Soft Sensor Development Research Articles

Building Optimal Multiresolution Soft Sensors for Continuous Processes

Data-driven models used in soft sensor applications are expected to capture the dominant relationships between the different process variables and the outputs, while accounting for their high-dimensional, dynamic, and multiresolution character. While the first two characteristics are often addressed, the multiresolution aspect is usually disregarded and confused with a multirate scenario: multiresolution occurs when variables have different levels of granularity, because of, for instance, automatic averaging operations over certain time windows; on the other hand, a multirate structure is caused by the existence of different sampling rates, but the granularity of the recorded values is the same. This has two major and immediate implications. First, current methods are unable to handle variables with different resolutions in a consistent and rigorous way, since they tacitly assume that data represent instant observations and not averages over time windows. Second, even if data is available at a single-resolution (i.e., all variables with the same granularity), it is not guaranteed that the native resolution of the predictors is the most appropriate for modeling. Therefore, soft sensor development must address not only the selection of the best set of predictors to be included in the model, but also the optimum resolution to adopt for each predictor. In this work, two novel multiresolution frameworks for soft sensor development are proposed (MRSS-SC and MRSS-DC) that actively introduce multiresolution into the data by searching for the best granularity for each variable. The performance of these methodologies is comparatively assessed against current single-resolution counterparts. The optimized multiresolution soft sensors are bounded to be at least as good as their single-resolution versions, and the results confirm that they almost always perform substantially better.

Quantum statistic based semi-supervised learning approach for industrial soft sensor development

Unlike process variables which can be easily measured online, quality variables are often hard to be collected. Therefore, only a small proportion of soft sensor inputs are attached with quality related output labels. The semi-supervised learning mechanism can elegantly incorporate unlabeled input samples for soft sensor improvement and hence has become popular. In this work, a novel mechanism called quantum statistic is incorporated with semi-supervised learning by quantum states. The quantum states are constructed by superposing conventional pure states with composite states and the extended state space as the complements could be more desirable for representing state uncertainties of incomplete labels. Based on that, a quantum statistical based semi-supervised soft sensor is developed. The quantum statistic based model is comprehensively compared with the conventional state-of-the-art method in a numerical example and an industrial process. Results demonstrate that the proposed soft sensor is more effective and stable than the traditional state-of-art method.

Identification of systems with slowly sampled outputs using LPV model

Identification of systems with slowly sampled output is studied. A linear parameter varying (LPV) model with multi-model structure is used to solve the problem. The output error (OE) method is used to estimate model parameters. Firstly, the local models and weighting functions are estimated separately using optimization methods. Then, a relaxation iteration method is developed to refine the parameters of the total model. For LPV model structure determination, an engineering approach is proposed that combines process knowledge with the so-called final output error criteria (FOE). The method is verified using both simulation data and industrial data. In the industrial case study, the LPV models give more accurate prediction of product qualities than that of a linear dynamic model and that of a static nonlinear model; the result also indicates the necessity of using test signals in soft-sensor development.

State Estimation Using Non-uniform and Delayed Information: A Review

The study and application of methods for incorporating nonuniform and delayed information in state estimation techniques are important topics to advance in soft sensor development. Therefore, this paper presents a review of these methods and proposes a taxonomy that allows a faster selection of state estimator in this type of applications. The classification is performed according to the type of estimator, method, and used tool. Finally, using the proposed taxonomy, some applications reported in the literature are described.

Deep Learning of Semisupervised Process Data With Hierarchical Extreme Learning Machine and Soft Sensor Application

Data-driven soft sensors have been widely utilized in industrial processes to estimate the critical quality variables which are intractable to directly measure online through physical devices. Due to the low sampling rate of quality variables, most of the soft sensors are developed on small number of labeled samples and the large number of unlabeled process data is discarded. The loss of information greatly limits the improvement of quality prediction accuracy. One of the main issues of data-driven soft sensor is to furthest exploit the information contained in all available process data. This paper proposes a semisupervised deep learning model for soft sensor development based on the hierarchical extreme learning machine (HELM). First, the deep network structure of autoencoders is implemented for unsupervised feature extraction with all the process samples. Then, extreme learning machine is utilized for regression through appending the quality variable. Meanwhile, the manifold regularization method is introduced for semisupervised model training. The new method can not only deeply extract the information that the data contains, but learn more from the extra unlabeled samples as well. The proposed semisupervised HELM method is applied in a high–low transformer to estimate the carbon monoxide content, which shows a significant improvement of the prediction accuracy, compared to traditional methods.

Soft Sensor Development for Multimode Processes Based on Semisupervised Gaussian Mixture Models

The Gaussian mixture models (GMM) is an effective tool for modeling processes with multiple operating modes that widely exist in industrial process systems. Traditional supervised version of GMM, namely the Gaussian mixture regression (GMR), for developing soft sensors merely relies on the labeled samples. However, labeled samples in the soft sensor application are usually very infrequent due to economical or technical limitations, which may lead the GMR to unreliable parameter estimation and finally poor performance for predicting the primary variable. To tackle this problem, a semisupervised GMM for regression purpose is proposed, where both labeled and unlabeled samples take effect, and the Gaussian parameters and regression coefficients are learned simultaneously based on the expectation-maximization algorithm. Two case studies are carried out using simulated dataset and real-life dataset collected from a primary reformer in an ammonia synthesis process, which demonstrates the effectiveness of the proposed method.

LWS based PCA subspace ensemble model for soft sensor development

Most regression approaches, such as principal component analysis (PCA), are based on an assumption that the process data follow a Gaussian distribution. However, the process data usually dissatisfy that assumption. Thus, the locally weighted standardization (LWS) method is employed for transforming data into an approximate Gaussian distribution. Furthermore, the LWS based subspace PCA ensemble modeling method is developed. The subspace PCA can select important variables in each subspace for ensemble modeling. As a result, the proposed method gives a weaker assumption constrain and a better regression performance. The effectiveness of this approach is testified by two study cases.

Adaptive Soft Sensor Development for Multi-Output Industrial Processes Based on Selective Ensemble Learning

Soft sensors are vital for online predictions of quality-related yet difficult-to-measure variables in process industry. In this paper, an adaptive soft sensing approach based on selective ensemble learning is proposed for multi-output nonlinear and time-varying industrial processes, which we refer to as the selective ensemble learning for multi-outputs (SEL-MO). Specifically, an adaptive localization approach is developed for dealing with the process nonlinearity based on the statistical hypothesis testing theory, which can construct redundancy-free local model set. At the online operation stage, these constructed local models are partially combined under an adaptive selective ensemble learning framework, where the weightings of local models are query-sample-oriented such that both gradual and abrupt changes in the process characteristics can be handled. In addition, an insensitivity strategy is proposed to enhance the online computational efficiency of the SEL-MO by avoiding the unnecessary search of the historical data set. Case studies are carried out on a simulated fed-batch penicillin process and a real-life industrial primary reformer, and the results obtained demonstrate the effectiveness of the proposed method.

Variable selection for nonlinear soft sensor development with enhanced Binary Differential Evolution algorithm

In this paper, two enhanced Binary Differential Evolution (BDE) algorithms are proposed to select variables for nonlinear process soft sensor development. Firstly, the Parallel BDE (PBDE) algorithm is presented to extract the optimal individuals of several parallel short evolution paths of basic BDE, where the spurious variables are effectively eliminated. And the most relevant variables are selected through a double-layer selection strategy with the validating Root Mean Square Error (RMSE) for evaluating criterion. Secondly, the Boosting BDE (BBDE) algorithm is proposed through applying the boosting technique to the parallel evolution paths. The performance of the previous path needs to be taken into account when conducting the current evolution path. The selected probabilities of variables are given through the weighted summation of the selection results of all paths. Also, a double-layer selection is conducted on BBDE algorithm. The feasibility and effectiveness of the proposed methods are demonstrated through a nonlinear numerical example and a real industrial process.

Soft Sensor Development in Fermentation Processes Using Recursive Gaussian Mixture Regression Based on Model Performance Assessment

Soft Sensor Development in Fermentation Processes Using Recursive Gaussian Mixture Regression Based on Model Performance Assessment

Classification of Diesel Fuel Using Two-Dimensional Fluorescence Spectroscopy

Air pollution is a serious problem, and to decrease the emission of pollutants, governments have established environmental rules to limit the concentration of sulfur in diesel fuel. Control of sulfur content in diesel streams demands online measurement of this component, with low pure time delay and easy operation. In this regard, two-dimensional (2D) fluorescence spectroscopy becomes a promising choice for soft-sensor development. Despite fluorescence qualities, translation of fluorescence spectroscopic data into process knowledge is not a simple task, demanding multivariate analysis. This work aims to evaluate the applicability of 2D fluorescence spectroscopy for monitoring ultralow sulfur diesel streams (Diesel S10) and the capability of differentiation between ULSD and diesel streams with a higher sulfur concentration (Diesel S100). The evaluation of the different sample groups’ fluorescence spectra made clear the ability of the technique to capture changes in the composition between Diesel S10 and Di...

Just-in-Time Correntropy Soft Sensor with Noisy Data for Industrial Silicon Content Prediction

Development of accurate data-driven quality prediction models for industrial blast furnaces encounters several challenges mainly because the collected data are nonlinear, non-Gaussian, and uneven distributed. A just-in-time correntropy-based local soft sensing approach is presented to predict the silicon content in this work. Without cumbersome efforts for outlier detection, a correntropy support vector regression (CSVR) modeling framework is proposed to deal with the soft sensor development and outlier detection simultaneously. Moreover, with a continuous updating database and a clustering strategy, a just-in-time CSVR (JCSVR) method is developed. Consequently, more accurate prediction and efficient implementations of JCSVR can be achieved. Better prediction performance of JCSVR is validated on the online silicon content prediction, compared with traditional soft sensors.

Development of Adaptive Soft Sensor Using Locally Weighted Kernel Partial Least Square Model

Abstract Locally weighted partial least square (LW-PLS) model has been commonly used to develop adaptive soft sensors and process monitoring for numerous industries which include pharmaceutical, petrochemical, semiconductor, wastewater treatment system and biochemical. The advantages of LW-PLS model are its ability to deal with a large number of input variables, collinearity among the variables and outliers. Nevertheless, since most industrial processes are highly nonlinear, a traditional LW-PLS which is based on a linear model becomes incapable of handling nonlinear processes. Hence, an improved LW-PLS model is required to enhance the adaptive soft sensors in dealing with data nonlinearity. In this work, Kernel function which has nonlinear features was incorporated into LW-PLS model and this proposed model is named locally weighted kernel partial least square (LW-KPLS). Comparisons between LW-PLS and LW-KPLS models in terms of predictive performance and their computational loads were carried out by evaluating both models using data generated from a simulated plant. From the results, it is apparent that in terms of predictive performance LW-KPLS is superior compared to LW-PLS. However, it is found that computational load of LW-KPLS is higher than LW-PLS. After adapting ensemble method with LW-KPLS, computational loads of both models were found to be comparable. These indicate that LW-KPLS performs better than LW-PLS in nonlinear process applications. In addition, evaluation on localization parameter in both LW-PLS and LW-KPLS is also carried out.

Robust soft sensor development using multi-rate measurements

Two different types of measurements are often available for the key quality variables in process industries - (a) an accurate “slow-rate” laboratory measurements, and (b) a less accurate “fast-rate” online analyser measurements. Also, the analyser measurements are prone to fail due to hardware issues. Therefore, the main objective of this work is to present a novel approach for developing an accurate, fast-rate, inferential model of quality variables which is robust to outliers. For this purpose, we present a maximum likelihood based approach to integrate the multi-rate output data in the model building task, using Expectation Maximization algorithm. The efficacy of the proposed approach is demonstrated using a simulation example.

Melt index prediction with a mixture of Gaussian process regression with embedded clustering and variable selections

ABSTRACTIn this study, a penalized mixture of the Gaussian process regression model was proposed for the prediction of melt index (MI) in industrial polymer production. MI plays an important role in detecting the grade of a product. It is difficult to measure directly and is characterized by a large number of variables and multigrades. Because of multigrade products, in the development of soft sensors for MI prediction, it is not valid to assume unimodal Gaussian distribution of the data. To this end, the proposed method is capable of the simultaneous identification of significant variables and determination of important clusters of multigrade products. It is based on the shrinkage methods that have been shown to provide stable models that are more interpretable. Case studies are presented to show the features of the proposed method and its applicability to industrial MI prediction. © 2017 Wiley Periodicals, Inc. J. Appl. Polym. Sci. 2017, 134, 45237.

Development of soft sensors for permanent downhole Gauges in deepwater oil wells

Downhole pressure is an important process variable in the operation of gas-lifted oil wells. The device installed in order to measure this variable is often called a Permanent Downhole Gauge (PDG). Replacing a faulty PDG is often not economically viable and to have an alternative estimate of the downhole pressure is an important goal. Using data from operating PDGs, this paper describes a number of issues dealt with in the development of soft sensors for several deepwater gas-lifted oil wells. Some of the tested models include nonlinear polynomials, neural networks, committee machines, unscented Kalman filters and filter banks. The variety of model classes used in addition to the diversity of oil wells considered brings to light some of the key-problems that have to be faced and reveal the strengths and weaknesses of each alternative solution. A major constraint throughout the work was the use of historical data, hence no specific tests were performed at any time. The aim of this work is to discuss the procedures, pros and cons of the tested solutions and to point to possible future directions of research.

A variable selection method for soft sensor development through mixed integer quadratic programming

Soft sensors are widely employed in industry to predict quality variables, which are difficult to measure online, by using secondary variables. To build an accurate soft sensor, a proper variable selection is critical. In this project, a method of selecting the optimal secondary variables for a soft sensor model is proposed. It is formulated as a nested optimization problem. In each iteration, a mixed integer quadratic programming (MIQP) is conducted with the Bayesian information criterion (BIC) to estimate the prediction error. A warm start (WS) technique is developed to speed up the convergence. The proposed method is evaluated using a number of instances from the UCI Machine Learning Repository. The computational results demonstrate that this method is well suited for finding the best variable subsets. The method is successfully applied to build soft sensors for an industrial distillation column. The results show that the proposed method can effectively select feature variables that will improve the model prediction performance and reduce the model complexity. Comparisons with other methods, including the traditional partial least square technique, are also presented.

Variational Bayesian Gaussian Mixture Regression for Soft Sensing Key Variables in Non-Gaussian Industrial Processes

In this brief, a variational Bayesian Gaussian mixture regression (VBGMR) method is developed for soft sensing key quality-related variables in a non-Gaussian industrial process. Traditional Gaussian mixture regression (GMR) is based on Gaussian mixture model (GMM) and can be easily stuck into the issue of model selection since the GMM model generally requires a large amount of local components so as to achieve desirable predicting performances. However, such assignation can be computationally extensive and may also result in some numerical issues since only a few components are positively adopted. In this brief, a fully Bayesian modeling method is proposed for GMR-based soft sensor development. Conjugate Bayesian prior is defined for each empirical parameter, while a Dirichlet process prior is further defined on the mixture component. The full Bayesian GMR is first validated on a numerical case study and then applied to the industrial hydrogen manufacturing units, both compared with GMR. The results demonstrate feasibility and reliability of the new soft sensor and show that VBGMR generally outperforms GMR with the requirement of only a few activated components.

Recent Progress in the Development of Soft Robots

Soft robots, are mobile machines largely constructed from soft materials and have received much attention recently because they are opening new perspectives for robot design and control. This paper reports recent progress in the development of soft robots, more precisely, soft actuators and soft sensors. Soft actuators play an important role in functionalities of soft robots, and dielectric elastomers have shown great promise because of their considerable voltage-induced deformation. We developed soft inflated dielectric elastomer actuators and their networks, with the advantages to be highly deformable and continuously controllable. When it comes to control of soft robots, soft sensors are of great importance. We proposed a methodology to design, analyze, and fabricate a multi-axis soft sensor, made of dielectric elastomer, capable of detecting and decoupling compressive and shear loads with high sensitivity, linearity, and stability.

A Soft Sensor Development for the Rotational Speed Measurement of an Electric Propeller

In recent decades, micro air vehicles driven by electric propellers have become a hot topic, and developed quickly. The performance of the vehicles depends on the rotational speed of propellers, thus, improving the accuracy of rotational speed measurement is beneficial to the vehicle’s performance. This paper presents the development of a soft sensor for the rotational speed measurement of an electric propeller. An adaptive learning algorithm is derived for the soft sensor by using Popov hyperstability theory, based on which a one-step-delay adaptive learning algorithm is further proposed to solve the implementation problem of the soft sensor. It is important to note that only the input signal and the commutation instant of the motor are employed as inputs in the algorithm, which makes it possible to be easily implemented in real-time. The experimental test results have demonstrated the learning performance and the accuracy of the soft sensor.