The rapid development of deep learning has inspired some new methods to solve the 3D reconstruction problem of tomographic particle image velocimetry. In order to solve the problem that the supervised deep learning method requires a large number of ground truth data and the synthetic dataset is consistent with the experimental scene, an unsupervised particle reconstruction algorithm based on 3D convolution encoding and MLP decoding was proposed, named MLOS-CNN-MLP. This method first used the MLOS algorithm to generate the initial three-dimensional light intensity distribution from multi-view 2D images, then extracted features and constructs neural encoders through 3D convolution encoding, and finally used MLP decoding to regress the accurate three-dimensional light intensity distribution. Unlike the supervised learning methods that use ground truth data, MLOS-CNN-MLP employed a projection function to render the reconstructed 3D distribution into multiple views and compared them with the original 2D images to compute the loss function, which made the whole network training without ground truth data to achieve unsupervised learning. The verification result on the synthetic data shows that the reconstruction accuracy of the proposed unsupervised particle reconstruction algorithm can reach 0.95 under typical experimental particle density, which is much higher than that of the traditional algebraic reconstruction technique. Further comparing the performance of different particle densities and noise robustness, the reconstruction quality of the proposed method is better than that of the traditional algebraic reconstruction technique, and the calculation speed is at least one order of magnitude faster.