Contrastive Learning vs Autoencoders for Unsupervised Representation Learning of Vibrations

Contrastive learning excels at learning separable, discriminative representations directly from unlabeled vibration data by pulling semantically similar samples (e.g., different augmentations of the same bearing recording) together and pushing dissimilar ones apart. For example, the SimCLR framework applied to the CWRU bearing dataset has demonstrated a 5-8% improvement in downstream fault classification accuracy compared to autoencoder baselines when only 10% of labels are available, because it explicitly models the decision boundary between healthy and faulty operational regimes.

Autoencoders take a different approach by compressing high-dimensional FFT or raw vibration signals into a low-dimensional bottleneck and reconstructing the input. This results in a latent space that captures the dominant variance of the signal, which is excellent for signal denoising and anomaly scoring via reconstruction error. However, the learned features often encode operational conditions like load and speed as much as fault signatures, leading to entangled representations that require more labeled data to separate fault classes effectively.

The key trade-off: If your priority is maximizing downstream fault classification accuracy with minimal labeled data and ensuring clear separation between different fault types (e.g., inner race vs. outer race defects), choose contrastive learning. If you prioritize a simpler, well-understood architecture for unsupervised anomaly detection, data compression, or denoising raw accelerometer streams where reconstruction fidelity is the primary metric, choose an autoencoder. Consider contrastive learning when building a foundation model for a fleet-wide condition monitoring system; choose autoencoders for lightweight, edge-based anomaly scoring on individual assets.

Contrastive Learning excels at learning separable, discriminative features because it explicitly optimizes for instance-level distinction. By pulling representations of augmented views of the same vibration sample together while pushing apart representations from different samples, it naturally disentangles operational regimes (e.g., RPM, load) from fault signatures. For example, in the CWRU bearing dataset, SimCLR-based representations have demonstrated a 5-10% improvement in downstream fault classification F1-score compared to autoencoders when training data is limited to less than 100 labeled samples per class. This is because the model is forced to learn invariances to data augmentation that mimic real-world noise, rather than simply compressing and reconstructing the signal.

Autoencoders take a different approach by learning a compressed latent bottleneck that captures the dominant variance of the vibration signal. This results in a representation that is excellent for signal reconstruction and denoising, but often entangles operational context with fault information. A variational autoencoder (VAE) might achieve a lower reconstruction MSE on normal operating data, making it highly sensitive to any deviation from the norm. However, its latent space is not explicitly structured for classification; a t-SNE visualization of autoencoder features often shows clusters that are separated by operating speed rather than by fault severity, requiring an additional discriminative fine-tuning step.

The key trade-off: If your priority is building a robust fault classifier with very few labels and you need the representation to be invariant to changing operational conditions, choose Contrastive Learning. If you prioritize high-fidelity signal reconstruction for anomaly scoring, denoising raw sensor streams, or generating synthetic vibration data, choose Autoencoders. Consider a hybrid approach where a contrastive objective regularizes an autoencoder if you need both reconstruction capability and a separable latent space.