U-net-based Segmentation Of Sensor Data For Detecting Important Events In Oil Wells

Detecting key events in oil well sensor data is essential for optimizing operations and preventing failures. This study presents a non-parametric, U-Net-based deep learning approach for segmenting time-series sensor data to identify significant changes. Unlike analytical methods that require manual tuning or machine learning models dependent on fixed thresholds and frequent retraining due to concept drift, our model adapts automatically to different wells without the need for manual adjustment. We modify the U-Net architecture to handle time-series segmentation, enabling automatic detection of sensor transitions without predefined thresholds. Trained on both synthetic and real-world data, the model is evaluated across diverse operational conditions. The segmented output supports downstream classification models in determining whether detected events correspond to regular human operations or previously unnoticed anomalies. To enhance reliability, we incorporate wavelet-based denoising, which effectively removes noise while preserving critical transitions. By combining U-Net segmentation with wavelet-based denoising, this approach offers a novel solution for event detection in industrial time-series data, advancing automation and anomaly detection in oilfield operations.