Rolling bearing is the core component of large rotating machinery in agricultural engineering. The composite fault is more harmful than the single fault in the process of operation. The source signals of composite faults are coupled with each other through the convolution in the process of propagation, which brings difficulties to fault detection. The maximum second-order cyclostationary blind deconvolution (CYCBD) can be used to reduce the influence of the transmission path using the deconvolution process. The mutual coupling between signals can be eliminated to effectively extract the periodic pulse signals. However, the CYCBD cycle frequency is directly related to the cycle stability of deconvolution signals. There is a great influence on deconvolution. The fault characteristic frequency depends mainly on the manual experience or optimization. It is a high demand to determine the composite fault diagnosis of rolling bearings in production practice. This study aims to extract the composite fault features of rolling bearings for the adaptive diagnosis of composite faults. An improved composite fault diagnosis was proposed for the CYCBD rolling bearings using RCC-NPH fusion index. Firstly, an investigation was made to comprehensively characterize the composite fault signals, then to integrate the ratio of cyclic content (RCC) and normalized proportion of harmonics (NPH) indexes. A new RCC-NPH fusion index was also proposed to consider the signal SNR, impact property, and harmonic components. As such, the CYCBD was independent of the prior knowledge to determine the cycle frequency covering all the fault frequency space. Secondly, the cycle frequency of CYCBD was set adaptively, according to the RCC-NPH fusion index. The cycle frequency dataset was also set to achieve the adaptive selection of CYCBD parameters. Thirdly, the parameter adaptive CYCBD served as the deconvolution on the input composite fault signals. The fault signals corresponding to different fault frequencies were then extracted to realize the effective separation of composite faults. Finally, the extracted single fault signal was demodulated by the Hilbert envelope to realize the fault identification. An experimental platform was developed to verify the improved model using the simulation signals and the experimental data. Experimental results show that the improved model with the RCC-NPH fusion index accurately and efficiently estimated the cycle frequency in the line with the characteristics of the signal. The CYCBD was also independent of the prior knowledge on the composite fault diagnosis. At the same time, the RCC-NPH fusion index effectively suppressed the interference frequency, in order to visually depict the composite fault features. An accurate extraction was realized for the fault components contained in the signals by systematic comparison with four indexes, including the RCC, NPH, autocorrelation spectrum, and multi-point kurtosis spectrum. The mutual coupling between signals was eliminated to successfully extract each single fault component after adaptive fault diagnosis for the rolling bearing composite faults. The comprehensive diagnosis of composite faults was realized to effectively avoid the misdiagnosis and missed diagnosis. Therefore, the composite fault adaptive diagnosis can be expected to effectively identify and separate each single fault feature in the composite fault, particularly for the adaptive diagnosis of rolling bearing composite faults. [ABSTRACT FROM AUTHOR]