Aiming at the subjectivity and blindness of different device feature selection in current non?intrusive load identification, a non?intrusive load identification algorithm based on Fisher?SVM feature selection is proposed. Firstly, the original data of household?side current and voltage are extracted based on the high frequency sampling device. Fourier transform is used to decompose the original signal into active power, reactive power and harmonic time series. Secondly, the load waveform is divided into four stages and the transient characteristics of the load waveform are calculated. Then, by utilizing the Fisher?SVM algorithm for feature selection among different classifiers, the optimal subset of classification features is obtained. Additionally, the results are calibrated using the Sigmoid function for probability calibration. Finally, different classifiers are integrated based on Bayesian theory to achieve identification of different loads. The algorithm is tested on a dataset consisting of 831 actual users from three different distribution areas. The results show that the algorithm effectively exploits the uniqueness of different electrical load imprints, overcomes the blindness in feature selection, and increases the load identification ability.