Non-Intrusive Load Monitoring (NILM) disaggregates appliance-level energy consumption from aggregate load data, offering a practical approach for analyzing household energy use and promoting efficiency. Event detection, a critical element in NILM, identifies state transitions in appliances. However, existing approaches often struggle with real-world data, where challenges arise from subtle load changes, high-frequency noise, and load-switching transients, all of which impact detection accuracy. Additionally, high computational demands can limit real-time applicability and reduce adaptability to varied appliance behaviors. In this paper, we present a multi-window cumulative sum (CUSUM) method featuring adaptive thresholds and variable-length windows, enhanced by a dynamic parameter adjustment mechanism driven by signal statistics. To address noise interference, the method integrates median filtering and exponentially weighted moving average (EWMA) preprocessing. Dynamic parameter adjustments enable adaptation to fluctuating appliance behaviors, while multi-window accumulation and rate-of-change analysis improve discrimination between genuine events and spurious fluctuations. Experimental results demonstrate the proposed method's superiority over conventional CUSUM and other improved methods on PLAID, BLUED, and private datasets, achieving F1 scores of 94.72%, 96.43%, and 91.98%, respectively. Its robustness across thresholds, practicality, and low computational requirements highlight its potential for real-world NILM applications on resource-constrained devices. Keywords: NILM, event detection, multi-window, adaptive threshold, variable-length