Path Loss Classification Using Time-Series Features in Open RAN Key Performance Metrics

Open RAN (O-RAN) enables data-driven optimization via exposed Key Performance Measurements (KPMs). We explore whether KPM-only time series data is sufficient to classify path-loss values without Physical layer probes. We used an Openairinterface5g (OAI/FlexRIC) testbed, multivariate KPM streams (throughput, delay, resource usage) in three fixed path-loss (20, 23, 26 dB) scenario. A MiniROCKET transform with ridge classifier evaluated window sizes W = 10-90 samples (2-18 s at 200 ms/sample) and strides S = 10-90. We observed increment in performance with increas in W and plateaus at W ≈ 80-90, achieving a best macro-F1= 0.881 at (W, S) = (90, 40). Errors are noted between 20 and 23 dB (due to overlapping KPM distributions), whereas 26 dB is trivially separable. Thus, the results indicate that standard O-RAN telemetry alone provides practical propagation awareness for RAN automation.