Detecting anomalies in wire-cutting torque time-series data using distance-based clustering, Dynamic Time Warping, FastDTW, and reconstruction-based deep learning with VAE-LSTM.
This project focused on predictive maintenance for a wire-cutting process in an industrial manufacturing setting. The main goal was to analyze torque measurements over time and identify unusual cutting behavior that may indicate failures or maintenance needs.
The project compared multiple unsupervised anomaly detection approaches for time-series data, including distance-based clustering with Dynamic Time Warping and reconstruction-based anomaly detection with a VAE-LSTM model.
Wire cutting is an essential manufacturing process where quality and precision directly affect the reliability of final components. By monitoring the torque signal during the cutting process, abnormal behavior can be detected before it becomes a larger production or quality issue.
Main objective: detect anomalies and support predictive maintenance in the wire-cutting process using torque/time measurements.
The cutter process includes a stator after wire cutting, a wire-cutting station, and a cutting knife. During the process, torque values are recorded over time to capture the mechanical behavior of the cut.
Each cutting cycle contains two rotations. Each rotation lasts approximately 1.5 seconds for a full 360° rotation. The first 360° rotation receives special attention because it corresponds to the main cutting process.
The data were collected from the wire-cutting process, stored in a database, and exported in JSON format. The key measured variable was torque, recorded at high frequency with timestamps.
The raw JSON files were consolidated into a single dataframe to make the data easier to analyze, visualize, and model. Unique identifiers were added so that angle and torque values belonging to the same timestamp could be handled separately and reshaped for efficient analysis.
The cleaned dataframe was saved in Parquet format to improve storage efficiency and enable faster loading for large-scale time-series analysis.
The project explored both distance-based and reconstruction-based anomaly detection methods. This was useful because industrial time-series anomalies can appear either as shape differences between cutting cycles or as high reconstruction errors in a learned sequence model.
| Method Group | Method | Purpose |
|---|---|---|
| Distance-Based | Time Series K-Means with DTW | Cluster torque curves based on temporal shape similarity. |
| Distance-Based | Time Series K-Means with FastDTW | Reduce the computational cost of DTW-based similarity comparison. |
| Reconstruction-Based | VAE-LSTM | Learn normal sequence structure and detect anomalies by reconstruction error. |
| Exploratory | K-Means, DBSCAN, GMM, Autoencoder | Initial experiments to compare clustering and representation learning approaches. |
Dynamic Time Warping was used because it compares time-series curves based on their shape, even when patterns are shifted or stretched in time. This is useful for torque signals, where the exact timing may vary between cycles while the overall shape remains meaningful.
The VAE-LSTM model was used as a reconstruction-based anomaly detector. The model learns to reconstruct normal torque sequences. If a sequence has a high reconstruction error, it is treated as potentially anomalous.
The VAE-LSTM model detected anomalies using reconstruction error. A 98th percentile threshold was used to identify high-error sequences as anomalous. This approach produced a compact set of anomaly candidates for inspection.
The anomaly detection threshold was set to approximately 0.02536 based on the 98th percentile of reconstruction errors. Sequences above this threshold were flagged as anomalies.
Distance-based clustering helped group similar torque curves and highlight unusual curve shapes. The VAE-LSTM reconstruction approach was useful for detecting deviations from learned normal behavior, especially where anomalies appear as reconstruction-error spikes.
The results suggest that reconstruction-based anomaly detection is a strong direction for predictive maintenance, while DTW-based clustering remains valuable for understanding shape-based groups and visually inspecting abnormal cutting patterns.
The main limitation is that anomaly detection was mostly unsupervised, meaning detected anomalies still need expert validation. Another limitation is that the project identified anomalous sequences but did not yet localize the exact time window inside each sequence where the anomaly occurs.
This project strengthened my experience with industrial time-series data, predictive maintenance, anomaly detection, sequence modelling, and data preparation for large-scale sensor datasets. It also showed the value of combining classical time-series distance methods with deep learning reconstruction models.