FR2689273A1 - Signal processing apparatus for evaluation of control data. - Google Patents

Abstract

Disclosed is a signal processing apparatus comprising a signal processor 3, 4, 5 having a neural network structure for automatically and rapidly analyzing and evaluating control data obtained by a defect detector or the like. A measured signal from the defect detector 1 is converted into digital parameters by an input arithmetic unit 2 and these parameters are discriminated as defect, deformations and deposits by a first neural network 3; for the signals discriminated as representing a defect, a second neural network 4 performs a quantitative evaluation of the defect (with classification according to: circumferential / axial direction, inner / outer surface, dimension, depth, etc.); for signals which have a small signal-to-noise ratio so that their discrimination is difficult, they are subjected to quantitative evaluation by a third neural network 5 which has been instructed to separate the noise; if discrimination is difficult for this third neural network, a gray signal is generated.

Description

The present invention relates to a signal processing apparatus for

  the evaluation of data obtained by non-destructive testing, including fault and other checks, such as periodic monitoring of the heat transmission pipe of a steam generator in a nuclear reactor facility. In the past, it has been a general practice

  to carry out a periodic inspection of the transmission pipe

  heat generation of a steam generator in a PWR type reactor by a current test method

of Foucault.

  This eddy current test method is designed to detect the presence of a crack, a defect or deposits on an object to be controlled, for example a heat transmission pipe, on

the following principle.

  More specifically, a winding of detection

  tion is put in place on the object to be checked and the

  rolling is excited with an alternating voltage.

  The alternating field generated from the excited winding passes through the object so that eddy currents are induced in the object (metal) since the magnitude of the eddy currents differs depending on the material constituting the object, the presence or absence of a crack or deposition of a foreign metal, the desired non-destructive testing can be performed by measuring the generation of eddy currents. Eddy currents are actually measured by detection. of a variation occurring in the current passing through the winding

  in such a way that, by analysis of the data,

  measured, it is possible to determine the

  in the measured part of the object, for example the presence of a crack, a defect

  or a deposit, or a change of

properties of the material.

  The analysis and evaluation of such measured data was done until now by a professional analyst, ie an oscilloscope was used to produce a Lissajous (impedance plane) impedance curve. so that the resulting curve is observed by the analyst to determine, for example, the presence and dimensions of a crack, a defect or a deposit. However, with this type of evaluation method, is difficult to distinguish between a signal representing a crack or a signal representing a deposit when this signal is mixed with noise and, until now, a process of analysis involving the difference (s) between or among measured values of two or three test frequencies in a multi-frequency eddy current test method

(MFECT).

  According to the MFECT method, when cumulative measurement data is accumulated, in addition to a main frequency determined mainly in relation to the characteristics of a control probe and an object to be controlled, at least one different secondary frequency is used. as a test frequency and the eddy current test operation

  is performed using several frequencies simultaneously

  test cases of this kind. The detailed process of evaluation

  according to this method of measurement is the following

efore.

  1) In the first place, taking into account the different test frequencies, the Lissajous curve obtained from a measurement signal with respect to the main frequency is observed and the signals

  significant amounts exceeding the noise level are sensed.

  2) For other different test frequencies

  of the main frequency, the measurement signal shall be analyzed to determine whether the significant signals referred to in paragraph (1) are attributable to a structural defect, deformation or deposits on the main frequency.

the object to control.

  3) The identification of the depth and the dimension (planar size) of the defect is carried out with respect to the signal determined as representing the

  defect and the identification of the magnitude of the distortion

  tion or deposition is made in relation to the signal

  determined to represent deformation or deposition.

  However, with the techniques of the prior art

  In addition, the evaluation of the defect control data obtained by the measurements is carried out by the individual analyst and furthermore the evaluation is based only on the Lissajous curve, which is indicative. deficiencies, thereby creating the following difficulties: 1 Since the assessment is carried out

  relationship only with the Lissajous curve of the

  In this case, it is difficult to discriminate a noise-impaired defect signal or a deposition signal, and it is also impossible to use various information contained in the measured signal (raw data). 2 The time required for evaluation

  of all measurement results obtained for the different

  Annuity test frequencies is huge.

  3 Evaluation results are functions

  the skill and experience of the analyst.

  4 Evaluation results also depend on

the condition of the analyst.

  A main object of the present invention is to overcome the aforementioned drawbacks which are encountered in the prior art and more particularly an object of the invention is to create an apparatus for

  signal processing that is capable of automatically evaluating

  the measured data and who is also capable of

  improve the efficiency and accuracy of evaluation.

  In order to achieve the aforementioned objectives, it is therefore provided according to the present invention a signal processing apparatus which uses a signal processor having a neural network structure so as to automatically perform the analysis and evaluation of control data. according to a mode

in real time.

  In accordance with one aspect of this

  In the invention, the signal processing apparatus comprises input means for receiving a measured signal that is produced as a result of controlling an object to be measured.

  by a control device, as well as a means of evaluation

  to analyze the measured signal received by the input means to evaluate the condition of occurrence of a defect and / or a deposit on the object to be measured, and the evaluation means comprises a processor of signals

  having a neural network structure and having a plurality of

  each processing unit fulfilling each of the individually predetermined input / output response functions as well as a plurality of connection means for connecting the processing units to each other in accordance with individual transmission characteristics of predetermined weighting factors of such that the processing units form a plurality of nodes of said neural network The signal processor has an instruction function for correcting said weighting factors in such a way that each of the processing units connected to an associated connection means produces an output optimal in

  response to an input data applied to it.

  According to a preferred embodiment of the present invention, the signal processor having the neural network structure comprises a layered structure comprising at least an input layer and an output layer and accordingly neural networks of the type such that the processing units located in the input layer and the processing units located in the

  output are linked by the connection means.

  According to another preferred embodiment of the present invention, the input means comprises arithmetic means for extracting a factor in relation to the defect and / or the deposit on the object from the measured signal received from the apparatus of control and for converting said factor into corresponding numerical parameters forming input data

  applied to the signal processor.

  The signal processing apparatus of the

  present invention is well suited for the analysis and evaluation of

  measurement of measured data from the aforementioned control apparatus including various non-destructive testers such as an eddy current tester, particularly a multi-frequency eddy current tester. It should be noted that, unless otherwise stated, the term "defect" should be interpreted as referring to all abnormal conditions such as

  cracks and cracks as well as deformations

  changes in the overall shape, corrosion, etc. In accordance with the present invention, the fact that the signal processor having the neural network structure is used to perform the evaluation

  measured data provides the following benefits:

  solvents. 1 The measured signal from the control device is converted parametrically and is used as one of the input signals applied to the processor

  of signals, with the result that the different

  previously unused in the measured signal can be exploited and therefore the analysis and evaluation can be carried out

  desired with greater precision than before.

  2 Since the signal processor having

  the neural network structure serves an instruc-

  it is quite possible that it has a higher analytical capacity than an analyst

  corresponding professional degree to his degree of

  tion. 3 By automating the analysis and evaluation of the measured data, it is possible

  to prevent the production of a human error.

  4 By automating the analysis and evaluation of the measured data, the evaluation time can be considerably reduced by comparison with

  the operation performed by the analyst.

  The analysis and evaluation of the data measured by the signal processor having the structure

  neural network can be performed by a computer

  and not by the use of a large computer

  special design so that, for example, a personal computer can be brought on

  place of cumulation of data in order to perform the analysis

  lysis and evaluation of measured data according to a

in real time.

  Other features and advantages of the invention will be highlighted in the following

  the description given by way of non-limiting example,

  with reference to the accompanying drawings in which

  Figure i shows the structure of a processing device

  signal signal corresponding to an embodiment of the present invention; Fig. 2A is a block diagram showing the structure of the direct-type neural network, Fig. 2B is a schematic representation showing an example of signal processing within the node; and

  Figure 3 is a block diagram showing the structure of

  neural network of interconnection type.

  An embodiment of the present invention will now be described with reference to FIG. 1. In the Figure, a signal processing apparatus corresponding to the present embodiment is designated

  as a whole by the reference number 7

  7 comprises an arithmetic unit 2 serving as an input means for receiving a measured signal from an eddy current fault detector 1 and a signal processor 6 having a neural network structure. In this embodiment, the processor 6 has three stages of neural networks

3, 4 and 5.

  The arithmetic unit 2 performs operations

  arithmetic meanings to convert a measured signal

  entry into parameters in the form of numerical values

  Since the parameters are obtained by these arithmetic operations, they can be selected as at least one combination of different parameters, such as domain-time parameters, domain-frequency parameters and parameters which are Contrast with Lissajous curves Arithmetic unit 2 selects some of the parameters that are appropriate for the purpose of the control, preferably in the form of a combination, and the measured signal is converted into numerical values corresponding to the selected parameters. The data signal which is produced by the arithmetic unit 2 is applied as an input signal to the signal processor 6 having the neural network structure. Specific methods of converting

  three types of numerical value parameters in relation to

  with the arithmetic operations of the arithmetic unit

  2 are, for example, the following: 1 Domain-time parameters With respect to its components of variation over time, the measured data is converted into numerical values in the form of parameters in agreement with the results, such as the time difference or the time difference.

  correlation coefficient of the measured signal.

  2 Frequency domain parameters In relation to the nature of its frequency components, the measured data is converted into numerical values as parameters in accordance with

  the results of a Fourier analysis or the like.

  3 Parameters in contrast to a Lissajous curve: The area of a closed surface defined by a

  Lissajous curve, the phase angles of the waves or ana-

  logues due to measured data are converted

  in numerical values as parameters.

  Further, in the present embodiment, the signal processor 6 having a neural network structure comprises first, second and third neural networks 3, 4 and 5 This signal processor 6 can be realized by means of software using a existing personal computer and as a result it is possible to transport this personal computer to the place of data collection to perform an analysis.

real-time data.

  The neural networks 3, 4 and 5 are identical in basic construction and this basic construction

will now be described.

  The neural network mentioned above is a model of a signal processing system simulating

  the neurons of a human being and, in this model,

  For example, a neural network of the direct type having the simplest layer structure is arranged as shown by way of example in FIG. 2A. Each of the signal processing units corresponding to the neurons is interconnected with given weighting factors. different treatment units

  u 1, u 2, u 9 corresponds to one of the nodes in the network.

  As a result, in the description that follows, the

  terms "processing unit" and "node" are considered

  as being synonymous with each other.

  In FIG. 2A, the neural network 3 (or 4 or 5) has for example a three-layer structure,

  comprising an input layer 9, an intermediate layer

  10 and an output layer 11, and the layers are composed of several nodes 8 (ul, u 2, ui 9) In these layers, the nodes 8 are interconnected by connecting means having transfer characteristics corresponding to factors predetermined weighting W (which will be described later) With this type of neural network, the number of nodes 8 in the input layer 9 is determined in correspondence with the number of input signal classes converted parametrically and also the number of nodes 8 in the output layer It is determined in correspondence with the number of classes of output signals (the number of data to be output) On the other hand, the number of layers and the number of nodes 8 are both variable in who

  relates to the intermediate layer 10.

  The signal processing in each node 8 is schematically highlighted in FIG. 2B. More specifically, as indicated by the following equation (1), the input signals (II I 2 In) from the previous layer (FIG. layer on the input side) are respectively multiplied by the weighting factors (Wi, W 2, Wi, Wn) corresponding to the previous layer and the total sum

  Z Wi x Ii of these weighted input values is determined

  Then a threshold value H is subtracted from the total sum so as to determine the value of XX = ú wi xi Hi (1) The value of X is substituted for example in a nonlinear sigmoid function f (X), as is indicated in Equation (2), so as to obtain

an output 0.

  J o = f (X) = {l + tg-1 (X / u)} / 2 (2) J It should be noted that the response function of nodes 8 is not always limited to the sigmoid function mentioned above. above and any other function

can be used.

  When the signal processor 6 having such a neural network structure is used, its operation in a mode of instruction is first

  executed In this description, what is meant

  by the word "instruction" is an adjustment of the weighting factors Wij between the nodes and the threshold values H so that the role of the information processing system J is filled. A specific method of instruction is such that a a large number of matched data corresponding to known inputs (questions) and outputs (responses) are prepared and that the weighting factors W 1 and the values

  H threshold between the nodes be adjusted and deter-

  in such a way that the relations between the inputs and the outputs of the processor coincide with the previously known known relationships with respect to all the input data pairs and

Release.

  Once the instruction has been completed, the neural networks 3 to 5 of the signal processor 6 operate by "intelligence" so that an output (response) is produced as a result of a calculation in

  each node in response to one or more unknown inputs

  (in this way), in each of the

  neural networks 3 to 5, the nodes 8 perform a

  the parallel process and the process of

  exits is done by simple arith-

  This increases the speed of treatment.

  Also the neural networks that have completed the instruc-

  may be subject to another instruction during

  their use (this has the effect of adjusting the

  weights and threshold values to more optimal values) and evaluation capacity can be progressively improved in response to results.

  repeated real performances.

  Although the direct-type neural networks involved in the layered structure have been described sufficiently, an interconnection model, which is closer to a human nervous system or a

  signal processing with inter-type neural networks

  connection (Figure 3) was designed and this system can be used to evaluate a measured signal While, in the direct-type neural network of Figure 2, the interconnection is established only between the respective layers and the nodes located in the same layer are not interconnected, on the contrary in the network

  neural-type interconnection, the nodes are inter-

  connected three-dimensionally as shown on the

Figure 3

  With reference again to FIG. 1, the data signals converted parametrically by the arithmetic unit 2 are first applied to the first neural network 3. In the first neural network 3, the signals

  data are subject to the discrimination of a defect

  tuosity (here defects, cracks, etc., of the object are called defects, to distinguish them

  deformations), deformations (including the

  sizing) and deposits (including the classification of copper and magnetite (Fe 304)) and the results are out of the outputs of the first

  neural network 3, the signals discriminated as representing

  both a defect are then applied to the second

  neural network 4 o the defect is evaluated quantitatively

  The quantitative evaluation of the defect

  involves the classification of circumferential directions

  erential or axial and interior or exterior surfaces

  as well as the identification of the dimensions and depths of the defect and the results are out. Also the data signals discriminated as being weak with respect to the S / N ratio (signal-to-noise ratio) and difficult to discriminate by the second neural network 4 are applied to the third neural network 5 whose function is to separate the noise

  so that the defect is evaluated quantitatively

  Also, if these data signals are determined to be difficult to discriminate in the third neural network 5, a gray signal is produced by this third neural network. When the gray signal is generated, the evaluation is performed.

  corresponding to a comparison of discrimination

  by the expert analyst and discrimination by

the signal processor 6.

  The inventors have made a signal processing apparatus having the above-mentioned structure and incorporated into a personal computer (its commercial designation is PC-9801, NEC Corporation), so as to perform the evaluation of control data.

  the defect and the optimal parameters were then

  are determined respectively for the first, second and third neural networks 3, 4 and 5 Then, using as instruction data a total of twenty four pairs of data (sets of inputs and outputs) relating to twelve defects, four deformations, four copper deposits and four deposits

  of magnetite, the system operated in a mode of instruc-

  the weighting factors W and the threshold values H between the respective nodes have

have been adjusted optimally.

  Then, a defect control was actually performed on a test sample (the steam generator heat transmission pipe) and the resulting measured signals from the defect detector 1 were evaluated by the test equipment. signals This resulted in 100 percent correct answers for a total of ten defects, deformations, and deposits as indicated in the following. Also, the time required for evaluation was less than 1 second.

  for each of the input data.

  1 A defect to a depth of 80% of the thickness in the circumferential direction

  of the inner surface of the sample.

  2 A defect to a depth

  80% of the thickness in the circumferential direction-

  the of the outer surface of sample.

  3 A defect to a depth of 80% of the thickness in the axial direction of

  the inner surface of sample.

  4 A defect to a depth of 80% of the thickness in the axial direction of the outer sample surface. A deformation A in the outer surface

sample.

  6 Deformation B in the outer surface

  sample re (which was different in size

deformation A).

  7 A copper deposit A on the outer surface

sample.

  8 A copper B deposit on the outer surface

  of the sample (which was different in size

deposit A).

  9 A magnetite A deposit on the surface

outside of the sample.

  A magnetite B deposit on the outer sample surface (which was different in

dimension of the deposit A).

  It should be noted that while in the

  As mentioned above, the signal processor 6 comprises the three neural networks 3, 4 and 5; it is needless to say that a single neural network can be

  used to perform the evaluation of the measured data.

  In addition, the signal processing apparatus of the

  present invention is applicable to the analysis and evaluation of

  Measurement of data measured by another non-destructive testing process than defect control

by eddy currents.

Claims (3)

  A signal processing apparatus for evaluating control data, comprising input means (2) for receiving, from a control apparatus (1), a measured signal produced as a result   control of an object to be controlled by means of that   control circuit (1), and evaluation means (6) for analyzing said measured signal received by said input means to evaluate a state of occurrence of a defect and / or a deposit on said object, characterized in that said evaluation means (6) comprises signal processors (3, 4, 5) having a neural network structure and comprising a plurality of processing units (8, ui u 2 u 9), each of which answer functions   predetermined input / output   sieurs means of connection (WI, Wi, W n) for connecting said processing units with each other with   individual transfer characteristics corresponding to   predetermined weighting factors (Wij) so that several nodes of the neural network are formed by said processing units, and in that said signal processors (3, 4, 5) have a function   instructions for correcting these weighting factors.   (W ij) in such a way that each of said processing units produces at its output an optimum signal in response to input data applied thereto. 2 A signal processing device according to   claim 1, characterized in that said   signal processors (3, 4, 5) have a layered structure comprising at least one inner layer (9) and one   outer layer (10) and that the treatment units   (8, u 1, u 2, u 9) located in said input layer (9) and the processing units (8, uai u 2, u 9) located in said outer layer (10) are respectively linked by said connecting means for   form a neural network of type-direct (Figure 2).   A signal processing apparatus according to claim 1, characterized in that said input means comprises arithmetic means (2) for extracting   said signal measured a factor relating to a defect   and / or a deposit on said object and for the purpose of   vertically into numerical values as data from   to said signal processors (3, 4, 5).