JPH03105229A - Abnormality detector for structural body - Google Patents

Abstract

PURPOSE:To enable detection of abnormality in a structural object even when an object to be inspected is changed by providing a judging circuit for judging a maximum of a pattern output of a calculation circuit with a learning means to adjust the weight of calculation so that abnormality in the structural object is obtained accurately. CONSTITUTION:A structural object to be inspected is struck with a hammer to detect vibration thereof with a vibration detection circuit 1 and an electric output signal corresponding to a vibration waveform is inputted into a frequency analysis circuit 2, which measures an intensity of (n) frequency components of an input signal and values of the components are inputted into an input calculation circuit 7 of a pattern sorting circuit 3 in parallel. The circuit 7 converts the component values inputted to non-linearity to be sent to an intermediate calculation circuit 8, which 8 multiplies the input signal by a specified value of weight for conversion to non-linearity to be sent to an output calculation circuit 9. The circuit 9 multiplies the results by a specified weight for conversion to non-linearity to be sent to a maximum judging circuit 11, with which 11 the maximum output node is specified among those of the circuit 9 to be fed to a display circuit 4 and an alarm circuit 5. When the results of judgment of the circuit 11 are different from an actual condition, a correct output value is inputted from the circuit 6 to perform a learning with a weight.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an abnormality detection device for a structure having a learning function.

[Prior Art] Generally, abnormalities in structures are detected by having an expert listen to the sound made when a steel structure is hit with a hammer, and making judgments based on the experience of the expert.

If you try to automate this, you can consider the following: In other words, abnormalities are detected by comparing the frequency analysis data of the sound when a structure is hit with a hammer with the frequency analysis data of various abnormalities that have been recorded in advance. Another possible method is to perform a detailed analysis of the frequency distribution during anomalies in advance and compare that data with observed data to detect anomalies. [Problem to be solved by the invention] However, in the former case, a large amount of frequency distribution data must be accumulated. If you change the inspection target,
The disadvantage is that the frequency distribution data must be stored again. In addition, the latter case has the disadvantage that detailed analysis of the frequency distribution is difficult. An object of the invention is to provide an abnormality detection device for a structure that can detect abnormalities in a structure even when the object to be inspected changes. [Means for Solving the Problems] The structure abnormality detection device of the present invention includes a circuit that detects vibrations of a structure, and a circuit that analyzes the frequency components of the detected vibrations and individually calculates n frequency division values. The pattern classification circuit includes an analysis circuit that outputs an output, a pattern classification circuit that receives each output of the analysis circuit and classifies it into one pattern output, and a circuit that inputs a comparison target value to the pattern classification circuit as necessary, The classification circuit includes a calculation circuit that weights and non-linearly transforms each output of the analysis circuit into one pattern output, a determination circuit that determines the pattern output that takes the maximum value among those pattern outputs, and a The present invention has a learning means for adjusting the weight of the calculation circuit so that the determination circuit obtains a correct result regarding the abnormality of the structure based on the comparative target value of . The calculation circuit of the pattern classification circuit described above consists of n input calculation circuits that perform nonlinear exchange, n intermediate calculation circuits that perform nonlinear exchange after weighting and summing the outputs of the input calculation circuits, and It is preferable to consist of one output calculation circuit that weights and sums the outputs of the calculation circuits, and then performs nonlinear exchange. This output calculation circuit can be provided as many as the number of abnormalities. [Operation] Information on the n frequency component values of the vibration analyzed by the frequency analysis circuit is input to the calculation circuit of the pattern classification circuit, and is output after being weighted and nonlinearly transformed. Here, nonlinear transformation means, if the input is X, for example, t (x) = 1/(1+l3Xl)(-X))
= Calculating the function of (1). The operation of the calculation circuit described above is explained as follows when the circuit is composed of three parts: an input calculation circuit, an intermediate calculation circuit, and an output calculation circuit, and the weighting is done in two parts: the intermediate calculation circuit and the output calculation circuit. become that way. In other words, the output xi of the input calculation circuit is xi=f(x) (2).Next, in the intermediate calculation circuit, the outputs of the input calculation circuit are weighted and summed, and then non-linearly transformed. Output.
That is, if the output of the intermediate calculation circuit is yj, then yj=f(gUji xi) (j=1.2,"
-11)...(3). However, Uji is the weighting parameter of the intermediate calculation circuit. Then, in the output calculation circuit, the outputs of these intermediate calculation circuits are weighted and summed, and then nonlinearly transformed and output. That is, if the output of the output calculation circuit is Zk, then Zk=f(ΣWkj yj) (k=1.2,...
・A) ♂ ...(4) becomes. However, Wkj is the weighting parameter of the output calculation circuit. The determination circuit determines and specifies the pattern output with the maximum output value among the outputs of this one output calculation circuit. The outputs of the J output calculation circuits directly correspond to the type of abnormality such as a failure or defect to be inspected, and the abnormality is detected. The input calculation circuit, intermediate calculation circuit, and output calculation circuit are a three-layer network (a multilayer network), taking into account the ellipse of paulownia wood, etc.
The above weighting parameters are set. Therefore, an abnormality is detected directly from the judgment result of the above-mentioned judgment circuit. If the structure to be inspected is changed, the comparison reference value is input from the input circuit to the input calculation circuit. A certain judgment result is obtained through the input meter X circuit, intermediate calculation circuit, and output calculation circuit that make up the multilayer network, and the intermediate calculation circuit, The weighting parameters of the output calculation circuit are automatically corrected, resulting in appropriate correction values. Therefore, it becomes possible to immediately detect abnormalities even in other elliptical objects, greatly improving the applicability of the device. [Examples] The present invention will be described below based on illustrated examples.

As shown in FIG. 1, this detection device includes a vibration detection circuit 1,
Frequency analysis circuit 2, pattern classification circuit 3, display times N 4
, a g-information circuit 5, and a target value input circuit 6. Each of the circuits 1 to 6 constituting the abnormality detection device is formed using ordinary electronic components such as a linear IC or a microcomputer.

The vibration detection circuit 1 is a circuit that detects the vibration that occurs when a 'Jff4 structure to be inspected is hit with a hammer, converts the detected vibration into an electrical signal, and outputs the signal. The frequency analysis circuit 2 is a circuit that receives the output of the vibration detection circuit #I1 and measures the intensity of several specific frequency components (n pieces) of the signal, and is equipped with an output terminal corresponding to each frequency component. ing. As shown in FIG. 2, the pattern classification circuit 3 includes n input calculation circuits FI @ 7, m intermediate calculation circuits 8, 1 output calculation circuit 1i 189, a coupling unit 10 connecting these circuits, and an output calculation circuit 9. Maximum value judgment time v1I connected to
It is composed of ll. This pattern classification circuit 3 can be constructed by an electronic circuit or a computer.

A target value input circuit 6 is provided on the input side of the input calculation circuit 7, so that the output of the target value input circuit 6 can be selected instead of the output of the frequency analysis circuit 2.

The input calculation circuit 7 has a function of nonlinearly converting each of the n input frequency partial values and outputting the result to all of the m intermediate calculation circuits 8. Each of the m intermediate calculation circuits 88 receives the individual outputs of the first to nth input calculation circuits 7, weights the outputs, sums them up, and non-linearly transforms them into one Output to all output calculation circuits 9. ．． l! Each of the output calculation circuits 9 is
The individual outputs of the first to first intermediate calculation circuits 8 are received, each of these outputs is further weighted, summed, nonlinearly transformed, and output to the maximum value determination circuit 11. Each output of these j output calculation circuits 9 is 1
Types of abnormal patterns observed at different frequencies,
In other words, it's a failure. It corresponds to the type of abnormality such as a defect or the candidate cause of the abnormality. The i large value determination circuit 11 identifies the node with the maximum output among the outputs of one output calculation circuit 9, and outputs it to the display time N4 and the alarm time N5. Now, when the tub structure to be inspected is struck with a hammer, the vibration is picked up by the vibration detection circuit 1, and an electrical output signal corresponding to the vibration waveform is applied to the frequency analysis circuit 2. The frequency analysis circuit 2 measures the intensity of each of the n frequency components of the input electrical signal, and inputs the respective frequency fraction values in parallel to the input calculation circuit 7 of the pattern classification circuit 3.

The input calculation circuit 7 nonlinearly transforms each input frequency component value and sends it to the intermediate calculation circuit 8. The intermediate calculation circuit 8 multiplies the input signal by a certain weight value, performs nonlinear transformation, and sends the signal to the output calculation circuit 9. The output calculation circuit 9 further multiplies this input signal by a certain weight value, performs nonlinear transformation, and sends the signal to the maximum value determination circuit 11 .

The maximum value determination circuit 11 identifies the node with the maximum output among each node (type of abnormality) of the output calculation circuit 9 and sends its output signal to the display circuit F#I4 and the alarm circuit 5.

The display circuit 4 displays the output value of the node determined to have the maximum output in the maximum value determination step #111 among the nodes of the one output calculation circuit 9. The alarm circuit 5 also generates an alarm when the node with the maximum output value is determined. Next, the node determined to have the maximum output by the maximum value determination circuit 11 is the known! If it is incorrect in light of the actual state such as an abnormality of the R structure, the correct output value is input from the target value input circuit 6 to perform learning to change the weighting of the calculation of the connection part in the pattern classification circuit 3. conduct. Let us explain the calculation contents of the calculation circuits 7, 8.9 in more detail. First, each of the n input calculation circuits 7 nonlinearly transforms the n frequency component values measured by the corresponding frequency analysis circuit 2, and outputs it to all of the m intermediate calculation circuits 8. If the frequency component value is F r (r=1.2,...n), the output of the input calculation circuit is x i = f (Fi)
-(2)' However, f (X) = 1/(1+l3XE
l(-X)). Next, each of the m intermediate calculation circuits 8 performs the following calculation and outputs the output yj (j=1.2, . . . 1) to all of the A output calculation circuits 9. yj=f(ΣUji xi) =13) However, Uji: weighting parameter in the intermediate calculation circuit. Each of the output calculation circuits 9 performs the following calculation and outputs an output Zk (k=1.2, . . . A).

Zk = f (each Wkj yj) ・(4
) However, Wkj is a weighting parameter in the output calculation circuit. The maximum value determination circuit 11 selects the maximum value among the outputs Zk of these one output calculation circuit, and this Zk is an abnormality pattern of the structure as seen from the frequency distribution, that is, a feature value indicating the type of abnormality of the structure. This corresponds directly to the above, and an anomaly will be detected.

The weighting parameters Uji and Wkj of each node in the calculations of the intermediate calculation circuit 8 and the output calculation circuit 9 give various frequency component information that is calculated as an expected value in the event of an abnormality based on the structure of the structure to be inspected. The settings are stored in advance so that the abnormality can be detected correctly when the abnormality occurs. By the way, if the structure to be inspected changes,
The above-mentioned preset function alone may not be enough to deal with the problem, resulting in incorrect abnormality detection. In such a case, the weighting parameters stored in each node of the intermediate calculation circuit 8 and the output calculation circuit 9 of the multilayer network can be inputted from the target value input circuit 6 by inputting the comparison target value regarding the modified Sugi relics from the target value input circuit 6. Modified so that anomalies can be detected even in new structures. This modification is automatically adjusted by a learning function based on the result obtained from the maximum value determination circuit 11 of the pattern classification circuit 3. To be more specific, if the structure to be inspected is changed,
The pattern classification circuit 3 is switched to select the output side of the target value input circuit 6. First, the target value input circuit 6 calculates an expected value (comparison target value) at the time of abnormality in the structure after the change, and the maximum value determination circuit 11 of the pattern classification circuit 3
is given to. Based on this comparison target value, the maximum value determination circuit 11 determines a target value for determining which node of the output calculation circuit 9 should be the network that takes the maximum value, that is, each node of the intermediate calculation circuit 8 and the output calculation circuit 9. Set the target value for the weighting parameter value of . On the other hand, the specific calculation content or function of each node of the intermediate calculation circuit 8 and the output calculation circuit 9 is determined by the value of the weighting parameter stored at that time. Therefore, temporarily, one of the output calculation circuits 9 takes the maximum value, and that node is recognized by the maximum value determination circuit 11. Therefore, the maximum value determination circuit 11 issues a correction instruction for the weighting parameter value to the multilayer network as long as the node of the output calculation circuit 9 that actually takes the maximum value is different from the target value set from the comparison reference value. Output and output calculation circuit 9. It continues to be given to each node in the order of intermediate calculation circuit 8. Each node of the intermediate calculation circuit 8 and the output calculation circuit 9 adjusts the internal weighting parameters in a certain direction based on this correction command, and the resulting judgment is again obtained by the maximum value judgment circuit 1l, and the Based on the modification command, the weighting parameters inside each node are further adjusted. Through such repetition or learning function, the maximum value determination circuit 11 finally obtains the same determination result as the target value, and the adjustment of the weighting parameters inside each node is completed. Since the adjustment here is based on the learning function, it does not matter whether the adjustment direction starts in the correct direction or is performed by trial and error; in either case, the correct correction will be made as a result. With such a learning function, the weighting parameters of each node of the intermediate calculation circuit 8 and the output calculation circuit 9 are automatically corrected to the expected values at all times. Ability to quickly respond to completely unexpected situations. This greatly improves the performance and range of application of the device. The detection device of the above embodiment uses a vibration detection circuit 1 to detect vibrations when a building is hit with a hammer, then frequency-analyzes the signal, and then uses a pattern classification circuit 3 to determine whether there is an abnormality in the structure. Determine whether or not. The pattern classification circuit has a learning function, so if it makes a mistake in judgment, it can be trained to make the correct judgment. It also has a learning function, so it can respond to various situations. In the above embodiment, the pattern classification circuit #I3 consists of only one set, but the range of application can be further expanded by providing a plurality of pattern classification circuits in parallel as shown in FIG. [Effects of the Invention] As described above, the structure abnormality detection device of the present invention has the following effects:
Since it has a learning function, it can automatically detect abnormalities and can also deal with changes in the structure being inspected. This greatly improves the performance and adaptability of the device.

[Brief explanation of drawings]

FIG. 1 shows an elliptical diagram of the present invention, FIG. 2 shows a configuration diagram of a pattern classification circuit, and FIG. 3 shows a configuration diagram of a pattern classification circuit in another embodiment of the invention. In the figure, 1 is a vibration detection circuit, 2 is a frequency analysis circuit, 3 is a pattern classification circuit, 4 is a display circuit, 5 is an alarm circuit, 6 is a target value input circuit, 7 is an input meter X circuit, and 8 is an intermediate calculation circuit. ,
Reference numeral 9 indicates an output calculation circuit, and 10 indicates a coupling section.

Claims (1)

[Claims] 1. A circuit that detects vibrations of a structure, an analysis circuit that analyzes frequency components of the detected vibrations and outputs n frequency component values individually, and each output of the analysis circuit. The pattern classification circuit includes a pattern classification circuit for classifying the received output into l pattern outputs, and a circuit for inputting a comparison target value to the pattern classification circuit as necessary, and the pattern classification circuit weights and classifies each output of the analysis circuit. A calculation circuit that performs nonlinear transformation to produce l pattern outputs, a judgment circuit that judges the pattern output that takes the maximum value among those pattern outputs, and a comparative target value from the input circuit. An abnormality detection device for a structure, comprising: learning means for adjusting the weights of the calculation circuit so as to obtain correct results regarding the abnormality. 2. n input calculation circuits in which the calculation circuit of the pattern classification circuit performs nonlinear exchange, and n intermediate calculation circuits that perform nonlinear exchange after weighting and summing the outputs of the input calculation circuits; 2. The abnormality detection device for a structure according to claim 1, further comprising l output calculation circuits that weight and sum the outputs of the intermediate calculation circuits and then perform nonlinear exchange. 3. The abnormality detection device for a structure according to claim 2, wherein the number of output calculation circuits is equal to the number of types of abnormalities.