Multimodal Root Cause Analysis for Cloud Microservices: Integrating Temporal Convolutional Networks with Causal Discovery and LLM Priors

Root cause analysis (RCA) in cloud microservice systems remains a critical challenge, with Mean Time to Identify Root Cause often exceeding 60 minutes, and downtime costs are becoming prohibitively expensive. Existing approaches either rely on rule-based systems that do not adapt to new failure modes or employ correlation-based methods that fail to distinguish causation from correlation, leading to misidentification of cascading failures. We present MuRCA, a multimodal deep learning system that integrates metrics, logs, and traces through a gated fusion mechanism, combining a principled fusion of statistical causal discovery (PCMCI)-based causal discovery with Large Language Model (LLM)-derived domain knowledge priors. Our architecture employs depthwise separable Temporal Convolutional Networks (DS-TCN) for efficient temporal modeling, achieving a 66% parameter reduction while maintaining a 127-timestep receptive field. Cross-service attention augmented with causal bias injection enables the model to distinguish root causes from cascading effects. Evaluated on the RCAEval benchmark, MuRCA achieves 88.9% AC@1 accuracy (ensemble) with 48% variance reduction over single models, representing a 37.0 percentage point improvement over the existing state-of-the-art. Individual model performance ranges from 70.4% to 92.6% across random seeds (mean: 81.5%, std: 10.1%), demonstrating that ensemble aggregation is essential for production deployment. With 3.3ms inference latency (272× faster than existing methods) and statistical significance (p < 0.001, Cohen's d = 1.89), MuRCA enables reliable, real-time fault diagnosis for production microservice environments.