The widespread application of wireless communication has led to increasingly complex electromagnetic environments, where spectrum abuse and malicious interference frequently cause abnormal signals. Radio anomaly detection technology has emerged to address these challenges. This paper focuses on the performance limitations of existing radio anomaly detection methods under low interference-to-signal ratio (ISR)conditions. We propose a fusion detection algorithm, LOF-OCSVM, integrating local outlier factor (LOF) and one-class support vector machine (OCSVM). Innovatively, we introduce three novel features: fluctuation entropy (FE), fluctuation mutual information (FR-MI), and lognormal distribution fitting parameters derived from signal fluctuation sequences. These features quantify the disorderliness, adjacent correlation, and statistical distribution characteristics of signal fluctuations, significantly enhancing the detection sensitivity for weak interference signals. Simulation experiments demonstrate that: feature effectiveness: the new features improve recall by >30% and F1-score by >23% at −20 dB ISR. Model superiority: the LOF-OCSVM fusion model achieves an F1-score of 0.8634 at −20 dB ISR through a hierarchical decision mechanism, outperforming single model and simple hybrid approaches. Robustness: compared to Deep SVDD and E-GAN, our method improves AUC by 3% under low ISR conditions while maintaining stronger robustness.