Comparative Analysis

Model Comparison

4 approaches tested on 87,813 industrial IoT datapoints

Architecture

CNN-Transformer Autoencoder

2,097,148 parameters

MLP Autoencoder

Baseline

Val Loss

0.6261

Feed-forward AE. Flattens temporal data, no time awareness. Good baseline but overfits with 25M params.

Params

24.9M

Epochs

✓ Advantages

· Fast training

· Low val loss

✗ Limitations

· No temporal patterns

· Overfitting risk

LSTM Autoencoder

RNN Classic

Val Loss

0.7904

Recurrent encoder-decoder with LSTM cells. Classic time-series approach, compact but sequential.

Params

600K

Epochs

✓ Advantages

· Temporal awareness

· Compact architecture

✗ Limitations

· Vanishing gradients

· Sequential processing

CNN-Transformer AE

Best ⭐

Val Loss

0.7073

Hybrid: CNN extracts local patterns, Transformer captures global dependencies via self-attention.

Params

2.1M

Epochs

✓ Advantages

· Local + global patterns

· Parallelizable

✗ Limitations

· Higher memory

· Requires GPU

Isolation Forest

Non-DL

Val Loss

0.4728

Tree-based outlier detection. No neural network — fast but can't learn complex temporal patterns.

Params

N/A

Epochs

—

✓ Advantages

· No GPU needed

· Interpretable

✗ Limitations

· No temporal learning

· Less sensitive

Multi-Model Dashboard

Latent Space (t-SNE)

Dimensionality reduction of encoder outputs — visualizing learned representations

GPU Compute

Training Pipeline

NVIDIA RTX 4070 Ti SUPER · CUDA 12.8