In root cause analysis under microservice failures, existing log analysis methods are constrained by data redundancy and format heterogeneity. Tracing methods focus on coarse-grained localisation, making it difficult to achieve metric-level fine-grained diagnosis. To address these issues, we propose a fine-grained root cause analysis method (TF-RCA) based on time-frequency fusion and prediction-driven analysis. First, by fusing time-frequency domain features, local feature expression is enhanced and periodic dependencies between metrics are captured. Second, it introduces a Granger causality test mechanism driven by prediction errors, quantifying causal strength via perspective of improved time series prediction accuracy to generate an initial causal graph. This is then combined with the Kruskal minimum spanning tree strategy to construct a directed acyclic causal graph with optimal similarity. Finally, it combines the PageRank algorithm with standardised anomaly scores to identify the root cause. Experiments show that TF-RCA outperforms existing state-of-the-art methods, achieving improvements of 19% and 30% in Avg@K on real microservice benchmark datasets and synthetic datasets, respectively.