Industrial buildings are responsible for a large share of the worldwide electricity consumption. Disaggregated information about electricity consumption enables decision- making and feedback tools to reduce and optimize the electricity consumption. In industrial settings, electrical load comes from a variety of equipment and machinery which can be awkward and expensive to monitor individually. We believe that Non-Intrusive Load Monitoring (NILM) can ease the burden of such a monitoring infrastructure. This hypothesis has been evaluated by collecting a rich data set from more than forty sensors measuring power consumption for six months at an industrial cold store. Their electrical equipments includes compressors, industrial fans, evaporators etc. which by earlier work have been hypothesised as too difficult to detect by NILM algorithms. This paper provides a detailed study of how industrial equipment and machinery challenge NILM algorithms. We consider how NILM can be used with different levels of sub- metering for providing breakdowns of the power consumption in an industrial setting. Our results show that changing the level of sub- metering increased the test accuracy(F1 score) with a third from 0.4 to 0.6. We introduce FHMM with day specific training, meaning having a model for each day in the week. The FHMM with day specific training reduced by half the mean normalized error from 0.7 to 0.3. These results thereby open up for the use of NILM in an industrial setting.