Pre-trained models have demonstrated exceptional generalization capabilities in time-series forecasting; however, adapting them to evolving data distributions remains a significant challenge. A key hurdle lies in accessing the original training data, as fine-tuning solely on new data often leads to catastrophic forgetting. To address this issue, we propose Replay Tuning (R-Tuning), a novel framework designed for the continual adaptation of pre-trained time-series models. R-Tuning constructs a unified latent space that captures both prior and current task knowledge through a frequency-aware replay strategy. Specifically, it augments model-generated samples via wavelet-based decomposition across multiple frequency bands, generating trend-preserving and fusion-enhanced variants to improve representation diversity and replay efficiency. To further reduce reliance on synthetic samples, R-Tuning introduces a latent consistency constraint that aligns new representations with the prior task space. This constraint guides joint optimization within a compact and semantically coherent latent space, ensuring robust knowledge retention and adaptation. Extensive experimental results demonstrate the superiority of R-Tuning, which reduces MAE and MSE by up to 46.9% and 46.8%, respectively, on new tasks, while preserving prior knowledge with gains of up to 5.7% and 6.0% on old tasks. Notably, under few-shot settings, R-Tuning outperforms all state-of-the-art baselines even when synthetic proxy samples account for only 5% of the new task dataset.

Encoding time series into tokens and using language models for processing has been shown to substantially augment the models' ability to generalize to unseen tasks. However, existing language models for time series forecasting encounter several obstacles, including aliasing distortion and prolonged inference times, primarily due to the limitations of quantization processes and the computational demands of large models. This paper introduces Apollo-Forecast, a novel framework that tackles these challenges with two key innovations: the Anti-Aliasing Quantization Module (AAQM) and the Race Decoding (RD) technique. AAQM adeptly encodes sequences into tokens while mitigating high-frequency noise in the original signals, thus enhancing both signal fidelity and overall quantization efficiency. RD employs a draft model to enable parallel processing and results integration, which markedly accelerates the inference speed for long-term predictions, particularly in large-scale models. Extensive experiments on various real-world datasets show that Apollo-Forecast outperforms state-of-the-art methods by 35.41% and 18.99% in WQL and MASE metrics, respectively, in zero-shot scenarios. Furthermore, our method achieves an acceleration of 1.9X-2.7X in inference speed over the baseline methods.

Deep learning has made significant strides in fault diagnosis when training and testing are conducted within domains of the same distribution. However, data collected from actual industrial equipment often originates from varying load operating conditions, leading to shifts in the manifestation of fault patterns. Additionally, non-stationary signals tend to exhibit similarities or weak fault frequencies in the frequency domain, making them difficult to distinguish. To address these problems, a prototypical normalized output domain adversarial adaptation network (PNOAN) with local and full frequency fusion perception is proposed. Initially, the vibration signal is transformed to time frequency graph by Morlet wavelet. Then, a band pooling module is employed to learn the long-range local context of axial time and frequency in time frequency graph to achieve local perception of fault signal, while a dual-part attention mechanism focuses on both high and low frequency information to achieve full frequency perception. Furthermore, an improved domain adversarial neural network is designed to standardize the conditional alignment of features and prediction probabilities for domain adaption fault diagnosis. Extensive results on a rotating bearing dataset illustrate the effectiveness and superiority of the proposed PNOAN.

Language models have been proven to handle time series data after tokenization and show generalization performance on unseen forecasting tasks. However, existing techniques for tokenizing time series data struggle to eliminate redundant information and noise, which can lead to signal aliasing and cumulative quantization errors, making it difficult to further improve prediction performance. In this paper, we propose an Adaptive Component Extraction and Tokenization (ACET) framework, which includes two key novelties to address these challenges: the Dynamic Component Extraction Module (DCEM) and the Time Series Tokenization Module (TSTM). The DCEM dynamically isolates the principal components from the interference in the original signal, eliminating the need for manual parameter tuning. This not only enhances the accuracy of signal tokenization but also mitigates the adverse effects of high-frequency noise. Then, the TSTM tokenizes continuous time series data while preserving long-term trend features, ensuring that critical information is retained for subsequent forecasting. Extensive cross-domain experiments on various real-world datasets demonstrate that, in zero-shot forecasting scenarios, ACET achieves improvements of 19.27% in WQL and 6.63% in MASE, compared to baseline methods.

Industrial time series prediction (ITSP) is critical to the predictive maintenance system of modern industry. However, time-varying conditions and complex industrial processes cause the distribution drift of industrial time series, raising the difficulty of prediction. This article proposes an ITSP model considering distribution information, namely MDEformer. First, the multivariate dynamic embedding (MDE) is designed to provide the property of the channel-binding dynamic distribution awareness. Specifically, a dynamic mode transition and selection module is adopted to exploit dynamic distribution features of time series, and the bidirectional dynamic residual connection integrates dynamic distribution information into embedding vectors to filter distribution change interference. Then, the vanilla Transformer encoder is used to achieve multivariate prediction. Finally, a generative pretraining and fine-tuning strategy is used to enhance the generalization ability in real production scenarios. Extensive results on a real-world zinc smelting dataset illustrate the superiority of MDEformer.

Industrial time series prediction (ITSP) is an indispensable part of predictive control in modern industry. Recently, supervised deep learning-based methods have provided solutions with sufficient annotated data. However, there is massive unlabeled data with complex temporal features in modern industrial production, resulting in poor performance of these methods. To address this problem, a self-supervised generative pre-trained model with a learnable mask network (SSGPM-LMN) is proposed in this paper. First, the multivariate time series are made into patches channel-independently. Then, these patches are fed into a Transformer encoder with the learnable mask-reconstruction paradigm, drawing mask indices with high temporal features by calculating the cosine similarity in low-dimensional feature space to better learn general representations. Furthermore, a two-step fine-tuning strategy, including linear probing and full fine-tuning, is adopted for various downstream scenarios. Finally, extensive experimental results on case studies of ITSP and transfer learning indicate that our SSGPM-LMN achieves superior performance.

At present, data-driven fault diagnosis methods have made excellent achievements. In industrial scenarios, it is difficult to obtain sufficient amount of fault data, which means intelligent fault diagnosis is often faced with imbalanced data problem. Moreover, the label noise is usually brought due to manual recording errors so as to seriously affect the diagnosis performance. To address these problems, this paper proposed a safe-domain generative adversarial network with Transformer (SDGAN). A safe domain selecting method is used to remove the noisy samples and construct a pure dataset which poses no risk to the training process of GAN. Therefore, GAN is able to generate high-quality minority samples to balance the original dataset. In addition, the Vision Transformer (ViT) is also applied as a classifier to recognize the global information for each fault sample and achieve high diagnostic accuracy. The experimental results show that SDGAN achieves great diagnosis performance on various imbalanced ratios and noise ratios cases. Furthermore, SDGAN outperforms other baseline methods on imbalanced fault diagnosis with label noise, which indicates that the SDGAN can effectively solve real-world industrial problems.