Abstract

We have recently developed a new method for surface characterisation combining light scattering and machine learning (ML). The method has been successfully applied to the binary classification (acceptable vs. defective) of micro-structured [1-4] and additive manufacturing (AM) surfaces [5-8]. The general idea is to shine laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. We have developed the following machine learning models: convolutional autoencoder (CNN) [5], autoencoder based on a multilayer, fully connected network (ANN) [6], classifier based on principal component analysis (PCA) [7], and classifier based on a oneclass support vector machine (SVM) [8]. All these methods share a common approach where the ML system is internally used to learn an optimal strategy to encode/decode only the scattering patterns of acceptable surfaces. Defected surfaces are then detected during operation, when poor encoding/decoding performance is observed on a new pattern. The single-class approach is advantageous for AM because the ML model can be trained without the need of examples from defective surfaces. In this paper, the performance of the ANN, CNN, PCA and SVM-based classifiers is compared by processing multiple scattering patterns experimentally acquired from two laser powder bed fusion surfaces (one acceptable, one defected). The performance indicators are the accuracy of the binary classifier; the time to classify each newly acquired pattern; and the internal separation of acceptable vs defective observations, measured by the difference in the RMS reconstruction errors of the encoding/decoding process associated with the acceptable (reference) and unacceptable surfaces. For the SVM, internal separation is measured as the difference of likelihood for an observation to belong to either one of the two classes. The results of the comparison are shown in Table 1 and Figure 1. The classification accuracy was high (>99%) for all methods. The PCA-based classifier was faster (1 ms) whilst the ANN method was the slowest (100 ms) using the same hardware (CPU Intel Xeon E5-2650 2.2 GHz, RAM 256G). The analysis of the internal separation between the classes, representative of discriminative power (Figure 1), appears to indicate better discrimination achieved by the SVM, though the results will need refining on larger experimental data.