Gearbox fault diagnosis under noisy operating conditions remains a critical yet unsolved challenge for industrial condition monitoring. The primary challenge originates from a fundamental conflict: fault signatures are present within certain frequency ranges, yet standard deep learning models process raw vibration signals without recognizing which frequencies are relevant. Moreover, noise affects the entire spectrum uniformly. Existing transformer-based methods for vibration analysis treat time steps as tokens and therefore fail to capture cross-sensor dependencies, while conventional denoising approaches apply fixed filters that cannot adapt to the varying spectral characteristics of different fault types and noise levels. We propose the Adaptive Frequency-Aware Inverted Transformer (AF-iTransformer), a lightweight transformer framework that lets the model learn which frequencies to attend to on a per-sample basis. In particular, we propose a learnable spectral filter that transforms the input signal to the frequency domain via FFT. Then it predicts a soft frequency mask conditioned on signal statistics and applies it before reconstructing the filtered signal through iFFT, allowing the model to suppress noise bands while preserving fault-relevant spectral content dynamically. After adaptive filtering, the architecture employs channel-level tokenization to uniformly represent heterogeneous channels as input tokens, relying on cross-channel attention to automatically learn their distinct contributions to fault diagnosis. Feature-wise linear modulation is introduced to inject signal-level statistics at every encoder layer. Furthermore, the framework utilizes residual attention propagation to stabilize deep training, and an auxiliary spectrum prediction head provides spectral regularization during training. On the UConn Gearbox dataset with nine fault categories, AF-iTransformer achieves 99.63% accuracy on clean data and maintains 95.1% at 0 dB signal-to-noise ratio, substantially outperforming all baselines under noisy conditions. On the SEU gearbox dataset with five fault categories, AF-iTransformer achieves 99.74% clean accuracy. On 2 GB edge GPUs, AF-iTransformer achieves a per-window inference latency of 7.3–13 ms with peak memory below 17 MB, and 1.4–4.7 ms on modern CPUs, confirming its viability for real-time industrial deployment.