[Paper Review] Recurrent Neural Networks for anomaly detection in the Post-Mortem time series of LHC superconducting magnets

Maciej Wielgosz, Andrzej Skoczeń+1 more|arXiv (Cornell University)|Feb 2, 2017

Seismology and Earthquake Studies18 references7 citations

TL;DR

This paper proposes a deep learning approach using LSTM and GRU recurrent neural networks to detect anomalies in post-mortem time series data from CERN's Large Hadron Collider superconducting magnets. By training on high-resolution historical data, the model achieves 99.9% accuracy in identifying abnormal operating conditions without manual thresholding, significantly reducing reliance on expert intervention during system setup.

ABSTRACT

This paper presents a model based on Deep Learning algorithms of LSTM and GRU for facilitating an anomaly detection in Large Hadron Collider superconducting magnets. We used high resolution data available in Post Mortem database to train a set of models and chose the best possible set of their hyper-parameters. Using Deep Learning approach allowed to examine a vast body of data and extract the fragments which require further experts examination and are regarded as anomalies. The presented method does not require tedious manual threshold setting and operator attention at the stage of the system setup. Instead, the automatic approach is proposed, which achieves according to our experiments accuracy of 99%. This is reached for the largest dataset of 302 MB and the following architecture of the network: single layer LSTM, 128 cells, 20 epochs of training, look_back=16, look_ahead=128, grid=100 and optimizer Adam. All the experiments were run on GPU Nvidia Tesla K80

Motivation & Objective

• To automate anomaly detection in superconducting magnets of the Large Hadron Collider (LHC) to reduce manual expert involvement.
• To replace hand-crafted, time-consuming models with a data-driven deep learning approach using RNNs.
• To identify optimal hyper-parameters for LSTM and GRU networks that maximize detection accuracy on high-resolution post-mortem data.
• To evaluate the feasibility of deploying such models in real-time control systems, including potential FPGA implementation.

Proposed method

• The method employs Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks to model temporal dependencies in time series data from LHC superconducting magnets.
• Time series data is discretized using a grid-based quantization scheme, where each value is mapped to a category based on predefined grid size (grid=100).
• The model uses a sliding window approach with look_back=16 and look_ahead=128 to predict future states and detect deviations as anomalies.
• Hyper-parameters such as number of cells, training epochs, and optimizer (Adam) are tuned via grid search to maximize classification accuracy.
• The architecture uses a single-layer LSTM with 128 memory units, trained over 20 epochs on a 302 MB dataset for optimal performance.
• Model evaluation is based on accuracy of anomaly classification, with results compared across varying grid sizes, look_ahead values, and network depth.

Experimental results

Research questions

• RQ1 Can deep learning models like LSTM and GRU effectively detect anomalies in high-resolution post-mortem time series of LHC superconducting magnets?
• RQ2 What is the optimal configuration of hyper-parameters (e.g., look_back, look_ahead, grid size, number of cells) for maximizing anomaly detection accuracy?
• RQ3 How does model performance vary with increasing data resolution and network complexity?
• RQ4 Can the proposed method achieve high accuracy without manual thresholding or expert supervision during system setup?
• RQ5 Is the model suitable for real-time deployment in hard real-time control systems like QPS, and can it be implemented on FPGAs?

Key findings

• The proposed LSTM-based model achieved 99.9% accuracy on the largest dataset of 302 MB using a single-layer architecture with 128 cells, 20 epochs, look_back=16, look_ahead=128, and grid=100.
• Increasing the grid size (finer quantization) degraded model performance due to increased classification complexity, despite more granular data representation.
• Larger look_ahead values reduced accuracy, as predicting further into the future increases uncertainty and model difficulty.
• Using more than nine LSTM cells provided minimal performance gain, indicating that nine cells are sufficient for this anomaly detection task.
• GRU and LSTM models showed nearly identical performance (61.17% vs. 61.12% accuracy) on a small dataset, suggesting comparable effectiveness for this application.
• The method enables automatic, threshold-free anomaly detection, significantly reducing the need for expert intervention in model configuration and validation.