LSTM Autoencoders
for Anomaly Detection

An LSTM autoencoder is a neural network that learns to reconstruct normal time-series data. When it encounters abnormal data — the early signatures of equipment degradation — its reconstruction error spikes. That spike is your anomaly signal, appearing 5 to 24 days before failure.

This guide explains how LSTM autoencoders work for industrial anomaly detection, why they outperform traditional threshold-based monitoring, and what you need to deploy one at your facility.

What Is an Autoencoder?

An autoencoder is a neural network with two parts:

• Encoder: Compresses the input data into a lower-dimensional representation (the "latent space"). This forces the network to learn the most important patterns in the data.
• Decoder: Reconstructs the original input from the compressed representation. If the network has learned the patterns well, the reconstruction should closely match the input.

The key insight: train the autoencoder only on normal data. It learns to reconstruct healthy patterns perfectly. When you feed it abnormal data — degrading equipment — it can't reconstruct what it's never seen before. The reconstruction error spikes. That's your anomaly signal.

Why LSTM (Long Short-Term Memory)?

Standard autoencoders process each input independently. But sensor data is time-series — the current reading depends on what happened before. A vibration reading of 4.2 mm/s might be normal at 8:00 AM but abnormal at 8:05 AM if it was 3.5 mm/s five minutes ago.

LSTM cells are designed for sequential data. They maintain a memory of previous inputs, allowing the network to learn temporal patterns — trends, cycles, and gradual changes over time. This is critical for equipment degradation, which unfolds over days and weeks, not in single readings.

How It Works in Practice

Step 1: Data Collection

Collect sensor data from healthy equipment. You need:

• Vibration data (accelerometers on bearing housing)
• Temperature data (bearing temperature, winding temperature)
• Pressure data (discharge pressure, suction pressure)
• 2 to 4 weeks of continuous baseline data

The sampling rate depends on the equipment type. For rotating equipment, vibration data should be sampled at least 10x the highest frequency of interest (typically 10-25 kHz for bearing fault detection).

Step 2: Preprocessing

• Normalization: Scale all sensor readings to [0, 1] or standardize to zero mean, unit variance
• Windowing: Slice the continuous data stream into overlapping windows (e.g., 60-second windows with 50% overlap)
• Feature extraction (optional): For vibration data, compute frequency-domain features (FFT, spectral kurtosis) alongside time-domain features

Step 3: Training

Train the LSTM autoencoder on the baseline (healthy) data only. The network learns to compress and reconstruct normal patterns. Use mean squared error (MSE) as the loss function — the model minimizes reconstruction error on healthy data.

Architecture typically includes:

• Input layer: windowed sensor data (e.g., 60 timesteps x N sensors)
• LSTM encoder: 1-2 LSTM layers with decreasing units (e.g., 64 → 32 → 16)
• Latent representation: 8-16 dimensions
• LSTM decoder: 1-2 LSTM layers with increasing units (e.g., 16 → 32 → 64)
• Output layer: reconstruction (same shape as input)

Step 4: Threshold Setting

After training, compute the reconstruction error across the entire baseline dataset. Set the anomaly threshold at a percentile of the error distribution — typically the 95th or 99th percentile. Any window with reconstruction error above this threshold is flagged as anomalous.

In our validation, this approach produced zero false alarms on cross-dataset transfer tests — meaning the threshold was robust enough to avoid crying wolf while still catching every real anomaly.

Step 5: Live Monitoring

Stream live sensor data through the trained model in real time. For each window:

• Compute reconstruction error
• If error exceeds threshold, flag as anomaly
• Track anomaly persistence — a single anomalous window might be noise; sustained anomaly across multiple windows indicates real degradation
• Generate alert with: confidence score, affected equipment, time of first detection, and recommended action

Why This Outperforms Threshold-Based Monitoring

What You Need to Deploy One

• Sensors: Vibration and temperature sensors on critical rotating equipment (accelerometers, RTDs/thermocouples)
• Data acquisition: DAQ system or IoT gateway that can stream sensor data to a database or data lake
• Compute: GPU for training (can be cloud-based), CPU sufficient for inference
• Software: Python with TensorFlow/PyTorch, or a platform like VLTHRLAB's that handles the ML pipeline
• Baseline data: 2-4 weeks of healthy equipment operation

Validation Results

Our LSTM autoencoder was validated on 8 public benchmark datasets:

The cross-dataset transfer result is the most significant. It means the model trained on one type of equipment successfully detected anomalies on completely different equipment — different machine, different sensors, different operating conditions. This demonstrates that the learned representation generalizes, which is essential for practical deployment where you can't collect years of failure data for every machine type.

Want to Assess Your Readiness?

Before deploying an LSTM autoencoder, you need the right data infrastructure, sensor coverage, and maintenance processes. Our free readiness assessment scores your facility across 4 dimensions in 3 minutes.

---

VLTHRLAB builds decision intelligence systems for complex operations. This guide is based on research validated on 8 public datasets. Contact us to discuss deployment at your facility.