High-speed robot-based manufacturing (RBM) systems require reliable and efficient maintenance strategies to minimize downtime and ensure operational continuity. Traditional predictive maintenance approaches rely on intrusive and costly sensing devices, such as vibration and torque sensors, which increase installation complexity and operational overhead. As a non-invasive alternative, acoustic sensing enables continuous condition monitoring. Current acoustic monitoring studies primarily depend on traditional single-sensor time-domain analysis and multisensory approaches that involve data-level fusion strategies. These methods often fail to adequately capture the complex spatial dynamics and noisy industrial sounds present in robotic environments. To overcome this challenge, this study introduces a Multichannel Acoustic Feature Fusion (MCAFF) framework to improve robotic condition identification for predictive maintenance. The proposed approach integrates signals from spatially distributed microphones to generate fused time-frequency representations using the Short-Time Fourier Transform (STFT). These representations are then used to train a Convolutional Neural Network (CNN) for classifying robot operating conditions. Experimental validation was conducted on a FANUC M-1iA/0.5A robot performing high-speed tasks at varying speeds under constant payload conditions. Results demonstrate that the proposed MCAFF framework improves classification accuracy by 15% compared to single-sensor approaches and by 10% compared to conventional data-level fusion methods. The findings highlight the effectiveness of multichannel acoustic feature fusion as a scalable, non-invasive solution for improving predictive maintenance and reliability in RBM systems.