Coverage is an important index to test the completeness of chip verification， especially functional coverage， which can measure whether the functional characteristics of the design are fully verified. At present， for the collection of function coverage， the general practice is to set function points in the coverage group， divide the test scene into a specific number of bins according to complexity， and then run the simulation to determine whether each bin is hit. Its implementation is relatively simple， but due to the existence of various factors， the hit situation of each bin in a function point is often unbalanced， resulting in insufficient coverage of some scenes. To solve this problem， a verification method based on machine learning algorithm to achieve uniform coverage distribution is proposed. By training neural networks， various excitation vectors can be accurately predicted. In this method， the reverse network prediction method and the forward network real-time fitting method are designed respectively for small and large number of covered bins， which can realize the balanced hit of each bin easily. The experimental results show that， compared with the case where the difference between the maximum and the minimum hit times of random test coverage points is several orders of magnitude， the small point bins can basically achieve the average distribution and the large point bins can reduce the extreme value ratio to less than 1.5 times， thus significantly reducing the verification risk of some cases.