Abstract: Isolation forest is an efficient anomaly detection algorithm, but it is difficult to detect anomalies in nonlinear separable data by the way of axial parallel line isolation. To solve this problem, a semi-supervised outliers detection method, Sparse deep Random Projection Positive isolation Forest (SRP-PiF), is proposed by using neural network random projection fusion isolation forest. Algorithm uses neural network to carry out nonlinear random projection, and subspaces of different sizes are divided nonlinearly to provide effective low-dimensional feature attributes for classification. The sparse rate of hyperparameter is set to adjust the weights in neurons and improve the efficiency of operation. Finally, the semi-supervised learning method is used for training and prediction. The method verified on 10 data sets in the real world. Compared with similar anomaly detection algorithms, SRP-PiF has a significant improvement in the anomaly detection performance index AUC-ROC. The accuracy of the algorithm is 3% - 20% higher than that of the isolation forest algorithm and the unsupervised outlier detection method of cumulative experience distribution function(ECOD) for large data sets.