The carbonate reservoirs in Central Asia have undergone complex sedimentary, diagenetic, and tectonic processes, resulting in the development of multiple reservoir spaces such as pores, vugs, and fractures. The heterogeneity of the reservoirs is strong, making it difficult to classify and evaluate them. The accuracy of logging identification and prediction is low, which poses challenges to rapid water breakthrough and serious decline in oil well production. Aiming at the difficult problem of classification and evaluation of complex carbonate reservoirs, different combinations of reservoir spaces were identified based on core and imaging logging data. These reservoirs were divided into five types: porous type, fractured-porous type, porous-ruggy-fractured type, porous-ruggy type, and fractured type. The logging response characteristics of different reservoir types were summarized, and the logging curves sensitive to reservoir types are optimized. By comparing and analyzing the application effects of different machine learning methods such as SOM, MRGC, KNN, and ANN, the reservoir types of coring wells are divided into machine learning training samples, and the reservoir types of non-coring wells are identified and predicted based on sensitive logging curves, which greatly improves the accuracy of logging interpretation, finally, a complete logging evaluation method for complex reservoir is obtained, which has important guiding significance for the efficient development of this kind of oilfield.