Lightweight CNN-Based Anomaly Detection for High Voltage Converter Modulators in the Spallation Neutron Source

This paper proposes a lightweight CNN-based anomaly detection framework for Spallation Neutron Source high-voltage converter modulators that leverages architectural inductive bias by strategically ordering temporal and cross-channel operations, achieving state-of-the-art performance in identifying fault precursors across multiple subsystems.

The Gist

The Big Picture: The "Heartbeat" of a Giant Machine

Imagine the Spallation Neutron Source (SNS) as a massive, high-speed train system. Its job is to shoot tiny particles (neutrons) at a target to help scientists study materials. To keep this train running, it needs a massive amount of power, delivered in short, intense bursts called "pulses."

The High Voltage Converter Modulators (HVCMs) are the engines that create these power bursts. Think of them as the heart of the machine. If the heart skips a beat or stutters, the whole train stops. When the train stops, scientists lose valuable time, and expensive parts can get damaged.

The problem is that these engines don't always break suddenly. Often, they give subtle "warning signs" (precursors) before they fail. The goal of this paper is to build a smart, lightweight computer program that can listen to the engine's heartbeat and say, "Hey, something is wrong," before the engine actually stops.

The Challenge: Listening to 14 Different Instruments

The engineers have 14 different sensors watching the engine. Some measure current (like blood flow), some measure voltage (like blood pressure), and some measure magnetic fields (like the rhythm of the heart).

The tricky part is that a "sick" engine doesn't always look the same.

• Sometimes, just one sensor goes crazy (like a spike in blood pressure).
• Sometimes, the sensors don't go crazy individually, but they start talking to each other strangely (like two heartbeats getting out of sync).

Previous computer programs tried to listen to all 14 sensors at once, but they were like a person trying to hear 14 different conversations in a noisy room all at the same time. They got confused about which conversation mattered.

The Solution: A New Way to Listen

The authors of this paper proposed a new way to organize the computer's "ears." They realized that to understand the engine, you need to do two things in a specific order:

1. Listen to the rhythm of each individual sensor (Time).
2. Compare the sensors to see how they relate to each other (Channels).

They tested three different ways to arrange these steps, using a technique borrowed from mobile phone cameras (which need to be fast and light):

1. The "Solo First" Approach (DS): Listen to each sensor's rhythm individually first, then compare them.
   • Analogy: Imagine a choir director asking every singer to practice their part alone first, and then having them sing together to see if they harmonize.
2. The "Mix First" Approach (PW-First): Mix all the sensors together first, then listen to the rhythm of the mix.
   • Analogy: Imagine blending all the singers' voices into one smooth smoothie first, and then listening to the rhythm of that smooth drink.
3. The "Mix First with a Spotlight" Approach (PW-First+SE): Mix the sensors, but add a smart "spotlight" that can instantly decide which voices are important for this specific moment and turn up the volume on them while turning down the noise.
   • Analogy: This is like a DJ at a party who mixes all the music but can instantly boost the bass or the vocals depending on what the crowd needs right now.

The Results: The "Spotlight" Wins

The team tested these three approaches on real data from the SNS, which includes four different types of engine setups (RFQ, DTL, CCL, SCL).

• The Winner: The "Mix First with a Spotlight" (PW-First+SE) approach was the best. It was the most accurate at spotting the warning signs.
• Why it won: It was flexible. Sometimes the problem was just one sensor acting up (so the spotlight focused on that one). Other times, the problem was a weird relationship between two sensors (so the spotlight helped the computer see the connection).
• The Score: It achieved a score of 0.816 (on a scale where 1.0 is perfect) for spotting these rare faults. This is better than any previous method tested on this specific data.

What the Computer Learned (The "Aha!" Moments)

By analyzing how the computer made its decisions, the authors found some interesting things:

1. Three Super-Sensors: Out of the 14 sensors, three were the most important: C-Flux (magnetic field), Mod-V (output voltage), and CB-I (capacitor current). If you turned off the other 11, the computer could still do a decent job. But if you turned off these three, the computer got lost.
2. The "Derivative" was Redundant: One sensor measured the change in voltage (how fast it was rising). The computer realized this was just a math copy of the voltage sensor itself. It didn't need both; one was enough.
3. Different Faults Need Different Strategies:
   • If a fault causes a huge jump in one sensor's value (like a loud scream), the simple "Solo First" approach works fine.
   • But if a fault is subtle and only shows up as a weird relationship between sensors (like a whisper), the "Mix First with a Spotlight" approach is essential. It's the only one that can catch the whisper.

The Bottom Line

This paper shows that for detecting faults in giant, complex machines, how you organize your data matters just as much as the data itself.

By building a lightweight computer model that can flexibly switch between listening to individual sensors and comparing them as a group, the researchers created a system that is better at predicting failures than the current state-of-the-art methods. This means the SNS (and potentially other similar machines) can run longer with fewer unexpected stops, saving time and money.

Technical Summary

Technical Summary: Lightweight CNN-Based Anomaly Detection for HVCM in SNS

Problem Statement

Unscheduled trips of high-power pulsed converters are a primary source of downtime at large accelerator facilities. At the Spallation Neutron Source (SNS), High Voltage Converter Modulators (HVCMs) are the second-largest contributor to lost beam time. Each HVCM pulse is recorded across 14 sensor channels (currents, voltages, magnetic fluxes, and derivatives) where mutual interactions encode the system's operating state.

The core challenge lies in the heterogeneity of fault precursors. Faults do not manifest uniformly: some alter the temporal structure of individual signals, while others change statistical dependencies among channels. Existing deep-learning approaches often employ standard convolutional pipelines that entangle temporal and cross-channel operations from the first layer. This design lacks an explicit mechanism to represent channel independence or structured inter-channel interaction, potentially hindering the model's ability to adapt to regime-dependent channel correlations that shift during pre-fault conditions. Furthermore, the data is characterized by severe class imbalance (faults are rare) and varying degrees of class separability across the four SNS subsystems (RFQ, DTL, CCL, SCL).

Methodology

The authors propose a factorized inductive bias for anomaly detection, decomposing the 1D convolutional operator into distinct temporal filtering and cross-channel mixing stages. Inspired by depthwise-separable convolutions used in mobile vision, the study operationalizes three lightweight CNN variants and compares them against a standard joint-mixing baseline:

1. DS (Depthwise-First): Extracts per-channel temporal features first using depthwise convolutions, followed by cross-channel mixing via pointwise convolutions. This order assumes discriminative signals reside primarily in the shape of individual channel traces.
2. PW-First (Pointwise-First): Reverses the order, mixing channels first via pointwise convolutions (learning linear combinations of raw signals) before applying per-channel temporal filtering. This targets scenarios where discriminative signals exist in cross-channel relationships.
3. PW-First+SE: Augments PW-First with a Squeeze-and-Excitation (SE) gate inserted between the mixing and filtering stages. This allows for per-pulse adaptive re-weighting of the mixed channels, enabling the network to dynamically shift reliance on specific channel combinations based on the input pulse.

All variants share an identical backbone (four convolutional blocks with widths 32, 64, 128, 64), classifier head, and training protocol. The models are trained in a supervised regime on the public HVCM dataset, utilizing a weighted binary cross-entropy loss to address class imbalance.

Key Contributions

• Architectural Factorization: The paper proposes and evaluates three lightweight CNN variants (DS, PW-First, PW-First+SE) that explicitly factorize convolutional operations into temporal and channel-mixing components, contrasting them with a standard joint-mixing baseline.
• Comprehensive Benchmarking: The variants are benchmarked against three families of existing HVCM-specific models (including supervised CNN+LSTM, unsupervised autoencoders, and multi-module CVAEs) and traditional machine learning baselines (KNN, RF, SVM) across all four SNS subsystems.
• Ablation and Sensitivity Analysis: The study includes extensive ablation studies (sensor-group and per-channel zero-out) to identify which input channels drive performance and how different architectures distribute reliance across physical sensor groups.
• Fault-Family Specific Insights: The analysis links performance variations to the nature of fault precursors, distinguishing between faults that manifest as amplitude shifts in individual channels versus those requiring joint channel representations.

Results

Evaluated on the public HVCM dataset across all four subsystems, the PW-First+SE variant achieved the best overall performance, with a pooled AUC-PR of 0.816 and AUC-ROC of 0.934.

• Performance Ranking: A monotonic improvement was observed across the variants: Standard < DS < PW-First < PW-First+SE. The greatest gains were seen in metrics sensitive to the minority class (Recall, G-Mean, AUC-PR).
• Subsystem Variations: PW-First+SE outperformed other variants on RFQ, DTL, and CCL. On the SCL subsystem, PW-First narrowly outperformed PW-First+SE, suggesting that the adaptive gating mechanism is most beneficial when channel-specific cues are weak (as in CCL) and the network must combine evidence from multiple signals.
• Comparison to State-of-the-Art: The proposed method outperformed the strongest published supervised baseline (Zhou et al.'s CNN+LSTM) on F1, Recall, and G-Mean across subsystems. It also surpassed unsupervised baselines (RAEs, CVAE) on AUC-ROC for the most difficult fault families (FLUX, Driver, IGBT, SCR).
• Channel Importance: Ablation studies identified C-Flux, Mod-V, and CB-I as the three dominant input channels. The DV/DT channel was found to be largely redundant. The DS variant relied more heavily on the Flux group, while joint-mixing variants relied more on the Power group.

Significance and Claims

The paper claims that the ordering of temporal filtering and cross-channel mixing is a critical architectural choice for multi-channel physical sensor data, rather than a neutral design decision. The results suggest that:

1. Inductive Bias Matters: Explicitly separating temporal and channel operations allows the model to better capture regime-dependent interactions.
2. Adaptive Gating is Beneficial: The SE block in PW-First+SE provides a significant advantage when fault precursors are subtle and require the integration of information across multiple channels, rather than relying on strong amplitude shifts in single channels.
3. Supervised vs. Unsupervised: The supervised, joint-channel approach used here achieves competitive or superior results compared to unsupervised reconstruction methods, particularly for fault families where the CVAE (which relies on single-channel reconstruction error) struggles.

The authors conclude that while the proposed lightweight CNNs offer a robust solution for HVCM anomaly detection, future work could explore spatial/temporal attention mechanisms conditioned on pulse trajectories and graph-based architectures that encode the physical wiring topology of the modulators.