Engine Sound Anomaly Detection with Autoencoders

Download the MIMII dataset (Malfunctioning Industrial Machine Investigation and Inspection) from Zenodo. It contains recordings of four machine types (fans, pumps, sliders, valves) under normal and anomalous conditions at multiple SNR levels. For this project, use the fan or pump subset — these are most analogous to turbomachinery. Alternatively, use the ToyADMOS dataset or collect your own recordings from a running engine or motor.

Split the normal recordings into train (80%) and validation (20%). Set aside all anomalous recordings for testing only — the model must never see anomalous data during training. This is the core constraint of anomaly detection: you learn normalcy, not faults.

Segment each recording into fixed-length clips (e.g., 2 seconds). For each clip, compute a 128-band mel-spectrogram using librosa: librosa.feature.melspectrogram(y=audio, sr=16000, n_mels=128, n_fft=1024, hop_length=512). Convert to log scale (dB) and normalize to [0, 1] range.

The result is a 2D image-like representation (128 frequency bins × T time frames) for each clip. Visualize several spectrograms from normal and anomalous samples side by side. Anomalies often appear as unusual spectral peaks, broadband energy changes, or periodic patterns at unexpected frequencies. These visual differences are what the autoencoder's reconstruction error will capture automatically.

Build a symmetric encoder-decoder architecture in PyTorch. The encoder: 4 convolutional layers with increasing channel counts (1→32→64→128→256), kernel size 3×3, stride 2, with ReLU and batch normalization after each. This progressively compresses the spectrogram to a small latent representation. The decoder: 4 transposed convolutional layers that mirror the encoder, reconstructing the spectrogram from the bottleneck.

The bottleneck dimensionality is a critical design choice. Too large, and the autoencoder memorizes everything (including anomalies). Too small, and it cannot reconstruct normal sounds well enough to distinguish them from anomalies. Start with a bottleneck that compresses the input by 16–32×. Use sigmoid on the output layer to match the [0, 1] normalized spectrograms.

Train the autoencoder using MSE reconstruction loss on the training set (normal data only). Use Adam optimizer with learning rate 1e-3 and a cosine annealing schedule. Train for 100–200 epochs with batch size 32. Monitor the validation loss (also computed on normal data) for overfitting.

After training, compute the reconstruction error (MSE per spectrogram) for all training samples and all validation samples. Plot the distribution of reconstruction errors — it should be a tight, roughly Gaussian distribution centered at a low value. This distribution is your model of "normal" — anything with reconstruction error in the tail is potentially anomalous.

Run the trained autoencoder on the test set containing both normal and anomalous samples. Compute reconstruction error for each sample. Plot the error distributions for normal and anomalous samples separately — the anomalous distribution should be shifted to higher errors, though overlap is common.

Select a threshold using the validation set error distribution: set the threshold at the 95th or 99th percentile of normal validation errors. Compute precision, recall, and F1 score at this threshold. Plot the full ROC curve and precision-recall curve by sweeping the threshold. Report the AUC-ROC — values above 0.85 are good; above 0.95 is excellent. The PR curve is often more informative because the class balance is skewed.

For several correctly detected anomalies, plot three images side by side: the input spectrogram, the reconstructed spectrogram, and the residual (absolute difference). The residual map highlights exactly which time-frequency regions the autoencoder failed to reconstruct — these are the spectral signatures of the anomaly.

This visualization is powerful for engineering interpretation: a bearing fault might appear as elevated energy in a specific frequency band, while a blade rub might show broadband impulse events at the rotation frequency. The autoencoder doesn't need to know what these faults are — it just needs to know they're not normal. Compare the residual patterns across different anomaly types to see if the autoencoder implicitly provides diagnostic information.

Test the model's robustness to varying signal-to-noise ratios. The MIMII dataset provides recordings at different SNR levels (6 dB, 0 dB, -6 dB). Evaluate detection performance at each SNR — performance typically degrades gracefully but may collapse below a critical SNR where background noise masks the anomaly signature.

Discuss the architecture's suitability for real-time deployment: measure inference time per spectrogram, estimate the computational requirements for continuous monitoring, and consider how you would handle concept drift (gradual changes in "normal" as the engine ages). A well-engineered anomaly detection system periodically retrains or uses adaptive thresholds to account for slow baseline shifts. Write a comprehensive report covering architecture decisions, training methodology, threshold selection rationale, and deployment considerations.