Industrial Soft Sensor Research Articles

A nonlinear industrial soft sensor modeling method based on locality preserving stochastic configuration network with utilizing unlabeled samples

Stochastic configuration network (SCN) is an emerging incremental randomized regression modeling technology with the advantages of adaptively determining the hidden layer parameters, and has been successfully applied to industrial soft sensor modeling field. However, the traditional SCN model is intrinsically a supervised learner, which has the underlying assumption that all the training samples are labeled. In fact, most of process samples are unlabeled and the labeled samples are relatively rare in real industrial scenarios. To handle this issue, this paper presents one modified SCN model, called locality preserving SCN (LPSCN), for semi-supervised industrial soft sensor modeling. In this method, all the training samples, including the labeled and the unlabeled, are fed into the soft sensor model, where the labeled samples are used to capture the modeling error, while the unlabeled samples help construct the local adjacency graph. Based on these two kinds of samples, the supervised optimization objective in the traditional SCN is improved to be a semi-supervised version by minimizing the modeling error and preserving the local data relationship simultaneously. Furthermore, the random parameter configuration mechanism is deduced under the modified semi-supervised optimization framework. A new inequality constraint condition with considering the unlabeled samples is obtained to generate the hidden layer nodes incrementally so that the LPSCN model structure is determined automatically. Experiments on two real industrial systems demonstrate that the proposed LPSCN method outperforms the SCN method in terms of the soft sensor prediction performance.

Multivariate Deep Reconstruction Neural Network for Multi-step-ahead Prediction of Industrial Process Quality Variables

In industrial processes, deep learning has been widely used to solve the soft sensing problem. Multi-step-ahead prediction is one of the most challenging problems in the field of soft sensors. Recently, N-Beats has been proposed as a promising deep neural architecture for multi-step prediction, but it can only be used for univariate time series prediction, not for industrial soft sensor modeling. Inspired by N-Beats, a novel deep learning model, multivariate deep reconstruction neural network (MDRNN), is proposed for multivariate time series prediction in this work. MDRNN is designed on the basis of a doubly residual structure with a deep stack of fully-connect layers. MDRNN inherits the merits of N-Beats and incorporates the “doubly residual stacking” idea into the industrial soft sensor modeling to improve the prediction accuracy. To evaluate the feasibility and effectiveness of the proposed MDRNN, it is applied to the quality prediction task and validated with two real-world industrial processes. The application results demonstrated that the proposed MDRNN can achieve higher prediction accuracy compared to the existing N-Beats and Multilayer Perceptron (MLP)-based methods.

FIGAN: A Missing Industrial Data Imputation Method Customized for Soft Sensor Application

Missing data is quite common in the industrial field, resulting in problems in downstream applications, as most data driven methods used in these applications rely on complete and high-quality dataset to build a high-quality model. Existing methods deal with missing data individually regardless of its downstream application, treating all variables equally without considering their different roles in the downstream application. This would affect imputation performance for key variables, thus deteriorating the accuracy of the downstream model. A considerable challenge is how to refine the missing data imputation task. In this paper, a new method termed fine-tuned imputation GAN (FIGAN) is designed to achieve customized data imputation for industrial soft sensor. The major contribution of the paper lies in two aspects: 1) different from the original imputation GAN (GAIN) which treats all variables equally, FIGAN is guided by a soft sensor module so as to achieve customized data imputation by performing improved data imputation on quality-related variables. Enhanced accuracy for the final industrial soft sensor would be possible; 2) in addition, since labels of the soft sensor might also have missing data, a soft sensor with pseudo labeling is designed to conquer the problem with data imputation and label prediction being optimized interactively. Case studies on a converter steelmaking process and a penicillin fermentation process show the feasibility of the proposed FIGAN. It is noted that such customized imputation could be readily transferred to other downstream applications with missing data. Note to Practitioners—Industrial data is often incomplete and needs proper treatment. Meanwhile, downstream applications with preprocessed data vary under different industrial scenes. The focus of this study is to develop a customized data imputation method for specific downstream applications such as soft sensing. A fine-tuned imputation GAN is designed with a soft sensor module so as to guide the data imputation for good imputation on key variables of the soft sensor. Considering that missing data exists not only in measurement variables but also in soft sensor labels, a semi-supervised soft sensor is designed to handle missing data in the labels, optimized together with the imputation model. The customized data imputation can thus improve the final performance of the downstream model which is a soft sensor in this work. The customization could be transferred to other applications such as an anomaly detection model as well.

Industrial Soft Sensor Optimized by Improved PSO: A Deep Representation-Learning Approach.

Soft sensors based on deep learning approaches are growing in popularity due to their ability to extract high-level features from training, improving soft sensors’ performance. In the training process of such a deep model, the set of hyperparameters is critical to archive generalization and reliability. However, choosing the training hyperparameters is a complex task. Usually, a random approach defines the set of hyperparameters, which may not be adequate regarding the high number of sets and the soft sensing purposes. This work proposes the RB-PSOSAE, a Representation-Based Particle Swarm Optimization with a modified evaluation function to optimize the hyperparameter set of a Stacked AutoEncoder-based soft sensor. The evaluation function considers the mean square error (MSE) of validation and the representation of the features extracted through mutual information (MI) analysis in the pre-training step. By doing this, the RB-PSOSAE computes hyperparameters capable of supporting the training process to generate models with improved generalization and relevant hidden features. As a result, the proposed method can generate more than 16.4% improvement in RMSE compared to another standard PSO-based method and, in some cases, more than 50% improvement compared to traditional methods applied to the same real-world nonlinear industrial process. Thus, the results demonstrate better prediction performance than traditional and state-of-the-art methods.

Multi time scale inception-time network for soft sensor of blast furnace ironmaking process

Time series(TS) forecasting has been widely applied in many fields and industrial soft sensor is one of them. Most time series modeling methods require that all inputs are sampled at equal intervals. However, in industry, the process variables are often sampled on different time scales with varying intervals. To address this problem, in this paper, we designed a framework using deep learning with time representation techniques to model the long temporal industrial data with multiple sampling frequency. Data are aggregated into different time scale by time representation and the network extracts the information on both temporal and spatial dimensions simultaneously using the bottleneck layer and one-dimensional filter. The proposed model has a significant improvement compared with other methods and has been deployed in the factory and updated every month.

A Data-Driven Self-Supervised LSTM-DeepFM Model for Industrial Soft Sensor

Soft sensor, as an important paradigm for industrial intelligence, is widely used in industrial production to achieve efficient monitoring and prediction of production status including product quality. Data-driven soft sensor methods have attracted attention, which still have challenges because of complex industrial data with diverse characteristics, nonlinear relationships, and massive unlabeled samples. In this article, a data-driven self-supervised long short-term memory–deep factorization machine (LSTM-DeepFM) model is proposed for industrial soft sensor, in which a framework mainly including pretraining and finetuning stages is proposed to explore diverse industrial data characteristics. In the pretraining stage, an LSTM-autoencoder is first unsupervised pretrained. Then, based on two self-supervised mask strategies, LSTM-deep can explore the interdependencies between features as well as the dynamic fluctuation in time series. In the finetuning stage, relying on pretrained representation, the temporal, high-dimensional, and low-dimensional features can be extracted from the LSTM, deep, and FM components, respectively. Finally, experiments on the real-world mining dataset demonstrate that the proposed method achieves state of the art comparing with stacked autoencoder-based models, variational autoencoder-based models, semisupervised parallel DeepFM, etc.

Semi-supervised echo state network with temporal–spatial graph regularization for dynamic soft sensor modeling of industrial processes

Echo state network (ESN) has been successfully applied to industrial soft sensor field because of its strong nonlinear and dynamic modeling capability. Nevertheless, the traditional ESN is intrinsically a supervised learning technique, which only depends on labeled samples, but omits a large number of unlabeled samples. In order to eliminate this limitation, this work proposes a semi-supervised ESN method assisted by a temporal–spatial graph regularization (TSG-SSESN) for constructing soft sensor model with all the available samples. Firstly, the traditional supervised ESN is enhanced to construct the semi-supervised ESN (SSESN) model by integrating both unlabeled and labeled samples in the reservoir computing procedure. The SSESN computes the reservoir states under high sampling rate for better process dynamic information mining. Furthermore, the SSESN’s output optimization objective is modified by applying the local adjacency graph of all training samples as a regularization term. Especially, in view of the dynamic data characteristic, a temporal–spatial graph is constructed by considering both the temporal relationship and the spatial distances. The applications to a debutanizer column process and a wastewater treatment plant demonstrate that the TSG-SSESN model can build much smoother model and has better generalization capability than the basic ESN models in terms of soft sensor prediction results.

Adversarial Attacks on Neural-Network-Based Soft Sensors: Directly Attack Output

Neural-network-based soft sensors are widely employed in the industrial process. Such models have great significance to smart manufacturing. Considering the strict requirements of industrial production, it is vital to ensure the safety and robustness of these models in their actual deployment. However, recent research has shown that neural networks are quite vulnerable to adversarial attacks. By imposing tiny perturbation to the original sample, the fabricated adversarial sample can cheat these models to make wrong decisions. Such a phenomenon may bring serious trouble to the practical application of soft sensors. This article focuses on the adversarial attacks on industrial soft sensors. For the first time, we verify and analyze the effectiveness and deficiencies of the existing attack methods in the industrial soft sensor scenario. Based on solving these defects, this article proposes a novel perspective for attacking soft sensors. We analyze the optimization mechanism behind this new idea and then design two algorithms to perform attacks. The proposed methods more conform to the actual situation. Besides, compared with the existing approaches, the proposed methods have potentials to cause severer damages since their attacks are not only more concealed but also more likely to cheat the technicians to execute wrong operations. The research and analyses of the proposed methods lay a solid foundation for more thorough defenses against various attacks, which is quite necessary for making the deployed soft sensors more robust and secure.

Dynamic historical information incorporated attention deep learning model for industrial soft sensor modeling

Due to the limitations of sampling conditions and sampling techniques in many real industrial processes, the process data under different sampling conditions subject to different sampling frequencies, which leads to irregular interval sampling characteristics of the entire process data. The dynamic historical data information reflecting the production status under irregular sampling frequency has an important influence on the performance of data feature extraction. However, the existing soft sensor modeling methods based on deep learning do not consider introducing dynamic historical information into the feature extraction process. To combat this issue, a novel attention-based dynamic stacked autoencoder networks (AD-SAE) for soft sensor modeling is proposed in this paper. First, the sliding window technology and attention mechanism based on position coding are introduced to select dynamic historical samples and calculate the contribution of different historical samples to the current sample, respectively. Then, AD-SAE combines obtained historical sample information and current sample information as the input of the network for deep feature extraction and industrial soft sensor modeling. The experimental results on the actual hydrocracking process data set show that the proposed method has better performance than traditional methods.

A Multi-Rate Probabilistic Slow Feature Regression Model for Dynamic Feature Learning and Industrial Soft Sensor Development

In practical process industries, the measurements coming from different sources are collected at different sampling rates, thereby soft sensors developed using uniformly sampled measurements may result in poor prediction performance. Besides, industrial processes are inherently stochastic and most of them present dynamic characteristic. To cope with these issues, a multi-rate probabilistic slow feature regression (MR-PSFR) model is proposed in this paper for dynamic feature learning and soft sensor development in industrial processes. In the MR-PSFR, both input and output observation datasets with different sampling rates are used to extract the slow features, which can separate slowly and fast changing features and have a better interpretation of the outputs. Then, the expectation-maximization algorithm is modified to derive the model parameters of MR-PSFR and the quality prediction strategy for multi-rate processes is constructed. Finally, the proposed method is investigated through a numerical example and a real industrial process. The simulation results show that the extracted slow features better represent the intrinsic characteristics of the processes and the proposed model has better prediction performance for multi-rate dynamic processes than other methods.

Collaborative Apportionment Noise-Based Soft Sensor Framework

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-CNN) 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. And 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 denoising (WT) and denoising autoencoders (DAE). In addition, the prediction performance of the proposed CAN-SSF is better than the traditional CNN and Stacked autoencoders (SAE) models to solve the industrial soft sensor problems.

Adversarial smoothing tri-regression for robust semi-supervised industrial soft sensor

In industrial processes, soft sensor techniques are used to predict the hard-to-measure quality variables under the classic supervised learning paradigm. However, the data challenges, i.e., the widespread noises and inadequate labeled samples, usually make the data-driven sensors weak for application. In this paper, a simple but effective regularization method termed as adversarial smoothing regularization (ASR) is proposed, which measures the local smoothness of prediction around each input sample. By minimizing the divergence between the predictions for noisy and clean inputs, the proposed ASR regulates the learning model to be robust against local perturbation in a semi-supervised manner. Theoretical analysis explains the smoothing capability of ASR, and shows that it is helpful for the improvement of generalization performance. We also design a tri-regression framework to further use the information of unlabeled samples with pseudo labels and present the adversarial smoothing tri-regression (ASTR) model for soft sensor. Based on two industrial processes, comprehensive soft sensor experiments and noise tests are performed to show the robust semi-supervised learning capability of the proposed ASR and ASTR.

An Improved JITL Method for Soft Sensing of Multimodal Industrial Processes for Search Efficiency

In the industrial process, due to product change, working condition switch, or controller adjustment, process data often presents multimodal characteristics. Data-driven approaches are often based on single-modal assumptions, which may fail to describe process characteristics. The traditional just-in-time learning (JITL) method can continuously update the model to describe the multimodal data, but it takes much time and cannot meet the real-time requirements. In this paper, an improved JITL method is proposed to find similar samples quickly. The new samples are divided into the main category first, and then find the similar samples to improve the search efficiency. The effectiveness of the method is proved by a case of an industrial soft sensor case combined with partial least squares (PLS). Compared with the basic JITL, the root mean square error (RMSE) of the proposed method is reduced by 0.09, and the running speed is increased by 8.8 times.

A Spectral AutoML approach for industrial soft sensor development: Validation in an oil refinery plant

Spectral AutoML is a platform for fast development of PAT soft sensors that considers the combined effect of pre-processing, band selection, band-wise resolution definition, hyper-parameter tuning and model estimation. Spectral AutoML was compared with models developed under the classic paradigm, and their performance assessed on an independent test set. The validation study regards the prediction of 12 different diesel fuels properties, using FTIR-ATR spectra. The proposed framework led to clearly better predictions in 8 out of the 12 properties, and minor improvements in 3 properties. The Spectral AutoML results were obtained overnight, without interfering in the daily work of the users, while the benchmark models resulted from several months of work and fine tuning of the methods. The results demonstrated the added value of the proposed Spectral AutoML approach in terms of prediction accuracy, development time of the models and reduced dependence on resident experts.

Bayesian network for dynamic variable structure learning and transfer modeling of probabilistic soft sensor

Bayesian network is a frequently-used uncertainty reasoning method, which systematically describes relations between random variables. Dynamic Bayesian network is an extension of Bayesian network, which contains the relations between variables at different times. Soft sensor is an important industrial application, in which feature variables are selected to predict the value of the target variables. For industrial soft sensor applications, dynamics is still a tough problem, particularly together with the uncertain feature of process data. In this article, DBN is employed for dynamic variable structure learning and transfer modeling to some strong regression models to build a soft sensor for dynamic industrial processes. At the beginning, a series of dynamic Bayesian networks are constructed on the training set with a sliding window. From these network structures we can find variables related to the quality variables. Then, in each time period, a sequence of data is compared with training data to select the most similar sequence by Dynamic Time Warping. Therefore, the structure of variables is built, i.e., the related feature variables of the quality variables can be determined by the network structure. In the regression step, the dynamic variable structure is transferred to some strong regressors, like Support Vector Regression and Adaboost for further regression. In case study, we use the debutanizer process and a low temperature transformer case to confirm the quality of the soft sensor method. The result reveals that, the prediction accuracy of the new method is much higher than the original commonly-used regression methods.

Gated Stacked Target-Related Autoencoder: A Novel Deep Feature Extraction and Layerwise Ensemble Method for Industrial Soft Sensor Application.

These days, data-driven soft sensors have been widely applied to estimate the difficult-to-measure quality variables in the industrial process. How to extract effective feature representations from complex process data is still the difficult and hot spot in the soft sensing application field. Deep learning (DL), which has made great progresses in many fields recently, has been used for process monitoring and quality prediction purposes for its outstanding nonlinear modeling and feature extraction abilities. In this work, deep stacked autoencoder (SAE) is introduced to construct a soft sensor model. Nevertheless, conventional SAE-based methods do not take information related to target values in the pretraining stage and just use the feature representations in the last hidden layer for final prediction. To this end, a novel gated stacked target-related autoencoder (GSTAE) is proposed for improving modeling performance in view of the above two issues. By adding prediction errors of target values into the loss function when executing a layerwise pretraining procedure, the target-related information is used to guide the feature learning process. Besides, gated neurons are utilized to control the information flow from different layers to the final output neuron that take full advantage of different levels of abstraction representations and quantify their contributions. Finally, the effectiveness and feasibility of the proposed approach are verified in two real industrial cases.

Soft sensor design using transductive moving window learner

Changes in characteristics of the industrial processes require implementation of adaptive mechanisms in soft sensors. In the current study, we propose combining two common adaptation methods, moving window (MW) and Just-In-Time-Learning (JITL), using transductive inference (MWtr). Transductive learning exploits the knowledge from the feature values of the query points in determining predictions of the response variable at these points. Here, we use JITL predictions for obtaining predictions for the query points, which are used in training the MW learner in a weighted Lasso regression setting. Tests of the proposed method on three publicly available industrial real datasets show that prediction accuracy of MWtr is higher than both MW, and JITL models, and MWtr is able to combat against various types of concept drifts via integrating the capabilities of MW and JITL methods. Ease of implementation, robustness and stability render MWtr a convenient adaptive method in industrial soft sensor design.

Deep Learning With Spatiotemporal Attention-Based LSTM for Industrial Soft Sensor Model Development

Industrial process data are naturally complex time series with high nonlinearities and dynamics. To model nonlinear dynamic processes, a long short-term memory (LSTM) network is very suitable for soft sensor model development. However, the original LSTM does not consider variable and sample relevance for quality prediction. In order to overcome this problem, a spatiotemporal attention-based LSTM network is proposed for soft sensor modeling, which can, not only identify important input variables that are related to the quality variable at each time step, but also adaptively discover quality-related hidden states across all time steps. By taking the spatiotemporal quality-relevant interactions into consideration, the prediction performance can be improved for the soft sensor model. The effectiveness and flexibility of the proposed model is demonstrated on an industrial hydrocracking process to predict the initial boiling points of heavy naphtha and aviation kerosene.

Adversarial Sample Based Semi-Supervised Learning for Industrial Soft Sensor

In industrial processes, soft sensor techniques are often utilized to predict the hard-to-measure quality variables. However, the labeled data which are obtained from the offline lab analysis can be quite rare. In the present work, a new divergence-based semi-supervised learning method is developed to exploit the unlabeled samples together with labeled ones for soft sensor application, namely adversarial tri-regression. First, the adversarial samples are generated based on the consideration of maximum disturbance, and through training on the combination of the adversarial samples and the original labeled samples, three regressors are initialized with divergence. Second, for each regressor, an unlabeled sample is labeled when the other two regressors agree on the labeling of this sample, which actually provides that regressor with some unknown information based on the divergence. As the three regressors label more and more samples for each other, the final regression model obtained by averaging the three base regressors presents increasingly more accurate prediction. The proposed method tackles a practical soft sensor problem for the industrial production process of cigarette.

Data supplement for a soft sensor using a new generative model based on a variational autoencoder and Wasserstein GAN

In industrial process control, measuring some variables is difficult for environmental or cost reasons. This necessitates employing a soft sensor to predict these variables by using the collected data from easily measured variables. The prediction accuracy and computational speed in the modeling procedure of soft sensors could be improved with adequate training samples. However, the rough environment of some industrial fields makes it difficult to acquire enough samples for soft sensor modeling. Generative adversarial networks (GANs) and the variational autoencoder (VAE) are two prominent methods that have been employed for learning generative models. In the current work, the VA-WGAN combining VAE with Wasserstein generative adversarial networks (WGAN) as a generative model is established to produce new samples for soft sensors by using the decoder of VAE as the generator in WGAN. An actual industrial soft sensor with insufficient data is used to verify the data generation capability of the proposed model. According to the experimental results, the samples obtained with the proposed model more closely resemble the true samples compared with the other four common generative models. Moreover, the insufficiency of the training data and the prediction precision of soft sensors could be improved via these constructed samples.