Achieving early warning of electrical fires is of great importance in preventing their occurrence. The overheating diagnosis of electrical equipment by detecting volatiles is an effective way to achieve early warning of electrical fires. This paper proposes a machine olfaction-based overheating diagnosis method for electrical equipment. First, the primary materials commonly used for non-metallic elements in electrical equipment are determined. Second, a semiconductor oxide sensor array is designed based on the volatiles generated by these materials during overheating. Next, the output curves of the sensor array at different heating temperatures for these materials are obtained through experiments. Subsequently, the output data of the sensor array at multiple moments in the period of each output curve entering the plateau are extracted to construct the output vectors of the sensor array. Then, the Principal Component Analysis (PCA) algorithm and Linear Discriminant Analysis (LDA) algorithm are used to extract the characteristic quantities in the output vectors to construct the feature vectors. Finally, a second-order BP neural network model is designed based on the feature vector to determine whether the electrical equipment is overheated and what the overheated material is. The experimental results show that the accuracy of the proposed overheating diagnosis method can reach 94.58%.