تطبیق داده های اندازه گیری شده توسط کارخانه به منظور بهبود عملکرد سخت افزار عیب یابی در راکتورهای آب تحت فشار

With the fairly recent adoption of digital control and instrumentation systems in the nuclear industry a lot of research now focus on the development expert fault identification systems. The fault identification systems enable detecting early onset faults of fault causes which allows maintenance planning on the equipment showing signs of deterioration or failure. This includes valve and leaks and small cracks in steam generator tubes usually detected by means of ultrasonic inspection. Detecting faults early during transient operation in NPPs is problematic due to the absence of a reliable reference to compare plant measurements with during transients. The distributed application of control systems operating independently to keep the plant operating within the safe operating boundaries complicates the problem since the control systems would not only operate to reduce the effect of transient disturbances but fault disturbances as well. This paper provides a method to adapt the plant measurements that isolates the control actions on the fault and re-introduces it into the measurement data, thereby improving plant diagnostic performance.

All online plant diagnostic approaches rely on the availability of a reference to compare the measured plant output to. With the use of plant training simulators in the nuclear industry since the 1980s and the continual improvement of computational speed and mathematical modelling methods, with increasing processing speed and improving model accuracies a full scope engineering simulator running in real time becomes the ideal reference to indicate faulty plant behaviour and degradation of plant hardware. With this availability and advances in nuclear power plant simulation technology, a research project was initiated to make use of simulators to provide a deterministic dynamic reference for in-transient fault detection (Cilliers et al., 2011). The primary objectives of the research were to: 1. Develop an early fault detection system by using real time simulators of nuclear power plants, continuously monitoring and comparing simulated measurement data and control outputs of the model reference adaptive control negative feedback system with the actual measured data and control outputs from the plant. The fault detection system should detect small faults that would normally go undetected as well as detect faults during plant operating transients. 2. Develop a fault characterisation method, making use of measured and simulated data together with the actual and simulated control system response. The fault characterisation system should provide information on the magnitude and location of the fault. 3. Develop a control and protection framework that allows NPP licensing within the existing licensing framework, but is still able to uncover the benefits of expert control and protection systems. This paper introduces a section of the first objective, with following papers presenting the second and third objectives. Petersen and McFarlane (2004) define the requirements for process fault detection and diagnosis as the following: • the availability of well defined, accurate first principles models (with any modelling errors also suitably represented), • clearly identified fault modes and models to represent these, and • appropriately located sensors (often with levels of redundancy). We have found these requirements to also be generic requirements applicable to the fault identification system developed as part of this research. Plant diagnostics in the Instrumentation and Control discipline has been researched and developed over the last 40 years with various methods of providing dynamic reference models. In 1976, Willsky (1976) examined statistical techniques for the detection of failures in dynamic systems revealing key concepts, similarities and differences in problem formulations, system structures, and performance. Specifically, they discussed the problem of detecting abrupt changes in dynamic systems. Chen and Howell (2001) notes that statistical and learning methods are fast and do not require a plant model, but are comparatively brittle because they cannot handle situations that are not explicitly anticipated. The ‘brittleness’ referred to makes the implementation of heuristic methods particularly problematic in the nuclear industry where deterministic proof of system dependability is required. Simani and Fantuzzi (2000) notes that Fault Detection Systems have been widely developed during recent years with model-based methods, fault tree approaches and pattern recognition techniques amongst the most common methodologies utilised in such tasks with neural networks used in Fault detection problems for model approximation and pattern recognition. Iserman (2004) presented an introduction into the fault detection field by describing how model-based methods of fault detection were developed by using input and output signals and applying dynamic process models. These methods are based on amongst others, parameter estimation, parity equations or state observers. Also signal model approaches were developed with goal is to generate several symptoms indicating the difference between nominal and faulty status. Based on different symptoms fault diagnosis procedures follow, determining the fault by applying classification or inference methods. The development of steady state Fault Detection Systems making use of the steady state references to detect control operation deviations in the nuclear industry has been done since 2000 with Chen and Howell (2001) commenting: Little has been written about distributing diagnostic tasks presumably because traditionally, the diagnostic engineer’s view of feedback control is that it complicates, rather than aids, diagnostic reasoning. Feedback control adds to the complexity of fault detection in process plants by masking measurement deviations that might indicate a fault. Also, Hamelin and Sauter (2000) realised that most of the developed algorithms make many idealised assumptions such as steady state conditions which are very often not satisfied, since in reality the system parameters may either be uncertain or time dependent, resulting in a mismatch between the actual system and the associated mathematical model used for reference. They state that even though the problem of uncertain parameters is of crucial importance to the industrial implementation of fault detection methods, it has however received little attention with only a handful of works so far devoted to it. Hamelin and Sauter (2000) also state that future developments should concern the incorporation of the proposed approach into a closed-loop system stating that: “Clearly, in a closed-loop system, the controller gain will also have an effect on the residual output.” This fault masking effect is compounded during expected plant transients when the various control systems are operating to the return the plant to steady state. The fault masking effect of the control system prevents makes plant diagnostics very difficult when small slow acting faults occur. To address the fault masking effect of the feedback control system, not only are the measurements and simulated measurements compared, but also the simulated and actual control operations. With the primary objective being to improve plant performance and fault detection dependability, a secondary benefit was achieved when early onset faults are detected while the control system manages to mask the fault completely. These faults often take a long time to develop large enough for conventional plant protection systems to detect the fault condition. The benefits from detecting early onset faults are the ability to execute online maintenance on the affective systems when possible or plan maintenance on these systems during outages. This results in increased uptimes as well as the reduction of efficiency losses due to faults. This was also used by Roy et al. (1998) who developed a fault characterisation system to address the need to have improved predictive maintenance techniques in an operating plant. Guidance into the methodology came from one of the earliest applications of state estimation based fault detection methods in nuclear plants. Roy et al.’s (1998) primary objective was to provide an early warning to the human operator regarding the failing health of control equipment, in the process averting major breakdown with its associated large plant downtime.

We have shown that having an accurate first principles real time simulator available aid in fault detection and characterisation. The methods developed in this research are applicable to all industries, unfortunately not all industries could benefit from this right away, as the nuclear industry are unique in a few instances: Plant operation simulators as a dynamic reference as Nuclear power plants can easily be implemented as nuclear power plants have rather slow dynamics, with control systems using transfer functions with time constants in the order of seconds or even minutes. This allows a first principles approach in plant operations modelling to be done to a high level of accuracy, and in real time. Other industries might find that f developing a model of a plant based on first principles to be used as a dynamic reference is expensive and time consuming (Chen and Howell, 2001). Nuclear plants, on the other hand are required to be equipped with engineering and training simulators. This can now be used to provide a first principles dynamic reference for a fault detection system. Fault detection systems based on neural networks are effective but brittle. The implementation of such systems will be difficult in the nuclear industry because of the uncertainties associated with such systems (Embrechts, 2004). The possibility of developing a fault detection system based on neural networks is eliminated, whereas this type of fault detection method could serve other industries very well. The problems addressed in this research are applicable to all industries, but especially effective in the nuclear industry as described above, the problems addressed were: The effectiveness of fault detection systems are diminished by the closed loop response of the control system to disturbances (Hamelin and Sauter, 2000). A method is developed in this research to measure the effect of the fault on the system without the closed loop response of the control system. During expected transients the effect of the fault is hidden further due to the operation of the control system on both the fault and the expected disturbance. A method is developed in this research to measure only the fault effect on the system outside the closed loop control system. Fault characterisation methods rarely make use of distributed control systems due to the fault hiding effect of the control system. Having a method that produces the full effect of the fault on each subsystem allows distributed control systems to be used in fault characterisation. It should prove effective in the sense that each subsystem responds differently to the fault disturbance and each should reveal a part of the nature and cause of the fault. The results in this research are made possible by having access to real time plant data and being able to perform manipulations on the data. The method of using a real time plant simulator to provide expected measurement and control data produce a clear picture of the component and system health is only one example of making effective use of available plant data.