Search

Experiment design for Identification is usually based on asymptotic theory, where an infinite number of samples is assumed. However, such an assumption does not hold in practical cases, and hence, the nonasymptotic properties of system Identification should be considered. This paper proposes a new method for experiment design for Identification based on the nonasymptotic confidence region of the system parameters calculated using the sign-perturbed-sums (SPS) method for multivariate autoregressive exogenous input (ARX) systems. The objective function based on the volume of the confidence region is introduced in the proposed method. Moreover, the proposed optimization problem is solved using Bayesian optimization because the proposed objective function is calculated from the data obtained only after the experiment. The validity of the proposed method was assessed in an experimental case study of a three-tank system, where the proposed method was compared with the existing D-optimal method. As a result, the model obtained using the proposed method reduced the mean squared control error of model predictive control by 22.9% from that of the existing method.

Chemical process monitoring is essential for product quality, plant efficiency, and safety. Conventional methods often prove inaccurate, particularly when dealing with nonlinear process behaviour. This paper presents a new approach that combines data-driven nonlinear system identification using smoothed L1regularisation and a state prediction method using a sequential importance resampling (SIR) particle filter to provide a basis for process monitoring. The results obtained from the polycondensation reaction in an operator training simulator (OTS) with real process conditions validate the effectiveness of the method in detecting anomalies, addressing challenges in nonlinear process modeling, and reliable state prediction for chemical process monitoring.

The predictive performance of soft sensors deteriorates over time which is called the performance change of a soft sensor. These changes occur due to differences between the current characteristics of the process or plant and the soft sensor model. The deviation is a type of plant-model mismatch (PMM). Initially, this mismatch may be acceptable. However, over time, the PMM can become so large that it affects the prediction quality of the soft sensor and may become unacceptable. This paper develops a new method to evaluate the impact of PMM on closed-loops with soft sensors. Using coprime factorisation and small-gain theory, a performance-change index is developed to characterise the PMM-induced performance degradation. Then, a performance-based online PMM detection method is proposed using this performance-change index. To validate the effectiveness of the proposed algorithm, we use a numerical example and a continuous stirred tank reactor (CSTR). It is shown that that the proposed index can detect the change of the PMM.

Causality is a fundamental relationship in the physical world, around which almost all activities of human life revolve. Causal inference refers to the process of determining whether an event or action caused a specific outcome, which involves the evaluation of cause-and-effect relationships in data. This paper presents a new approach to discover latent causal representations of crucial variables in easy-to-obtain data. The proposed method takes a form of trade-off between compression of input data and the causality between the learnt latent variables and critical variables, thereby removing the irrelevant information contained in input data and obtaining the decoupled, strongest causal factors. By introducing variational bounds and specific configurations, the optimisation objective is relaxed to a tractable problem. The approach compacts causal discovery and inference into one model, which is flexible to downstream tasks and parsimonious in the parameters. A case study on an exhaust-emission dataset shows that the proposed method improves the predictive performance over the baseline model, which is a variational information bottleneck model with the same hyperparameters.

In modern process industries, elaborate monitoring and isolation of various disturbances and faults are needed for reliable and efficient system operation. The classic process-monitoring and fault-diagnosis methods can grasp the correlation between variables, and thus, only take care of abnormal situations caused by the corruption of the correlation relationship. However, dynamics anomalies are even more noteworthy as they reflect more internal details of the system dynamic behaviour under specific situations, and more importantly, can cause severe failures and spread to a broader range of areas while evolving over time. In this paper, a monitoring-and-isolation strategy is proposed to concurrently detect and isolate faults of static deviations and dynamic anomalies. The natural sparsity of the faulty variables is used to overcome the limitations of unknown fault directions and insufficient erroneous measurements, thereby translating the isolation problem into a quadratic programming problem with a sparsity constraint and solved by the least absolute shrinkage and selection operator (LASSO). The case study shows the advantages of the proposed method in monitoring and isolating static deviations and dynamic anomalies.

With the increase in the available data, it becomes increasingly important to develop automatic methods that can extract valuable nuggets of information from the dregs of uninformative and useless information for use in system identification. This paper presents an overview and summary of this field's current state of the art. A MATLAB programme is presented that can implement data quality assessment. A brief tutorial is presented using industrial kerosene freeze-point data to partition the data set into good and bad regions for system identification. A model is developed using the partitioned data. It is shown that the resulting four models can accurately predict the kerosene freeze point in their respective regions and across the data set.

As processes become more complex and the need to measure each and every variable becomes more critical, the ability of physical sensors to always provide the sufficient accuracy and sampling time can be difficult. For many complex systems, such as nonideal mixtures, multiphase fluids, and solid-based systems, it may not be possible to even use a physical sensor to measure the key variables. For example, in a multiphase fluid, the concentration or density may only be able to be accurately estimated using a laboratory procedure that can only produce a limited number of samples. Similarly, the quality variables of steel may only be determinable once the final steel product has been produced, which limits the ability to effectively control the process with small time delays. In such cases, recourse has to be made to soft sensors, or mathematical models of the system that can be used to forecast the difficult-to-measure variables and allow for real-time process monitoring, control, and optimisation. Although the development of the soft-sensor model is well-established, the various applications and use cases have not been often considered and the key challenges examined. It can be seen that soft sensors have been applied to a wide range of processes from simple, chemical engineering systems to complex mining processes. In all cases, major improvements in the process operations have been observed. However, key challenges remain in updating the soft-sensor models over time, combining laboratory measurements, especially when they are infrequent or of uncertain quality, and the development of soft sensors for new conditions or processes.

Automated driving systems (ADS) have the potential to offer a safe and efficient future for mobility. At the beginning of the design process of an ADS, the operational limits have to be defined using the operational design domain (ODD). Nonetheless, the field of ODD has only become popular recently, and the necessary regulations, standards, and development approaches are still emerging. Current research contributions in the domain of ODD have fragmented in recent years and have not followed a concrete direction. This survey examines a large proportion of the recent research in the domain of ODD by systematically identifying subject areas and categorising relevant publications, thereby integrating the approaches and showing an overview of the emerging topic area. Furthermore, it identifies existing gaps in the ODD research that need to be considered. Finally, the paper suggest relevant future development of ODD.

Recently developed fault classification methods for industrial processes are mainly data-driven. Notably, models based on deep neural networks have significantly improved fault classification accuracy owing to the inclusion of a large number of data patterns. However, these data-driven models are vulnerable to adversarial attacks; thus, small perturbations on the samples can cause the models to provide incorrect fault predictions. Several recent studies have demonstrated the vulnerability of machine learning methods and the existence of adversarial samples. This paper proposes a black-box attack method with an extreme constraint for a safe-critical industrial fault classification system: Only one variable can be perturbed to craft adversarial samples. Moreover, to hide the adversarial samples in the visualization space, a Jacobian matrix is used to guide the perturbed variable selection, making the adversarial samples in the dimensional reduction space invisible to the human eye. Using the one-variable attack (OVA) method, we explore the vulnerability of industrial variables and fault types, which can help understand the geometric characteristics of fault classification systems. Based on the attack method, a corresponding adversarial training defense method is also proposed, which efficiently defends against an OVA and improves the prediction accuracy of the classifiers. In experiments, the proposed method was tested on two datasets from the Tennessee–Eastman process (TEP) and steel plates (SP). We explore the vulnerability and correlation within variables and faults and verify the effectiveness of OVAs and defenses for various classifiers and datasets. For industrial fault classification systems, the attack success rate of our method is close to (on TEP) or even higher than (on SP) the current most effective first-order white-box attack method, which requires perturbation of all variables.

In this paper, a data-driven or model-free approach is presented to design a fault detection system of continuous-time linear time-invariant (LTI) systems based on input and output data in the time domain. The main idea is to directly identify the subspaces and their related matrices relevant for parity-space-based residual generation based on a modulated output equation by use of modulation functions and their properties. Therefore, the explicit model identification of the process for a model-based approach in a conventional two-step procedure can be avoided saving design effort especially for large-scale systems. A simulation of the resulting fault detection system is provided showing the effectiveness of the design approach.

Industrial processes are measured and controlled using high-dimensional process variables, but its overall operation is usually characterised by low-dimensional patterns. The changes in the pattern are dominated by three features: free motion, controlled motion, and uncertainty. In this paper, all three features are taken into consideration to propose a new probabilistic dynamic-controlled latent variable (PDCLV) model structure using a dynamic Bayesian network for process modelling in the pattern space. To this end, the linear dynamic system characterised by control inputs is introduced, and the expectation maximisation algorithm is specially designed for learning the PDCLV model. Benefitting from the dynamic causality between control inputs and the explicit modelling of the pattern, a method for pattern-based stochastic model predictive control (SMPC) is implemented successfully to realise process optimisation. A case study on an industrial boiler combustion process demonstrates the benefits of the proposed PDCLV structure for pattern-space modelling and pattern-based SMPC.

Slow feature analysis (SFA) has attracted much attention as a method for dynamic modelling. However, SFA has an inherent limitation in that it assumes that the dynamic behaviour is linear. In this paper, a new criterion for SFA in general dynamic systems is defined based on the motivation of maximising the information retained during system evolution, which is called EVOLVE·INFOMAX. The theoretical properties of this new criterion are rigorously justified, the optimisation function under EVOLVE·INFOMAX is proposed, and a tailored algorithm based on neural networks is designed. The case study on a simulated data set and the Tennessee Eastman process benchmark shows that the proposed method has better performance to extract slow features of nonlinear dynamical systems.

An online model-based fault detection and isolation method for salient-pole permanent magnet synchronous motors over a finite horizon is proposed. The proposed approach combines parity-space-based residual generation and modulation-function-based filtering. Given the polynomial model equations, the unknown variables (i.e.the states, unmeasured inputs) are eliminated resulting in analytic redundancy relations used for residual generation. Furthermore, in order to avoid needing the derivatives of measured signals required by such analytic redundancy relations, a modulation-function-based evaluation is proposed. This results in a finite-horizon filtered version of the original residual. The fault detection and isolation method is demonstrated using simulation of various fault scenarios for a speed controlled salient motor showing the effectiveness of the presented approach.

Industrial processes are large scale, highly complex systems. The complex flow of mass and energy, as well as the compensation effects of closed-loop control systems, cause significance cross-correlation and autocorrelation between process variables. To operate the process systems stably and efficiently, it is crucial to uncover the inherent characteristics of both variance structure and dynamic relationship. Long-term dependency slow feature analysis (LTSFA) is proposed in this paper to overcome the Markov assumption of the original slow feature analysis to understand the long-term dynamics of processes, based on which a monitoring procedure is designed. A simulation example and the Tennessee Eastman process benchmark are studied to show the performance of LTSFA. The proposed method can better extract the system dynamics and monitor the process variations using fewer slow features.

The satellite sea surface temperature (SST) measurement is based on the detection of ocean radiation using microwave or infrared wavelengths within the electromagnetic spectrum. The radiance of individual wavelengths can be converted into brightness temperatures for using in SST determination. The calibration and validation of the determined SST data require reference measurements from in-situ observations. These in-situ observations are from various platforms such as ships, drifters, floats and mooring buoys and require a high measurement accuracy. This paper presents an investigation about the possibility of using a glider as in-situ platform. A glider is a type of autonomous underwater vehicle (AUV) which can log oceanographic data over a period of up to one year by following predetermined routes. In contrast to buoys, a glider allows a targeted investigation of regional anomalies in SST circulations. To assess the quality of SST observations from a glider, logged data from a glider mission in the Atlantic Ocean from 2018 to 2019 and corresponding satellite SST data were used. The influence of variables (e.g. measurement depth, latitude, view zenith angle, local solar time) of the bias between satellite and glider SST data was investigated using sensitivity analysis. A new and efficient distribution-based method for global sensitivity analyzes, called PAWN, was used successfully. Interested readers will find information about its operation principle and the usage for passive observations where only “given-data” are available.

Switched reluctance motors (SRM) are an inherent part in robotics and automation systems where energy and cost efficiency is required. This motor type has no windings and permanent magnets on the rotor which results in a simple and robust structure. However, SRMs require a complex electronic control system to generate a specified number of voltage pulses for each motor phase. This paper presents the signal generation of multiple phases using only one current sensor in an asymmetric half bridge (AHB). In addition to maintain the predetermined phase voltages, sufficient current measurement windows and a minimal current ripple for the individual phases are further optimization criteria for signal generation. The generation of a state vector which controls the individual semiconductor for each motor phase to achieve a required phase voltage and simultaneously fulfill the multi-objective optimization criteria is challenging. Due to the vast number of possible solutions, a genetic algorithm (GA) was used to find state combinations that are suitable for the formulated optimization criteria. The results were discussed and recommendations about the genotype representation and the used genetic operators were given. Interested readers will find detailed information about the software technical implementation using the Global Optimization Toolbox from MATLAB.

As the amount of data stored from industrial processes increases with the demands of Industry 4.0, there is an increasing interest in finding uses for the stored data. However, before the data can be used its quality must be determined and appropriate regions extracted. Initially, such testing was done manually using graphs or basic rules, such as the value of a variable. With large data sets, such an approach will not work, since the amount of data to tested and the number of potential rules is too large. Therefore, there is a need for automated segmentation of the data set into different components. Such an approach has recently been proposed and tested using various types of industrial data. Although the industrial results are promising, there still remain many unanswered questions including how to handle a prioriknowledge, over- or undersegmentation of the data set, and setting the appropriate thresholds for a given application. Solving these problems will provide a robust and reliable method for determining the data quality of a given data set.

An online model-based fault detection and isolation method for salient permanent magnet synchronous motors is proposed using the parity-space approach. Given the polynomial model equations, Buchberger’s algorithm is used to eliminate the unknown variables (e.g. states, unmeasured inputs) resulting in analytic redundancy relations for residual generation. Furthermore, in order to obtain the derivatives of measured signals needed by such a residual generator, robust exact differentiators are used. The fault detection and isolation method is demonstrated using simulation of various fault scenarios for a speed controlled salient motor showing the effectiveness of the presented approach.

Difficult-to-obtain variables in industrial applications have led to the rise of soft sensors, which use prior system information and measurements to estimate these difficult-to-obtain variables. In real systems, the measurements that need to be estimated by a soft sensor are often infrequently measured or delayed. Sometimes, these delays and sampling time are variable in time. Though there are papers considering soft sensors in the presence of time delays and different sampling times, the variation of those parameters has not been considered when evaluating the adequacy of the soft sensors. Therefore, this paper will evaluate the impact of such variations for a data-driven soft sensor and propose modifications of the soft sensor that increase its robustness. The reliability of its estimate will be shown using the Bauer-Premaratne-Durán Theorem. Furthermore, the soft sensor will be simulated applying it to a continuous stirred tank reactor. Simulation showed that the modified soft sensor gives good estimates, whereas the traditional soft sensor gives an unstable estimate.

This paper proposes a new data-driven method for addressing fault tolerant control problems. Unlike existing model-based or data-driven methods, the proposed method realizes fault tolerant control without knowing any system model parameter or performing any model identification. Q-learningsevers as a key tool in this procedure. In addition, unlike conventional Q-learningalgorithms used in the control community, the new proposed one can be applied in weakly stochastic environments, which facilitates the application of the fault tolerant control method in real industrial occasions. A DC motor simulation example demonstrates the effectiveness of the proposed method.

This paper presents a new solution for the design of a controller for a process using exclusively input/output (I/O) data, which enables the overall system to exhibit high control performance including tracking performance and system robustness in case of deterministic additive disturbances. First, a data-driven model is constructed based on deadbeat diagnostic observer/residual generator. Next, a new state, which is equivalent to the integral of the system tracking error, combined with process input and output information is used to define the performance criterion. Minimising this performance criterion using an online control policy gives an optimal feedback and feedforward controller. The effectiveness of the proposed approach for industrial implementations is shown using a DC-motor simulation.

In this paper, two common statistics, the T2and the Q statistics, for fault detection and process monitoring are compared. Specifically, the geometric relationship between the T2statistic and 3 common forms of the Qstatistics is analysed. Furthermore, using the false alarm rate (FAR) and the fault detection rate (FDR), the fault detection performance of both statistics is quantified and compared. The results show that, for a given significance level, the T2statistic has the best overall FDR.

Traditionally, closed-loop system identification in the absence of external excitation has focused on determining the identifiability of plant model based on the interplay between the orders of the different polynomials present. However, due to the presence of the controller, it is possible that the system may not be globally identifiable at a given complexity, but may be locally identifiable given certain restrictions or relationships between the individual parameters present in the system. In order to obtain parameter-specific solutions to the problem, many different approaches can be taken. In this paper, the focus will be primarily on an expectation-based analysis of the Fisher information matrix to determine parameter-based constraints on closed-loop identification. Additionally, a method for determining an analytical expression for the expectation operation will be presented. The proposed approach will be illustrated using a first-order autoregressive model with exogenous input controlled by a lead–lag controller. Monte Carlo simulations are used to validate the resulting constraints. [ABSTRACT FROM AUTHOR]

Time delay and sampling rate are two conditions that can make the design and implementation of soft sensors difficult due limiting the amount of information that can be obtained about the true process. Another wrinkle is the fact that neither of these two parameters is necessarily constant, with the value changing due to changes in the process. One such area of concern is the hot steel mill rolling process, where variable time delay can lead to problems with controlling the process. This paper seeks to examine the appropriate design of the bias update term of the soft sensor in light of variable, but known, time delays for a process operating with constant sampling rate and for both open- and closed-loop conditions. Using a mathematical approach to the problem, the previous results are extended to consider such a case. It is shown that if the process has variable time delay, then designing the bias update term in the soft sensor is very important, as it can have implications for process stability and trackability. Simulations are included to show the impact of different bias update terms.

This paper proposes two different approaches to the problem of modelling the key performance indicators (KPIs) in a hot strip mill process (HSMP). The first approach takes into consideration the inherent dynamics of the process and develops a nonlinear iterative dynamic partial least squares (PLS) model. The second approach takes into consideration the multibatch property of HSMP to develop an appropriate multibatch modelling framework. Both approaches can be integrated into the dynamic model framework. The proposed framework is examined using industrial data obtained from a steel plant, where it provides a decent model for this process. However, there exists the potential for improvements including dealing with interbatch variability.

In the development of a rigorous and complete procedure for automating the performance of data-driven process identification, there is a need to consider data quantisation. Such an issue can arise when the sensors have not been properly calibrated for the range of values experienced in the actual process. Through a detailed mathematical analysis of the problem, it is shown that the ratio between the variance of the signal and the gap between quantisation levels strongly influences the ability to identify a process. Using this criterion, a data quantisation index is proposed that allows for the effect of data quantisation on the data system to be quantified. Monte Carlo simulations of a closed-loop system with different system properties is examined to show that the proposed index can accurately distinguish between good and bad data quantisation.

In industry, system identification is a time-consuming exercise that impacts the profitability and safety of the plant. One way to avoid this problem is to use stored historical process data in estimating the required process models. Given the large amount of data available, which is often corrupted due to process disruptions, loss of information, and poor data quality, automated segmentation of the data set would be an invaluable asset. Recently, two different methods have been proposed to accomplish this task: one based on Laguerre models and one based on autoregressive with exogenous input (ARX) models. In this paper, the Laguerre approach will be analysed and it will be shown that the results are dependent on selecting appropriate Laguerre model parameters and input signals, while relatively insensitive to the variance thresholds. Furthermore, this approach has a tendency to overpartition the data set based on the smallest changes in the process. Therefore, in order to decrease the number of models identified it is proposed to couple this method with an entropy-based metric for determining similar models. Based on simulations that include this entropy metric, it is shown that a reduction in model partitions is obtained.

In industry, in order to store the reams of data that are collected from all the different flow, level, and temperature sensors, the fast-sampled data are very often downsampled before being stored in a data historian. This downsampled or even compressed data are, then, used by process engineers to recover the appropriate process parameters. However, little has been written about the effects of the sampling on the quality of the model obtained. Therefore, in this paper, the effects of sampling time are investigated from both a theoretical and practical perspective using results that come out of the theory of closed-loop system identification with routine operating data. It is shown that, if the ratio between the time delay and sampling time is sufficiently large, then it is possible to recover the true system parameters. The most common industrial processes that fulfill this constraint are temperature control loops. On the other hand, for processes, such as flows, pressures, or levels, with almost no time delay, then the sampling time must be extremely small in order to identify the process parameters. These results suggest that the sampling time has an important bearing on the quality of the model estimated from routine operating data. Using an experimental set-up with a heated tank, the effect of time delay on the identification of the true continuous time parameters was considered for different sampling times. It was shown that increasing the sampling time above a given threshold resulted in identifying an incorrect model.

The pioneering multivariate statistical process monitoring (MSPM) methods use the Q -statistic as an alternative for the T 2 -statistic to detect faults occurring in the residual subspace spanned by the process variables, since directly using T 2 for this subspace can lead to numerical problems. Such use has also spread to current work in MSPM field. However, substantial improvement of computational resource has sufficiently mitigated the numerical problem, which, thus, leads to a need to assess their detectability when using in the same position. This paper seeks to solve this historical issue by examining the two statistics in light of the fault detection rate (FDR) index to assess their performance when detecting both additive and multiplicative faults. Theoretical and simulation results show that the two statistics have different impacts on computing the FDR. Furthermore, it is shown that, the T 2 -statistic performs, in terms of the FDR, better at detecting most additive and multiplicative faults. Finally, based on the achieved results, a remedy to the interpretation of traditional MSPM methods are given. [ABSTRACT FROM AUTHOR]