MFPT Committee Meeting, Virginia Beach, VA, May 2000

An Internet-based Machinery Health Monitoring System
Rolf Orsagh, Mike Roemer, and Ben Atkinson Impact Technologies, LLC 125 Tech Park Drive Rochester, NY 14623

Abstract: An Internet-based machine condition monitoring system has been developed that allows off-site maintenance personnel to access machinery diagnostic information in real-time. The Internet provides an inexpensive vehicle for the collection and assessment of condition data from machines in distant plants that leads to improved maintenance scheduling. The On-line health monitoring system is designed to efficiently and reliably transfer acquired data to a remote server for processing and analysis then publish the results on a secure Web-site on a minute by minute basis. Sophisticated machine condition assessment techniques including sensor validation, statistical trending, and fuzzy logic algorithms are implemented to detect incipient forms of machine performance degradation. An Internet-based condition monitor for a supercritical steam turbine is presented to illustrate the features, operation, and architecture of the system. Key Words: Condition monitoring, Internet-based, health monitoring, diagnostics, statistical trending, real-time analysis Introduction: Web-based machinery performance monitoring implemented over the Internet is an inexpensive vehicle to deliver real-time performance and diagnostic information to key off-site personnel. Industries that could benefit from this technology include electric power generation, petroleum suppliers, chemical producers, or any other owner/operators of complex equipment in remote areas. On-line machinery condition monitoring merges the latest Internet communication technology with advanced diagnostic and prognostic algorithms to deliver sophisticated health monitoring information in a highly accessible format. Intelligent diagnostic algorithms continuously running on a web-server identify anomalies, and diagnose incipient performance faults, which are then published on a secure web-site. From the web-site, authorized personnel can easily view graphs showing performance trends, and tables of performance parameters, anomalies, and diagnosed faults A web-based health monitoring system for a high pressure (HP) section of a supercritical 700 MW steam turbine is used to illustrate the details of this information technology application. The unit was chosen because of its criticality to plant operation and its constantly varying load posed more of a condition monitoring challenge to the operators and maintenance personnel. System Architecture: The web-based performance monitoring system consists of three components as shown in Figure 1. First, the turbine performance data from the plant’s

mdb FTP Input. Browser Server Web-Site Output. Third. a data correction routine calculates the equivalent sensor readings under standard operating conditions. fuzzy logic algorithms diagnose the most likely cause of the performance degradation. and diagnose the most likely cause of the performance degradation. statistical trending analysis. Second. and fuzzy logic based diagnostics. data correction. active server pages (ASP) use VB Script programs and active data objects (ADO) to update the web-site with real-time data from the output database.db Analysis Algorithms Plant NewData. The appropriate set of parameters includes a sufficient level of 2 . data validation tests are then applied to detect abnormally low or high readings. First. VA. May 2000 data acquisition system must be stored in a database on the utility’s WAN (Wide Area Network) then transmitted using file transfer protocol (FTP) across the internet to a remote file server. Performance Parameters An optimal set of performance parameters must be selected during the design of an online monitoring system. performance parameters. A complete diagnostic record is written to an output database on the server. intelligent algorithms on the server validate the sensor data. data validation. detect performance anomalies.db Figure 1 Architecture of Web-Based Monitor Monitoring and Diagnostic Algorithms: Key monitoring and diagnostic algorithms on the web-site server consist of five components. Finally. Virginia Beach. in the event of a statically significant and severe deviation from the baseline values. Second.MFPT Committee Meeting. Third. statistical trending algorithms are used to detect incipient performance faults by identifying shifts in the mean values of recent corrected and calculated parameters. Authorized users can access the information on the web-site from remote locations via a user name and password.

as shown in Figure 2. Figure 2 Monitored Parameters Data Validation To provide reliable information. deposits. which are not used by the diagnostic system but may provide additional insights into the operational status of the unit. In the event of a sensor failure. VA. leakage. May 2000 redundancy to support sensor validation algorithms and ensure reliable results without becoming overly complex.Proceedings of the MFPT Conference. 3 . a health monitoring system must begin with a rigorous examination of the measured parameters to identify sensor malfunctions before they contaminate the diagnostic information. For the steam turbine demonstration. and HP efficiency) for HP turbine diagnostics. sensor recovery is possible through the use of artificial intelligence algorithms that can provide proxy data until the malfunctioning sensor can be repaired. a set of performance parameters were chosen to detect the onset of the most common causes of performance degradation in HP turbines. and blockage. Virginia Beach. The web-based monitoring system also displays other parameters. mass flow. solid particle erosion. cold reheat pressure. Sensor problems such as ground loop faults or sensor drift are often misinterpreted as the onset of performance or vibration faults. Diagnosis of these faults is based on techniques developed by Cottoni and Beebe ii and relies on five commonly used parameters (1st stage pressure. 1st stage efficiency.

Virginia Beach. 4 . To accurately trend the performance of variable speed or variable load equipment while it is in use. The sensor validation status is shown for each monitored parameter on the web page shown in Figure 2. When a signal goes outside these bands. Figure 3 shows examples of uncorrected and corrected data. the performance parameters must be corrected to their equivalent values at a standard speed or load condition. Corrections are performed using polynomials that represent baseline mean values of performance parameters over the operational range while the equipment is in a healthy condition. while others remain within. In the steam turbine monitoring example. and translated to an equivalent value at the design conditions. Corrections are based on a-priori knowledge of the relationship between the independent and dependent performance parameters in a baseline model. data validation algorithms test each diagnostic parameter to identify gross deviations from the expected operating range. The discontinuity in the slope of the uncorrected data (at mass flow = 1700 units) is due to the fact that the unit operates in two distinct modes. trained neural networks. Interruptions of the data transfer from the plant are also detected to avoid corruption of the statistical database. The variation of each diagnostic parameter from its baseline (expected) value is calculated in real-time. These parallel algorithms are combined in a data fusion process that determines the final confidence levels that a particular sensor has either failed or has suspect operation. and fuzzy logic analysis. VA. Care must be used when selecting normative data to ensure that it is unbiased by periodic or seasonal effects. May 2000 The sensor diagnostic process should be performed using multiple and collaborative techniques that offer advantages for isolating and detecting specific sensor failure modes. an anomaly is detected associated with those specific sensors.MFPT Committee Meeting. Data Correction Variable speed or variable load machinery presents a significant health monitoring challenge due to the difficulty involved in trending performance parameters. The fuzzy logic based sensor analysis continuously assesses the “normal” bands associated with each sensor signal at the current operating condition. Upper and lower bounds for the each parameter were established from normative (baseline) data supplied by the utility. The neural network operates by comparing the physical relationships between signals as determined from either a baseline model or equivalent computer model of the machine’s performance parameters. Correction curves were developed for the steam turbine monitor by fitting polynomials to the normative data supplied by the utility. Some available techniques that have been implemented with success include.

The load on the turbine is calculated by correcting the electric power generated for the non-linear electrical and mechanical power loss in the generator.Proceedings of the MFPT Conference. Performance trends may then be identified and evaluated under any operating conditions. mass flow is corrected to its equivalent value at the 5% over pressure condition using the corrected load. Finally. and the efficiencies are all corrected to their equivalent values at the 5% over pressure equivalent design mass flow from the turbine manufacturer’s heat balance diagram for the unit. Each record of data that is transferred from the plant may be collected under a different load condition. A statistical T-test is used to detect subtle shifts in the corrected performance parameters. cold reheat pressure. 5 . May 2000 Figure 3 Uncorrected and Corrected Data The steam turbine on-line monitor required corrections to the five diagnostic parameters and load. Statistical Trending Analysis Performance anomaly detection algorithms are designed to statistically detect the manner in which machinery performance parameters are shifting over time. First stage pressure. Before the data is analyzed. Mean and standard deviation values are calculated for the corrected performance parameters that comprise the baseline model. VA. Significant deviations of the corrected performance parameters from their baseline values indicate the presence on an anomaly and possibly the onset of performance degradation. it must first be corrected to eliminate the effects of the varying operating conditions. The mean values of a recent sample of performance parameters are compared against the mean values in the baseline model. The confidence value reported by the T-test represents the probability that the observed shift is not attributable to normal random fluctuations as shown in Figure 4. Virginia Beach. Small deviations during operation may be due to normal random fluctuations in the selected sample or due to real performance changes.

parametric trendline deviations from baseline performance are fed to a diagnostic module. i.MFPT Committee Meeting. Figure 5 Anomaly Detection Diagnostic Classification The total combination of performance parameter trends is used to provide a diagnosis of the most probable performance fault. May 2000 Baseline mean µ1 Current mean µ2 t = | µ1 − µ 2 | s1 s − 2 n1 n2 Figure 4 T-test is Used to Detect Incipient Faults An anomaly was detected shortly after the implementation of the steam turbine on-line monitor. Once an anomaly has been detected. Virginia Beach.e. 6 . Consultations with the utility revealed a data processing error in which data from a different turbine was transmitted to the on-line monitor. but to corrupt data. Figure 5 shows the corrected data and corresponding distributions for the baseline and anomaly conditions. sensor validation alarms indicated that the problem was not due to performance degradation. the performance error patterns. However.. VA. The T-test indicated a nearly 100% confidence that a 13% shift had occurred in the parameter shown. Artificial intelligence-based pattern recognition algorithms within the diagnostic module compare the observed error pattern with those of known faults to determine the most likely cause of the anomaly.

and gaging on the HP stator and rotating rows (Figure 6). If.Proceedings of the MFPT Conference. May 2000 The error patterns associated with known faults must be determined through modeling or analysis of the failure modes. This is performed through altering a combination of chord length and surface roughness for various levels of SPE or Deposit damage and their respective efficiency effects. seal leakage. It is deemed an “intelligent” tool because it can operate in gray areas just as a human evaluator would. Figure 6 Modeling Degradation of Turbine Blades Figure 7 SPE Diagnostic Results Within the diagnostic module. HP-section. Figure 7 shows a simulated normalized performance error pattern for the case of SPE (top left) versus ideal fault error patterns. The model has the ability to simulate SPE or Deposit damage in the HP control stage. the 1st stage pressure is a little high (Figure 15). or flow blockage anywhere in the unit. we find the CRH 7 . Virginia Beach. rather than in black or white. Fuzzy logic is an artificial intelligence technique that has the ability to recognize error patterns that are not necessarily identical to that of a known fault. For the web-based steam turbine monitor. VA. computer simulations of a similar unit were run using a through-flow. performance model. For instance. in the case of solid particle erosion (SPE). in addition. fuzzy logic is used to examine error patterns and determine the most likely cause of anomalous conditions.

. Machine Condition Monitoring. Engineering Publications. Cotton Fact Inc.MFPT Committee Meeting. Australia. References: i ii Cotton. May 2000 pressure is ok. plots of performance trends. Evaluating and Improving Steam Turbine Performance. Posting the information on a web-site makes it readily available to all authorized personnel. and the 1st stage efficiency is a little low we can say with confidence that the unit is probably experiencing the onset of (SPE) in the HP control stage based on its similarity to the known error pattern for SPE. Users can ‘dig down’ to learn more about the condition of their equipment by displaying logs of past diagnostic results. and automated fault diagnosis that are useful to both technical and non-technical plant personnel. Figure 8 Fuzzy Logic Classification Results form the anomaly detection and diagnostic algorithms are immediately presented on the web site along with supporting information. K. Conclusion: This information rich condition monitoring system offers access to current plant data. the mass flow is high.. 1988. 1993 Beebe. Access to current performance and machinery health information allows risk managers. NY. C. Rexford. Virginia Beach. The authors wish to specifically acknowledge Peter Millet. and plant operation planners to improve maintenance scheduling and reduce the machinery life cycle costs. Victoria. and Bill Reuland of EPRI for their support and technical assistance. 8 . VA. and trend plots of key raw and corrected parameters. R. Acknowledgements: The on-line monitoring demonstration was developed under a grant from the Electric Power Research Institute (EPRI).