In industrial process monitoring of large-scale equipment and complex systems, conventional fault detection methods are significantly impaired by noise interference and non-stationary data characteristics, which leads to unsatisfactory diagnostic performance. To address these challenges, an innovative fault detection framework was proposed, integrating global envelope preprocessing with an improved partial least squares (PLS) algorithm. First, global envelopes were constructed through moving average calculation and moving standard deviation analysis to suppress noise while preserving legitimate signal fluctuations. Subsequently, a standard deviation-weighted term was incorporated into the PLS formulation to enhance transient feature sensitivity. Finally, dual-domain statistical metrics were developed in both residual and original feature spaces to enable real-time online monitoring. The method's efficacy was validated through experiments on both simulated fault datasets and penicillin fermentation processes, demonstrating zero detection latency; furthermore, comparative experiments on the Tennessee Eastman process significantly improved the detection rate of quality-related faults, confirming the method's superiority. The research demonstrates that this method provides an effective technical solution for fault detection in complex environments with high noise and non-stationary processes, offering significant engineering application value.