16 November 2021 in IEEE Sensors Journal
Challenges of process industry, which is characterized as hugeness of process variables in complexity of industrial environment, can be tackled effectively by the use of soft sensor technology. However, how to supplement the dataset with effective data supplement method under harsh industrial environment is a key issue for the enhancement of prediction accuracy in soft-sensing model. Aimed at this problem, a SVAE-WGAN based soft sensor data supplement method is proposed for process industry. Firstly, deep features are extracted with the stacking of the variational autoencoder (SVAE). Secondly, a generation model is constructed with the combination of stacked variational autoencoder (SVAE) and Wasserstein generative adversarial network (WGAN). Thirdly, the proposed model is optimized with training of dataset in industrial process. Finally, the proposed model is evaluated with abundant experimental tests in terms of MSE, RMSE and MAE. It is shown in the results that the proposed SVAE-WGAN generation network is significantly better than that of the traditional VAE, GAN and WGAN generation network in case of industrial steam volume dataset. Specially, the proposed method is more effective than the latest reference VA-WGAN generation network in terms of RMSE, which is enhanced about 9.08% at most. Moreover, the prediction precision of soft sensors could be improved via the supplement of the training samples.
23 February 2024 in Measurement Science and Technology
The use of dynamic soft sensor modeling methods to mine the time-varying and dynamic characteristics of industrial process data is of great significance for improving production efficiency and quality, given the rapid development of industrial processes and the increasing prominence of dynamic changes in the production process. However, existing dynamic soft sensor methods have limited long-term memory capacity, making it difficult to capture the long dynamic dependence, which can severely affect the results of the soft sensor model. To address this issue, we propose a dynamic soft sensor model based on local perception transformer, where the transformer is applied to achieve global perception of the variables. Through the application of the self-attention mechanism in the transformer encoder, the dynamic tracking and prediction of parameters can be realized by assigning different weights to the process variables and quality variables at different time steps, thereby adapting to the time-varying nature of the process. Additionally, convolution is used to generate a Query and Key in the self-attention mechanism, thereby enhancing local information learning. The proposed dot product self-attention calculation method effectively utilizes local information, thereby reducing the potential impact of abnormal data at a certain moment. Furthermore, by utilizing LSTM to extract time series information, the final prediction result was obtained. In soft sensor modeling experiments of the sulfur recovery unit and debutanizer tower, our proposed model demonstrated higher prediction accuracy compared to other methods, such as SVR, MLP, LSTM, CNN + LSTM, and vanilla transformer.
18 August 2022 in IEEE Transactions on Instrumentation and Measurement
Recently, feature extraction-based soft sensor techniques have developed rapidly in the control, optimization, and detection processes of industrial production. However, the raw data obtained from the complex industrial processes are often contaminated by noise, which significantly impacts the results of soft sensor models. We introduce the collaborative apportionment noise (CAN) method based on the density peaks clustering (DPC) theory, based on which we have proposed a CAN-based soft sensor framework (CAN-SSF) and designed an example model called the CAN-based convolutional neural networks (CAN-CNNs) model for industry data prediction. In the CAN method, we determined the magnitude and direction of the noise by the bias degree and deviation of the data. Then, the noise is collaboratively apportioned by the credibility degree of the data. Finally, to further explore the feasibility of the CAN method, we added a hyperparameter called reduction degree and conducted two groups of independent experiments for the example model CAN-CNN. The results have shown that the adaptability and stability of the CAN method are higher than the traditional wavelet transform (WT) denoising and denoising autoencoders (DAEs). In addition, the prediction performance of the proposed CAN-SSF is better than that of the traditional CNN and stacked autoencoders (SAEs) models to solve the industrial soft sensor problems.
14 March 2024 in IEEE Sensors Journal
Soft sensor technology has become one of the effective means to solve the problems that the key variables in complex industrial processes are not easy to measure. Single supervised or unsupervised modeling methods alone cannot sufficiently represent the data, leading to suboptimal results when using the acquired soft sensor model for critical variable prediction. Therefore, we have proposed an ensemble learning soft sensor model (SCAE-ACNN), which integrates stacked convolutional auto-encoder and attention mechanism. Firstly, the stacked convolutional auto-encoder based learner constructed by stacking multiple pretrained convolutional auto-encoder is called SCAE, in which unsupervised learning of stacked convolutional auto-encoder can extract deep features of data. Secondly, in order to further enhance the prediction ability, this model uses convolution attention module to remove irrelevant and redundant information from the original features, and constructs a convolution attention module base learner called ACNN. The homogeneous base learners stated above are trained, and the final key variables are obtained through the weighted average combination strategy. The proposed method based on ensemble learning is simulated in the debutanizer of steam thermal power generation and petroleum refining. The results show that the proposed SCAE-ACNN soft sensor model has better prediction performance than the traditional SVM algorithm, multilayer neural network and advanced methods.
23 August 2023 in Measurement
7 July 2025 in Engineering Applications of Artificial Intelligence