A Physics-Informed Self-Supervised Learning-Based Blind Denoising Method for Distributed Fiber Optic Temperature Sensing Logging Data

To address the reliance of conventional deep learning–based denoising methods on large labeled datasets and long training times, a Physics-informed Self-Supervised Blind Denoising Network (PSSDN) for Distributed Temperature Sensing (DTS) logging data is proposed. The approach eliminates the need for manually generated labels and large-scale annotated datasets by leveraging the spatiotemporal coherence of downhole DTS measurements to suppress incoherent noise from unknown sources and by improving training efficiency via a spatiotemporal downsampling strategy. Multiple physics-constrained loss functions are introduced to enhance the model's ability to capture the true evolution of the temperature field. To accentuate structural differences between signal and noise, an Anisotropic Density Convolution (ADConv) module is designed, which applies non-uniform, physics-informed directional weights to convolutional kernels along the depth and time dimensions, thereby strengthening feature extraction and mitigating background noise interference. Experiments on both synthetic and field DTS datasets demonstrate a 25.92 dB improvement in signal-to-noise ratio. Compared with existing denoising algorithms, PSSDN is more efficient for DTS data processing, requiring only 7.22 s per pass. The method holds significant promise for DTS data processing and well-logging interpretation.