The invention discloses a novel diamond press anvil crack online detection method. By using a backtracking search optimization algorithm (BSA), a support vector machine model parameter is optimized and selected to avoid blindness of parameter selection; in the application of a K-SPWVD method, a template matching method and K-S inspection are utilized to avoid complicated time-frequency analysis of each signal to be detected and enhance the efficiency of the algorithm; and good time-frequency analysis performance of smooth pseudo Wigner-Ville distribution is also sufficiently unitized. The novel diamond press anvil crack online detection method disclosed by the invention has high accuracy and high computing speed and is beneficial for online detection of diamond press anvil cracks.