Design and implementation of distributed measurement systems using ...

8 downloads 100 Views 79KB Size Report
ligent sensors, an implementation of intelligent frequency-output sensors attributed to distributed systems is presented. The intel- ligent eddy-current sensor that ...
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001

1197

Design and Implementation of Distributed Measurement Systems Using Fieldbus-Based Intelligent Sensors Gui Yun Tian, Member, IEEE

Abstract—This paper presents a new topology of distributed measurement systems, which apply fieldbus-based intelligent sensors. Based on the framework, a novel intelligent measurement system, which provides output of measured data and data uncertainty simultaneously, is invented. The distributed measurement system is developed to connect intelligent sensors, intelligent actuators, and other intelligent control units together, where their intelligent functions are given by the availability of internal computing power and digital communications. With the help of fieldbus, intelligent sensors applying novel advanced methods of measurement technology are described. As an example to illustrate the proposed distributed system and advantages offered by fieldbus-based intelligent sensors, an implementation of intelligent frequency-output sensors attributed to distributed systems is presented. The intelligent eddy-current sensor that applies computing algorithms for separating sensor data errors and predicting sensor status is introduced. The contribution of blind source separation is highlighted. In the end, the test results and conclusions are provided. Index Terms—Distributed measurement, fieldbus, frequency output eddy-current sensor, intelligent sensor, self-calibration, self-compensation, self-validation, signal separation.

I. INTRODUCTION

C

URRENTLY in the monitoring and control of large-scale systems, a hierarchical centralized controller is normally used to coordinate a naturally dispersed system. With the advent of distributed control technologies, an open and extendible approach to system control is available, the most recent advancement in control and communication networks being fieldbus technology. While the concept and implementation techniques of fieldbus have been on-going for a number of years, only recently has there been serious industrial development of fieldbus systems. The market for fieldbus relating products is also expanding with products already installed in a number of plants. Simplified wiring and the resulted cost-saving in installation and maintenance are the immediate benefits, but the potential for creating added-value through integration is likely to be much more significant. Fieldbus, now becoming an IEC standard, takes advantages of modern communication capabilities, which in this system are optimized for measurement systems. Based on the emerging technology, the distributed measurement and control systems could be modernized. In this paper, we will introduce Manuscript received March 3, 2000; revised June 15, 2001. The author is with the School of Engineering, University of Huddersfield, Queensgate HD1 3DH, U.K. (e-mail: [email protected]). Publisher Item Identifier S 0018-9456(01)08744-7.

Fig. 1. Fieldbus integration for control, measurement, and management.

a topology of distributed measurement systems based on fieldbus. Following the infrastructure, the paper will discuss the key measurement system design aspects and their intelligent functionality. Fig. 1 illustrates an infrastructure of fieldbus for distributed measurement and control systems. II. TOPOLOGY OF DISTRIBUTED MEASUREMENT SYSTEMS The availability of low-cost, mobile access to telecommunications and computing networks and the sharp decrease in the cost of storing and processing digital information have revolutionized the way that organizations and individuals think about capturing, accessing, and using information. Businesses are recognizing that change and interaction are now endemic in industry, and that sharing information (particularly the control of shared information) can add more value than withholding it. Organizational change has moved from business process reengineering through enterprise transformation toward fully virtual organizations. To enhance information sharing between enterprises, the distributed information systems e.g., control and measurement, are rapidly being developed. Fieldbus is being accepted by industry as we accept Internet. Based on fieldbus, we will reengineer or optimize measurement systems from functionality and specification. As illustrated in Fig. 1, a new intelligent modular approach for distributed systems is presented. The components in the systems, e.g., sensors, actuators, and other controllers, are intelligent cells to integrate information from the fieldbus or capture information from other media to fieldbus. In contrast to traditional measurement systems, new distributed measurement systems are able to provide

0018–9456/01$10.00 © 2001 IEEE

1198

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001

Fig. 2. Infrastructure for intelligent sensor based on fieldbus.

the uncertainty of measured data in runtime. As an example, we will discuss the details of fieldbus-based intelligent sensors. The current status of the intelligent sensor has evolved from derision to acceptance in less than two decades. The technologies such as microcomputers and neural networks have promoted the progress considerably. As sensor systems get more intelligent, the distinction between a sensor and an instrument, between an intelligent sensor and a smart sensor, is becoming more blurred. Although the demand for high sensor accuracy, high reliability, low cost, and compactness has been constantly increasing for the last two decades, much research is still to be done. A generic hardware configuration of an intelligent sensor system was presented by Gerard and Meijer [1]. A definition of an intelligent sensor, which modifies its internal behavior to optimize its data from the physical world and then communicates it in a responsive manner to a host system, was given by Brignell [2]. Most papers on intelligent sensors have considered their performances and information processing from sequential processing to parallel and distributed processing [3], [4]. However, to date there have been no implementation and applications of intelligent sensors for distributed measurement. The current trend toward intelligent and smart sensors is to integrate 1) a sensing element that can be made in a standard process, 2) electronic circuits to fully and periodically calibrate and compensate the sensor, and 3) circuits for generating a bus-compatible output with not only the sensing data but also the sensor health status. This paper presents an infrastructure of distributed measurement systems based on fieldbus and intelligent sensors. The functionality of intelligent sensors with self-calibration, selfcompensation, self-validation, data fusion, and comprehensive communication is described and simulated. The intelligent functions are given by the availability of internal computing and digital communication, which are not provided by the existing distributed measurement system. Measurement is often afflicted with errors; i.e., there is often a difference between the measured value and the true value of the physical quantity under observation. The exact value of certain error may never be known, but it can be estimated with predefined tolerance. For example, statistical process and higher order neural networks can be used for the blind source separation problems [5], [6]. Measurements can be corrected to a certain degree only if the magnitude of a measurement error is given with the measured value to constitute meaningful mea-

surement results. This means that a measurement result that does not indicate its error range will not provide enough information. The most important feature of intelligent sensors based on fieldbus is that not only should the sensors output the measured data, but they should also provide the uncertainties of the data. A basic infrastructure of intelligent sensor and its functions are shown in Fig. 2. The intelligent sensor has adopted the integration of microcontrollers and neural networks. It has intelligent functions of self-calibration, self-compensation, self-validation, data fusion, and comprehensive communication by two-wire digital output or input through fieldbus communication protocols. The data output contains measured data, data quality index, sensor health status, and diagnostic information; the input data contains measurement accuracy and environmental condition such as temperature, humidity, and data for sensor fusion. The main feature of the intelligent sensor is realized through embedded microprocessors with fieldbus techniques. The proposed intelligent sensor has the functions of self-calibration, self-compensation, self-validation, auto-data fusion, and comprehensive output through fieldbus communication. These functions can be specified as follows. Self-Calibration: The conventional industrial practice corrects (recalibrates) sensors and measuring instruments according to a fixed schedule (calibration interval). It can be time consuming when the schedule is too tight or even can provide a false control when the schedule is too relaxed. It has no information about the history data and recalibration and overhaul suggestion. In the intelligent sensor, the embedded microcontrollers or digital signal processors (DSPs) have data storage such as RAM and E PROM that can store the sensor information and their history readings. The self-calibration means that the sensor can monitor the measuring condition by a confidence test to decide whether a new calibration is needed or not. For a real-time confidence test, a stimulus is looped back with a measurement and gets the level of confidence. The confidence is used to judge whether the system is still performing satisfactorily or not. If not, a calibration procedure is required to recover the sensor performance. Self-Compensation: A compensator is a measurement network, which makes use of the compensation methods, so as to achieve a high accuracy. Today compensator networks are not often used. Most applications use microprocessor-embedded in-

TIAN: DESIGN AND IMPLEMENTATION OF DISTRIBUTED MEASUREMENT SYSTEMS

Fig. 3.

1199

Fieldbus-based intelligent frequency output sensor.

struments to perform the measurement, in which the compensation method is internally based on software rather than hardware compensators. In the proposed intelligent sensor, the software of compensation is based on embedded neural networks. The compensation is divided into three types: nonlinear compensation, cross-sensitivity compensation and time based drift compensation. The nonlinear compensation focuses on the sensor relationships between input and output, based on independent component analysis (ICA). Cross-sensitivity compensation is one of the most complex issues in sensor compensation. It is affected by various factors such as offset, gain, temperature and others that can not be easily specified. The most common cross-sensitivity compensation is with the factor of temperature, for which temperature changes in the surrounding environment are an essential element for observation. In our experimental systems, error source separation, error compensation and fault detection are implemented with good results (Section III). For analog sensors, the processes of noise reduction, e.g., filter and ground loop integrity, will be applied. Self-Validation: Sensor self-validation or self-diagnosis can lead to fewer accidents and spillages, thereby enhancing safety, providing better product quality, reducing unscheduled shutdowns, and improving plant efficiency and availability. Self-validation applies mathematical modeling error propagation and error isolation or knowledge-based techniques [7] such as sources of error for describing sensor faults and status, as well as predicting the sensor failure and maintenance supervision. The intelligent sensor can monitor internal signals for evidence of faults. It provides for economical maintenance (highly skilled person is not required for sensor maintenance). Diagnostic and measurement data can both be communicated via a digital communication link of fieldbus [8]. III. EXAMPLE OF IMPLEMENTATION OF MEASUREMENT SYSTEMS USING INTELLIGENT SENSOR A prototype intelligent frequency output sensor is built to demonstrate the viability of the concept [9]. The prototype is comprised of a displacement frequency output sensor, a local operating network node (LonWorks node) and neural networks embedded in the Echelon neuron chip MC143 150 [10]. The system is configured as shown in Fig. 3.

A. Frequency Output Displacement Sensors and Data-Acquisition The intelligent sensor is composed of a frequency output displacement sensor. The sensor is based on eddy current effects. It is a noncontact device consisting of a coil of wire with a high-frequency current. The displacement is transferred into pulse frequency by direct digital methods [11], [12]. It is a digital sensor where A/D converters are not required for its data acquisition. The data-acquisition interface is implemented by programmable Timers/Counters 8254IC. The Neuron Chip MC143 150 has two programmable Timer/Counters for the data-acquisition. B. Using LonWorks Node for Simulation LonWorks from Echelon (Lon stands for local operating network) is a complete system that incorporates a communication standard. It also includes management and control. The neuron chip contains three 8-bit processors. The functionality of the three parallel processors is shown in Fig. 3. The application processor executing user code is written in a variant of “C” language with powerful input/output functions. A network processor handles addressing, routing, authentication of packets and the presentation of data to the application processor. The MAC processor is responsible for encoding I/O, importing the measurement, calculating, calibration and transmission of packets of data to the network. Most of the intelligent functions are implemented by embedded neuron networks stored in EEPROM. These two processors comply with the six layers of the ISO reference model. This allows the transparent use of the network to pass information between the different programs in the application processors. More details about LON communication technology can be found in [10], [13], [14]. In our systems, the application program can perform initialization, data acquisition, calibration, measurement uncertainty calculation, neural network-based compensation, and data management. C. Embedded Computing Algorithms for the Intelligent Eddy-Current Sensor Eddy-current displacement sensors are very sensitive to many environment factors and especially to temperature vari-

1200

Fig. 4.

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001

Sensor raw data drift on working time and temperature variation.

ation [15]. In the intelligent eddy-current sensor, several computing algorithms of data error separation and sensor status prediction are embedded in the microcontroller system, where DSPs are applied to implement the intelligent functions of selfcalibration, self-compensation, self-validation and self-data fusion. The information processing of ICA, Kalman filtering and the neural network is used. The Kalman filter is a set of mathematical equations that provides an efficient computational (recursive) solution of the least-squares method. It supports estimations of past, present, and even future status of sensor health from the noisy observations. For the neural networks, three processes for building intelligent algorithms are required. They are neural training, testing and construction. In this research, a hybrid system of PC and microcontroller is used for studying structures of neural networks. Fig. 3 shows the experiment system. The communication protocol from the intelligent sensor is LonTalk. The intelligent sensor is connected to a gateway where the LonTalk is transferred to the traditional RS232 protocol or vice versa. Therefore, a dual communication between a PC-based system and the fieldbus-based intelligent sensor is built. The PC-based system is used for neural network design, training and comparison study of neural software. The training data is loaded from microcontroller-embedded intelligent sensor through fieldbus, and the trained neural networks is then output and embedded in the memory of an intelligent sensor device such as EEPROM. To compensate the various errors (assuming they are independent from each other), a crucial step is to separate different errors from measurement data. A fast algorithm for ICA is used for the blind source separation and parameter calculations [6], [17]. transmit certain signals, which, Assume that sources after transmission through an arbitrary medium, are measured by sensors . The measured signals are related to the transmitted signals by function , which will be referred to as the mixing relation, with added measurement noise (1)

When

constitutes a linear relationship (2)

matrix referred to as the mixing matrix. In where is an signal sources are displacement, temperour sensors, the ature and sensor working time. We assume the mixing relation is linear and ambient temperature is quasistationary, i.e., it is constant during the estimating time. Fig. 4 displays the sensor raw data drift on the variation of working time and ambient temperature while the displacement is zero. Basically, the different signals have different spectral properties. Using the neural network approach is to separate the signal source and identify the materials of measured targets. Then, an appropriate compensation strategy is applied. The blind source separation proceeds in two steps: first, obtain the original sources from the mixed signals (the mixed must be inverted); and second, find the linear transformation that ensures independence up to the second order (i.e., decorrelation of the dataset). Principal component analysis or eigenvalue decomposition is applied in the sensors. The second step however, ensuring independence of a higher order, is very involved and requires the estimation of higher order moments or cumulates. The ICA of a random vector is used to search for a linear transformation that minimizes the statistical dependence between its components. In order to define suitable search criteria, the expansion of mutual information is utilized as a function of cumulating increasing orders. The solutions to the ICA problem for separating the different measurement errors such as nonlinearity and stability use the fourth-order cumulant or kurtosis of the measurement, defined for a zero-mean random variable as (3) For a Gaussian random variable, kurtosis is zero; for densities peaked at zero, it is positive, and for flatter densities, it is negative. Note that for two independent random variables and and for a scalar , it holds that and .

TIAN: DESIGN AND IMPLEMENTATION OF DISTRIBUTED MEASUREMENT SYSTEMS

1201

Fig. 5. Signal flow in the intelligent sensor. TABLE I THE INFRASTRUCTURE OF BACK PROPAGATION NEURAL NETWORKS FOR COMPENSATION

Assume that a sample of the sphered random vector (original measured data) is collected, which in the case of blind source separation is a collection of linear mixtures of independent source signals. Following the derivation of the preceding section [6], the fixed-point algorithm for ICA is as follows. Step 1. Take a random initial vector of norm 1. Let . , Step 2. Let expectation can be estimated using a vectors (say, 2000 large sample of points). by its norm. Step 3. Divide is not close Step 4. If , and go back enough to 1, let to Step 2. Otherwise, output the vector . Based on the above algorithms, the neural network software can be used to implement the intelligent processing. Neural networks are also implemented by an embedded DSP to improve sensor performance, versatility and reliability. In this study, neural networks are used for nonlinear compensation, cross-sensitivity compensation, and time dependence compensation and fault diagnosis as illustrated in Table I. After comparing several studies from different neural networks (trading-off the training time and the complexity of network structures), a 3-layer back propagation neural network is applied to the implementation of the intelligent functions. Learning rules by ICA are used for the weights of a neural layer; the neurons have signal separation capabilities. A frequency-based eddy current displacement sensor with a Lonworks is used in the experiment test. The signal flow chart is shown in Fig. 5. The sensor contains an operating time clock and a temperature monitor to input operating time and ambient temperature to the neural network. Based on the source separation and parameter calculation from the neural network, the sensor values are

compensated where a sensor measured different materials and is applied with different formulation for thermal and temporal compensation [18]. In our test rigs, the various measured materials such as Cu-alloy, Al-alloy, and Fe-alloy are classified correctly. The data uncertainty from error computing and error propagation has been used for real-time output. When the data uncertainty is higher than the accuracy from its measurement requirement, the data-acquisition interface is switched to test the internal oscillator signal to calibrate the system or indicate the fault of the sensor automatically. In the multicycle synchronic counter, the sensor has two channels to measure the frequency by hardware (8254 counter) and software counting. The redundancy is used for the self-validation. Other sensor faults such as oscillators and the 8254 status are detected by the I/O interface. The measurement requirements such as the compensation formulation and measuring accuracy, particularly when a sensor is changed and algorithms are updated, can be added into the sensor through the fieldbus interface. More details about the sensor specification can be found in [9]. It is much more flexible and accurate than the traditional compensation of “look-up table (LUT)” [16]. IV. CONCLUSIONS An intelligent modular distributed measurement system is formalized and implemented by fieldbus and intelligent sensors. The intelligent sensors based on fieldbus and its features are investigated. The functionality of the sensor is described and simulated. With the help of fieldbus technology, the intelligent sensor can output not only the measurement data but also the data quality index that considers sensor data confidence and sensor health information for maintenance. Fieldbus provides a comprehensive communication ability to meet the requirement of intelligent functions and off-line programming such as neural network training and testing. A LonWork fieldbus and its microcontrollers are introduced and used for implementation of the sensor. The neural networks of back propagation are used to achieve intelligent functions for enhanced sensor performance

1202

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 50, NO. 5, OCTOBER 2001

and reliability. The embedded software also attains the real-time measuring error source separation, error calculation and fault prediction. As a case study, a frequency output eddy current sensor and its intelligent functions such as blind source separation have been presented to demonstrate the proposed intelligent distributed measurement system [19]. Results presented in this paper show that source separation on an eddy-current sensor, ICA in particular, can be a suitable method to gain some robustness against temperature and target material effects. More research will be done to optimize algorithms of ICA, Kalman filters, and their implementation by using DSPs [20]. REFERENCES [1] G. C. M. Meijer, “Concepts and focus point for intelligent sensor systems,” Sens. Actuators A, vol. 41-42, pp. 183–191, 1994. [2] J. Brignell, “The future of intelligent sensors: A problem of technology or ethics?,” Sens. Actuators A, vol. 56, no. 1-2, pp. 11–15. [3] M. Ishikawa, “Robot sensors with parallel-processing capabilities,” Int. J. Jpn. Soc. Precision Eng., vol. 29, no. 3, pp. 201–204, 1995. [4] B. Squires, “Fieldbus trials at BP,” Comput. Contr. Eng., pp. 254–258, Dec. 1995. [5] P. Comon, “Independent component analysis, A new concept,” Signal Process., vol. 36, pp. 287–314, 1994. [6] A. Hyvarinen and E. Oja, “A fast fixed-point algorithm for independent component analysis,” Neural Comput., vol. 9, pp. 1483–1492, 1997. [7] Y. J. Kim, “A frame for an on-line diagnostic expert system with intelligent sensor validation,” KSME Int. J., vol. 11, no. 1, pp. 10–19, 1997. [8] M. P. Henry, “Sensor validation and fieldbus,” Comput. Contr. Eng., pp. 263–269, 1995. [9] G. Y. Tian, “The design of frequency output sensors and their signal processing metrologies,” Ph.D. dissertation, Univ. Derby, U.K., 1998. [10] . [Online]. Available: Http://www.fieldbus.org/business/information/ [11] S. Middelhoek, P. J. French, J. H. Huijsing, and W. J. Lian, “Sensors with digital or frequency output,” Sens. Actuators, vol. 15, pp. 119–133, 1988. [12] G. Y. Tian et al., “The research of the frequency modulated displacement sensor,” Sens. Actuators A, vol. 55, pp. 153–156, 1996.

[13] Neuron C Programmers Guide. Palo Alto, CA: Echelon Corp., Mar. 1995. [14] “LonTalk response time measurements,” in Engineering Bulletin 0050010-01. Palo Alto, CA: Echelon Corp., Mar. 1995. [15] P. Vasseur and A. Billat, “Contribution to the development of a smart sensor using eddy currents for measurement of displacement,” Meas. Sci. Technol., vol. 5, pp. 889–895, 1994. [16] F. X. Huang and G. Y. Tian, “A way of verticality measurement from holes to planes,” J. Sci. Instrum., vol. 15, no. 1, pp. 109–112, Feb. 1994. [17] J. F. Cardoso and P. Comon, “Tensor-based indendent component analysis,” Signal Process. Theories Applicat., vol. 5, pp. 673–676, 1994. [18] G. Y. Tian, Z. X. Zhao, and R. W. Baines, “The research of inhomogeniety in eddy current sensors,” Sens. Actuators A, vol. 58, pp. 153–156, 1998. [19] G. Y. Tian, “Eddy current frequency output sensors for precision engineering,” Insight, vol. 43, no. 5, May 2001. [20] G. Y. Tian, Z. X. Zhao, and R. W. Baines, “A miniaturised displacement sensor for deep hole measurement,” J. Precision Eng. Amer. Soc., vol. 23, pp. 236–242, 1999.

Gui Yun Tian (M’01) received the B.Sc. degree in metrology and instrumentation and the M.Sc. degree in precision engineering from the University of Sichuan, Chengdu, China, in 1985 and 1988, respectively, and the Ph.D. degree from the University of Derby, Derby, U.K., in 1998. From 1988 to 1995, he was with a member of the academic staff at the University of Sichuan. He was a Research Fellow and Senior Research Fellow in color imaging at the Color and Imaging Institute, University of Derby, and in the School of Information System, University of East Anglia, U.K., from 1998 to the end of 1999. He is currently a Senior Lecturer in Multimedia and Electronic Engineering in the School of Engineering, University of Huddersfield, Queensgate, U.K. He maintains a diverse and active research program in the area of sensors, instrumentation, distributed systems, computer vision, and signal processing. He has wide academic links and industrial collaborators. He has published over 60 papers in English and Chinese in the area of computer science and engineering.

Suggest Documents