Adaptive Soft Sensor Modeling Research Articles

Adaptive soft sensor modeling of chemical processes based on an improved just‐in‐time learning and random mapping partial least squares

AbstractThe just‐in‐time learning‐based partial least squares (JIT‐PLS) has been extensively applied to adaptive soft sensor modeling of complex nonlinear processes. However, it still has the problems of unreasonable relevant samples selection and unsatisfactory local modeling. Aiming at these problems, this paper proposes an improved just‐in‐time learning‐based random mapping partial least squares (IJIT‐RMPLS), including an improved relevant samples selection strategy and a random mapping PLS (RMPLS) model. On the one hand, considering the different correlation degrees between input variables and output variable, this method applies mutual information to evaluate the importance of each input variable and designs a variable‐weighted Euclidean distance to select relevant samples for local modeling. On the other hand, in order to prompt the prediction precision of local soft sensor models, this method combines the idea of nonlinear random mapping in extreme learning machines with PLS and builds a RMPLS with multiple activation functions. Applications on a numerical example and a real chemical process show that the proposed IJIT‐RMPLS has smaller prediction error compared with traditional JIT‐PLS.

Research on adaptive soft sensing modeling method of photovoltaic power generation process based on online semi-supervised selective ensemble learning

Soft sensors often estimate product quality variables that are difficult to measure in real time. However, strong nonlinearity and dynamic and time-varying characteristics lead to poor prediction performance of soft sensors for the photovoltaic power generation process. This paper proposes an adaptive soft sensor model construction strategy based on extreme learning machine online semi-supervised selective ensemble learning (SEMI-SEL-ELM). Firstly, a local learning strategy based on the Spatio-temporal criterion is proposed to limit the impact of nonlinear dynamic characteristics. The unlabeled samples are then examined to provide a semi-supervised sample set reconstruction. After that, using the extreme learning machine method, a local soft sensor model is created. Furthermore, the statistical hypothesis test is used to create the statistical information of the test samples, and the t-test is utilized to acquire the quantitative local model satisfaction. Additionally, the adaptive computation of mixed weights of distinct sub-models is performed using the improved information entropy. The black hole algorithm determines the parameters of this approach automatically. Finally, the method is applied to the historical data set of a photovoltaic power station in Australia. The results show that the method effectively deals with nonlinear, dynamic, and time-varying regression problems in PV power prediction.

Optimization of Distribution Control System in Oil Refinery by Applying Hybrid Machine Learning Techniques

In this research, prediction of crude oil cuts from the first stage of refining process field is laid out using rough set theory (RST) based adaptive neuro-fuzzy inference system (ANFIS) soft sensor model to enhance the performance of oil refinery process. The RST was used to reduce the fuzzy rule sets of ANFIS model, and its features in the decision table. Also, discretisation methods were used to optimise the continuous data’s discretisation. This helps to predict the two critical variables of light naphtha product: Reid Vapor Pressure (RVP) and American Petroleum Institute gravity (API gravity), which detect the cut’s quality. Hence, a real-time process of Al Doura oil refinery is examined and the process data of refining crude oil from these two sources improve the knowledge provided by the data. The response variables represent the feedback measured value of cascade controller in the top of the splitter in crude distillation unit (CDU) in the rectifying section, which controls the reflux liquid’s flow towards the splitter’s head. The proposed adaptive soft sensor model succeeded to fit the results from laboratory tests, and a steady-state control system was achieved through an embedded virtual sensor. The predictive control system has been employed using cascade ANFIS controller in parallel with the soft sensor model to keep the purity of the distillate product in the stated range of the quality control of oil refinery. The results obtained from the proposed ANFIS based cascade control have no over/undershoots, and the rise time and settling time are improved by 26.65% and 84.63%, respectively than the conventional proportional-integral-derivative (PID) based cascade control. Furthermore, the results of prediction and control model are compared with those of other machine learning techniques.

Adaptive Soft-Sensor Modeling of SMB Chromatographic Separation Process Based on Dynamic Fuzzy Neural Network and Moving Window Strategy

Adaptive Soft-Sensor Modeling of SMB Chromatographic Separation Process Based on Dynamic Fuzzy Neural Network and Moving Window Strategy

Adaptive Soft Sensor Modeling Method for Time-varying and Multi-Dimensional Chemical Processes

Adaptive Soft Sensor Modeling Method for Time-varying and Multi-Dimensional Chemical Processes

An adaptive soft sensor method of D-vine copula quantile regression for complex chemical processes

Non-linear and non-Gaussian properties are challenging topics in the soft sensor modeling of chemical processes, and fluctuations in the environmental conditions of chemical plants will also affect the accuracy of soft sensor models. This paper proposes an adaptive soft sensor method of D-vine copula quantile regression (aDVQR). In the modeling process, a sparse vine model is established using the Bayesian information criterion. Then, the conditional quantile function value of the specified quantile can be obtained via the recursive nesting method by the h function. An online model updating system based on the aDVQR model is also proposed, and an adaptive soft sensor model is established. The proposed adaptive soft sensor method can successfully approximate the non-linear and non-Gaussian relationships between variables and adapt to unstable environments. Finally, a numerical example and an example of the ethylene industry are used to verify the effectiveness of the proposed method.

Adaptive Hybrid Soft-Sensor Model of Grinding Process Based on Regularized Extreme Learning Machine and Least Squares Support Vector Machine Optimized by Golden Sine Harris Hawk Optimization Algorithm

Soft-sensor technology plays a vital role in tracking and monitoring the key production indicators of the grinding and classifying process. Least squares support vector machine (LSSVM), as a soft-sensor model with strong generalization ability, can be used to predict key production indicators in complex grinding processes. The traditional crossvalidation method cannot obtain the ideal structure parameters of LSSVM. In order to improve the prediction accuracy of LSSVM, a golden sine Harris Hawk optimization (GSHHO) algorithm was proposed to optimize the structure parameters of LSSVM models with linear kernel, sigmoid kernel, polynomial kernel, and radial basis kernel, and the influences of GSHHO algorithm on the prediction accuracy under these LSSVM models were studied. In order to deal with the problem that the prediction accuracy of the model decreases due to changes of industrial status, this paper adopts moving window (MW) strategy to adaptively revise the LSSVM (MW-LSSVM), which greatly improves the prediction accuracy of the LSSVM. The prediction accuracy of the regularized extreme learning machine with MW strategy (MW-RELM) is higher than that of MW-LSSVM at some moments. Based on the training errors of LSSVM and RELM within the window, this paper proposes an adaptive hybrid soft-sensing model that switches between LSSVM and RELM. Compared with the previous MW-LSSVM, MW-neural network trained with extended Kalman filter(MW-KNN), and MW-RELM, the prediction accuracy of the hybrid model is further improved. Simulation results show that the proposed hybrid adaptive soft-sensor model has good generalization ability and prediction accuracy.

A spatial-temporal LWPLS for adaptive soft sensor modeling and its application for an industrial hydrocracking process

Locally weighted partial least squares (LWPLS) is a widely used just-in-time learning (JITL) modeling algorithm for adaptive soft sensor development. In LWPLS, spatial variable distance is used to measure similarity and assign weights for historical samples, which is very effective to handle process time-varying problems of abrupt changes. However, the gradual process changes are not effectively handled in traditional LWPLS. To cope with this problem, a novel similarity is proposed for temporal distance measurement by introducing a temporal variable of sampling instant, in which newest sampled data can get large weights since they represent the more recent process state. Then, both spatial and temporal similarities are considered to construct a spatial-temporal adaptive LWPLS modeling framework in this paper. The effectiveness of the proposed algorithm is validated on an industrial hydrocracking process.

Streaming parallel variational Bayesian supervised factor analysis for adaptive soft sensor modeling with big process data

Time-varying and state shifting are two of the main process factors that cause poor prediction performance of soft sensors. Adaptive soft sensor is commonly an alternative practice to ensure high predictive accuracy. However, the large scale of process data often leads to inefficiency of model updating. In this paper, a streaming variational Bayesian supervised factor analysis (S-VBSFA) model is first proposed to capture the process time-varying and state shifting features through online updating of the posterior of model parameters. During the updating process, the symmetric Kullback–Leibler (SKL) divergence is utilized to determine priors of the next variation Bayesian inference. To improve the modeling efficiency for large-scale process data, the parallel computing strategy is further applied to the streaming model. As a result, the proposed streaming parallel VBSFA (SP-VBSFA) algorithm not only relieves the computing pressure of modeling big process data, but also improves the prediction accuracy and further reduces the tracking time delay for process variations. Two case studies demonstrate the superiority of the proposed method, compared to conventional methods.

An adaptive soft sensor deterioration evaluation and model updating method for time-varying chemical processes

Due to the time-varying nature of chemical processes, soft sensor models deteriorate, and data prediction accuracy decreases. To address this problem, an adaptive soft sensor modeling method is proposed that not only evaluates the model deterioration by an adaptive moving window-constrained statistical hypothesis test, but also adaptively updates the modeling samples using moving window-cosine similarity. First, this method evaluates the model deterioration via positioning by constrained statistical hypothesis testing based on the differences between the prediction performance evaluation index data obtained from moving window stepping and the original prediction performance evaluation indexes. Additionally, the dynamic temporal variation in chemical processes causes changes in the impacts of the auxiliary variables on the dominant variable, and this effect limits the improvement in the prediction accuracy of the soft sensor model by updating only the auxiliary variable data. The moving window-cosine similarity method is combined to propose a strategy that updates both the modeled auxiliary variables and the auxiliary variable data. Finally, the parameters of the soft sensor model are optimized via particle swarm optimization (PSO) to improve the fitting performance. Simulated data of a continuous stirred tank reactor (CSTR) and actual data from a debutanizer column process (DCP) are used for model verification to evaluate the performance of the proposed adaptive soft sensor modeling method, and the results show its effectiveness.

Ensemble regularized local finite impulse response models and soft sensor application in nonlinear dynamic industrial processes

This paper proposes a novel adaptive soft sensor modeling method based on ensemble of regularized local finite impulse response (FIR) models, which are estimated by using the stable kernel-based regularized least squares (SKRLS), for quality prediction of nonlinear dynamic industrial processes, referred to as the ELFIR-SKRLS. Meanwhile, the proposed ELFIR-SKRLS approach provides an effective way to extend the FIR-SKRLS for nonlinear dynamic system under the ensemble learning framework, and the tricky question of model order selection can be avoided. To deal with the process nonlinearity, an improved adaptive local domain partition method is employed to adaptively partition the historical data set into multiple local domains, where the optimal benchmark window selection mechanism is proposed to ensure the samples within the benchmark window belong to the same process state. In addition, a novel prediction performance-based adaptive ensemble method is proposed to combine the outputs of local FIR models, which is based on the weighted Euclidean distance-based similarity criterion and a strategy of reducing the online computation burden. The effectiveness of the proposed ELFIR-SKRLS adaptive soft sensor is demonstrated through a sulfur recover unit process application and a ladle furnace steel refining process application.

A quick condition adaptive soft sensor model with dual scale structure for dissolved oxygen simulation of recirculation aquaculture system

Dissolved oxygen (DO) is critical to recirculating aquaculture system (RAS), which is holistically affected by many nutritional, physiological, chemical, and fish behavioral processes. Modeling a DO soft sensor is an important means of providing broader sensing coverage for optimized DO control and DO sensor equipment validation. The main problem of DO soft sensor modeling is that stable long-term simulation parameters are usually insensitive to condition changes, resulting in slow adaptation. The second problem is that DO change has multiple influences. To improve condition adaptation performance and decouple multiple influence factors, a dual scale DO soft sensor modeling method is proposed to achieve dynamic condition adaptation during long-term simulation. The upper scale is a direct computer-mapping-based programmable structure (DCM-PS) model that represents the kinetics and equilibrium determined dynamic balance model for long-term simulation. The lower scale consists of recurrent neural networks (RNN) regression sub-models of aeration, feeding, and water flow. The RNN regression sub-models integrate the full connection network with RNN regression, which outputs frequency domain distributions of different conditions and predicts short-term DO expectations caused by condition change. To realize quick condition adaptation of the dual scale soft sensor model, a parameter update procedure is proposed to adapt upper scale simulation parameters to new conditions by comparing DCM-PS simulated result with RNN sub-model expectations. The instance with a sudden condition change proved that the dual scale model was able to implement quick adaptation to the changed operational condition of the RAS. The proposed method shows promise as a quick self-validation for DO sensor diagnosis, as well as for real time optimized control of DO and other parameters.

Dynamic soft sensing of organic pollutants in effluent from UMIC anaerobic reactor for industrial papermaking wastewater

With the rapid development of paper industry, the pressure of environmental pollution is going more and more serious. Recently, resource utilization of wastewater by anaerobic digestion has become a feasible way to solve this problem. In order to maintain the safe and efficient production of the process, a novel adaptive soft sensor model was developed to infer the chemical oxygen demand (COD) of paper mill effluent in this paper. First, the principal component analysis technique was performed in this model so as to eliminate the col-linearity between the process variables and accordingly obtain the low-dimensional feature principal component. Then, the least square support vector machine method was used to construct a quantitative regression model between principal component and the effluent COD. Along with it, particle swarm optimization was implemented to search for the best value of the LSSVM model parameters, namely the kernel parameters and the regularization factor. Finally, an online calibration strategy was designed to adapt to the process dynamic changes in an adaptive iterative manner. When the constructed model tested for performances in a full-scale factory, the average relative deviation and maximum deviation are 1.80% and 6.26%, respectively. The experimental results show that this proposed soft sensor model is featured with high accuracy and strong dynamic stability, and it can provide good guidance for COD prediction and optimal control of paper mill wastewate treatment.

Adaptive Soft Sensor Modeling Based on Weighted Supervised Latent Factor Analysis with Selectively Integrated Moving Windows

An adaptive soft sensor modeling method based on weighted supervised latent factor analysis is proposed. In conventional moving window based adaptive soft sensor, predictive model is constructed only with the latest process information. To fully take advantage of the past windows, a set of recent local models are integrated by the Bayes’ rule for quality estimation. However, the former built models may contain similar information about the process, and the redundancy would increase the calculation with a low-efficient accuracy improvement. Then a selecting method is proposed through a statistical hypothesis testing to determine whether a window dataset should be retained or not. In this way, the mostly informative models are left to integrate an efficient predictive model. A real industrial case demonstrates the feasibility and efficiency of the proposed adaptive soft sensor.

Double locally weighted principal component regression for soft sensor with sample selection under supervised latent structure

Just-in-time learning is an important framework for adaptive soft sensor modeling. However, the prediction accuracy strongly depends on the appropriate selection of relevant samples. Most traditional sample selection methods only consider the input information of samples, without any reference to the output information. In this paper, a new sample selection method in the supervised latent structure is proposed, in which the latent variables are highly related to the output. After that, to enhance the performance of the JITL local model, both sample importance and variable importance are taken into consideration for feature extraction in principal component regression (PCR), which is termed as double locally weighted principal component regression (D-LWPCR) in this paper. The effectiveness of the D-LWPCR model with sample selection in supervised latent structure is validated through a numerical example and an industrial blast furnace ironmaking process.

Dual learning-based online ensemble regression approach for adaptive soft sensor modeling of nonlinear time-varying processes

Soft sensors have been widely used to estimate difficult-to-measure variables in the process industry. However, the nonlinear nature and time-varying behavior of many processes pose significant challenges for accurate quality prediction. Thus a novel adaptive soft sensor, referred to as dual learning-based online ensemble regression (DLOER), is proposed for nonlinear time-varying processes. To deal with process nonlinearity, just-in-time (JIT) learning is used to build local domains and local models simultaneously while statistical hypothesis testing is employed to remove redundant local models. As a result, multiple diverse local models are constructed for characterizing various process states. Then the posterior probabilities of each test sample with respect to different local models are estimated through Bayesian inference and further set as adaptive weights to combine local predictions into a final output. Moreover, DLOER is equipped with incremental local learning and JIT learning for model adaptation, which enables recursive adaptation and online inclusion of local models, respectively. Therefore, process nonlinearity can be well handled under the local learning framework while both gradual and abrupt changes of processes can be efficiently addressed using the dual learning-based adaptation mechanism. The effectiveness of the DLOER approach is demonstrated through a fed-batch penicillin fermentation process.

Practical Use of Savitzky-Golay Filtering-Based Ensemble Online SVR

As a result of collaboration between Mitsui Chemicals, Inc. and the University of Tokyo, a soft sensor tool was developed and implemented in several plants in Mitsui Chemicals, Inc. A soft sensor is an inferential model constructed between process variables that are easy to measure (X) and process variables that are difficult to measure (y). y-values can be estimated in real time by inputting X-values into a soft sensor. To maintain predictive ability of a soft sensor to be high, we employ ensemble online support vector regression (EOSVR) model as an adaptive soft sensor model, which can adapt to both nonlinear changes and time-varying changes. Additionally, to reduce noise in estimated y-values, Savitzky-Golay (SG) filtering is used for estimated y-values. Our proposed method is called EOSVR-SG and implemented as a soft sensor tool. In this paper, we show our soft sensor tool used in real chemical plants and its execution results in which the EOSVR-SG model could estimate y-values accurately and smoothly.

Spatio‐temporal adaptive soft sensor for nonlinear time‐varying and variable drifting processes based on moving window LWPLS and time difference model

AbstractIndustrial plants often undergo different kinds of changes like variable drifts and time‐variant problems, which may cause the degradation of soft sensors. In this paper, a spatio‐temporal adaptive soft sensor modeling framework, which is based on the moving window and just‐in‐time learning (JITL) techniques, is proposed for nonlinear time‐varying processes. The JITL (locally weighted partial least squares) can adapt the model by spatial weighting technique, and the moving window technique can adapt the soft sensor model to the new process state. Furthermore, time difference (TD) model is utilized to handle the process change of variable drifts. Case studies are carried out on a numerical example and two industrial processes. The results show the effectiveness and flexibility of the proposed soft sensor framework. © 2015 Curtin University of Technology and John Wiley & Sons, Ltd.

Multi-model adaptive soft sensor modeling method using local learning and online support vector regression for nonlinear time-variant batch processes

Batch processes are often characterized by inherent nonlinearity, multiplicity of operating phases, and batch-to-batch variations, which poses great challenges for accurate and reliable online prediction of soft sensor. Especially, the soft sensor built with old data may encounter performance deterioration due to a failure of capturing the time-variant behaviors of batch processes, thus adaptive strategies are necessary. Unfortunately, conventional adaptive soft sensors cannot efficiently account for the within-batch as well as between-batch time-variant changes in batch process characteristics, which results in poor prediction accuracy. Therefore, a novel multi-model adaptive soft sensor modeling method is proposed based on the local learning framework and online support vector regression (OSVR) for nonlinear time-variant batch processes. First, a batch process is identified with a set of local domains and then the localized OSVR models are built for all isolated domains. Further, the estimation for a query data is obtained by adaptively combining multiple local models that perform best on the similar samples to the query point. The proposed multi-model OSVR (MOSVR) method provides four types of adaptation strategies: (i) adaptive combination based on Bayesian ensemble learning; (ii) online offset compensation; (iii) incremental updating of local models; and (iv) database updating. The effectiveness of the MOSVR approach and its superiority over traditional adaptive soft sensors in dealing with the within-batch and between-batch shifting dynamics is demonstrated through a simulated fed-batch penicillin fermentation process as well as an industrial fed-batch chlortetracycline fermentation process.

Adaptive Soft Sensor Based on a Moving Window Just-in-time Learning LS-SVM for Distillation Processes

In order to measure the distillation processes compositions under the time-varying conditions, an adaptive soft sensor model is proposed in this paper for composition quality prediction. In the traditional approach, the least squares-support vector machine regression (LS-SVM) method is used to build the composition quality prediction model. In order to implement on-line soft sensing of distillation compositions, the just-in-time learning strategy is used to update the dynamic model. But as more and more historical data is stored in the database, the just-in-time learning updating strategy will be very time consuming. To improve the calculation efficiency, the moving window just-in-time learning least squares-support vector machine regression (MWJIT-LS-SVM) by using a moving window with selective moving window size algorithm is proposed in this paper. The simulation results show that the proposed method achieved higher predictive accuracy than traditional methods when time-varying changes in chemical distillation processes.