Extraction of Fault Patterns on SLS Part Surfaces Using the Karhunen-Loève Transform

To gain a thorough understanding of the fault mechanisms in SLS machines, we decompose SLS profile signals into independent features using a novel tool called Karhunen-Loeve (KL) transform. These individual features can then be studied separately to monitor the occurrence of fault patterns on manufactured parts and determine their nature. Analytical signals with known fault patterns, simulating profile measurement signals from SLS parts, are used to determine the suitability of the proposed method. Multi-component patterns are assumed to manifest on SLS part surfaces, resulting from faults in the machine, for example, the roller mechanism. The results of this work determine the suitability of the KL transform for condition monitoring and extraction of fault-indicating patterns. Fault Patterns on Manufactured Parts Detecting and quantifying faults that occur during manufacturing is necessary to ensure the efficient production of accurate parts. The field of fault detection and diagnosis in manufacturing aims at eliminating the occurrence of faults by continuously monitoring the process, detecting faults, and taking corrective action. In this paper, the focus is on monitoring the condition of surface quality. The surface is measured at regular intervals for the purpose of detecting any degradation on part surface quality. Faults or deviations in the dynamics of the manufacturing machine or its submechanisms are assumed to leave a "fingerprint" on the surface of the part being manufactured, which manifest as fault patterns. These are the fault patterns that we seek to detect, quantify, and diagnose, in order to take remedial action if necessary. A fault is defined as the inability of a system to perform in an acceptable manner [9]. Faults typically manifest themselves as deviations in observed behavior from a set of acceptable behaviors. Fault detection is the recognition of an unacceptable behavior; and fault diagnosis is the identification of a component or set of components in the system that causes the fault [9]. As part of fault detection, analysts collect data, extract relevant features, and compare these extracted features to a specification of correct or incorrect data [5, 9]. The method of feature extraction and selection is a critical factor in detecting the correct faultindicating features from manufacturing signals. Complex signals are best analyzed and processed using signal processing tools. In this research, we aim to develop a unified method to detect and diagnose the correct features in an accurate fashion. The most reliable methods of fault diagnosis utilize signal processing algorithms to extract fault features. Poorly performing fault monitoring and diagnosis systems are common in industry, and they result in frequent false alarms or insensitivity to a legitimate failure condition [10]. It is often difficult to detect the proper features in the presence of random effects and nonstationarities. The