This article proposes a method based on the deep feature fusion of multi-sensor data for accurate motor fault diagnosis under varying speed condition. First, vibration, acoustic, and leakage magnetic signals are sampled from the data acquisition node. The accumulative rotating angle of the motor rotor is calculated from the leakage magnetic signal. Then, the order analysis is conducted on the vibration and acoustic signals based on the angle curve. Finally, the features of the pre-processed signals are extracted and fused by using the double-layer bidirectional long short-term memory (DBiLSTM) networks for fault pattern recognition. Experimental results show that the proposed method can identify 10 types of working conditions including high-resistance connection, eccentric, broken wire of the Hall sensor, interphase short circuit, and bearing faults with the accuracy of 99. 86% , by extracting and fusing of 8 channels of motor vibration and acoustic signals. The method is promising to be deployed into the internet of things edge computing node for remote online condition monitoring and fault diagnosis.