Abstract: To improve the recognition accuracy of vibration monitoring and diagnosis algorithms in transformer vibration monitoring points with displacement and low signal-noise ratio environment, a transformer condition recognition method based on multi-point vibration detection and improved MobileNet is proposed. Analyzed the vibration characteristics of different measuring points on the side wall of the transformer box and studied the classification performance of conventional artificial intelligence algorithms in the environment with measuring point offset and low signal-noise ratio. Proposed an improved MobileNet algorithm that adapts to multi-point detection: preprocessing vibration signals from multiple measuring points to generate a three-dimensional data structure. The first level algorithm uses a pointwise-depthwise-pointwise convolution architecture combined with a dilated convolution kernel to extract multi-scale features from the data stream. The second-level algorithm uses a fully connected network to identify samples. The research results show that the improved MobileNet under the multi-point detection scheme performs well compared with conventional algorithms. When the measuring point offset is less than 0.3m, the recognition accuracy decreases by less than 2%. When the signal-noise ratio is 10dB, the recognition accuracy remains at 97%. In addition, this algorithm has a smaller model footprint and shorter recognition and prediction time. Deploying it on mobile devices such as Raspberry Pi can significantly improve computing speed and reduce running power consumption.