In order to accurately detect the looseness position and degree of railway fasteners during service, a looseness detection algorithm of railway fasteners based on wavelet packet energy spectrum and improved BP neural network is proposed. First of all, referring to the current specifications, seven different loosening conditions of railway fasteners are designed on the spot, and the corresponding vertical vibration acceleration signals of rail under each condition are collected in turn; Then, the collected vertical vibration acceleration signal of the rail is decomposed by 7 layers of dB40 wavelet packet to obtain the corresponding wavelet packet node energy ratio data information under each working condition. On this basis, a vector dissimilarity coefficient index (VDC) is designed from the perspective of distance and dimension to realize the positioning of loose fasteners; In addition, according to the characteristics of the data, the particle swarm optimization algorithm is used to improve the BP neural network, build the PSO-BP railway fastener looseness detection model, and conduct parameter sensitivity analysis. The research results show that the minimum value of VDC of railway fastenings in healthy state is 0.17, and the maximum value is 0.41, which is significantly less than the VDC value of railway fastenings in each loose state. Based on this, accurate positioning of single and multiple loose fastenings can be achieved; The constructed PSO-BP railway fastener looseness detection model can achieve accurate detection of fastener looseness. When the number of neurons in the hidden layer is set to 20, the model detection effect is the best, and the corresponding recognition accuracy is 98.66%.