Compositional Vector Representations for Similarity-Based Monitoring of Heterogeneous Multivariate Telemetry

Autonomous and intent-driven network management requires telemetry representations that are comparable and useful for downstream reasoning. This thesis explores the use of Vector Symbolic Architectures (VSA) for representing heterogeneous multivariate time-series telemetry as fixed-dimensional vectors for similarity-based monitoring. The focus is not on introducing a new feature set, but on evaluating how VSA operations can structure and compose different views of the same telemetry window. A sensing-layer representation pipeline is developed in which raw telemetry is segmented into observation windows and encoded using residual-monitoring descriptors, catch22 features, learned TS2Vec representations, and aggregated combinations between them. Scalar descriptors are normalised, quantised, interpolated, bound to structural roles, and bundled into compact vectors. Learned representations are evaluated directly and in combination with descriptor-based representations. The main evaluation is performed across 28 entities of the Server Machine Dataset using reference-bank anomaly scoring, where test windows are compared to nominal training windows by cosine similarity. A secondary validation evaluates nearest-neighbour classification on five UEA time-series datasets. Robustness is evaluated using Gaussian noise and block missingness applied to SMD test telemetry. Residual-monitoring descriptors achieve the strongest standalone SMD performance, with a mean AP of 0.579 and mean AUROC of 0.844. TS2Vec is the second strongest standalone representation, while catch22 provides a weaker but still useful statistical view. The role-binding ablation shows that structural identity matters: value-only encoding performs substantially worse than representations that bind values to metric, descriptor, and family roles. Hierarchical aggregation improves residual+catch22 composition compared to flat aggregation, increasing mean AP from 0.487 to 0.578. Residual+TS2Vec gives a small average improvement over residual monitoring and improves 16 of 28 entities, while also showing stronger robustness under degraded telemetry. The results show that VSA-based encoding can compose heterogeneous time-series representations into compact vectors that support similarity-based reasoning. They also show that representation structure is not incidental: role binding, hierarchical aggregation, and branch selection affect the usefulness of the final representation space. This supports VSA as a structured composition mechanism for telemetry-state representation, rather than as a standalone anomaly detector or feature-engineering method.