Non-intrusive load monitoring (NILM) is a key component of intelligent electricity consumption behavior recognition. Due to the wide variety of simultaneous load accesses in medium-voltage distribution networks, and the fact that many of them have frequency conversion functions and do not have constant power characteristics, it is difficult to apply the existing load recognition methods focusing on households directly to complex equipment environments. In this paper, a load identification experiment is conducted with an elevator as a typical load in the complex equipment environments, IEC 61000-4-30 measurement data is used as input to identify whether the elevator is in operation. In order to eliminate the computational pressure caused by irrelevant features, a differential feature extraction method based on correlation coefficient of Pearson is proposed, which is combined with a convolutional neural network to realize the elevator load state identification in a real environment with multiple unknown loads. The results using the measured data show that the proposed method requires only a small number of samples to identify the elevator operating state with complex changes in operating power consumption, and the computational accuracy is higher than that of traditional machine learning methods.