0050] Probe selection example:

[0051] An Olympus 10MHz high-frequency linear array probe with 64 elements, an element spacing of 0.3mm, an excitation pulse voltage of 40V, and a sampling rate of 100MHz was selected. The probe contacts the weld surface through a water bladder (with deionized water as the coupling agent), and the water bladder is 5mm thick to ensure that the sound beam is incident perpendicularly.

[0052] Synchronization control:

[0053] The encoder and laser vision sensor share a moving guide rail, with a resolution of 0.01 mm and a trigger interval of 0.5 mm. For example, when the laser sensor scans to the 5 mm position of the weld, the encoder triggers the probe to acquire the ultrasonic full matrix data (FMC) at that position. The data format is a 64×64 matrix (transmitter element × receiver element).

[0054] FMC Data Acquisition Extension:

[0055] Each array element acts as a transmitter in sequence, and all array elements simultaneously receive echo signals. The single-point acquisition time is 0.1s, and the data format is a 64×64 matrix (transmitter array element × receiver array element). For example, when welding a 10mm wide weld, FMC data needs to be acquired at 20 locations, with a total data volume of approximately 20×64×64=81,920 echo signals.

[0056] FMC data provides raw acoustic field information for SAFT imaging, while the ultrasonic detection position corresponds one-to-one with the laser scanning position through the encoder, ensuring that the data space of the surface feature morphology and the internal structure is aligned.

[0057] Function: High-frequency probes improve the resolution of minute defects in the penetration direction (such as 0.1mm-level cracks), and full-matrix data retains complete information on sound field propagation, providing a basis for SAFT imaging.

[0058] III. Laser Point Cloud Preprocessing and Surface Feature Extraction

[0059] Detailed explanation of the preprocessing workflow:

[0060] Noise filtering: A bilateral filtering algorithm is used, and the spatial domain standard deviation is set. (Control neighborhood range), grayscale standard deviation (Controlling grayscale similarity) This algorithm removes random noise while preserving edge details. For example, burrs on weld edges can be effectively smoothed using this algorithm.

[0061] Coordinate correction: Based on the weld centerline fitting results (least squares method), the point cloud data is transformed into a local coordinate system with the weld centerline as the x-axis. The correction formula is as follows:

[0062] ;

[0063] in The x-coordinate of the weld centerline in the global coordinate system (obtained by fitting point cloud data).

[0064] Feature extraction example:

[0065] 1. Weld width calculation: In the local coordinate system, take... Within ±5mm of the direction The value, the weld width W, is:

[0066] ;

[0067] in The height of the base material surface (determined by the lowest point of the point cloud). This is the high threshold value. For example, the threshold value of a certain weld seam. =0mm, Then the melt width is Within range The area width is >0.2mm, and the actual weld width W=8mm.

[0068] 2. Calculation of weld reinforcement height: The weld reinforcement height H is the height difference between the highest point of the weld surface and the base metal surface.

[0069] ;

[0070] For example, the highest point of a weld Therefore, the remaining height H = 1.5 mm.

[0071] 3. Indentation Recognition: Based on grayscale value threshold segmentation, a grayscale threshold is set. The concave area satisfies and area For example, if all gray values ​​in a certain area are below 100 and the number of consecutive pixels exceeds 50 (corresponding to an area of ​​approximately 1.2 mm²), it is determined to be a depression.

[0072] Surface geometric features (such as weld width W=8mm and excess height H=1.5mm) provide structural references for ultrasound imaging. For example, weld width helps to locate the lateral range of ultrasound imaging, and the location of the depression guides the local analysis of acoustic features (such as echo energy).

[0073] Function: Preprocessing improves data quality, quantifies weld morphology by surface geometric features, and provides intuitive physical parameters for weld penetration prediction.

[0074] IV. Ultrasonic Full-Matrix Data Processing and Acoustic Feature Extraction

[0075] SAFT imaging processing:

[0076] 1. Delay superposition: For each transmit-receive array element pair (i,j) in the full matrix data, calculate the propagation time of the sound beam from transmission to reception. Acoustic images are reconstructed by time-delay overlay:

[0077] ;

[0078] in Let (i,j) be the echo signal of the array element pair. (x, z) represents the theoretical propagation time at the focal point (x, z) (calculated using the speed of sound c = 5900 m / s), 64 represents the number of array elements (array units) in the ultrasonic probe, t is the time variable representing the time axis of the echo signal, and i and j represent the indices of the transmitting and receiving array elements, respectively; for example, at a certain focal point (x = 0 mm, z = 1.2 mm), calculate the propagation time of all array element pairs. The corresponding echo signal is superimposed to obtain the sound intensity value I=250 (gray value 0-255) at that point.

[0079] 2. Image Enhancement: Non-local mean filtering is used to remove imaging noise. The filter window is set to 5×5, and the similarity weight h=0.1. For example, speckle noise in SAFT images can be effectively suppressed by this algorithm.

[0080] Example of acoustic feature extraction:

[0081] 1) Echo energy at the fusion interface: Extract the echo energy E at the fusion interface along the fusion depth direction (z-axis) in the SAFT image:

[0082] ;

[0083] Where d is the depth coordinate, representing the direction perpendicular to the weld surface in the weld cross-section (i.e., the penetration direction). The acoustic image grayscale value (sound intensity) at position (x, d) is obtained by SAFT imaging processing.

[0084] For example, if E=500 (unit of energy) at the fusion interface of a weld, it indicates that the sound beam is strongly reflected at this depth and the fusion is good.

[0085] 2) Sound beam propagation time difference: Calculate the sound beam propagation time difference between the weld area and the base material area. :

[0086] ;

[0087] in The sound beam propagation time in the weld area. For the base material region time. For example, the time for a certain weld. This indicates that the sound beam propagates more slowly in the weld (because the sound velocity decreases after the molten pool solidifies).

[0088] 3) Feature gradient value: Calculate the gray-level gradient of the SAFT image along the melting depth direction. :

[0089] ;

[0090] Take the maximum gradient value ,max is used as a feature. For example, the maximum value of a weld. The value of max=50 (grayscale / mm) indicates a significant change in acoustic impedance in the melting depth direction.

[0091] Acoustic characteristics (e.g., E=500, Multimodal feature vectors are formed by feature fusion of surface features (e.g., W=0.5μs) and surface features (e.g., W=8mm, H=1.5mm), providing comprehensive information for model prediction.

[0092] Function: SAFT imaging improves the resolution of acoustic images (up to 0.1mm level), and acoustic features reflect the melting depth from the perspective of internal structure, complementing surface features.

[0093] V. Multimodal Feature Fusion and Melting Depth Prediction

[0094] Feature fusion:

[0095] Feature vector construction: Surface geometric feature vectors Acoustic feature vector For example, a certain weld seam , .

[0096] Weighted splicing: setting acoustic feature weights Surface feature weights fused feature vector For example, the features after fusion are: .

[0097] Model training example:

[0098] Standard sample preparation: Twenty sets of samples were prepared using laser power (1500-2500W), welding speed (20-40mm / s), and defocusing amount (-1-1mm) as variables. For example, sample 1 parameters: power 2000W, speed 30mm / s, defocusing amount 0mm, actual penetration depth... (Measured by metallographic dissection).

[0099] Dataset construction: For each group of samples, the entire process of “laser visual scanning and data acquisition, ultrasonic phased array detection and full matrix data acquisition, laser point cloud preprocessing and surface feature extraction, ultrasonic full matrix data processing and acoustic feature extraction and multimodal feature fusion” was executed sequentially to obtain 1000 sets of feature vector-melt depth pairs (50 measurement points for each group of samples).

[0100] Model selection and training: Gradient boosting decision tree (GBDT) was used, with a tree depth of 5, a learning rate of 0.1, and 100 iterations. The loss function was mean squared error (MSE).

[0101] ;

[0102] Hyperparameters were optimized using 5-fold cross-validation, resulting in a final model MSE of 0.05 mm² on the test set. For example, it predicts the melt depth at a certain test point. =1.18mm, actual value =1.2mm, with an error of only 0.02mm.

[0103] Model-predicted melt depth It provides a quantitative basis for quality judgment, and the continuously detected penetration sequence can provide feedback to adjust welding parameters (such as power and speed), forming a closed-loop control.

[0104] Function: Feature fusion is used to mine the correlation of multi-source information, and the GBDT model realizes nonlinear mapping to improve the accuracy of melt depth prediction.

[0105] VI. Continuous Testing and Quality Assessment

[0106] Testing process extension:

[0107] Repeat the steps "laser vision scanning and data acquisition" sequentially at 2mm intervals along the weld length.

[0108] "Ultrasonic phased array detection and full matrix data acquisition, laser point cloud preprocessing and surface feature extraction, ultrasonic full matrix data processing and acoustic feature extraction, and multimodal feature fusion and melt depth prediction" to generate melt depth prediction sequences. For example, when welding a 100mm long weld, 50 tests are required to generate 50 penetration depth values, forming a penetration depth curve.

[0109] Quality assessment example:

[0110] Set the melting depth threshold range (Corresponding to the requirements for non-penetration welds in 304 stainless steel). Compare the predicted value with the threshold:

[0111] like It is marked as "insufficient penetration";

[0112] like It is marked as "over-penetration".

[0113] For example, predicting the melting depth at a certain location. If the penetration depth is 0.7mm, it is marked as out of tolerance, and the coordinate x=25mm is recorded. At the same time, the system alarm prompts "Insufficient penetration depth, welding parameters need to be adjusted".

[0114] Function: To enable real-time quality monitoring of the entire weld length, and to accurately locate out-of-tolerance areas to provide a basis for subsequent repair.

[0115] In summary, this embodiment integrates laser visual scanning and ultrasonic phased array detection coaxially, simultaneously acquiring high-temperature weld surface morphology and internal full-matrix acoustic data within the optimal post-weld time window. After extracting geometric features through point cloud preprocessing and acoustic features through SAFT imaging, a weighted fusion method is used to construct a multimodal feature vector, which is then input into a prediction model trained based on machine learning. Ultimately, continuous, online, quantitative, and non-destructive high-precision detection and real-time quality assessment of weld penetration depth are achieved.

Claims

1. A non-destructive testing method for the penetration depth of laser-welded non-penetration welds, characterized in that, Includes the following steps: S1. After the laser welding process is completed, a laser vision sensor coaxially integrated with the welding laser head is used to immediately scan the weld surface that has not yet cooled, and obtain three-dimensional point cloud data containing the three-dimensional shape and grayscale information of the weld surface. S2. Ultrasonic phased array detection is adopted. The full matrix capture FMC data is acquired above the weld area by encoder positioning triggering to obtain ultrasonic full matrix data in the direction of weld cross section. S3. Preprocess the laser point cloud data obtained in step S1 to extract key geometric feature parameters of the weld surface. The feature parameters include at least the weld centerline, weld width, weld height and surface depression features identified based on grayscale information. S4. Perform synthetic aperture focusing SAFT imaging processing on the ultrasonic full matrix data obtained in step S2 to reconstruct a high-resolution acoustic image of the weld cross-section and extract acoustic feature parameters related to the weld depth from it. acoustics The characteristic parameters include: the energy distribution of the fusion interface echo, the time difference of the acoustic beam propagation in the weld area, and the characteristic gradient value in the direction of the weld depth based on the gray-scale distribution of the SAFT image. S5. The geometric feature parameters of the weld surface extracted in step S3 and the acoustic feature parameters extracted in step S4 are fused in the feature layer to form a multimodal feature vector. S6. Input the multimodal feature vector into the pre-trained weld depth prediction model, which outputs the weld depth prediction value corresponding to the current weld measurement point; the feature layer fusion is a fusion method based on feature weighted splicing, wherein the weight coefficient of acoustic features is higher than the weight coefficient of surface geometric features; S7. Repeat steps S1 to S6 along the weld length to achieve continuous, non-destructive testing of the entire weld penetration depth.

2. The method for non-destructive testing of laser-assisted non-penetration weld penetration depth according to claim 1, characterized in that, In step S1, the laser vision sensor is coaxially integrated with the welding laser head, and its scanning path coincides with the center line of the weld. The scanning time is within 0.5-3 seconds after welding, taking advantage of the weld metal being at a high temperature and having good thermal radiation characteristics for measurement.

3. The method for non-destructive testing of laser non-penetration weld penetration depth according to claim 1, characterized in that, In step S2, the ultrasonic phased array detection uses a high-frequency linear array probe with a center frequency of not less than 10MHz. The probe moves synchronously with the welding head or vision sensor through an encoder to ensure that the ultrasonic detection position corresponds to the laser vision scanning position.

4. The method for non-destructive testing of laser non-penetration weld penetration depth according to claim 1, characterized in that, The melting depth prediction model mentioned in step S6 is a regression model based on machine learning algorithms, which is pre-trained through the following steps: A. Prepare a series of standard specimens with different welding parameters and known penetration depth values, wherein the known penetration depth values ​​are obtained by destructive dissection measurement; B. For each standard sample, perform steps S1 to S5 to obtain multiple sets of multimodal feature vectors and their corresponding true melt depth values ​​to form a training dataset. C. Using the training dataset, train and validate the selected machine learning algorithm to obtain the final melt depth prediction model.

5. The method for non-destructive testing of laser non-penetration weld penetration depth according to claim 4, characterized in that, The machine learning algorithm is Gradient Boosting Decision Tree (GBDT), Support Vector Regression (SVR), or Artificial Neural Network (ANN).

6. The method for non-destructive testing of laser non-penetration weld penetration depth according to claim 1, characterized in that, It also includes step S8: comparing the continuous penetration depth prediction value obtained in step S7 with the threshold range required by its design, determining the quality of the weld to be qualified, and locating and marking the out-of-tolerance parts.

Images