{"title":"An unsupervised online anomaly detection method for metal additive manufacturing processes via a statistical time-frequency domain algorithm","authors":"Alvin Chen, Fotis Kopsaftopoulos, Sandipan Mishra","doi":"10.1177/14759217231193702","DOIUrl":null,"url":null,"abstract":"Anomalies often occur in metal additive manufacturing from processing inconsistencies and uncertainty. A robust fault detection system that uses sensor measurements such as melt pool imaging has the potential to improve part quality and save production time by anticipating print failure. Toward this goal, we develop and validate a fault detection technique using melt pool geometry-related measurements from an in situ near-infrared optical camera. This method is unsupervised and is trained on a small dataset, mitigating human error in classifying fault types, and reducing lead times for preparing training datasets. Furthermore, this method uses learned geometry-informed nominal behavior of the melt pool signal to make informed decisions on the process health. There are spatial-temporal characteristics embedded in the melt pool images, caused by the periodicity in the geometry-dependent raster pattern. These characteristics can be captured in the frequency domain using the signal spectrogram, a representation of the frequency content over time. Defects will appear in the spectrogram, disrupting the healthy spectral response. To quantify healthy spectrograms, we use principal component (PC) decomposition to extract the features of these spectrograms as a set of nominal basis vectors. Anomaly detection is then performed by calculating the error between the original and reconstructed spectrogram vector by projection of the spectrogram PCs onto the nominal basis. The reconstruction error for anomalous signals is larger than that from healthy signals, which is then used for fault detection. A one-tailed statistical test is used to determine the fault detection threshold for the reconstruction error signal. This method is tested on three raster patterns and performs better than a comparative time-series thresholding method. We demonstrate that this time-frequency algorithm can detect both temporal faults (which occur at a single time instant) and spatial faults (such as those introduced by an improper sintering), differentiating them from nominal operation.","PeriodicalId":51184,"journal":{"name":"Structural Health Monitoring-An International Journal","volume":null,"pages":null},"PeriodicalIF":5.7000,"publicationDate":"2023-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Structural Health Monitoring-An International Journal","FirstCategoryId":"1085","ListUrlMain":"https://doi.org/10.1177/14759217231193702","RegionNum":2,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"ENGINEERING, MULTIDISCIPLINARY","Score":null,"Total":0}
引用次数: 0
Abstract
Anomalies often occur in metal additive manufacturing from processing inconsistencies and uncertainty. A robust fault detection system that uses sensor measurements such as melt pool imaging has the potential to improve part quality and save production time by anticipating print failure. Toward this goal, we develop and validate a fault detection technique using melt pool geometry-related measurements from an in situ near-infrared optical camera. This method is unsupervised and is trained on a small dataset, mitigating human error in classifying fault types, and reducing lead times for preparing training datasets. Furthermore, this method uses learned geometry-informed nominal behavior of the melt pool signal to make informed decisions on the process health. There are spatial-temporal characteristics embedded in the melt pool images, caused by the periodicity in the geometry-dependent raster pattern. These characteristics can be captured in the frequency domain using the signal spectrogram, a representation of the frequency content over time. Defects will appear in the spectrogram, disrupting the healthy spectral response. To quantify healthy spectrograms, we use principal component (PC) decomposition to extract the features of these spectrograms as a set of nominal basis vectors. Anomaly detection is then performed by calculating the error between the original and reconstructed spectrogram vector by projection of the spectrogram PCs onto the nominal basis. The reconstruction error for anomalous signals is larger than that from healthy signals, which is then used for fault detection. A one-tailed statistical test is used to determine the fault detection threshold for the reconstruction error signal. This method is tested on three raster patterns and performs better than a comparative time-series thresholding method. We demonstrate that this time-frequency algorithm can detect both temporal faults (which occur at a single time instant) and spatial faults (such as those introduced by an improper sintering), differentiating them from nominal operation.
期刊介绍:
Structural Health Monitoring is an international peer reviewed journal that publishes the highest quality original research that contain theoretical, analytical, and experimental investigations that advance the body of knowledge and its application in the discipline of structural health monitoring.