Speaker
Description
The ATLAS Detector Control System (DCS) ensures the safe and reliable operation of the detector by providing a standardized interface to all subsystems and infrastructure. The hardware status is currently assessed against static, vendor-defined thresholds and operational status, enforced in the SCADA back-end. The current system is therefore blind to gradual degradation and correlated hardware behavior. To overcome this limitation, the Embedded Monitoring Processor (EMP), a multiprocessing System-on-Chip platform, is being leveraged to prototype a hardware-accelerated anomaly detector. To this end, two approaches, Seasonal-Trend decomposition and a recurrent autoencoder, are trained and evaluated on EMP commissioning data.
Summary (500 words)
The Detector Control System (DCS) ensures the coherent, safe, and reliable operation of the ATLAS experiment. The system provides a standardized interface to all subsystems and infrastructure, thereby facilitating the continuous monitoring, control, and archiving of operational parameters across the entire detector. In its present implementation, the Supervisory Control and Data Acquisition (SCADA) back-end evaluates system health primarily through static, vendor-defined thresholds. While this approach has proven effective at identifying hard failures, it is inherently limited in detecting gradual performance degradation, seasonal variations, or subtle correlations across multiple sensor channels that may indicate emerging issues before thresholds are exceeded.
To address these limitations, the Embedded Monitoring Processor (EMP) is being used as the platform for prototyping a hardware-accelerated anomaly detector. The EMP serves as the back-end processing platform of a novel DCS interface solution developed for the High-Luminosity era. It is an embedded platform designed around a Zynq UltraScale+ multiprocessing System-on-Chip (MPSoC), featuring an ARM-based Processing System (PS) and a Programmable Logic (PL) fabric. The integration of embedded processing and reconfigurable logic makes it a suitable candidate for implementing a data-driven anomaly detector.
Two anomaly detection approaches are being developed and evaluated independently. The first is a univariate method based on Seasonal-Trend decomposition using LOESS (STL), which decomposes each time series into trend, seasonal, and residual components. Anomalies are identified in the residual component using an Isolation Forest model. Applied to individual channels such as temperature, voltage, and current, this approach is effective at detecting local variations, including transient spikes, level shifts, and deviations from expected periodic behavior.
The second approach is a multivariate Long Short-Term Memory (LSTM) autoencoder. This model is trained to learn the nominal temporal evolution and correlation structure of the full sensor data vector. By reconstructing input sequences and evaluating the reconstruction error, it identifies anomalous behavior such as gradual drifts, slow degradation, or inconsistencies across multiple channels, which are not observable through independent per-channel analysis.
Deploying both approaches on an embedded platform introduces several key design constraints. The anomaly detector must coexist with other DCS applications running concurrently on the EMP, requiring careful resource management to avoid saturating the PS or PL fabric. The pipeline must therefore scale efficiently across the monitored channels. Additionally, minimizing the false-positive rate is critical, as spurious alerts in a large-scale detector environment would undermine operational confidence in the system.
Currently, the models are trained offline using data collected during EMP commissioning, consisting of thermal and operational parameters from multiple devices under realistic conditions. This dataset serves as a benchmark for assessing detection performance and system behavior. The offline-trained models are deployed on the EMP, using the PS for STL-LOESS preprocessing and PL-based acceleration of inference for the Isolation Forest and LSTM models, enabling fully integrated anomaly detection within the DCS architecture. The resulting anomaly score is published as an OPC UA node and consumed by WinCC OA as an advisory signal, with per-EMP masking applied.