Raman optical time-domain reflectometers (ROTDRs) are a crucial branch of distributed optical fiber sensing (DOFS) and are widely used for online monitoring of distributed temperature variations. ROTDR relies on temperature-sensitive Raman scattering, but the weak intensity of Raman backscattered light makes it highly susceptible to noise from optical and electrical components, making it challenging to obtain clean scattering signals in practical measurements. To address this issue, machine learning methods have been explored for signal denoising. However, training denoising networks typically requires labeled data, and due to the complexity of noise, discrepancies exist between simulated and real noise. In the absence of prior knowledge about the real noise characteristics, networks trained on artificially generated labels may suffer from poor generalization to real-world data. In this paper, we propose a blind denoising autoencoder (BDAE) to overcome the challenge of training without pure signal labels. A label construction and training strategy is introduced, where noisy data are used to generate input-label training pairs. Comparing to conventional training with simulated data, we find that this approach allows unlabeled Raman scattering data to be directly incorporated into network training, enabling the model to learn real noise characteristics more effectively. Without requiring modifications to the network structure or explicit noise parameter estimation, BDAE can enhance denoising performance, preserve temperature-induced dynamic variations, and mitigate the risk of overfitting.

We leverage the retrievable nature of optical sensing signals to design a strategy for generating input-label data pairs for scattering signal denoising. Unlike conventional denoising autoencoders that require pure signals as labels, this approach only requires two noisy signals sampled from the same underlying pure signal to form input-label training pairs. Two consecutive samplings from a distributed temperature sensing (DTS) system, taken within a short time interval, are assumed to represent the same underlying signal, meeting this requirement. The structure of the convolutional autoencoder used in this paper is illustrated in Fig. 1. During experiments, it is observed that uneven intensity distribution in the distance domain can hinder the network’s ability to capture signal trends. To address this, additional training data (termed labeled data) are generated by applying noise addition to the same set of labeled data, forming input-label pairs suitable for the blind denoising strategy. This ensures a robust evaluation of network performance. The loss functions play a critical role in the training effectiveness of BDAE. The composite loss functions used in this paper consist of three components: mean square error (MSE), energy loss, and total variation loss (TV Loss). By incorporating these components, the loss function encapsulates prior knowledge of noise characteristics in Raman scattering signals, leading to a more targeted and effective denoising performance.

The experiment is designed to validate the network’s denoising performance on dynamic temperature variations under different light source intensities. The sensing fiber outside the DTS is immersed in a water bath to observe temperature trends over time. The temperature changes compared to reference values are illustrated in Fig. 4. The denoising performance of wavelet denoising (WD) and BDAE is compared to fiber temperature scattering signals (Fig. 5). The results demonstrate that BDAE effectively learns noise characteristics from unlabeled Raman scattering data, leading to more efficient noise reduction in scattering signals. Temperature variations over the measurement period are shown in Fig. 6, while Table 1 presents the average deviation of temperature measurements. The denoising model reduces the average temperature error from 2.21 to 1.71 °C, outperforming WD. To further assess the effectiveness of the proposed blind denoising strategy, experiments are conducted under three different signal-to-noise ratio (SNR) conditions: 35 dB, 30 dB, and 25 dB. The convolutional autoencoder optimizes the denoised signals to a level comparable to that achieved with labeled training data, demonstrating the effectiveness of the approach.

In this paper, we propose a blind denoising training strategy based on a convolutional autoencoder for denoising Raman scattering signals in DTS systems. The method effectively addresses the challenge of acquiring pure scattering signals by enabling network training without explicit labels, using noisy data to construct input-label training pairs. The composite loss function, incorporating MSE, energy loss, and TV Loss, further enhances the model's performance. In the experiments, variations in the intensity of the DTS light source are simulated to represent three different noise scenarios, while measuring the dynamic heating and cooling processes. The trained BDAE effectively reduces temperature measurement errors when denoising dynamic scattering signals under these three noise conditions, outperforming WD methods. The experimental results validate that this method achieves training performance comparable to that obtained with labeled data. However, due to the use of convolutional layers in autoencoder-based denoising, excessive smoothing at signal inflection points is observed, which slightly influences the spatial resolution of the measurement results. Future research will focus on improving this aspect.