Echo State Network Based Soft Sensor for Monitoring and Fault Detection of Industrial Processes
Introduction
The management of industrial processes has been carried out in the last decades based mostly on corrective and preventive maintenance programs (CM/PM). These strategies, even though well established in the industrial environment, are quite expensive, since, in most cases, they require the operation to be stopped in order to perform the maintenance tasks, compromising the process availability. With the increasing complexity of the industries in the 21st century, inserted in an ever-growing competitive scenario, where the operation must be clean, safe and of high productivity, more modern and efficient technologies for of industrial assets management must be adopted (Jardine et al., 2006).
This is particularly important for the oil and gas industry, where ecological issues related to global warming and the consequent need in reducing CO2 emissions, with forecasts showing the increasing replacement of oil by other cleaner energy sources. One can also cite the operational safety issues, in which faults can put both the health of operators and the process itself at risk. Finally, there is a strong geopolitical influence in this sector, which is prevalent in the well-known and cyclical oil crises (Hwang et al., 2018).
To achieve such results, companies are investing in proactive maintenance schemes, through the continuous prognosis of the processes condition, so that the monitoring can provide conditions for decision making about maintenance in a more reasonable time frame, before or as soon as the faults happen (Vachtsevanos et al., 2007; Muller et al., 2008). These strategies are generally referred to as conditional-based monitoring (CBM) and can maximize the useful life of the equipment, reduce maintenance costs, avoiding faults and increasing the operational availability of the process. In this context, digital monitoring tools based on data and machine learning are increasingly occupying the application space in industrial environments. In general, this monitoring approach is based on transforming the data collected by sensors distributed throughout the process into statistical indicators for the integrity of the monitored process (Karpenko et al., 2001; Wang, 2007; Du et al., 2013).
It is common to build monitoring systems based on the use of Non-linear AutoRegressive Moving Average with eXogenous input (NARMAX) models, using delayed inputs to consider the dynamic behavior of the systems, or on recurrent neural networks (RNN), known as universal approximators of dynamic systems (Funahashi & Nakamura, 1993; Du et al., 2013). However, the use of RNN in real-time applications is often hampered by training methods, based on gradient descent, generally implemented as backpropagation-through-time technique (BPTT) (Werbos, 1990), which presents very slow convergence features and no guaranteed global convergence, suffering from the known problems of bifurcations (Vega et al., 2008) and vanishing gradient (Hochreiter, 1998).
Reservoir Computing (RC), when only the output layer is fitted, with the internal synaptic weights of the network model being selected at random, was developed to overcome these training problems (Lukoševičius & Jaeger, 2009). A particular type of RC, the Echo State Networks (ESN) (Jaeger, 2001) have become popular due to its the ability to capture the dynamic relationships among the data with closed solutions for training, making them simpler and computationally cheaper than traditional RNN (Antonelo et al., 2017; Jordanou et al. 2019; Dias et al., 2019).
As usual with other types of neural networks, the major limitation on the use of ESN is the lack of methods to select the large number of hyperparameters. The most used methods for adjustment of hyperparameters (Grid Search and Random Search) suffer with the multidimensionality of the search space and often provide models with poor fit quality, generated without the appropriate selection of the hyperparameters, discouraging the application of ESN to solve real-world problems (Jaeger, 2005; Behar et al., 2013).
Some works published in different areas proposed the use of ESNs for monitoring and fault detection. Nevertheless, this area is still incipient, as most of these works have not been implemented in real world industrial environments, using synthetic data or controlled data collected in pilot experimental units. For example, Morando et al. (2013, 2015) used ESNs to create a fuel cell aging model, while Fan et al. (2016) used ESNs to predict faults in air compressor equipment based on synthetic data. Xie & Zhang (2017) applied ESNs for fault detection in rotating machinery and showed that the technique was quite robust and can provide good results even when the available data set is small. Particularly, Ribeiro et al. (2018) applied ESNs to detect leaks from a pilot unit piping. More recently, ESNs were used to monitor the performances of wind turbine gearboxes (Wu et al., 2019) and the transmission condition of 3D printers (Zhang et al., 2019).
In the scenario of chemical and petrochemical industries, the implemented CBM generally uses more traditional techniques, such as Principal Component Analysis (PCA) and Canonical Correlate Analysis (CCA) (Thorsen & Dalva, 1995; Ly et al., 2009).
Particularly, the use of ESNs in this industrial segment is even more incipient. These applications involve the use of ESN models to compose predictive control structures or to predict some relevant output variable of the process, which frequently cannot be measured regularly in line due to the aggressiveness of the environment or even the inexistence of appropriate sensors. Antonelo et al. (2017) used ESN models to estimate the downhole pressure using real data obtained from sensors available in an operating oil well. The measurements of this pressure are useful to map regions of instability (slugging) in the oil lift process. Jordanou et al. (2019) employed an online adaptive controller based on ESNs in diverse scenarios for control of an oil production platform. As a matter of fact, ESN models proved useful for derivation of model-based control laws, having been used effectively for control of complex dynamic systems. For instance, Dias et al. (2019) showed that ESN models are able to model the gas lift process in oil wells very closely and reported the use of ESNs as the internal models of predictive controllers that were able to deal with measurement noise, unmeasured disturbances and complex dynamic transitions. However, it must be emphasized that, to the best of our knowledge, published reports have not described the used of ESN models for purposes of monitoring and fault detection in chemical processes using real data.
Based on the previous paragraphs, in the present manuscript an ESN-based system is developed and implemented for process monitoring and fault detection, and some characteristics and relative advantages of using these models are discussed. The system uses known procedures for data cleaning, variable and hyperparameters selection to build ESN models that are employed as soft sensors to describe and monitor the regular operation of the analyzed process. For the hyperparameters selection, the Tree Parzen Estimator (TPE) (Bergstra et al., 2011) is used, which is a Bayesian-inspired sequential optimization method that reduces the amount of objective function evaluations in order to reduce the computational effort of the method, making real-time implementations possible.
Based on the formulation of the soft sensors, the proposed methodology, in addition to monitoring and detecting faults, also monitors the reliability of the model predictions, attesting, for example, whether the data predicted by the model in a faulty situation can be used to replace the measured data. In order to do that, the internal states of the ESN reservoir are tracked, so that the modification of the internal state values provides valuable information about the occurrence of process changes while the stability of the internal state values provides information about the reliability of the operation. Additionally, tracking of internal state values also provides indications about the need to retrain the model, which can be very important since the complex and real industrial processes tend to evolve naturally to other operational conditions, without necessarily indicating a fault behavior, either due to aging of equipment or changes in the quality of raw materials, among many other possible reasons (Kruger & Xie, 2012).
The proposed monitoring system is initially tested with synthetic time series datasets generated by the Mackey-Glass Equation (Mackey & Glass, 1977), which is a widely used benchmark in the field of nonlinear dynamic analysis, more specifically in time series forecasting, being especially used in the context of ESNs (Jaeger & Haas, 2004; Wang & Yan, 2015; Løkse et al., 2017; Liu & Zhang, 2020). In addition, ESNs are used here for the first time to detect faults in real assets of the oil and gas industry, monitoring the operation conditions of equipment installed in a plant of Petrobras (Petróleo Brasileiro S.A.). As described in previous studies (Clavijo et al., 2019, 2021), the monitored pieces of equipment are key components of an oil and gas fiscal metering station and are indeed very important, because the measurements they provide are related to allocation and custody transfer of product streams of the platform.
The proposed methodology is presented in Section 2, as well as the ESN structure and TPE Optimization procedure, which are the key parts of the proposed monitoring procedure. In addition, the monitored systems, MGAB model and Fiscal Metering Station are presented in detail. The obtained monitoring results are discussed in Section 3, while the conclusions are presented in Section 4.
Section snippets
Echo State Network (ESN) and Tree Parzen Estimators Optimization (TPE)
ESNs (Jaeger, 2001) are composed of three layers, similar to other popular neural network: an input layer, where the patterns are fed into the network; a hidden layer, called dynamic reservoir, consisting of a large number of sparsely connected neurons with a high degree of recurrence; and an output layer, with the values predicted by the network, as one can see in Fig. 1.
Given a vector of inputs , where n is the discrete time, a vector of reservoir states and an output vector ,
Mackey-Glass Anomaly Benchmark
The selected model presented the following hyperparametric configuration: N = 316, SR 4.3955.10−1, = 3.4749.10−4 and term = True. The selection process performed with the TPE procedure evaluated the objective function 4000 times; however, no significant improvement could be detected after iteration number 1414, with MSE = 1.0066.10−4. Fig. 5A shows a comparison of the values of the MG time series (originals and predicted by the ESN model), as well as the values of the SPE and
Conclusions
The present manuscript described a system for monitoring and identifying processes in real time through the construction of Echo State Networks-based soft sensors. This development is relevant and can assist the identification of processes and support decision making in industrial environments, thus allowing to obtain safer and more efficient processes. The system can be implemented in common computing frameworks without the need to use sophisticated numerical techniques, where the critical
Funding
This study was financed in part by the Coordenação de Aperfeiçoamento de Pessoal de Nível Superior-Brasil (CAPES)-Finance Code 001. The authors also thank CNPq-Conselho Nacional de Desenvolvimento Científico e Tecnológico, and Petrobras (Petróleo Brasileiro SA), for the financial support of this work.
CRediT authorship contribution statement
Tiago Lemos: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – original draft, Writing – review & editing, Visualization, Project administration. Luiz Felipe Campos: Validation, Formal analysis, Investigation, Data curation, Visualization, Project administration. Afrânio Melo: Validation, Formal analysis, Investigation, Visualization. Nayher Clavijo: Validation, Formal analysis, Investigation, Visualization. Rafael Soares: Validation,
Declaration of Competing Interest
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; nor in the decision to publish the results.
References (58)
- et al.
Echo State Networks for data-driven downhole pressure estimation in gas-lift oil wells
Neural Networks
(2017) - et al.
Actuator and sensor fault isolation of nonlinear process systems
Chemical Engineering Science
(2013) - et al.
Predicting Air Compressor Failures with Echo State Networks
PHME 2016 Proc. Third Eur. Conf. Progn. Heal. Manag. Soc.
(2016) - et al.
Online learning control with Echo State Networks of an oil production platform
Eng. Appl. Artif. Intell.
(2019) - et al.
Using echo state networks for classification: A case study in Parkinson's disease diagnosis
Artif. Intell. Med.
(2018) - et al.
Nonlinear process modelling using echo state networks optimised by covariance matrix adaption evolutionary strategy
Comput. Chem. Eng.
(2020) - et al.
Prediction for noisy nonlinear time series by echo state network based on dual estimation
Neurocomputing
(2012) - et al.
Recent advances in physical reservoir computing: A review
Neural Networks
(2019) - et al.
Use of bifurcation analysis for development of nonlinear models for control applications
Chem. Eng. Sci.
(2008) - et al.
Effects of spectral radius and settling time in the performance of echo state networks
Neural Networks
(2009)
Optimizing the echo state network with a binary particle swarm optimization algorithm
Knowledge-Based Syst.
An Echo State Neural Network for Foetal ECG Extraction Optimised by Random Search
Nips
Uncertainty study of fiscal orifice meter used in a gas Algerian field
Flow Meas. Instrum.
Algorithms for hyper-parameter optimization
Making a science of model search: Hyperparameter optimization in hundreds of dimensions for vision architectures
Development and application of a data-driven system for sensor fault diagnosis in an oil processing plant
Processes
Variable Selection for Fault Detection Based on Causal Discovery Methods: Analysis of an Actual Industrial Case
Processes
Extracting valuable information from big data for machine learning control: An application for a gas lift process
Processes
Pruning and regularization in reservoir computing
Neurocomputing
Approximation of dynamical systems by continuous time recurrent neural networks
Neural Networks
Natural Gas Measurement Handbook
Hands-on machine learning with Scikit-Learn, Keras and TensorFlow: concepts, tools, and techniques to build intelligent systems
The Selection of Variables in Multiple Regression Analysis
J. Educ. Meas.
The vanishing gradient problem during learning recurrent neural nets and problem solutions
Int. J. Uncertainty, Fuzziness Knowlege-Based Syst
A study of the development of a condition-based maintenance system for an LNG FPSO
Ocean Eng
The “echo state” approach to analysing and training recurrent neural networks
GMD Rep
Harnessing Nonlinearity: Predicting Chaotic Systems and Saving Energy in Wireless Communication
Science (80-.)
Reservoir riddles: Suggestions for echo state network research (extended abstract)
A review on machinery diagnostics and prognostics implementing condition-based maintenance
Mech. Syst. Signal Process.
Cited by (16)
Development of a CNN-based fault detection system for a real water injection centrifugal pump
2024, Expert Systems with ApplicationsDeep feature representation with online convolutional adversarial autoencoder for nonlinear process monitoring
2024, Journal of the Taiwan Institute of Chemical EngineersThe role of artificial intelligence-driven soft sensors in advanced sustainable process industries: A critical review
2023, Engineering Applications of Artificial IntelligenceArtificial intelligence modeling of ultrasonic fatigue test to predict the temperature increase
2022, International Journal of FatigueCitation Excerpt :The random search algorithm is an alternative intended to address the computational cost issue but has been found unreliable for training complex models [35]. One of the main issues of hyperparameter optimization is that each time a new set of hyperparameters is evaluated, it is necessary to call the objective function that trains a model on the training data, make predictions on the validation data and then calculate the validation score [36]. Bayesian optimization suits well for hyperparameter tuning problems, since it keeps track of past results and proposes better candidate hyperparameters, leading to fewer overall evaluations of the objective function [37].
Evaluation of neural network models and quality forecasting based on process time-series data
2024, Canadian Journal of Chemical Engineering