Conventional fire detecting systems lack the ability to detect fire in the early stages since they trigger the alarm by the high density of smoke or the high air temperature. In this paper, a new electronic nose (EN) system is proposed as an alternative way to detect various sources of fire from the burning smells. The EN is added with a mechanism to reduce the effect of the temperature and the humidity. Consequently, the time series data from the same smell in every repetition data are highly correlated. We have selected only a single data from each source of smell that has the highest average similarity index (SI) value to be a training data for the error back-propagation neural networks (BPNN). Generally, the time series data can be used as the input data for the BPNN directly. However, it will consume a lot of time for training due to the huge dimensional data. A new method called slope max mean (SMM) is applied to reduce the dimension of the input data. By using the SMM data, the training time is reduced and the overall classification rate of 99.8% is achieved which shows the high feasibility to apply the EN as a precision fire detecting system.