The error backpropagation mechanism adjusts the synaptic weights by feeding back the total error signal layer by layer in a neural network, which is unreasonable in biology. In this paper, the supervised learning algorithm based on spike back reconstruction is proposed, which gets rid of the concept of total error, and derives the local error signal of each layer of the network by generating the desired spike signal of each neuron. The reconstruction spike signal is directly fed back to the presynaptic local neurons, thus the training of the multi-layer neural network can be realized through local training mode with biological interpretation. We derive the spike back reconstruction learning algorithm for deep spiking neural networks, in which the learning rule is expressed by the inner product operation of spike trains. Additionally, the spike train learning experiments are applied to verify the effectiveness of the proposed algorithm, and the impact of relevant factors on the learning performance of the algorithm is also analyzed. Experimental results demonstrate that the learning ability of spike back reconstruction method is obviously superior to the traditional error backpropagation mechanism.