The emergence of new operating modes due to changes in manufacturing demands or production technologies leads to the occurrence of multimode industrial processes. To address the multimode problem, the dictionary learning method is widely utilized due to its excellent sparse representation technique capability. Despite the abundance of offline data enabling dictionary learning to model various modes, consolidating dictionaries from individual models considerably increases model complexity. Therefore, a method called adaptive importance coding dictionary learning is proposed in this study on the basis of dictionary learning. This approach utilizes dictionary learning techniques to determine the importance of dictionary atoms by encoding frequencies and selecting significant dictionary atoms to form a new dictionary, remarkably reducing the complexity of the dictionary model. In accordance with this idea, a comprehensive framework for monitoring multimode processes, including online fault detection, mode classification, fault isolation, and new mode updates, is introduced. Experimental validations are conducted through numerical examples, simulations of continuous stirred tank reactors, and application in an actual petrochemical process. The experimental results demonstrate the effectiveness and feasibility of the proposed method in monitoring multimode processes.

Generalized Zero Shot Learning (GZSL) is proposed to solve the training data missing problem by transferring the knowledge learned in seen classes to unseen classes. Many methods project the visual features into semantic space and find their nearest neighbours among the pre-defined attributes, which has achieved significant success. However, there are two problems involved in this type of methods, one is that the projection is a many-to-one mapping, which cannot maintain the diversity of features in semantic space, the other is that searching within all classes cannot well utilize the knowledge learned in the seen classes. In this paper, we propose a novel method named Dual Prototype Relaxation (DPR) by relaxing the projection from many-to-one to many-to-many. Specifically, we add noise to the semantic prototype in response to the projection of multiple features within a class, and reconstruct the visual features with the same relaxed prototype. Besides, in order to make better use of the knowledge learned in the seen classes, an Out-of-Domain (OoD) based method is employed to first classify the feature to seen or unseen domains, and then the same DPR model is applied to recognize its category within each domain. Extensive experiments on four popular datasets are conducted and the results show that our method can outperform many linear and deep state-of-the-art methods although our method is a linear one.

Abstract This article presents a hybrid model for predicting the temperature of molten steel in a ladle furnace (LF). Unique to the proposed hybrid prediction model is that its neural network-based empirical part is trained in an indirect way since the target outputs of this part are unavailable. A modified cuckoo search (CS) algorithm is used to optimize the parameters in the empirical part. The search of each individual in the traditional CS is normally performed independently, which may limit the algorithm’s search capability. To address this, a modified CS, information interaction-enhanced CS (IICS), is proposed in this article to enhance the interaction of search information between individuals and thereby the search capability of the algorithm. The performance of the proposed IICS algorithm is first verified by testing on two benchmark sets (including 16 classical benchmark functions and 29 CEC 2017 benchmark functions) and then used in optimizing the parameters in the empirical part of the proposed hybrid prediction model. The proposed hybrid model is applied to actual production data from a 300 t LF at Baoshan Iron & Steel Co. Ltd, one of China's most famous integrated iron and steel enterprises, and the results show that the proposed hybrid prediction model is effective with comparatively high accuracy.

This paper presents a data-driven fault diagnosis and prognosis method for multi-dimensional process faults. Fault detection is carried out using a principal component analysis (PCA) model of the normal process operation data. Fault diagnosis is carried out using the fault reconstruction approach. A method for formulating the fault direction matrix for process faults is proposed. The first loading vector from the PCA model of the fault data is used to construct the fault direction matrix. The reconstructed fault magnitudes are then used to develop data-driven fault prognosis models. Both linear autoregressive models and extreme learning machine (ELM) models are developed for fault prognosis. However, linear autoregressive models fail to give acceptable long range prediction. ELM models can give accurate long range predictions of fault magnitudes and can be used in process fault prognosis. The proposed methods are demonstrated on a simulated continuous stirred tank reactor process.

With industrial production processes becoming more and more sophisticated, traditional fault diagnosis systems may be insufficient to meet current industrial diagnostic performance requirements. In order to improve fault diagnosis performance, this paper proposes an enhanced neural network based fault diagnosis system by integrating Andrews plot and Autoencoder. Features are first extracted from on-line measurements by Andrews plot and the high-dimensional features are compressed by autoencoder to an appropriate dimension, which are then fed to a neural network for fault classification. Application to a simulated CSTR process demonstrates that the proposed method can give more reliable and earlier diagnosis than the traditional neural network based fault diagnosis method.

Slow feature analysis is a technique that extracts slowly varying latent variables from a dataset. These latent variables, known as slow features, can capture underlying dynamics when applied to process data, leading to improved generalisation when a data-driven model is built with these slow features. A method utilising slow feature analysis with neural networks is proposed in this paper for improving generalisation in nonlinear dynamic process modelling. Additionally, a method for selecting the number of dominant slow features using changes in slowness is proposed. The proposed method is applied to creating a soft sensor for estimating polymer melt index in an industrial polymerisation process to validate the method’s performance. The proposed method is compared with principal component analysis-neural network and a neural network without any latent variable method. The results from this industrial application demonstrate the effectiveness of the proposed method for improving model generalisation capability and reducing dimensionality.

Heat integrated distillation columns (HIDiC) can save around 50% of energy consumption compared to conventional distillation columns. To reach this amount of energy saving, the HIDiC should be properly controlled. The high thermally coupled nature of HIDiC leads to high loop interactions making HIDiC difficult to control. In order to improve control performance, inferential active disturbance rejection control (ADRC) for HIDiC is proposed in this paper. The proposed control scheme integrates ADRC strategy with inferential feedback control. Tray temperatures are used as secondary measurements to estimate the product compositions. Due to the strong collinearity in tray temperatures, principal component regression (PCR) is used to build the soft sensor from process operation data. The proposed technique is applied to a simulated HIDiC for separating Benzene-Toluene mixture. Simulation results demonstrate the effectiveness of the proposed approach.

This paper presents an integrated disturbance observer based control and inferential control of a binary distillation column. Disturbance observers have been shown being capable of providing excellent disturbance attenuation. Research into the application of inferential control methods in distillation columns have achieved a much higher quality of control in comparison to the feedback control using composition analyzers which possess substantial measurement delays. This paper takes the novel approach of integrating two frequency domain disturbance observers with multiple tray temperature based inferential estimations of top and bottom product compositions. In order to cope with colinearities in tray temperature measurements, principal component regression (PCR) and partial least square (PLS) regression are used in the soft sensor development. The number of principal components in the PCR model and the number of latent variables in the PLS model are determined using cross validation. The frequency domain disturbance observers have been designed by optimizing the value of the filter time constants within a first order filter, in order to maximize disturbance rejection. The proposed control approach is applied to a binary distillation column for separating methanol and water mixture.

This paper presents a multivariable inferential active disturbance rejection control (ADRC) method for product composition control in distillation columns. The proposed control strategy integrates ADRC with inferential feedback control. In order to overcome long time delay of gas chromatography in measuring product compositions, static and dynamic estimators for product compositions have been developed. The top and bottom product compositions are estimated using multiple tray temperatures. In order to overcome the colinearity issue in tray temperatures, principal component regression is used to build the estimator. The proposed technique is applied to a simulated methanol-water separation column. It is shown that the proposed control strategy gives good setpoint tracking and disturbance rejection control performance.

A new iterative learning control method is presented for the trajectory tracking control of linear time-variant systems. This new method does not need detailed knowledge of the controlled system. However, a reference batch is designed, in which some small change of the input trajectory in the current batch is applied to the controlled system, and then its output trajectory is obtained. The ratio of the input change and corresponding output change in the current and reference batch is used as the gain in the learning law for calculating the inputs of the next batch. Thus the method can track the desired trajectory successfully while the iteration is running on. The convergence of the proposed method is analyzed, and the method is validated on a typical linear time-variant system.

This paper presents an iterative learning control strategy for a fed-batch fermentation process using linearised models identified from process operational data. Off-line calculated control policies for batch fermentation processes may not be optimal when implemented on the processes due to model plant mismatches and/or the presence of unknown disturbances. In order to overcome the effect of model plant mismatches and unknown disturbances, a batch to batch iterative learning control strategy is developed to modify the control actions for the next batch using the information obtained form current and previous batches. The control policy updating is calculated using a model linearised around a reference batch. In order to cope with process variations and disturbances, the reference batch can be taken as the immediate previous batch. After each batch, the newly obtained process operation data is added to the historical process data base and an updated linearised model is re-identified. Since the control actions during different stages of a batch are usually correlated, it is proposed in this paper that the linearised model can be identified from partial least square regression. The proposed technique has been successfully applied to a simulated fed-batch fermentation process.

A recurrent neuro-fuzzy network based strategy for batch process modeling and optimal control is presented in this paper. The recurrent neuro-fuzzy network allows the construction of a "global" nonlinear long-range prediction model from the fuzzy conjunction of a number of "local" linear dynamic models. In this recurrent neuro-fuzzy network, the network output is fed back to the network input through one or more time delay units. This particular structure ensures that predictions from a recurrent neuro-fuzzy network are long-range or multi-step-ahead predictions. Long-range predictions are particularly important for batch processes where the interest lies in the product quality and quantity at the end of a batch. To enhance batch process control and monitoring, a model capable of predicting accurately the product quality/quantity at the end of a batch is required. Process knowledge is used to initially partition the process nonlinear characteristics into several local operating regions and to aid in the initialization of the corresponding network weights. Process input output data is then used to train the network. Membership functions of the local regimes are identified and local models are discovered through network training. An advantage of this recurrent neuro-fuzzy network model is that it is easy to interpret. This helps process operators in understanding the process characteristics. The proposed technique is applied to the modeling and optimal control of a fed-batch reactor.

This paper presents two soft-sensing models for predicting the product yields profile and the cracking degree of an ethylene pyrolysis furnace. The model based on single neural network with only one hidden layer trained by Levenberg-Marquardt algorithm with regularisation was first developed. It was found that the single neural network lack generalisation capability in that they can give undesirable performance when applied to unseen data. To improve the generalisation capability of the soft-sensing model, multi-model soft-sensors based on bootstrap aggregated neural networks with sequential training are used. In the sequential training of bootstrap aggregated networks, the first network is trained to minimise its prediction error whereas the rest of the networks are trained not only to minimise their prediction errors but also minimise the correlation among the trained networks. The overall output is obtained by combining all the individual networks. Application results show that the multi-model soft-sensors possess good generalisation capability in that they give good performance when applied to unseen data.

A single neural network model developed from a limited amount of data usually lacks robustness. Thus combining multiple neural networks can enhance the neural network model performance. In this paper, a Bayesian combination method is developed for nonlinear dynamic process modelling and compared with simple averaging. Instead of using fixed combination weights, the estimated probability of a particular network being the true model is used as the combination weight for combining that network. A nearest neighbour method is used in estimating the network error for a given input data point, which is then used in calculating the combination weights for individual networks. The prior probability is estimated using the SSE of individual networks on a sliding window covering the most recent sampling times. It is shown that Bayesian combination generally outperforms simple averaging.

A type of recurrent neuro-fuzzy network is proposed in this paper to build long-term prediction models for nonlinear processes. The process operation is partitioned into several fuzzy operating regions. Within each region, a local linear model is used to model the process. The global model output is obtained through the centre of gravity defuzzification which is essentially the interpolation of local model outputs. This modeling strategy utilizes both process knowledge and process input output data. Process knowledge is used to initially divide the process operation into several fuzzy operating regions and to set up the initial fuzzification layer weights. Process input output data are used to train the network. Network weights are such trained so that the long-term prediction errors are minimized. Through training, membership functions of fuzzy operating regions are refined and local models are learn. Based on the recurrent neuro-fuzzy network model, a novel type of nonlinear model-based long range predictive controller can be developed and it consists of several local linear model-based predictive controllers. Local controllers are constructed based on the corresponding local linear models and their outputs are combined to form a global control action by using their membership functions. This control strategy has the advantage that control actions can be calculated analytically avoiding the time consuming nonlinear programming procedures required in conventional nonlinear model-based predictive control. The techniques have been successfully applied to the modeling and control of a neutralization process.