Faults such as bending and breaking of the bearing of the sintering machine trolley will cause the wheels to fall off，which in turn will affect the normal operation of the sintering machine trolley. Aiming at such problems as large signal interference，poor detection environment and low real-time performance in the current bearing fault detection of trolleys，a wheel stuttering and swing detection algorithm based on multi-trajectory gated recurrent unit ( GRU) and multi-head attention mechanism is proposed to determine whether the bearing has failed. Firstly，a wheel stuttering and swing detection platform of sintering machine trolley is built，and the actual motion trajectory of the bolt is drawn by using the object detection algorithm. Then，the obtained bolt trajectory coordinates are clipped into trajectory coordinate fragments with the same number of frames according to continuous frames，and divided into normal trajectory fragments，stuttering trajectory fragments and swing trajectory fragments. Finally，the constructed multi-trajectory unit and multi-head attention mechanism model are used to classify the normal trajectory，stuttering trajectory and swing trajectory，the classification results of other recurrent neural network models are compared，and the optimal model is selected for bearing fault detection. In order to verify the feasibility of the method，it is tested on the trolley wheel detection platform. Experiments show that the accuracy of the multi-trajectory unit and multi-head attention mechanism model proposed in this paper is 90. 38% ，which is better than other recurrent neural network models constructed，and can accurately detect the wheels that stutter and swing， providing an effective solution for the automatic detection of bearing failures in trolleys.