·Table of Contents
·Methods and Instrumentation
Ultrasonic Pulse-Echo Method for Failed Rod Detection Based on Neural NetworkW.C.A.Pereira, Z.D.Thomé, J.M.Seixas, M.C.Bossan, *
COPPE/UFRJ - Federal University of Rio de Janeiro
P.O. Box 68510, Rio de Janeiro 21945-970, Brazil
Email : firstname.lastname@example.org
W. Soares Filho
IPqM- Brazilian Navy Research Institute
R. Ipiru,2 - Rio de Janeiro 21931-090- Brazil
*Partially supported by Eletrobras Termonuclear S.A., FUJB and CNPq (Brazil).
Nuclear Reactor Power Plant is a well-known technology for large-scale energy generation. They are based on the uranium nuclear fission. Among them, the Pressurized Water Reactor (PWR) is one of the most used . For operation the uranium is kept in zircaloy rods (3 m height) as a stack of pastilles. One of the most important issues in routine PWR operation is the detection of failed fuel rods [2,3]. It has implications not only on the reactor performance itself but mainly from the safety point of view. The rods have a gap between the cladding and the nuclear fuel. In normal conditions, this gap contains pressurized helium. The rods are immersed in water and, when a failure occurs, water enters into the gap, filling them partially. So the presence of water inside the gap is a sign of a rod failure. Its detection of can be made by ultrasonic techniques .
This work presents an ultrasonic (US) inspection method (based on pulse-echo technique), which is developed for determining possible failed rods in each assembly (matrix formed by 16x16 fuel rods) of Angra I (Brazilian Nuclear Power Plant).
The pulse-echo technique is based on the fact that the US wave is able to distinguish different acoustic impedances when there is gas or water inside the rod gap. This property is used as the physical principle for the inspection.
|Fig 1: Cross section of a fuel rod showing the interfaces reached by the US waves. The gap contains gas or water. The rod is immersed in water, as well as the US sensor.|
The US pulse propagates in the water and reaches the water/cladding (zircaloy) interface (Figure 1). Part of the pulse is back propagated to the transducer's face and the other part of the US signal goes forward and reaches the next interface: cladding/gas (or cladding/water, according to the contents of the gap). The wave inside the cladding shell is reverberated between its internal and external interfaces (cladding/gap and cladding/outside water, respectively). This signal eventually goes back to the transducer and is recorded for the analysis. A typical echo signal is presented in Figure 2. From the ultrasonic point of view the reverberation signal has different characteristics depending on the contents of the gap (whether it is gas or water). In general for failed fuel rods (water inside the gap) the reverberation signal carries less energy compared to the intact ones. This is due to the ultrasonic impedance difference at the cladding/gap water interface, compared to the cladding/gap gas interface. More energy passes through the water layer than through gas layer.
|Fig 2: Typical envelope signal received by the US sensor. The 4 peaks in the middle represent the reverberation.|
A neural discriminator was used to classify the US signals obtained in the acquisition phase . These signals were time aligned to compensate for the significant fluctuations on the time delay required for signal arrival. The correlation of each signal with a reference signal, selected from the database, was used to make the alignment. The neural discriminator, was a three-layer, feedforward and fully-connected neural network, trained with the backpropagation algorithm . The input nodes of the neural network was fed with the 135 samples that followed the fully decay of the first peak of the time aligned echo signal envelopes. This fixed time window contains the echo signals corresponding to the reverberations inside the rod internal wall. Ten neurons were used in the hidden layer and one single neuron in the output layer. The number of neurons in the hidden layer was chosen to give a good classification performance.
The set of echo signals (4,568 signals corresponding to air-filled rods and 284 corresponding to failed rods) was split into two, so that half of the signals for each rod formed the training set. The other half did not participate of the training phase and formed the testing set, which was used to evaluate the generalization capability of the network.
The identification of a given failed rod was performed by analyzing the network response to all signals, in the testing set, that were accumulated for the rod under consideration. For such identification, the neural network has to classify the majority of the reflected signals as belonging to the class of failed rods.
The neural detection of failed fuel rods was compared to two non-neural methods that have currently been used for classifying ultrasonic signals in this environment: the exponential method and the Fisher linear discriminant method. The exponential method explores the absorption process features of the ultrasonic pulse inside a rod. The idea is to identify failed rods by using an exponential decay model for the peaks of the received echo signal. To have the signal peaks better defined, the received signal was low-pass filtered using a seven-point moving average filter. On the other hand, the orthogonality condition between the ultrasonic source and the external surface of the rods has been proved to be quite important to define a well-behaved decaying exponential. The required orthogonality condition can be evaluated by examining the amplitude of the first peak of the echo signal. If this peak is above a certain threshold (240 arbitrary units in this work), the orthogonality condition can be satisfied and the corresponding received echo signal may be validated. For the data set under analysis, only 182 signals corresponding to failed rods and 2703 to normal rods were accepted, according to this method.
The second method used for comparison refers to the Fisher linear discriminant , W, which is a mapping that aims at reducing the dimensionality of the signal vectors v from N to d=M-1, where M is the number of classes involved, while preserving the optimal separability among classes. The separation is obtained by analyzing the inner product "s" of the linear discriminant "W" and the signal "v".
In case of two classes, W can be computed as:
where m1 and m2 are the averages of the signals corresponding to the classes C1 and C2 and SW is the intra-class spreading matrix, defined by
To obtain a typical signal like the one presented in Figure 2, the sensor and the rod must be well aligned. This is not always the case. Figure 3 presents the highest peak of the echo for several positions of alignment. It is possible to see from the figure a plateau where the amplitudes are less sensitive to the alignment and, therefore, there is enough energy in the signal to be processed. This establishes a range of alignment positions from where the echoes can be obtained and processed. Outside this region the echo amplitude decreases substantially.
|Fig 3: Highest peak of the echo signal for several positions of alignment between the rod and the sensor. Vertical scale is the amplitude (arbitrary units) of the 1st echo from the outside water/cladding interface.|
Table 1 shows quantitatively the performance of the neural classifier for the testing set. The neural network correctly classifies 80% of the signals from rods with water and 95% from rods with air. The overall efficiency for the testing set was 94%. The corresponding efficiency for the identification of failed rods was found to be 93% (from a set of 14 failed rods analyzed, 13 rods were correctly classified), with a false alarm detection probability of 2% (from a set of 191 rods with air, 4 rods were misclassified as failed rods). The overall efficiency for the classification of rods was 98%, using the neural classifier.
|Decision||Signals (%)||Rods (%)|
|Table 1: Classification performance for signals and rods.|
Table 2 shows a summary of the performance obtained by the different methods, which indicates that the neural classifier outperforms the non-neural methods.
|Method||Signals (%)||Rods (%)|
|Table 2: Performance of different detection methods.|
Figure 4 shows in details the neural detection of failed rods by using an efficiency map for the entire assembly. Here, the cells represent the original position of the rods in the prototype assembly, and the detection efficiency is coded using a gray scale, for which lighter colors represent higher efficiency. According to this, the lighter is the map, the better is the discriminator performance.
|Fig 4: An efficiency map for the neural classifier. GT are guiding tubes that belong to the rod matrix structure.|
Figure 5 shows the efficiency maps for the non-neural methods. It can be seen that the exponential method suffers dramatically from the fluctuations of the ultrasonic signals, which translates into poor failed rod identification performance. In this map, rods with poor statistics due to the required signal validation and which could not be analyzed are marked with XX label. In a practical inspection procedure, these rods would need a second inspection for their classification.
Fig 5: Efficiency maps for the exponential (a) and Fisher discriminant (b) methods. GT are guiding tubes that belong to the rod matrix structure.
It should be noticed that the neural classifier proved to be insensitive to the orthogonality condition of the ultrasonic probe and consequently did not require any signal validation. Thus, the neural classifier performs much better and considers all acquired signals from the continuous emission of the ultrasonic probe, which keeps high the statistics for signal detection and allows a faster inspection. The Fisher discriminant method improves quite a lot the performance of the exponential method, but does not reach the level of the neural identification of failed rods.
These results were obtained using a laboratory prototype with the sensor moving through one single face of the rod assembly. Measurements with the sensor scanning through an orthogonal face (of the same assembly) could be used for improving the efficiency of the analysis. These would mean data acquired in a different relative geometry (from the sensor point of view) supplementing the experimental data information about the rods.
In this paper we have described a methodology to identify failed fuel rods in PWR nuclear reactors. This was possible by the combination of the use of a specially designed US transducer and a specific methodology for signal processing. The accuracy achieved points out for its implementation in the real case of testing FAs in the Brazilian PWR Angra I.
|© AIPnD , created by NDT.net|||Home| |Top||