Autoencoder

Autoencoders are trained exclusively on normal operational data—healthy DGA readings, standard thermal profiles, and routine load cycles. This eliminates the need for labeled fault data, which is historically scarce in substation environments. The network learns to compress and reconstruct the high-dimensional feature space of normal behavior, creating a model of the asset's expected state without prior knowledge of failure modes.

The core diagnostic mechanism relies on reconstruction error—the mean squared error (MSE) between the input vector and its reconstructed output. When a transformer begins to develop a fault:

• Thermal faults produce gas ratios outside the learned manifold
• Partial discharge introduces signal patterns the decoder cannot reproduce
• The resulting high reconstruction loss triggers an alert

This continuous anomaly score provides a graded severity indicator rather than a binary classification, enabling trend analysis over time.

The bottleneck layer forces the network to learn a compressed, lower-dimensional representation of transformer state. This latent space captures the essential correlations between:

• Dissolved gas concentrations (H₂, CH₄, C₂H₂, C₂H₄, C₂H₆)
• Load current and ambient temperature
• Oil moisture content and dielectric breakdown voltage

By visualizing the latent space with t-SNE or PCA, asset managers can identify clusters of similar operating regimes and detect subtle shifts preceding failure.

Standard autoencoders produce a deterministic reconstruction. Variational Autoencoders (VAEs) extend this by learning a probability distribution over the latent space. This enables:

• Reconstruction probability instead of raw error—a statistically rigorous anomaly metric
• Sampling from the learned distribution to generate synthetic normal profiles
• Robust handling of sensor noise common in online DGA monitors

VAEs provide a principled framework for setting dynamic alarm thresholds that adapt to seasonal load variations.

Modern transformer monitoring generates heterogeneous data streams. Autoencoders can be designed with multi-modal input layers to fuse:

• Time-series data: Hourly DGA readings and top-oil temperature
• Spectral data: Frequency response analysis (FRA) signatures
• Contextual features: Transformer age, kVA rating, and maintenance history

The shared latent representation captures cross-modal correlations—for example, linking rising acetylene levels with specific FRA deviations to confirm arcing faults.

Once trained, the encoder-decoder architecture is computationally lightweight during inference. The model can be deployed directly on substation edge gateways or intelligent electronic devices (IEDs) supporting IEC 61850. This enables:

• Sub-millisecond anomaly scoring without cloud latency
• Continuous monitoring during communication outages
• Local data privacy—raw DGA readings never leave the substation

Only aggregated anomaly scores and alerts are transmitted to the central SCADA system.

The autoencoder consists of two symmetrical sub-networks. The encoder compresses high-dimensional DGA and thermal sensor data into a low-dimensional latent bottleneck, forcing the model to learn the most salient features of normal operation. The decoder then attempts to reconstruct the original input from this compressed representation. The bottleneck acts as an information filter—only patterns present in the training data survive compression. When a transformer develops a fault, the anomalous gas ratios and temperature profiles cannot be efficiently compressed, resulting in a high reconstruction error that triggers an alert.

The core diagnostic mechanism: the mean squared error (MSE) or mean absolute error (MAE) between the original input vector and the decoder's output. During training on healthy transformer data, the autoencoder minimizes this error. In inference, a spike in reconstruction error indicates the model has encountered a pattern it cannot faithfully reproduce—a deviation from normal behavior. Asset managers set a threshold on this error signal, often using statistical methods like the Mahalanobis distance or percentile-based limits on the training error distribution, to distinguish benign noise from incipient faults.

A variant trained to reconstruct clean inputs from deliberately corrupted versions. During training, Gaussian noise or dropout is applied to the input DGA readings. The model learns to recover the underlying signal, making it robust to sensor drift, calibration errors, and transient noise common in substation environments. For transformer monitoring, a DAE prevents false positives caused by temporary load fluctuations or communication artifacts, focusing the anomaly score on genuine deviations in gas generation rates rather than measurement noise.

Unlike a standard autoencoder that learns a deterministic latent representation, a VAE learns a probability distribution over the latent space. The encoder outputs mean and variance vectors parameterizing a Gaussian distribution, from which the latent vector is sampled. This probabilistic framing provides a more nuanced anomaly metric: the reconstruction probability, which considers both the reconstruction error and the likelihood of the latent variable. For transformer diagnostics, VAEs better capture the stochastic nature of gas diffusion and thermal dynamics, offering superior separation between normal variability and true fault signatures.

A specialized architecture where both encoder and decoder use Long Short-Term Memory (LSTM) layers instead of dense feed-forward layers. This variant captures temporal dependencies in streaming DGA monitor data—critical because the rate of change of gas concentrations (e.g., acetylene rising over hours) is often more diagnostic than absolute values. The model learns the normal trajectory of gas evolution and hot-spot temperature fluctuations. A sudden deviation from the predicted temporal sequence signals a developing fault, enabling earlier detection than static threshold-based alarms.

The compressed bottleneck representation can be projected into 2D or 3D using t-SNE or UMAP for visual diagnostics. In a well-trained model, normal operating states cluster tightly, while fault conditions—arcing, overheating, partial discharge—form distinct, separable clusters. This provides explainability to asset managers: a new data point falling into a known fault cluster offers immediate diagnostic context. Latent space trajectories over time can also visualize the progression of degradation, showing how a transformer's condition drifts from the healthy cluster toward a failure region.