Automatic modulation classification (AMC) is one of the most crucial technologies for identifying the modulation formats of unknown wireless signals in cognitive radio. However, few works focus on collaborative AMC (Co-AMC) performance under unknown varying environments. Such differences of received signals under these conditions make it hard for the existing fusion strategies to exhibit their advantages, even probably leading to the performance degradation due to the inappropriate fusion. To address these issues, we design a signal-level Co-AMC framework, where an adaptive signal-level fusion (ASF) method is proposed to achieve high classification performance. The proposed ASF method is capable of generating the fused spectral quotient (FSQ) signal possessing the maximum signal-to-interference ratio (SIR). Furthermore, we construct the signal distribution property vector (SDPV) from the FSQ signal to concisely represent the modulation-related information. At last, the SDPVs are sent to a convolutional neural network (CNN) to accomplish the final classification. The simulations are conducted to prove that the proposed Co-AMC framework provides significant classification performance improvement compared to existing AMC methods under different scenarios.