Semi-Supervised Approach for Early Stuck Sign Detection in Drilling Operations

Introduction

Drilling operations in the oil and gas industry face numerous challenges, one of which is the risk of a stuck pipe. Stuck pipe scenarios can cause severe operational delays and financial losses – in the worst cases, forcing wellbore abandonment. Traditional supervised methods for early stuck pipe detection rely on labeled data, which can be sparse or uncertain. This paper proposes a semi-supervised methodology that focuses on identifying anomalous “early signs” of stuck pipe before the event actually occurs, using minimal assumptions and extensive normal-operation data.

By adopting a depth-domain representation of drilling parameters, this approach captures the localized behavior of drilling data in a given well. The paper demonstrates how autoencoders (AEs) and variational autoencoders (VAEs), trained exclusively on normal data, effectively spot anomalous signals (potentially indicative of a stuck pipe) by comparing real-time values to learned “normal” patterns.

Depth-Domain Representation

Instead of using a simple time-series format, the paper converts drilling parameter data into two-dimensional histograms based on depth. Each histogram has a vertical axis corresponding to “localized” bit depth ranges, and a horizontal axis for one of the drilling parameters (e.g., hook load, top drive torque, rotational speed, standpipe pressure).

Two histograms are generated for each data clip:

Signal Histogram: captures the most recent 2 hours of data.

History Histogram: captures data from 7 days to 2 hours before the current time.

Semi-Supervised Learning Framework

Normal vs. Anomalous Data

A key insight is that stuck pipe incidents are rare and diverse, making them difficult to label accurately. Instead, the methodology trains a model on “normal” segments that are safely removed (in time) from stuck events. The model thus learns to reconstruct typical depth-histogram signatures. During evaluation, large reconstruction errors suggest abnormal drilling conditions – potentially early stuck signs.

Autoencoders (AE) and Variational Autoencoders (VAE)

Both AEs and VAEs map the input histograms (signal + history) to a lower-dimensional latent representation and then attempt to reconstruct the original input.

Autoencoder (AE): Learns a deterministic encoding and decoding function to minimize the distance between original and reconstructed histograms.
Variational Autoencoder (VAE): Applies a probabilistic extension to AE, creating a continuous latent space that enforces smooth variations in the encoding.

Reconstruction Error and Anomaly Detection

Once the model is trained, each new histogram is reconstructed. The difference between original and reconstructed histograms is the reconstruction error. A high error indicates that the new data deviates from “learned normal” patterns. Summaries of these errors over a chosen time window (e.g., up to 6 hours) form a stuck score. If the score surpasses a threshold, an early stuck alert is triggered.

Results

The dataset consisted of actual drilling logs from multiple wells, including 43 stuck events. After filtering and labeling normal windows, 68,345 normal instances were used for training. The models were then tested on eight distinct stuck cases. Key findings:

Detection Rate: The proposed semi-supervised models (AEs and VAEs) detected early signs for all eight stuck cases. In contrast, a supervised baseline from prior research only caught five of the eight.

Hyperparameter Tuning: Optimal kernel sizes, loss functions, and time windows varied by stuck mechanism. VAEs commonly performed better in the presence of more subtle or varied anomalies.

Individual Parameters: Interestingly, single-parameter models (e.g., only using Hook Load histograms) were sufficient for some incidents, indicating that one strongly deviant parameter can provide ample early warning.

Early vs. Late Alerts: In certain cases, anomalies appeared hours prior to the actual pipe sticking event. This can lead to “false positives” in a purely supervised sense, but field observations suggest such long-horizon alerts can be valid “early signs” of deteriorating conditions.

Discussion

The results emphasize the robustness and adaptability of a semi-supervised approach in drilling anomaly detection. Because it only needs normal-operation logs, the method bypasses many inaccuracies associated with labeling stuck events or uncertain early signs. A trade-off is the potential for heightened false alarms, which in practice might be mitigated by integrating additional context or domain knowledge.

Hyperparameter selection (e.g., kernel size, choice of reconstruction loss, VAE vs. AE) remains crucial. Results show that a single kernel size or single reconstruction error function may not work best for all stuck mechanisms, suggesting a multi-model or ensemble strategy. Longer time windows often produce more robust detection of slow-evolving anomalies, but they can also conflate other operational transitions and increase false positives.

Conclusion

This paper introduced a semi-supervised pipeline for early stuck detection in drilling operations. By training autoencoders (AEs) and variational autoencoders (VAEs) solely on normal segments, the system flags unusual behavior as high reconstruction error. Depth-domain histograms of drilling parameters enhance localized data comparisons, crucial for capturing subtle shifts in drilling conditions.

The approach effectively detected all stuck events in the test dataset, outperforming a previous supervised method that only caught 63% of cases. Future research avenues include refining kernel sizes, leveraging additional drilling parameters, and further reducing false positives by merging multiple parameter-based models into an ensemble for heightened reliability.

In conclusion, transitioning to a semi-supervised model for stuck pipe detection shows promise for improving drilling safety, minimizing non-productive time, and reducing operational costs.