Introduction
Modern Cyber-Physical Systems (CPS), such as smart grids and large-scale sensor networks, are characterized by high complexity and interdependency. Monitoring these systems effectively requires identifying invariants—stable relationships between sensor variables that persist during normal operation. Deviations from these invariants serve as critical indicators of system anomalies or physical faults.
However, traditional monitoring techniques often struggle with the indirect nature of many physical relationships. In a microgrid, for instance, the correlation between two distant sensors might be driven by hidden environmental factors or global control logic rather than a direct physical link. Standard Autoregressive Exogenous (ARX) models, which focus on direct linear correlations, frequently fail to capture these "hidden" invariants, resulting in a sparse and incomplete system model.
Methodology: Latent Factor Regression (LFRX)
To address the limitations of direct modeling, this research introduces Latent Factor Regression (LFRX). This framework enhances traditional autoregressive models by incorporating latent variables that represent unobserved system drivers.
The LFRX process involves several key stages:
- Latent Factor Extraction: Factor analysis is applied to the observed time-series data to identify $k$ latent factors ($h_1, \dots, h_k$) that capture underlying commonalities across sensors.
- Model Augmentation: The standard ARX model is extended. Instead of predicting a sensor's state solely from its own history and one other peer sensor, the LFRX model includes the extracted latent factors as exogenous inputs.
- Invariant Search and Selection: The system evaluates pairs of sensors to determine the most robust relationship. A "superiority threshold" ($\Delta$) is employed to favor simpler ARX models unless the LFRX model provides a significantly better fit, preventing over-parameterization.
- Graph Construction and Anomaly Detection: Discovered invariants are represented as an "invariant graph." During operation, the system monitors these edges. A significant prediction error indicates a "broken" invariant. By analyzing the topology of broken edges, the system can localize faults to specific nodes.
Experimental Evaluation
The LFRX framework was validated across three diverse datasets:
- Synthetic Data: Used to verify the algorithm's ability to recover known ground-truth dependencies amidst noise.
- Intel Berkeley Research Lab: A dataset of 54 sensors measuring temperature, humidity, light, and voltage.
- NEC Microgrid Dataset: A real-world dataset comprising 84 features, including solar PV output, battery status, and weather data.
Results and Discussion
The experimental results demonstrate that LFRX significantly outperforms traditional methods in both coverage and accuracy. In the Intel Berkeley dataset, LFRX consistently reduced estimation errors across all sensor types compared to ARX.
Notably, the inclusion of latent factors dramatically increased the number of discovered invariants. For temperature sensors, LFRX identified 183 invariants compared to 145 for ARX. In the complex Microgrid dataset, the integrated approach uncovered 2,285 invariants, nearly 35% of which were indirect relationships that would have been invisible to standard models. Furthermore, in anomaly detection tests, LFRX achieved a precision/recall score of 0.68, a substantial improvement over the 0.51 achieved by ARX-only methods.
Conclusion
Latent Factor Regression provides a more comprehensive lens for monitoring complex cyber-physical systems by uncovering hidden invariants. By modeling the "hidden glue" of latent variables, LFRX creates a more robust invariant graph that improves both anomaly detection and fault localization. Future research will focus on reducing the computational complexity ($O(Cn^2t^3)$) for real-time deployment in massive sensor networks and exploring non-linear latent models to capture even more intricate system dynamics.
Summary updated