Search

Failure prognostics can improve industrial systems’ availability and reliability by determining the occurrence of failure and estimating the system’s remaining useful life (RUL) before deterioration. This paper presents a prognostic method based on data-driven and degradation model approaches to accurately predict the RUL of a scale replica system under different operating conditions. For this purpose, vibration data collected off a faulty system over many days is analyzed in both time and frequency domains to excavate data features related to the degradation and then identify the health indicator (HI) and the failure mechanism. At that point, the constructed HI, thanks to the feature importance ranking and principal component analysis (PCA) techniques, is used as a mathematical input for an exponential degradation model. Finally, the results are validated using real-world vibration-based degradation information leading to an effective predictive strategy for industrial systems.

In this paper, a numerical procedure to obtain the pseudo-polynomial characteristic equation of a commensurate time-delay system is proposed. The method is formulated in terms of an interpolation problem, and it is based on the generation of a suitable set of random numbers. Some results, showing how the random number generation method affects the procedure convergence are also reported. This aspect is investigated in the paper and a suitable strategy to guarantee the problem solution is proposed. A numerical example is reported to outline the method behavior.

In this paper, a black-box model, suitable for being used as a Soft Sensor for an Ionic Polymer-Metal Composite, is introduced. Applications of such materials, in many fields, such as bio-inspired robotics, medicine, and aerospace, have been reported. The proposed Soft Sensor is a data-driven Nonlinear Finite Impulse Response model, implemented by using a Deep Neural Network, where the output is estimated only on the basis of input past samples. The model has been selected with the aim of eliminating any hardware position sensor and, therefore, reducing the complexity of the positioning system.

Deep Neural Network (DNN) based Soft Sensors (SSs) have been demonstrated as successful alternatives to other data-driven structures. Here, a dynamic DNN based SS is proposed for the estimation of the Research Octane Number (RON) for a Reformer Unit in a refinery. The SS is required to estimate the RON when the plant operates in two different working conditions. Nonlinear Finite Inputs Response (NFIR) models have been investigated. The regressors in the models have been selected according to a cross-correlation analysis between candidate inputs and the RON value. The performance of the proposed SSs has been compared with previously designed deep structures, based on different dynamic first level models, coupled with a fuzzy algorithm.

In this paper the design of a Soft Sensor to estimate the sulphur concentration in a desulphuring unit of a refinery operating in Sicily is described. In particular the problem of the input variables selection is addressed by comparing two different methods. The first method is based on the generalization of the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm to nonlinear models implemented by using Multi-Layer Perceptron (MLP) neural networks. The second one is based on the Lipschitz's quotient analysis. A comparison between the performance and the computational complexity exhibited by the two methods is discussed. The results show that the LASSO-MLP algorithm allows to construct a model with a low number of input variables, thus reducing computational complexity and measuring costs.

In this paper a new approach for safety monitoring of dangerous gases in the industrial plants is proposed. A single artificial neural network is used for determination of the gas concentrations based on sensor array measurements, performing at the same time compensation of the temperature and humidity influence on the sensor outputs. The obtained results show good accuracy in gas concentration estimation, enabling efficient risk warning.

In this paper evolutionary computation algorithms are applied to select optimal parameters in model order reduction for linear systems. In particular a parameterized set of reduced order model is obtained by using a Parametric Generalised Singular Perturbation Approximation of a balanced realization. The optimization algorithm is then used to select the parameter set that minimize a suitable performance index. Numerical examples are reported in comparison with other model order reduction methods.

After the fabrication of several devices showing memristive switching behavior, recently a growing interest to the realization of dynamical nonlinear circuits based on memristors has been manifested. Currently, many memristor circuits have been mostly conceived on the basis of theoretical memristor models. However, in order to analyze the dynamical behavior of memristor circuits with real components and to implement them, the characteristics of the fabricated devices have to be included in the models used. To this aim, a compact data-driven model is proposed in this paper. The model is based on neural networks and is derived starting from experimental measurements performed on printed TiO2 memristors.

In this paper a new class of orthonormal functions which includes as particular case the Laguerre filters are introduced. These functions are defined as the product of a fixed transfer function of order n and of an all-pass filter of order n × h for any n and h. The orthogonality of these functions is proven in the general case. Moreover, the singular values of the sum of the first Nh members of this class are shown to be all equal to the orthonormalization constant.

In this paper different strategies to model Ionic Polymer-Polymer Composite (IP2C), used as actuator, are compared. Starting from some previous results regarding the ionic polymer metal composites (IPMCs) modeling, a linear gray-box model has been determined for an IP2C actuator. Moreover linear and nonlinear black-box models have been identified from experimental data. A comparison among developed models has been performed in order to determine the model that better describes the actuator behaviour.

The paper deals with the design of a data driven soft sensor, able to estimate propylene percentage in the bottom flow of a Propylene Splitter showing seasonal variations. Experimental data have been collected in a refinery in Sicily. The soft sensor is intended to replace the online analyzer during maintenance, in order to guarantee the desired plant performance. In order to take into account seasonal variations, two models have been designed and implemented by using MLP neural networks. Seasonal variations are mainly related to the temperature of seawater used in the plant for cooling that shows significant variations along the year. A set of fuzzy rules has been designed in order to allow a soft transition between the winter and the summer models. A comparison is performed with a neural model working on the whole data set, i.e. covering both winter and summer collected data.

A new strategy to realize an FPGA implementation of a soft sensor for an industrial process is proposed. In order to cope with the problem of small data sets in the identification of a non linear model the proposed approach is based on the integration of bootstrap re-sampling, noise injection and stacked neural networks (NNs), using the Principal Component Analysis (PCA). The aggregated final NN-PCA system has been implemented on Field Programmable Gate Array (FPGA). The proposed method has been applied to develop a soft sensor for the estimation of the freezing point of kerosene in an atmospheric distillation unit (topping) working in a refinery in Sicily, Italy.

In this paper the problem of regressors selection in Virtual Instruments (VI) design is addressed The VI is designed to replace the on line analyzer of a Sulfur Recovery Unit (SRU) of a large refinery located in Sicily during maintenance operations. It is designed by using nonlinear MA models implemented by a MLP neural network. The use of a set of cross-correlation functions, proposed by Billings and Voon to evaluate the performance of nonlinear models is used to select the regressors of the discrete-time NMA model by implementing an automatic regressor selection algorithm. The designed Soft Sensor has been implemented at the refinery to be tested on line.

In this paper we describe, test, and compare the performance of a number of techniques used for outlier detection to improve modeling capabilities of soft sensors on the basis of the quality of available data. We analyze methods based on standard deviation of population, on residuals of a linear input-output regression, on the structure correlation of the data, on principal components and partial least squares (both linear and nonlinear) in multi dimensional space (2D, 3D, 4D), on Q and T2 statistics, on the distance of each observation from the mean of the data, and on the Mahalanobis distance. We apply techniques for outlier detection both on a fictitious model data and on real data acquired from a sulfur recovery unit of a refinery. We show that outlier removal almost always improves modeling capabilities of considered techniques.

In this paper we compare a number of strategies to cope with the problem of small data sets in the identification of a nonlinear process. Four methods are analyzed: expansion of the training set by adding zero-mean fixed-variance Gaussian noise, expansion of the training set by adding zero-mean gaussian noise variance variable according with signal amplitude, integration between bootstrap method and stacked neural networks, and a new method based on the integration of bootstrap method, of the noise injection method, and of stacked neural networks. Such methods have been applied to develop a soft sensor for a thermal cracking unit working in a refinery in Sicily, Italy.

In the paper the Soft Sensor design strategy for an industrial process, via neural NMA model, is described. In details, the hydrogen sulphide (H2S percentage) in the tail stream of a Sulfur Recovery Unit (SRU) of a refinery located in Sicily, Italy, is estimated by a Soft Sensor, that was designed to replace the online analyzer during maintenance operations. A general design strategy, based on the automatic selection of regressors of a NMA model is proposed. It is based on the minimization of the Lipschitz numbers by a Genetic Algorithms (GA) approach. A comparative analysis with an empirical model, developed on the basis of suggestions given by plant experts, is included to show the validity of the proposed procedure.

A microcontroller based system has been developed in order to characterize, in real time, magnetoencephalographic (MEG) signals. The signal are analyzed by using nonlinear data analysis method to extract the dinfin parameter. The signals under consideration are acquired on the whole-head MEG acquisition system and are analyzed in order to characterize spatiotemporal pattern. The microcontroller based implementation represents a low cost stand-alone solution for the analysis of any physiological signals (MEG, EEG, ECG, etc.) suitable for first aid approaches and early diagnostics. The real time feature of the proposed approach could thus allow, in the future, a clinical applications of such methodology

This work deals with the design and implementation of soft sensors designed to perform the what-if analysis on a Sulfur Recovery Unit (SRU) in a refinery. Soft sensors are mathematical models able to emulate the behaviour of existing sensors on the basis of available measurements. This paper deals with the design of a soft sensor by using a nonlinear one step ahead model, eventually usable on-line. The realized models represented the building blocks for the design of the Virtual Instrument based on a multi steps ahead predictor to be used in the what-if analysis. The instrument allows the estimation the consequences that variations of the input trends produces on system outputs, to eventually improve control policy effectiveness.

In this paper the problem of regressor selection in virtual sensor design is addressed. In particular nonlinear models designed by experimental data are used to estimate relevant process variables of an industrial plant. The plant considered is a Sulphur Recovery Unit of a large refinery settled in Sicily. The proposed approach is used to face with the problem of input regressor selection of NMA models. The approach is based on a recursive evaluation of the cross correlation function between input variables and model residuals. The obtained results are compared with corresponding estimation obtained by using a reference model. Significant improvements in the model estimation capability show the suitability of the proposed method

The paper proposes a comparison of different strategies of regressors selection for the design of a Soft Sensor for a Sulfur Recovery Unit of a refinery. The Soft Sensor is designed to replace the on line analyzer during maintenance and it is designed by using nonlinear MA models implemented by a MLP neural network. A number of strategies for the automatic choice of influent input variables and regressors selection, on the basis of available experimental data, are compared with a strategy based on a trial and error approach, guided by the knowledge of the experts, both in terms of their performance and their computational complexity.

In the paper a soft sensor designed to estimate the acid gases hydrogen sulfide (H2S) in the tail stream of a Sulphur Recovery Unit (SRU) in a refinery located in Sicily, Italy, is described. In particular a model stacking approach is proposed to improve the estimation performance of the Soft Sensor. Neural Networks, Principal Component Analysis (PCA) and Partial Least Squares (PLS) approaches are used for the realization of the first level's models and Simple average, Neural Networks and PLS are used as combination approaches. The validity of the proposed aggregation strategies has been verified by a comparison with the performance of a neural model. The obtained soft sensor will be implemented in the refinery in order to replace the measurement device during maintenance and guarantee continuity in the monitoring and control of the plant.

The capability of neural networks with complex neurons to approximate complex valued functions is investigated. A density theorem for complex multilayer perceptrons (MLPs) with a nonanalytic activation function and one hidden layer is proved. The backpropagation algorithms for MLPs with real, complex analytic and complex non-analytic activation functions are compared with a numerical example. >

The level of nitrogen oxides in atmosphere has been increasing in the last century, mainly due to human activities. Unfortunately nitrogen oxides have a number of negative effects on air quality: they contribute to photochemical smog, visibility, acid rain can also have a negative impact on human health. In the paper a novel strategy to improve the estimation of nitrogen oxides emissions produced by chimneys of refineries is proposed. In particular nonlinear models, obtained by using MLPs neural networks, which are being a commonly used tool in processing data acquired in petrochemical processes, are proposed. The performance of the proposed model with respect to both traditional heuristic models and linear models are described.

The choice of the optimal topology for a multilayer perceptron neural network is considered by using genetic algorithms (GAs). The proposed strategy is intended both to select the number of neurons in a structure with one hidden layer and to choose the number of layers into which a fixed number of neurons should be optimally arranged to solve a given problem. The proposed GA has shown its suitability in determining efficiently the optimal topology of a neural network. The procedure is not time consuming and is able to easily take into account all the constrains eventually included in the problem. >

This paper introduces a still image segmentation technique based on an active contour obtained via single-layer CNNs. The contour initially laid on the frame of the image shrinks, deforms and multiplies until it matches the edges of each of the objects present in the scene. The shape of each object in the image is accurately extracted and nested objects, if any, are correctly detected. Experimental measures of the accuracy of the segmentation were carried out using the Hausdorff distance

In this work a new method for the analysis of circuits interconnected with lossy transmission lines, is presented. At first a lumped model is obtained by adopting suitable approximation techniques. Then classical model order reduction algorithms are applied to obtain a low order model of the circuit, suitable for computing a time domain response to arbitrary input signals. An example is reported, showing the suitability of the proposed approach. >

In this paper a Neuro-Fuzzy approach to obtain the relationships between the parameters involved in the characterization of noise pollution in urban traffic and the 'equivalent' number of vehicles has been developed. A Fuzzy Neural Network which implements a fuzzy reasoning with constant consequences has been chosen. Such a method allow one to overcome some difficulties which appear both in neural and fuzzy modelling and gives a satisfactory performances with a reasonable computational cost.

In the paper it is proposed the use of a complex valued multi-layer perceptron neural network (MLP) with complex activation functions and complex connection strengths in order to perform the estimation of chaotic time series. In particular, the Ikeda map is taken into consideration. A comparison between the behavior of the real MLP and the complex one is also reported, showing that the complex valued MLP requires a smaller topology as well as a lower number of parameters in order to reach comparable performance. >

In this paper genetic algorithms are applied as an optimization strategy in a controller order reduction procedure. GAs capability to get out local minima is used to determine a reduced-order controller which optimally approximate the full-order one. Particular attention has been devoted to guarantee a good stability margin and a suitable index is minimized, by using GAs, in order to compute the parameters of a reduced-order controller which guarantee the stability of the closed loop systems. An example taken from literature has been reported to confirm the suitability of the strategy. >

In this paper is proposed a cellular neural net (CNN) based circuit for secure communication applications. An hyperchaotic Saito oscillator (1990) is designed using a configuration of cellular neural network named state-controlled CNN. A secure communication system, based on chaotic inverse system synchronization, is described and results for a prototype circuit realization are reported.

The attitude control problem of a rigid body, such as a spacecraft, in three-dimensional space is approached by introducing two new control strategies developed in hypercomplex algebra. The proposed approaches are based on two parallel controllers both derived in quaternion algebra. The first one is a feedback controller of PD type, while the second is a feedforward controller, which is implemented either by means of a hypercomplex multilayer perceptron (HMLP) neural network or by means of a hypercomplex radial basis function (HRBF) neural network. Several simulations show the performance of the two approaches. The results are also compared with a classical PD controller and with an adaptive controller, showing the improvements due to the use of the neural networks, especially when an external disturbance acts on the rigid body.

This paper presents a two-dimensional conveyor belt controlled by reaction-diffusion cellular neural network (RD-CNN) to generate waves propagation phenomena. They can propagate on the conveyor belt plane (an elastic membrane), moving an abject between two points by a new kind of actuators - the Nitinol wires. A neural identification process of the membrane is illustrated to allow a suitable choice of the Nitinol dimensions. Moreover, it is shown both thermal and timing Nitinol behaviors are very similar to the slow-fast dynamics exhibited by a RD-CNN.

A sensor validation strategy based on soft computing techniques to isolate and classify some faults occurring in the measurement system of a Tokamak fusion plant is described. Particular attention is focused on the system used to measure vertical stress in the mechanical structure of a Tokamak nuclear fusion plant during fusion experiments. The strategy adopted is based on a modular structure comprising two stages. The first stage consists of a neural network which acts as a symptom model able to estimate directly some suitable features of the expected sensor responses, thus allowing the most frequently occurring sensor faults to be isolated. The second stage consists of a fault classifier implemented via a fuzzy inference system, in order to exploit the knowledge of the experts. The proposed strategy was validated at the Joint European Torus (JET), on several experiments. A comparison was made with both traditional sensor monitoring techniques and validation performed manually by experts. A great improvement was achieved, in terms of both fault detection and classification capabilities, and the degree of automation achieved.

Natural circulation loops represent important elements of many technologically relevant systems. For this reason, their instability represents a major concern, as it consists of oscillations leading to flow reversal. The analysis of these processes was addressed in various theoretical works, mainly based on mathematical approaches to the problem. The models obtained in this way suffers a poor correspondence between simulated and experimental data. To solve this problem, the identification of the system was adopted in this paper, and a neural network model was obtained by means of input-output measurements detected on an experimental natural circulation loop. Moreover, the neural model was used in a predictive scheme, in order to allow long term prediction of the birth of unstable behaviors.

A novel neural network based fault detection strategy to isolate and classify faults occurring in a tokamak fusion plant is described. In particular, attention is focused on measurements of vertical stresses during plasma disruptions. The strategy is based on a neural model which estimates suitable features of the expected sensor response, allowing to isolate the most frequently occurring faults. The proposed strategy has been validated at JET, the Joint European Torus, on several disruptions, and is currently used for fault detection purposes, providing great accuracy in detecting sensor faults, together with a high degree of automation.