With the construction and use of large radio telescopes such as Five-hundred-meter Aperture Spherical radio Telescope (FAST), pulsar survey data has entered the PB era. To solve the problem of scalar data mining with such a large number of high-speed sampling and promote the discovery of new astronomical phenomena, this paper proposes a pulsar candidate sifting scheme based on unsupervised clustering. This scheme uses a hybrid clustering algorithm based on density hierarchy and division method, combined with MapReduce/Spark parallel computing model and a sliding window-based grouping strategy, thereby improving the efficiency of screening a large number of candidate signals. Comparative experiments on the data set HTRU2 (High Time Resolution Universe) show that the algorithm can achieve higher accuracy and recall rates, which are 0.946 and 0.905, respectively. And when parallel nodes are sufficient, the time complexity of the algorithm is significantly reduced compared to the serial execution method. It can be seen that this method provides a feasible idea for the analysis and mining of big data pulsar observation.