Second, one needs representa- expert systems for fault diagnosis has been considered tive patterns against which the similarity of other by Becraft and Lee Modifications have been suggested to the standard back-propagation network for the problem of fault diagnosis also. It has been argued that basis functions generating bounded decision regions could be better suited to the problem of fault diagnosis. For example, Leonard and Kramer suggested the use of radial basis function networks for fault diagnosis applications.
In classification, the decision boundary is often not unique and this requires a means of saying whether a Fig. Characteristics of nodes used in neural networks. On the literature is the K-means clustering algorithm Duda the other hand, if there is only a single sensor indicating and Hart, K-means clustering pre-supposes the whether the system is normal or faulty, then nothing can number of clusters needed and would cluster the data be diagnosed including the proper functioning of the accordingly. It utilizes all the cluster centers so that each sensor itself. Effectiveness of any diagnostic procedure is of the clusters is guaranteed at least one pattern.
Given a particular choice of measurements, updates this cluster center and all its topological these methods can specify a whether a fault is neighbors. In unknown faults that we do not consider? This leads from other faults we intend to identify? In answering to the problem of gravity where all the cluster centers these questions these methods provide design schemes in migrate towards dense regions leaving less dense regions which the effects of unknown disturbances can be unrepresented.
To avoid this, a fuzzy clustering ap- minimized. Whiteley and Davis demon- strate the use of ART2 network for the interpretation of E is the distribution matrix for the disturbances d t sensor data. Chen, Wang, Yang, and Mcgreavy which in the general case includes both structured and and Wang, Chen, Yang, and Mcgreavy discuss unstructured uncertainties. The term Ed t characterizes the integration of wavelets with ART networks for the the unknown input or disturbance and represents all development of diagnostic systems.
For a collection of uncertainties acting upon the system. The major pro- papers on the application of neural networks in chemical blem in this approach is in this simplistic approximation engineering problems, an interested reader is referred to of the disturbances that include modelling errors. In the Venkatasubramanian and McAvoy and Bulsari design of the nonlinear unknown input observer, more In practice, however, severe modelling uncertainties occurring due to parameter drifts come in the form of multiplicative uncertainties.
This is a general limitation 4. A comparison of various approaches of all the model-based approaches that have been developed so far. In addition to difficulties related to So far in this three part series, we have reviewed the modelling they do not support an explanation facility three conceptually different frameworks for process owing to their procedural nature.
The type of models the fault diagnosis. In this section, we provide a compara- analytical approaches can handle are limited to linear tive evaluation of these different frameworks against a and some very specific nonlinear models. For a general common set of desirable characteristics for a diagnostic nonlinear model, linear approximations can prove to be system that we proposed in part I. The evaluations are poor and hence the effectiveness of these methods might summarized in Table 2. When a large-scale process is Quantitative model-based methods, such as parity considered, the size of the bank of filters can be very space and observer-based approaches, have several large increasing the computational complexity.
If one has Rule-based expert systems can be used where funda- complete knowledge of all inputs and outputs of the mental principles are lacking, where there is an abun- system, including all forms of interactions with the dance of experience but not enough detail is available to environment, fault diagnosis would be a well-defined develop accurate quantitative models.
Causal models are also a very good alternative when the quantitative models are not avail- able but the functional dependencies are understood. Abstraction hierarchies help to focus the attention of the diagnostic system quickly to problem areas. One of the advantages of qualitative methods based on deep- knowledge is that they can provide an explanation of the path of propagation of a fault.
This is indispensable when it comes to decision-support for operators. They can also guarantee completeness in that the actual fault s will not be missed in the final set of faults identified. However, they suffer from the resolution problems resulting from the ambiguity in qualitative reasoning. When quantitative information is available partially, one could use the order-of-magnitude analysis or interval-calculus to improve the resolution of purely qualitative methods. Pattern recognition approaches or classifiers are Fig.
- Fault detection and diagnosis in engineering systems.
- A Guide to Fault Detection and Diagnosis.
- 1st Edition.
Location of multiple faults in the output space. Neural network architectures, such as radial basis function measurement space shown in Fig. Three fault classes networks and ellipsoidal unit networks, have been are shown as shaded circles at three corners of the cube. There are some considered in the measurement space for generalization. It is the limitation of their generalization capability outside of the training data. This allows the network to detect unfamiliar situations arising from novel faults.
This is, in general, beyond the scope of a classifier and Besides its lack of ability to generalize to unfamiliar cannot be guaranteed. If we consider classifiers which regions of measurements space, networks also have a are linear transformations, then we should expect a difficulty with multiple faults. This brings out a crucial similar structure of the data in both the input and point of distinction between model-based approaches output space.
In the case that the structure in the input space has strong implica- of qualitative model-based approaches, the combinator- tions on what output one can expect. Because of the combinatorially many fault diagnosis of x1 properly. However, this cannot be multiple fault combinations, the search for multiple ensured for the measurement space. One needs to faults by specifying them explicitly as different classes determine a proper feature space that has this property.
If one uses as input the residuals from the parity Given this limitation, the ability of neural networks to equations of a parity-space approach, then the input generalize to multiple faults has been tested with some would be as expected in Fig. This suggests that, subramanian et al. Particular architectures when model information is available, one should use a showed a greater ability to generalize than certain bank of filters to generate the inputs. For example, use of a mixture of Gaussians or Filter banks are computationally expensive.
To explain this, consider the 3-dimensional some of the methods discussed in this three part series. In favorable, then multiple fault diagnosis is possible. This can be contrasted with for comparison. A check mark would indicate that the an observer-based approach where one could explicitly particular method column satisfies the corresponding include multiple fault identifiability in the design desirable property row. A cross would indicate that the procedure.
We used similar basis for generating the property is not satisfied and a question mark would entries for other methods given in Table 2. This is due to the fact that process history properties of neural networks have been demonstrated based approaches are easy to implement, requiring very through the use of many case studies by different little modelling effort and a priori knowledge. Neural networks that generate bounded decision lized towards fault diagnosis applications. Due to the procedural very few industrial applications in published literature nature of neural network development, they lack the that deal with parametric failures.
Among the process explanation and adaptability properties. Further, gen- history based approaches, statistical approach seems to eration of classification error estimates is difficult using have been well studied and applied. The reason for this the neural network approach. Since neural networks might be that with the current state-of-art in applica- predominantly work with process history data, the tions, detection seems to be a bigger concern than modelling requirements are minimal. Further, once a detailed diagnosis.
Hence, statistical approaches that neural network is trained, the on-line computations are are easy to build and which do very well on fast simple function evaluations and hence the on-line detection of abnormal situations have been successfull computational complexity is minimal. Finally, regarding in industrial applications. Once a diagnostic system is deployed, it should be QTA approach specializes more on diagnosis than able to adapt with minimal effort as new situations are detection and hence might be a useful tool where encountered and the scope of the system is expanded.
QTA approach seems to be These issue need to be investigated further for successful robust to routine variations in process operations; deployment of diagnostic systems in industrial setting. There are some observers have made an impact in mechanical and general guidelines based on experience on the economic aeronautical engineering applications, they have had impact due to abnormal situations, but there are no case very little impact in process industries.
This might be studies that analyze the specific benefits that can be due to the following reasons: attained through the implementation of diagnostic systems.
- Fault Detection and Diagnosis in Engineering Systems.
- Computational text analysis for functional genomics and bioinformatics.
- Fault Detection and Diagnosis in Engineering Systems (Electrical Engineering and Electronics).
- AutoCAD Civil 3D 2015: Essentials (Autodesk Official Press Series);
- Fault Detection and Diagnosis in Engineering Systems | Taylor & Francis Group.
- Machining of titanium alloys and composites for aerospace applications?
- Sarabande in G Major?
More research is needed on this issue in line i Chemical processes are inherently nonlinear in with the work that has been carried out analyzing the nature. While the theory of linear quantitative benefits of implementation of advanced control systems.
Hybrid methods open issue. Though all the methods are restricted, in the approach such as PCA might be minimal. Hence, sense that they are only as good as the quality of it is easier to implement a PCA-based detection information provided, it was shown that some methods approach than a model-based approach. It iii Model-based approaches have been predominantly is our view that some of these methods can complement restricted to sensor and actuator failures.
Integrating these complementary features is one way The impact of qualitative model-based approaches to develop hybrid methods that could overcome the such as QSIM and QPT in terms of applications has limitations of individual solution strategies. Hence, been minimal. Many of the academic demonstrations of hybrid approaches where different methods work in these models have been on very simplistic systems and conjunction to solve parts of the problem are attractive. Graph-based ap- fault isolation might be very difficult using digraphs due proaches have been researched upon quite extensively to the qualitative ambiguity and analytical model-based and they have been applied in safety studies, such as methods might be superior.
Hence, hybrid methods HAZOP analysis and tools are being developed for might provide a general, powerful problem-solving using these types of models in real-time decision making. There has already been some work on hybrid In general, literature on industrial applications of architectures. The two-tier approach by Venkatasubra- diagnostic systems are not many. This could be due to manian and Rich using compiled and model- the proprietary nature of the development of in-house based knowledge is one of the earliest examples of a systems.
Also, there seems to be a general lack of overall hybrid approach. Process specific knowledge, also called penetration of diagnostic systems in process industries. This and industrial practice. Two of the most important compiled knowledge can be acquired or learned during considerations from an industrial viewpoint such as the the operations. This can reduce the time of search for adaptability of the systems and ease of deployment are frequent faults.
Frank advocates the use of seldom addressed in academic research. Most of the knowledge-based methods to complement the existing techniques would do poorly on the issue of ease of analytical and algorithmic methods of fault detection. Contrast this with, for example, the ease with ent kinds of knowledge in one single framework for which PID controllers can be deployed. The other issue better decision making. The resulting overall fault of adaptability is crucial too from industrial perspective.
This is a task that mework where different diagnostic methods work in should be performed at the design stage of the plant. The wamy and Venkatasubramanian and Vedam, detectability index characterizes the ability of the system Dash, and Venkatasubramanian A blackboard in detecting specified faults, whereas, separability index architecture, called Dkit, in which different diagnostic characterizes the ease of separation of faults as a methods analyze the same problem, and a scheduler function of the minimum angle between the functional which regulates the decision-making of these methods is subspaces of the faults.
Observability index imposes full the central concept in this framework. The utility of rank conditions on the system matrix. Sensors are such a hybrid framework for solving real-time complex located based on the minimization of all three indices. This framework was adopted by the Honeywell so on. In their work, they have shown how DG and 5. Role of fault diagnosis in design and other process SDG can be used for deciding the location of sensors. These include such tasks as identifying the occurrence of events outside of normal operation, 5.
Data reconciliation diagnosing the root cause and finally synthesizing and implementing a corrective action.
Data reconciliation is an important activity per- Fault diagnosis shares with other process operations formed in the continuous process industries. Data the realization that with powerful knowledge represen- reconciliation is essentially a quantitative fault diagnosis tation schemes, one can capture the expertise of approach with the focus on detecting sensor faults and operators and control engineers that was gained over sensor biases.
Also, another important goal is the years of experience with process plants.
Process specific reconciliation of measurement data. Data reconciliation knowledge can be used to improve general purpose usually consists of three parts: i identification of the methodologies. There is a close coupling between biased parameter; ii estimation of the bias; and iii diagnosis and process operations and design of chemical rectification of the sensor measurements. Proper design of a chemical plant can reduce the Data reconciliation can be performed in both steady- burden on the task of diagnosis. Also, the information state and transient conditions.
Steady-state data recon- from diagnosis can be used to continuously improve the ciliation is the problem of removing errors from sensor performance of process operations. The information variables given a collection of data points. It is usually from fault diagnosis can be incorporated into the handled using linear least squares estimation techniques traditional solution paradigms of other process opera- and has been shown to give reliable results.
In contrast, tions. The aim of this section is to provide a brief dynamic data reconciliation is the problem of removing overview of various design and operations modules that errors under time evolutions of sensor variables. This is would particularly share information with fault diag- a significantly tougher problem due to the presence of nosis module and also outline the nature of interaction differential constraints and the pronounced effect of that one can expect. Instead of the purely quantitative approaches to data 5.
Optimal sensor location reconciliation one can use a combination of qualitative approaches and quantitative approaches to better solve Much of the information one gets about the state of the data reconciliation problem. The diagnostic quali- the system is from its sensors. The usefulness of such equipment malfunction the process could have drifted an approach has been demonstrated by Vachhani, towards a new steady state.
There could be other Rengaswamy, and Venkatasubramanian There could be changes in the process 5. Supervisory control parameters due to external influences, which could call for a different control configuration, or different set Supervisory control is an activity that falls in scope points and gains.
Planning, Hence, fault diagnosis is an important module that typically, is the task of creating schedules and making can help with information for supervisory decision- operating decisions on the time-frame of months. It making. At a lower level, ideas from diagnosis can be becomes imperative to have a decision layer at a lower used to perform controller diagnostics.
Kraus and level than planning to coordinate various individual Myron discuss an Expert Adaptive Controller control loops and to do exception handling on a much Tuning EXACT Controller using ideas from pattern smaller time-frame. Plant behavior in most cases is not recognition approach. Gertler discusses an in- perfectly known.
Even rigorous models are not adequate telligent supervisory control which supplements a basic to predict plant behavior with satisfying accuracy. There feedback loop with an outer adaptation loop, consisting is a natural variability in the process due to raw material of an identifier and a tuner. Identification means the variability and due to unsteady environmental condi- estimation of the plant model on the basis of a sequence tion. Tuning would require control strategy. Once a model is obtained and the both monitoring the system and adaptation based on the controller designed, changes in the plant operation can tests performed.
The monitoring part of the algorithm occur that can render the controller ineffective. For does excitation monitoring, stability monitoring and many situations this information on changes cannot be trend filtering. Tuning would then be performed to anticipated at the design stage, but the controller may be improve process operations or restore stability as the expected to perform over a different operating region or case may be.
When this happens, the controller may exhibit undesirable behavior which requires some action either 6. Conclusions and future directions by the operator or the control engineer. First some re- tuning may be performed on-line. If this does not The basic aim of this paper is to organize, classify, correct the problem the loop would be taken off line review and compare various approaches to fault diag- and a complete redesign performed. As the number of nosis from different perspectives. This is where we model-based methods; and iii process history based need a supervisory control system to assist in the methods.
We also present a framework that shows how decision process. The supervisory control system would these different approaches relate to and differ from each use the information available from the fault diagnosis other regarding the transformation of information from system to check and monitor the loops in the regulatory the measurement space to the decision space. The control system.
If there has been changes in the control important components in the transformation are the loops, the supervisory control system would then look transformation from measurement to feature space and for different control configurations or set points that the transformation from feature to decision space. A would improve the process operations.
This led to the discussion on mathematical models of the process and the controller, diagnostic systems in terms of the typology of a priori but they are also crucially dependent upon whether the knowledge used. In many ways the assumptions may be that one would like a diagnostic system to possess. We violated.
This terms of these characteristics. This comparative study is one of the key assumptions underlying the models. If a identifies the relative strengths and weaknesses of the sensor is faulty, the controller action may become different approaches. It also reveals that no single ineffective or it may even cause adverse process beha- method has all the desirable features we stipulated for vior. The model could have been linearized near the a diagnostic system. Integration framework.
Integrating these complementary broadly classified the future directions into three cate- features is one way to develop hybrid systems that gories, they are, however, not insulated from each other. We discuss strategies. We think that hybrid systems are an im- below the key issues involved in these three future portant future direction for research and development in directions of research briefly. Two other areas of equal importance As noted before, the drawbacks of single-method for future research are: i integration of diagnostic based diagnostic systems are serious enough to limit methods with other process operations for a more their applications to small case studies and render them comprehensive and effective intelligent supervisory con- unsuitable for large-scale industrial situations.
This trol system; and ii implementational issues for large- makes the design and development of hybrid systems scale industrial applications. Even though we have important. The design and For similar reasons, the Honeywell ASM Con- implementation of hybrid systems as well as integration sortium adopted the Dkit architecture as its AEGIS of diagnosis with other process operations face several prototype, a next-generation intelligent control system technical challenges.
Without going into detail, we for operator support. Our experience in designing such would like to list some key ones here: systems suggests that a practically successful hybrid system is likely to have at least three diagnostic i Ability to reason about process operations without components: i a quantitative method such as PCA assuming accurate models.
QTA, wavelets for explicitly assessing process information about the process. Despite such promising starts towards hybrid systems, iv Ability to make assumptions about a process when much work still remains to be done. One has to ensure the Another important direction for research is the validity and consistency of these assumptions. Rengaswamy, The overall problem of process vi Ability to maintain global database and global operations management involves several subproblem management of process knowledge.
For effective compression.
Fault Detection and Diagnosis in Engineering Systems (Electrical Engineering and Electronics)
In the Developing solutions to overcome these difficult case of data reconciliation, traditionally one does not issues towards the design of intelligent supervisory consider parameter drifts and structural faults as part of control systems will set the pace of research and the problem.
However, an integrated view is necessary development for the coming decade and beyond for for reconciliation of measured data in the presence of engineers in academia and industry. Indeed, we see the process faults. Low-level events such as sensor failure or successful design and implementation of intelligent some other equipment malfunction, can have a signifi- supervisory control systems for operator support in a cant impact on the higher level plans by calling for the variety of large-scale process applications as the next revision of previous schedules.
Likewise, higher level grand challenge problem for process control engineers. Thus, while these operational tasks may be intrinsically different from each other, they are, however, closely related to each other and cannot be treated as isolated tasks. References Hence, one needs an approach wherein all these different tasks can be integrated into a single unified Anderson, T.
An introduction to multivariate statistical framework so that the operational decision-making can analysis. New York: Wiley. Bakshi, B. Multiscale PCA with application to multivariate be made more comprehensively and more effectively.
ISBN 13: 9780824794279
Temporal representation of academia, have largely tended to treat them as separate process trends for diagnosis and control. In IFAC symposium on problems. This Bakshi, B. Wave-net: a multi- creates great research opportunities for the traditional resolution, hierarchical neural network with localized learning. A model-object based Garcia, C. In control. Basseville, M. Detection of abrupt Gertler, J.
Intelligent supervisory control. Information and system sciences J. Davis Eds. Prentice Hall. Becraft, W. Analytical redundancy methods in fault deduction system approach for fault diagnosis. Computers and Chemical and isolation. Bhushan, M. Design of sensor network Gertler, J. All rights reserved. Remember me on this computer.
Cancel Forgot your password?
Fault Detection and Diagnosis in Engineering Systems by Janos Gertler
Some features of WorldCat will not be available. Create lists, bibliographies and reviews: or. Search WorldCat Find items in libraries near you. Advanced Search Find a Library.