2.3. Data Processing

Weld depth measurements are performed by OCT data and the metallographic evaluation of probes. Scanned OCT measurements follow the same data processing procedure: OCT sensor signal transformation to depth information, noise reduction, segmentation in y-direction, segmentation in z-direction, and feature extraction for the measurement of weld penetration depth (Figure 2). The OCT sensor signal transformation to depth information is carried out by the k-linearization of the sensing data and Fourier transformation of the measurement information.

Data processing steps for the identification of weld depth with OCT measurements. Colored data processing blocks are varied.

Noise reduction is performed by background subtraction with averaged background information. Remaining noise artifacts from lenses in the optical setup are identified by Hough transform and set to zero.

Afterwards, a segmentation in the y-direction is required to distinguish between the workpiece surface and the keyhole opening. First, the workpiece surface is detected by identification of the image row with maximum accumulated signal intensity as the signal intensity on the workpiece surface is higher than the signal intensity in the keyhole. The workpiece surface determines the reference position for weld depth measurement in the keyhole. Second, the keyhole is determined by identification of columns in the y-direction with reduced signal intensity by analysis of the falling and rising edge intensity at workpiece surface level.

Only the keyhole opening is considered for further processing steps. The segmentation of the keyhole opening in the z-direction aims for the isolation of measurement information from the keyhole bottom by separating the workpiece surface, the keyhole region, and the region below the keyhole. Two different methods are applied for this segmentation process: the moving average method [19] and the maximum intensity projection (MIP) method [20] (Figure 3).

Segmentation in the z-direction. Schematic of the moving average (MA) and maximum intensity projection (MIP) approach.

The moving average method calculates the mean signal intensity in a window with a depth of 40 pixels in the z-direction and a width that corresponds to the width of the keyhole opening in the y-direction. Starting at the workpiece surface, the window is moved in the z-direction after each calculation up to a distance of 5 mm. This limit is set according to the thickness of the workpiece. Reflections from the keyhole bottom lead to an increase in signal intensity. In consequence, we expect windows not to contain measurement information from the keyhole bottom if they fall below the 75th percentile of all calculated mean values. A window is chosen as the lower limit of the keyhole if the percentile condition is fulfilled, and no neighboring window above the 75th percentile can be found in the negative z-direction. All windows below this window are discarded.

In contrast to the moving average method that is applied to each 2D OCT image frame, the MIP method considers the measurement information from all 2D frames along the measurement time. Here, each coordinate of a 2D OCT image frame in z- and y-direction are compared with the identical coordinate in all other taken 2D OCT image frames regarding the intensity value. Maximum intensity values of a coordinate are considered for the resulting projected image frame. The highest pixel intensity region within the projected image frame defines the keyhole region, while measurement information below this limit is discarded. The maximum width of the keyhole opening defines the outer limit in the y-direction.

The measurement of weld penetration depth is tested with three different feature extraction algorithms: intensity accumulation, max-value, and kernel density estimation. These approaches for feature extraction differently consider the dimensionality of the measurement information (Figure 4).

Feature extraction for weld depth identification. Schematic of the applied three different feature extraction methods (KDE, max value, intensity accumulation (IA)).

The application of KDE is an adaption from the literature [1]. Here, each measurement point in the y-direction within the keyhole region is considered as a singular keyhole mapping measurement point and calculates the KDE and the modality index with the help of Hartigan’s dip test. According to the literature, only single-modal measurement points are considered for further weld penetration depth extraction. Finally, measurement points that fulfill the single-modal condition are analyzed by an 80th percentile filter along the z-direction resulting in the pixel value at the expected weld penetration depth.

The accumulation of intensity differs from the application of KDE as the lateral measurement information is seen as a whole in the y-direction. The accumulation of intensity approach follows two steps: thresholding and the accumulation of intensity. Thresholding is performed by the identification of each maximum intensity value within each measurement point in the y-direction and subsequent application of an 80th percentile. This percentile filter is chosen as measurement points that are directed towards keyhole walls show lower intensity values than a measurement point directed in the keyhole bottom. Afterwards, the weld penetration depth is extracted by an accumulation of intensity. Here, the intensity values for all lateral measurement points in the y-direction are added at each z-coordinate in the keyhole. This process is repeated for the subsequent 2D-OCT image frames. The z-position with the maximum accumulated intensity is considered the weld penetration depth. Finally, the max-value approach is the simplest feature extraction method. Here, the weld penetration depth is determined by extraction of the maximum signal intensity value within the keyhole. In the end, the extracted pixel coordinate in the z-direction is transformed to the weld penetration depth by considering the axial pixel size of 11.7 µm.

In this study, a comparison of the different segmentation and feature extraction algorithms is performed. The analysis of the different data processing methods is possible by comparing measured weld penetration depths with weld penetrations depths that are determined via metallographic cross-sections. The comparison between measured weld penetration depth by OCT and metallographic analysis are performed by root mean square error (RMSE).

The RMSE can be calculated as follows:

where n is equal to the total number of in-process frames, y_i is the algorithm’s calculated weld seam depth, and y^ the mean weld seam depth of the metallographic analysis. The higher the RMSE value, the higher the deviations from the metallographic weld seam depth of an OCT measurement. The maximum deviation allowed is the standard deviation of the metallographic weld seam depth of the corresponding welding parameters.