💻 computer science

Anomaly Detection for Electro-Hydrostatic Actuators using LSTM Autoencoder

This study proposes an offline anomaly detection framework for Electro-Hydrostatic Actuators (EHAs) using a reconstruction-based LSTM autoencoder that achieves high accuracy and low false-alarm rates in identifying sensor anomalies in temperature and pressure data, outperforming conventional statistical and machine learning methods by effectively capturing temporal dependencies.

Original authors: Nehal Afifi, Abdelmonem Elhendawi, Felix Leitenberger, Nadine Piat, Sven Matthiesen

Published 2026-06-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nehal Afifi, Abdelmonem Elhendawi, Felix Leitenberger, Nadine Piat, Sven Matthiesen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Smart Watch" for Machine Parts

Imagine an Electro-Hydrostatic Actuator (EHA) as the muscle of a heavy-duty machine, like the hydraulic arm on a construction crane or a flight control surface on an airplane. These machines are powerful and compact, but they work hard. To keep them safe, they are covered in sensors (like thermometers and pressure gauges) that constantly shout out numbers: "I'm at 40 degrees!" or "Pressure is 200 PSI!"

The problem is that these sensors talk very fast and very loudly. They produce a massive stream of data. If a sensor starts acting weird (an "anomaly"), it could mean the machine is about to break. But because the data is so fast and complex, old-school methods of checking for problems often miss the signs or scream "Help!" when everything is actually fine (false alarms).

This paper introduces a new, smarter way to listen to these machines using a type of Artificial Intelligence called an LSTM Autoencoder.

The Old Way: The "Rulebook" vs. The "Story"

The authors explain that traditional methods (like simple math rules or basic clustering) are like a rulebook.

How it works: They say, "If the temperature goes above 50 degrees, it's a problem."
The flaw: Machines are dynamic. Sometimes the temperature spikes for a split second because the machine is working hard, not because it's broken. The rulebook doesn't understand the story or the rhythm of the machine. It just sees a number that is too high and panics. This leads to many false alarms.

The New Way: The "Memory-Keeping Detective"

The authors propose a new method using a Long Short-Term Memory (LSTM) Autoencoder. Think of this as a detective with a perfect memory who has studied the machine for a long time.

The Training Phase (Learning the Normal Rhythm):
The detective is shown only videos of the machine working perfectly (normal data). The detective learns the machine's natural rhythm. They learn that "Usually, when the pressure goes up, the temperature follows slowly after a second." They memorize the "song" the machine sings when it is healthy.
The Autoencoder (The "Copycat" Game):
The core of this system is an Autoencoder. Imagine a game where the detective tries to replay the machine's signal from memory.
- If the machine is healthy, the detective can replay the signal perfectly because they know the pattern.
- If the machine is broken or a sensor is glitching, the signal will sound "off-key." The detective tries to replay it, but the result is messy and inaccurate.
The Score (The "Reconstruction Error"):
The system measures the difference between the real signal and the replayed signal.
- Small difference: "This sounds normal. No alarm."
- Big difference: "This sounds weird! The machine is doing something I've never seen before. ALARM!"

What They Did (The Experiment)

The researchers built a test bench (a mini-machine lab) to test this idea.

The Setup: They ran the machine under different loads (like carrying a heavy weight or pushing against a spring).
The Test: They intentionally broke things or injected "faults" (like sudden spikes in temperature or pressure) to see if their detective could catch them.
The Comparison: They compared their "Memory Detective" against the old "Rulebook" methods and other basic AI models.

The Results: A Clear Winner

The results were impressive. The LSTM Autoencoder was like a sharp-eyed supervisor:

Accuracy: It was right about 99% of the time.
False Alarms: It almost never cried "Wolf" when there was no wolf (very low false alarms).
Sensitivity: It caught almost all the actual problems (high recall).

In contrast, the old methods were terrible at catching the actual faults (low recall) or they flagged everything as broken (low precision). The new method understood the context and the timing of the signals, not just the raw numbers.

The Limitations (What They Didn't Do)

The authors are honest about what this study didn't do:

Offline Only: Right now, this detective works in "slow motion." It analyzes data after the machine has stopped. It cannot yet listen to the machine while it is running in real-time (though they plan to fix this next).
One at a Time: The detective only listens to one sensor at a time (like just the thermometer). It doesn't yet know how the thermometer and the pressure gauge talk to each other.
Simple Faults: The "faults" they tested were obvious glitches. They didn't test for slow, creeping degradation that happens over years.

Summary

This paper shows that teaching an AI to learn the "normal rhythm" of a machine and then listening for "off-key" notes is a much better way to find problems than just setting simple number limits. It's like hiring a musician to listen to a band: they know exactly when a note is out of tune, whereas a simple volume meter just screams when the music gets loud.

The authors conclude that this method is highly effective for checking machine health, and their next step is to make the detective fast enough to work in real-time while the machine is running.

Technical Summary: Anomaly Detection for Electro-Hydrostatic Actuators using LSTM Autoencoder

Problem Statement
Electro-Hydrostatic Actuators (EHAs) are critical components in aerospace and industrial systems, requiring reliable monitoring to ensure safety. However, the high sampling frequencies and large volumes of time-series data generated by EHA sensors pose significant challenges for anomaly detection. Conventional statistical methods (e.g., Z-score, Interquartile Range, Median Absolute Deviation) and classical machine learning approaches (e.g., Isolation Forest, Gaussian Mixture, k-means) often fail to capture the complex temporal dependencies and nonlinear dynamics inherent in EHA signals. These limitations result in reduced detection accuracy and elevated false-alarm rates. Furthermore, systematic evaluations of data-driven anomaly detection specifically for EHA systems under varying operational conditions remain scarce, particularly given the scarcity of labeled fault data in practical settings.

Methodology
The study proposes an offline, reconstruction-based anomaly detection framework tailored for univariate EHA sensor signals, specifically focusing on temperature and pressure data. The core of the approach is a Long Short-Term Memory (LSTM) Autoencoder.

Architecture: The model utilizes an LSTM encoder-decoder structure designed to capture long-range temporal dependencies. The encoder maps input sequences to a latent representation, while the decoder reconstructs the original sequence.
Training Strategy: The model is trained exclusively on nominal (fault-free) data. The objective is to minimize the reconstruction loss, defined as the Mean Absolute Error (MAE), between the input and the reconstructed output. MAE was selected over squared-error loss to preserve physical units and improve stability against occasional large spikes.
Anomaly Scoring: Anomalies are detected by calculating the reconstruction error for each sliding window of the time series. A decision threshold ( $\tau$ ) is calibrated based on the distribution of reconstruction errors from the nominal validation set (e.g., using a high quantile or conservative maximum-error rule).
Data Preprocessing: The workflow involves normalizing signals using statistics from nominal training data, segmenting the data into overlapping windows, and applying a sigma-based filtering rule ( $\mu + k\sigma$ ) to ensure the training set is free of outliers. The dataset was split 70/30 for training and testing, with the test set containing both nominal and fault-injected segments.

Key Contributions

Framework Development: The paper presents a specific offline framework for univariate EHA sensor anomaly detection that leverages the temporal modeling capabilities of LSTM autoencoders, addressing the gap in systematic studies for this specific actuator type.
Reliability-Oriented Evaluation: The study provides a rigorous experimental evaluation using high-resolution data from a controlled EHA test bench. This includes fault-injection scenarios across different load conditions (inertial and spring loads) and operating regimes.
Benchmarking: The proposed LSTM-AE is benchmarked against a wide range of baselines, including statistical methods, classical machine learning algorithms, and other deep learning architectures (standard Autoencoders and CNN-AEs).
Performance Analysis: The work demonstrates that modeling temporal dependencies significantly improves the balance between sensitivity (recall) and false-alarm suppression (precision) compared to non-temporal baselines.

Results
The LSTM-AE framework demonstrated superior performance across multiple metrics compared to the baselines:

Accuracy: Achieved an average accuracy of 99.0% (ranging from 99.5% to 99.9% in specific comparisons).
Precision: Reached up to 100% (ranging 96%–100%), indicating a very low false-alarm rate.
Recall: Achieved between 90.2% and 99.6%, showing high sensitivity to actual faults.
F1-Score: Ranged from 93.1% to 99.8%.
ROC-AUC: Values were close to unity (0.95–0.99), indicating strong separability between nominal and anomalous states.

In contrast, statistical and classical machine learning methods achieved high nominal accuracy but exhibited near-zero precision and F1-scores due to class imbalance and a lack of temporal modeling. Qualitative analysis confirmed that the LSTM-AE detects coherent anomalous segments rather than isolated points, effectively suppressing spurious alarms caused by normal fluctuations.

Significance and Limitations
The paper claims that the proposed reconstruction-based LSTM autoencoder is a feasible and effective solution for offline anomaly detection in EHA systems. It highlights that capturing temporal structure is essential for distinguishing true faults from normal operational variations, offering a practical pipeline for condition monitoring where labeled fault data is unavailable.

However, the authors maintain a modest stance regarding the current scope and future implications:

Offline Nature: The current framework is strictly offline; future work is required to adapt it for online, real-time environments with minimal latency.
Labeling Constraints: The evaluation labels are derived from a transparent sigma-rule ( $\mu + k\sigma$ ), which may not fully reflect fault criticality or slow degradation processes that remain within nominal bounds.
Univariate Scope: The analysis is limited to univariate signals, meaning it does not currently exploit cross-channel dependencies between sensors.
Threshold Sensitivity: Performance relies on threshold calibration, which may require retuning under significant regime shifts or sensor drift.

The study concludes that while the method shows high potential for offline EHA monitoring, extending the approach to multivariate sensor fusion, real-time deployment, and broader fault coverage remains necessary for robust industrial application.