Academia.eduAcademia.edu

Figure 1 – uploaded by Carl Galy

See full PDFdownloadDownload figure
Figure 1 — Hierarchy of Prognostic Approaches In general, health management technologies will observe features associated with anomalous system behavior and then relate these features to useful information about the system’s condition. In the case of prognostics, this information relates to the condition at some future time. Inherently probabilistic or uncertain in nature, prognostics can be applied to system/component failure modes governed by material condition or by functional loss. Like diagnostic algorithms, prognostic algorithms can be generic in design but specific in terms of application. Various approaches to prognostics have been developed that range in fidelity fom simple historical failure rate models to high-fidelity physics-based models. Figure 1 illustrates the hierarchy of potential prognostic approaches in relation to their applicability and relative costs. This paper will discuss some generic prognostic implementation approaches and provide some specific

Figure 1 — Hierarchy of Prognostic Approaches In general, health management technologies will observe features associated with anomalous system behavior and then relate these features to useful information about the system’s condition. In the case of prognostics, this information relates to the condition at some future time. Inherently probabilistic or uncertain in nature, prognostics can be applied to system/component failure modes governed by material condition or by functional loss. Like diagnostic algorithms, prognostic algorithms can be generic in design but specific in terms of application. Various approaches to prognostics have been developed that range in fidelity fom simple historical failure rate models to high-fidelity physics-based models. Figure 1 illustrates the hierarchy of potential prognostic approaches in relation to their applicability and relative costs. This paper will discuss some generic prognostic implementation approaches and provide some specific

Related Figures (14)

H u a ealth management system architectures must allow for the integration of anomaly, diagnostic, and prognostic (A/D/P) technologies from the component level all the way p through the vehicle or platform level. In general, A/D/P technologies observe features associated with anomalous system behavior and relate them to useful information bout component or system condition. Before getting into some specific examples of diagnostic and prognostic techniques applied to different aspects of an air vehicle, a brief description of some of the more common technical approaches are given. The software modules are currently in prototype development stage. Figure 2 - PEDS Plug and Play "Toaster" Model
H u a ealth management system architectures must allow for the integration of anomaly, diagnostic, and prognostic (A/D/P) technologies from the component level all the way p through the vehicle or platform level. In general, A/D/P technologies observe features associated with anomalous system behavior and relate them to useful information bout component or system condition. Before getting into some specific examples of diagnostic and prognostic techniques applied to different aspects of an air vehicle, a brief description of some of the more common technical approaches are given. The software modules are currently in prototype development stage. Figure 2 - PEDS Plug and Play "Toaster" Model
Figure 3 - Outline of the OSA/CBM Architecture MIMOSA is a not-for-profit trade association founded in 1994 and incorporated in December of purpose is the development and pu conventions for information exchange machinery maintenance information systems. The core o the MIMOSA development activity is t (Common Relational Information Schema). The second 996. Their genera blication of open between plant and he MIMOSA CRIS version of the CRIS (CRIS V2.1) was re and is publicly available at the MIMOSA website [http://www.mimosa.org/]. The CRIS defines a relationa eased in May 2000 database schema for machinery maintenance information. The schema provides broad coverage of the types of data that need to be managed within the CBM domain.
Figure 3 - Outline of the OSA/CBM Architecture MIMOSA is a not-for-profit trade association founded in 1994 and incorporated in December of purpose is the development and pu conventions for information exchange machinery maintenance information systems. The core o the MIMOSA development activity is t (Common Relational Information Schema). The second 996. Their genera blication of open between plant and he MIMOSA CRIS version of the CRIS (CRIS V2.1) was re and is publicly available at the MIMOSA website [http://www.mimosa.org/]. The CRIS defines a relationa eased in May 2000 database schema for machinery maintenance information. The schema provides broad coverage of the types of data that need to be managed within the CBM domain.
Figure 4 - Experienced-Based Approach has little or no sensed parameters associated with it. In this case, the prognosis of when the component will fail or degrade to an unacceptable condition must be based solely on analysis of past experience or OEM recommendations. Depending on the maintenance complexity and criticality associated with the component, the prognostics system may be set up for a maintenance interval (i.e. replace every 1000+/-20 Effective Operating Hrs) then updated as more data becomes available. Having an automated maintenance database is important for the application of experience-based prognostics.
Figure 4 - Experienced-Based Approach has little or no sensed parameters associated with it. In this case, the prognosis of when the component will fail or degrade to an unacceptable condition must be based solely on analysis of past experience or OEM recommendations. Depending on the maintenance complexity and criticality associated with the component, the prognostics system may be set up for a maintenance interval (i.e. replace every 1000+/-20 Effective Operating Hrs) then updated as more data becomes available. Having an automated maintenance database is important for the application of experience-based prognostics.
Figure 6 - Feature/AI-Based Prognostics Utilizing known transitional or seeded fault/failure degradation paths of measured/extracted feature(s) as they progress over time is another commonly utilized prognostic approach. In this approach, neural networks or other AI techniques are trained on features that progress through a failure. In such cases, the probability of failure as defined by some measure of the “ground truth” is required as a-priori information as described earlier. This “ground truth” information that is used to train the predictive network is usually obtained from inspection data. Based on the input features and desired output prediction, the network will automatically adjusts its weights and thresholds based on the relationships between the probability of failure curve and the correlated feature magnitudes. Figure 6 shows an example of a neural network after being trained by some vibration feature data sets. The difference between the neural network output and the “ground truth” probability of failure curve is due to error that still exists, after the network parameters have optimized, to minimize this error. Once trained, the neural network architecture can be used to intelligently predict these same features progressions for a different test under similar operating conditions.
Figure 6 - Feature/AI-Based Prognostics Utilizing known transitional or seeded fault/failure degradation paths of measured/extracted feature(s) as they progress over time is another commonly utilized prognostic approach. In this approach, neural networks or other AI techniques are trained on features that progress through a failure. In such cases, the probability of failure as defined by some measure of the “ground truth” is required as a-priori information as described earlier. This “ground truth” information that is used to train the predictive network is usually obtained from inspection data. Based on the input features and desired output prediction, the network will automatically adjusts its weights and thresholds based on the relationships between the probability of failure curve and the correlated feature magnitudes. Figure 6 shows an example of a neural network after being trained by some vibration feature data sets. The difference between the neural network output and the “ground truth” probability of failure curve is due to error that still exists, after the network parameters have optimized, to minimize this error. Once trained, the neural network architecture can be used to intelligently predict these same features progressions for a different test under similar operating conditions.
The purpose of this investigation was to identify features tha could be incorporated in an automated system for diagnosing clogged fuel nozzles. The main focus was diagnostics, but there is potential for incorporating a prognostics element giver sufficient clogging progression data. This example focuses largely on a feature-based approach. Figure 8 — Clean Nozzle and One with Severe Clogging Clogging reduces the efficiency of the combustion process and can create potentially damaging hot spots in the combustor and turbine sections. At startup, this is especially true to the extent that “hot starts” or “no starts” may be produced. For this project, Impact Technologies used test data from several start-ups of an Allison 501-K17, taken at NAVSEA Philadelphia, comprising both clean and fouled fuel nozzles.
The purpose of this investigation was to identify features tha could be incorporated in an automated system for diagnosing clogged fuel nozzles. The main focus was diagnostics, but there is potential for incorporating a prognostics element giver sufficient clogging progression data. This example focuses largely on a feature-based approach. Figure 8 — Clean Nozzle and One with Severe Clogging Clogging reduces the efficiency of the combustion process and can create potentially damaging hot spots in the combustor and turbine sections. At startup, this is especially true to the extent that “hot starts” or “no starts” may be produced. For this project, Impact Technologies used test data from several start-ups of an Allison 501-K17, taken at NAVSEA Philadelphia, comprising both clean and fouled fuel nozzles.
Figure 7 - Physics-Based Prognostics
Figure 7 - Physics-Based Prognostics
Table 1 — Recorded Parameters from the Digital Control System When a compressor undergoes fouling, several key performance factors are affected. The most sensitive of these factors is the compressor capacity or referred mass flow. (Peltier et al, 1995) This is because loss of capacity comes from throat blockage and increases in roughness on the suction side of the blading. Unfortunately, in most practical naval applications, compressor capacity is not reliably determinable. The compressor inlet temperature (CIT), outlet temperature (CDT), inlet total pressure (CIP;) and discharge total pressure (CDP;) can typically be used to find compressor efficiency. (Boyce 1995) However CDT, CDP; are not standard sensors in most Naval platforms.
Table 1 — Recorded Parameters from the Digital Control System When a compressor undergoes fouling, several key performance factors are affected. The most sensitive of these factors is the compressor capacity or referred mass flow. (Peltier et al, 1995) This is because loss of capacity comes from throat blockage and increases in roughness on the suction side of the blading. Unfortunately, in most practical naval applications, compressor capacity is not reliably determinable. The compressor inlet temperature (CIT), outlet temperature (CDT), inlet total pressure (CIP;) and discharge total pressure (CDP;) can typically be used to find compressor efficiency. (Boyce 1995) However CDT, CDP; are not standard sensors in most Naval platforms.
‘igure 9 — Feature Diagnostic Scalars at Baseline and under Degraded (Fouling) Conditions
‘igure 9 — Feature Diagnostic Scalars at Baseline and under Degraded (Fouling) Conditions
Figure 10 - Borescopic Image of Salt deposits: 1* stage Blading
Figure 10 - Borescopic Image of Salt deposits: 1* stage Blading
Figure 11 - Prognostic Model Visualization To manipulate the data into the form of this model, the time dependency of the test results had to be removed because of the unrealistic fouling rates. The percent changes in static pressure ratio, fuel flow, and CDT were recast in terms of 4 % pseudo-efficiency drops. This increment was chosen because it was the highest resolution that still permitted statistical analysis. With the assimilation of the data into these discrete bands, the statistical parameters (e.g., mean and standard deviation) can be ascertained for use in the prognostic model. Figure 11 shows the evolution of the compressor degradation for the LM-2500 test at 1% pseudo-efficiency drops (for visual clarity). The top two plots illustrate the distributions of pressure ratio and fuel flow respectively while the bottom two provide the joint probability distributions. Once the statistical performance degradation path is realized along with the capability to assess current degradation severity, the final step was to implement the predictive capability. The actual unit-specific fouling rate is combined with historical fouling rates with a double exponential smoothing method. This time series technique weights the two most recent data points over past observations. The following equations give the general formulation. (Bowerman, 1993). Figure 12 shows how this technique can give significantly different results than standard regression.
Figure 11 - Prognostic Model Visualization To manipulate the data into the form of this model, the time dependency of the test results had to be removed because of the unrealistic fouling rates. The percent changes in static pressure ratio, fuel flow, and CDT were recast in terms of 4 % pseudo-efficiency drops. This increment was chosen because it was the highest resolution that still permitted statistical analysis. With the assimilation of the data into these discrete bands, the statistical parameters (e.g., mean and standard deviation) can be ascertained for use in the prognostic model. Figure 11 shows the evolution of the compressor degradation for the LM-2500 test at 1% pseudo-efficiency drops (for visual clarity). The top two plots illustrate the distributions of pressure ratio and fuel flow respectively while the bottom two provide the joint probability distributions. Once the statistical performance degradation path is realized along with the capability to assess current degradation severity, the final step was to implement the predictive capability. The actual unit-specific fouling rate is combined with historical fouling rates with a double exponential smoothing method. This time series technique weights the two most recent data points over past observations. The following equations give the general formulation. (Bowerman, 1993). Figure 12 shows how this technique can give significantly different results than standard regression.
Figure 13 - Gear Model-Based Prognostics Under the Phase I SBIR effort a physics-based model for geartooth failure was developed. It illustrates the physics- based model approach. This model was chosen because it could be validated and calibrated on transitional (run-to- failure) data from the MDTB (Mechanical Diagnostics Test Bed) at the Penn State Applied Research Lab. This prognostic module is a near real-time, self-calibrating. physics-based statistical RUL predictor of gear tooth failure due to tooth spalling or low cycle fatigue (LCF)
Figure 13 - Gear Model-Based Prognostics Under the Phase I SBIR effort a physics-based model for geartooth failure was developed. It illustrates the physics- based model approach. This model was chosen because it could be validated and calibrated on transitional (run-to- failure) data from the MDTB (Mechanical Diagnostics Test Bed) at the Penn State Applied Research Lab. This prognostic module is a near real-time, self-calibrating. physics-based statistical RUL predictor of gear tooth failure due to tooth spalling or low cycle fatigue (LCF)
Figure 12 - Prediction of Degradation Rates The test data made two essential contributions to the development of this prognostic model. First, it provided a means by which to validate an analytical model of how performance parameters change as a function of compressor fouling. Secondly, they gave insight into the sensitivity and statistical distributions of performance parameters as a function of load. Hence, having been developed and validated on real data, a large amount of knowledge is “built in” to the prognostic model.
Figure 12 - Prediction of Degradation Rates The test data made two essential contributions to the development of this prognostic model. First, it provided a means by which to validate an analytical model of how performance parameters change as a function of compressor fouling. Secondly, they gave insight into the sensitivity and statistical distributions of performance parameters as a function of load. Hence, having been developed and validated on real data, a large amount of knowledge is “built in” to the prognostic model.
Related topics:
Naval EngineeringHistoryAlgorithmsIndependent Component AnalysisMarine EngineeringGAS TURBINELife Cycle ManagementAircraft MaintenancePain MedicineCondition Assessment
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved