The time series data in modern industrial production have multiple working conditions and long-term drift of key parameters, which leads to the lack of generalization ability of time series prediction models and the continuous decline of prediction performance. Therefore, this paper proposes a two-stage multi-source adversarial fusion and Maximum Mean Discrepancy guided adapter injection method (MAFGAI), which can take into account the situation of multi-condition information fusion and long-term distribution shift adaptation, and solve the engineering problem that traditional methods are difficult to deal with multiple domain differences. In the first stage of the method, the domain confrontation strategy was used to realize the cross-domain fusion of multi-condition information, and an adaptive gradient controller was designed to dynamically adjust the confrontation intensity, so as to construct a pre-trained model with strong generalization ability. In the second stage, the method accurately identified the network layer with large distribution shift in the target domain, and injected lightweight time adapters for targeted training to correct the distribution shift generated by the layer and improve the prediction effect on the target domain. In this paper, comparative experiments are carried out on the ETT industrial time series dataset and the grinding process dataset, and all models and experiments are based on the PyTorch framework. On the three target domains of the ETT dataset, the three indicators are increased by 2.54%, 0.87%, and 0.74% on average compared with the optimal comparison model. On the three target domains of the grinding process dataset, the three indicators are increased by 16.26%, 18.01%, and 18.01% on average. The results show that the prediction performance of the proposed MAFGAI method is better than the existing mainstream transfer learning methods on the two datasets.

Long-term prediction of key parameters is essential for the control and optimization of industrial processes with time-delay characteristics. Such prediction remains challenging as industrial time series typically exhibit complex nonlinearity, sparse information, imprecision, and nonstationarity due to complex processes dynamics and environmental instability. Existing modeling paradigms require enhancement to effectively decouple confounding factors and adapt to unknown uncertainties. To this end, we propose a multifrequency decoupling network with dual-stage shift segmental modeling (DS3-MFDNet). The network begins by decomposing the industrial data into distinct frequency components to isolate and manage its complex nonlinearity. Then, we propose a dual-stage shift segmental modeling method for mid-frequency component. This method effectively reduces information sparsity and mitigates the impact of random noise by leveraging contextual information. Furthermore, a position-weighted directional alignment loss is designed to facilitate the training process during the second shift stage. Finally, we design a reversible dual-band condition adaptive normalization mechanism to address the data nonstationarity. Across different prediction lengths, DS3-MFDNet achieves an average improvement of 3–4% over the best baseline on the industrial power distribution process dataset and 6%–8% on the practical loosening and conditioning process dataset, demonstrating the effectiveness of the proposed method.

Compared with the optimization technology of retraining models, fine-tuning pretrained neural networks have been widely used in industrial process monitoring because of their high precision and low training cost. However, data privacy protection in industry makes it impossible to improve the upstream pertaining model to avoid negative transfer when the target domain and the source domain are too different or the pertaining model is maliciously attacked. To solve the above problem, we proposed a fine-tuning method assisted by target domain knowledge for industrial time-lag parameter prediction. This method uses an Auto-Encoder to learn the representation of black-box and time-lag knowledge in the target domain. Then, the black-box and time-lag knowledge are used as auxiliary information to update the high-level weights of the pretrained network. At the same time, we proposed an auxiliary learning method that can dynamically update weights without introducing new parameters and provide training methods for different neural network optimizers. The experimental results on the heating furnace temperature prediction and wind condition prediction of wind farms demonstrate that the prediction performance can be effectively improved, and the defective inheritance of the pretrained model can be effectively reduced to avoid negative transfer.Note to Practitioners—Due to the complexity of industrial processes, the detection of key parameters is time-consuming and it is difficult to obtain labeled samples. With the widespread use of large models in computer vision and natural language processing, fine-tuning pretrained models can alleviate these challenges. However, the problem that comes with it is safety. The industrial prediction model cannot be built when the publicly available pretrained model is attacked. Moreover, the pretrained model could not be modified because the pretrained data could not be obtained. Therefore, this paper used the time-lag prior knowledge of industrial data to modify the pretrained model in the fine-tuning stage, and helps improve the pretrained accuracy and model robustness by mining the knowledge of existing data. Two groups of experiments show that the auxiliary knowledge-based fine-tuning mechanism can significantly improve the prediction accuracy and alleviate the defect inheritance of the pretrained model, which is valuable for the application of large models in the field of industrial parameter prediction.

Deep learning models have gained significant attention and application in recent years to improve the accuracy and efficiency of industrial time series prediction. However, the dynamic changes in industrial processes present a key challenge for data-driven models. Specifically, the performance of deployed models deteriorates over time and fails to adapt to new operating conditions. Currently, two common update methods exist: Retraining the model using historical and new operating data, which incurs high computation and storage costs, or incrementally fine-tuning the model solely using new data, which leads to catastrophic forgetting of learned patterns. To address these issues, this article proposes an adaptive continual learning method for nonstationary industrial time series prediction. Our approach tackles the problems by hint-based network parameter learning to retain the dark knowledge from previous tasks and avoid catastrophic forgetting of accumulated knowledge. In addition, we design a soft buffer to aid memory and learning of key patterns under the current operating condition. Lastly, a time-sensitive activation function is proposed to endow the neural network with time-evolving properties, thereby enhancing the model's generalization ability. Compared with other update methods and different continual learning methods, the superiority of our method is validated on solar power generation data and real data of grinding and grading process.

With the advancement of technology, the field of multi-channel time series forecasting has emerged as a focal point of research. In this context, spatio-temporal graph neural networks have attracted significant interest due to their outstanding performance. An established approach involves integrating graph convolutional networks into recurrent neural networks. However, this approach faces difficulties in capturing dynamic spatial correlations and discerning the correlation of multi-channel time series signals. Another major problem is that the discrete time interval of recurrent neural networks limits the accuracy of spatio-temporal prediction. To address these challenges, we propose a continuous spatio-temporal framework, termed Recurrent Spatio-Temporal Graph Neural Network based on Latent Time Graph (RST-LTG). RST-LTG incorporates adaptive graph convolution networks with a time embedding generator to construct a latent time graph, which subtly captures evolving spatial characteristics by aggregating spatial information across multiple time steps. Additionally, to improve the accuracy of continuous time modeling, we introduce a gate enhanced neural ordinary differential equation that effectively integrates information across multiple scales. Empirical results on four publicly available datasets demonstrate that the RST-LTG model outperforms 19 competing methods in terms of accuracy.

Forecasting tasks involving multi-channel time series data pervade numerous practical applications and have attracted significant attention. Spatio-temporal graph neural network models for multi-channel time series forecasting have recently gained traction, owing to their ability in capturing both spatial and temporal features. A common practice is the integration of graph convolutional networks with recurrent neural networks. However, the discrete intervals of recurrent neural networks pose limitations on the temporal resolution of time series forecasting, impeding the model's ability to capture subtle changes in the data. To address this challenge, we introduce a continuous spatio-temporal framework, termed Graph Ordinary Differential Equation Recurrent Network (GODERN). GODERN incorporates continuous recurrent neural networks with a learnable and directed graph convolution layer to model the spatio-temporal dynamics in latent space. Furthermore, given the actual time representation in GODERN, we propose a novel augmented method of neural ordinary differential equation with fast-slow dynamics, thus allowing the encapsulation of multi-scale information. Through our experiments, we demonstrate that GODERN achieves superior accuracy on four real-life datasets, outperforming 13 baseline models. The code is available athttps://github.com/Fei-u/GODERN.

Industrial time series, as a kind of data that responds to production process information, can be analyzed and predicted for effective monitoring of industrial production processes. There are problems of data shortage and algorithm cold start in industrial modeling process caused by complex working conditions, change of data acquisition environment, and short running time of equipment. As a result, the accuracy of the existing data-driven industrial time series prediction algorithm is greatly limited. To address the aforementioned problems, we propose a new time series prediction method for industrial processes under limited data based on dynamic transfer learning in this work. This method aims to effectively use historical data of similar equipment or working conditions rather than discard them to help establish an industrial time series prediction model with limited target data. In this method, first, historical data are divided into multiple batches, and then a new multisource transfer learning framework with dynamic maximum mean difference loss is established according to the distribution distance between each batch of historical data and the limited target data at the current moment. The framework also combines multitask learning methods to establish multistep prediction model for online learning in industrial processes. Compared with other commonly used methods, experiments on two real-world datasets of solar power generation prediction and heating furnace temperature prediction demonstrate the effectiveness of the proposed method.

Ocean temperature prediction has always occupied an important position in the research of ocean-related fields. The current studies are mostly based on the temperature of the sea surface, but the prediction of ocean internal temperature is more important in practical applications. At present, most of the research studies on the prediction of ocean internal temperature are based on time series, few of which consider the dual characteristics of time and space. Therefore, the accuracy is insufficient, especially for the prediction of thermocline and deep-sea locations. This letter proposes the stereo spatial and temporal 4-D convolution model (SST-4D-CNN) to predict the temperature in the ocean, which fully considers the dual characteristics of time series and oceanic spatial relationship to improve the prediction accuracy. The model includes 4-D convolution module, residual module and recalibration module to predict the horizontal and profile temperature changes from the sea surface to 2000-m underwater. In this letter, the prediction experiment is carried out using the real-time analysis data-temperature dataset from National Marine Data Center. The results show that the accuracy of this method in horizontal and profile prediction is above 98.02%, and most of them are more than 99%.

The electro-fused magnesia furnace (EFMF) has complex characteristics, such as strong nonlinearity and multimodes. In this paper, the between-mode process modeling and monitoring method of the EFMF is proposed. In the original methods, the data are handled in a single mode matrices, the influence from one mode to another tends to be ignored. However, the hidden effect could be useful in process analysis and control. New method is proposed for between-mode part to establish an integrated monitoring system, which would simplify the monitoring model structure and enhance its robustness. The manifold is learned to extract the common part of between-mode transition and the monitoring performance of the between-mode is significantly improved. From the between-mode viewpoint, the multimodes processes behaviors are separated into two subspaces. In the common subspace, the underlying process-relevant variation stays invariable, showing the common contribution to multimodes processes. The specific subspace changes with the alternation of modes and has the different influences on multimodes processes modeling and monitoring. Based on subspace separation, process information is captured across modes and between-mode transition regions are distinguished from two modes. Two modes and between-mode transition models are developed respectively for multimodes processes monitoring. Experiment results show effectiveness of the proposed method.