The objective of this work is to detect in situ the occurrence of lack-of-fusion defects in titanium alloy (Ti-6Al-4 V) parts made using directed energy deposition (DED) additive manufacturing (AM). We use data from two types of in-process sensors, namely, a spectrometer and an optical camera which are integrated into an Optomec MR-7 DED machine. Both sensors are focused on capturing the dynamic phenomena around the melt pool region. To detect lack-of-fusion defects, we fuse (combine) the data from the in-process sensors invoking the concept of Kronecker product of graphs. Subsequently, we use the features derived from the graph Kronecker product as inputs to a machine learning algorithm to predict the severity (class or level) of average length of lack-of-fusion defects within a layer, which is obtained from offline X-ray computed tomography of the test parts. We demonstrate that the severity of lack-of-fusion defects is classified with statistical fidelity (F-score) close to 85% for a two-level classification scenario, and approximately 70% for a three-level classification scenario. Accordingly, this work demonstrates the use of heterogeneous in-process sensing and online data analytics for in situ detection of defects in DED metal AM process.