In order to detect the turbopump fault short of fault samples, the spectrums of turbopump vibration signals were analyzed, and the frequency band energy ratio was selected as the fault feature of those signals. After SOM competitive learning theory and U matrix description of clustering results were discussed, the frequency-band-energy-ratio-based SOM algorithm for turbopump fault detection is presented, and the selection of the best matching unit (BMU) and the adaptive upgrade of their weight vectors are also realized in this algorithm. With a liquid rocket engine (LRE) historical test data, this algorithm is validated. These results show that there is only one class when the algorithm is used to healthy turbopump vibration data, and the distance between the neighboring neuron is less than 0.1; while there are two or more classes when the algorithm is used for faulty turbopump vibration data, and the distance between the neighboring neuron is greater than 0.1. Therefore the algorithm can effectively detect the turbopump fault.