KR101967633B1 - Apparatus For Making A Predictive Diagnosis Of Nuclear Power Plant By Machine Learning - Google Patents

Abstract

The present invention relates to a machine learning type prediction and diagnosis apparatus for a nuclear power plant, which comprises: a vibration signal receiving module (3) to store a vibration signal received from a fan; a signal preprocessing module (5) to preprocess the vibration signal; a characteristic extraction module (7) to calculate statistical and morphological characteristic values from the vibration signal; a characteristic selection module (9) to select a plurality of characteristic values suitable for the fan among the characteristic values extracted in the characteristic extraction module (7); and a diagnosis module (11) to learn the state of a fan based on the characteristic value selected in the characteristic selection module (9) and to diagnose the fan based on a learned result. The signal preprocessing module (5) comprises a standardization unit (13) to standardize the vibration signal received by the vibration signal receiving module (3) and a defect signal extraction unit (15) discriminating the vibration signal of the fan, which is standardized in the standardization unit (13), by time and frequency division domains after being divided into arbitrary sections, and outputting the discriminated vibration signal to the characteristic extraction module (7). Accordingly, defect signal data is processed to predict and diagnose a probable defect in a fan, thereby largely improving reliability and performance in an automatic prediction and diagnosis result for the fan. Moreover, machine learning is introduced, thereby largely increasing utility of the prediction and diagnosis apparatus.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a nuclear power plant,

The present invention relates to a machine learning type predictive diagnosis apparatus for a nuclear power plant, and more particularly, to a machine learning type predictive diagnostic apparatus for a nuclear power generation facility, And a machine learning type predictive diagnostic apparatus of a nuclear power generation facility capable of diagnosing a vibration state of a fan part by a learned machine after learning a machine according to a selected characteristic value.

Diagnostics technology that precisely grasps the operating status of a nuclear power plant is a series of activities to maintain the intended function and standard performance of the facility, It is a key technology in management technology (terotechnology). Equipment health monitoring and prediction The defect diagnosis technology is a knowledge-based service technology used for maintenance and repair of equipment. It is used to monitor the state of critical equipment such as fans, and each decision stage A new type of knowledge-based process monitoring technology that fuses data to deduce reliable results in the detection and detection of defects is actively being applied and developed in industrial facilities.

However, in the case of the nuclear power plant, the above-mentioned defect diagnosis technology stage is at the level of the preventive maintenance (ie, periodically checking the facility), and therefore, a reliable result for predicting and diagnosing defects is derived There was a problem that can not be done.

In addition, the individual components of the nuclear power plant are based on preventive maintenance (PM), which is based on regular maintenance plans, based on operating hours and so on. Equipment condition monitoring sets the alarm and emergency stop value according to the vibration evaluation standard of RMS (Root Mean Square) or PP (Peak to Peak) value of vibration and sets vibration value for safety facility operation until the next preventive maintenance period (trend) management.

However, in the above-described method, excessive maintenance may occur due to unconditional maintenance when a certain period of time occurs, and maintenance of a facility that has been stably operated by unnecessary maintenance may cause problems that have not occurred before. In addition, since the conventional technology only monitors the vibration trend for stable operation of the facility, it is difficult to predict when the facility (fault) occurs and there is a problem that a detailed analysis of the expert is required when the fault occurs.

Also, in selecting (selecting) the characteristic values that can represent the vibration state by collecting the vibration signals from the respective parts of the nuclear power generation facility, the state of the parts including the normal state is divided into different states according to the respective defects , Which means that the characteristics of vibration selected by state are different. As described above, since the features to be selected are different according to the states, there arises a problem that it is required to be divided and analyzed according to the type of the expected defect. Therefore, the multiple management for the facility may cause confusion of the user of the diagnosis apparatus, Complicating the analysis procedure, and consequently increasing the diagnostic time.

KR 10-1060139 KR 10-1607047

The present invention has been proposed in order to solve the problems inherent in the conventional power plant predictive diagnosis apparatus. The present invention acquires a vibration signal from each component of a nuclear power plant such as a fan, The object of the present invention is to extract statistical characteristic values of a signal, to learn these characteristic values on a machine, and then to predict and diagnose the vibration state of the fan by a machine.

Another objective is to replace the predictive diagnostics performed by manual maintenance with automatic predictive diagnostics for each part of the power generation facility, such as fans, to further enhance the equipment integrity and to better operate the nuclear power plant.

Another object of the present invention is to enable efficient automatic predictive diagnosis of a nuclear power plant by extracting a defective component of a vibration signal caused by a defect that may occur in the component as described above from a vibration signal.

In order to achieve the above object, the present invention provides a vibration signal receiving module for receiving and storing a vibration signal from a fan; A signal preprocessing module for preprocessing the vibration signal received by the vibration signal receiving module; A characteristic extraction module for calculating statistical and morphological characteristic values from the vibration signal pre-processed in the signal preprocessing module; A characteristic selection module for selecting a plurality of characteristic values suitable for the pan among the characteristic values extracted from the characteristic extraction module; And a diagnostic module that learns the state of the fan with a property value selected by the property selection module and diagnoses the fan through the learning result, wherein the signal preprocessing module is configured to receive An averaging unit for averaging the vibration signals; And a defect signal extracting unit for dividing the vibration signal of the fan that has been averaged by the averaging unit into an arbitrary section and then dividing the divided signal into a time division area and a frequency division area and then outputting the divided signal to the characteristic extraction module Provides a machine learning predictive diagnostic system for nuclear power plant.

In addition, the characteristic selection module may include a classifier for classifying the characteristic values extracted from the characteristic extraction module according to classes; A normalization unit that matches the data ranges of the characteristic values stored in the classifying unit with respect to each of the classes and makes the distributions similar; And a selection unit for selecting a suitable characteristic value by performing a genetic algorithm with the characteristic value normalized by the normalization unit as a candidate.

The selection unit may include: an initial solution setting sector for generating a plurality of chromosomes by randomly assigning binary random numbers to initial solutions of respective characteristic values; A fitness evaluation sector for evaluating a fitness of an initial solution set in the initial solution setting sector or a progeny solution of chromosomes after the second generation; And a descendant that transmits a certain percentage of the excellent chromosome evaluated as excellent in the fitness evaluation sector to the next generation as it is and a descendant that generates a descendant solution of the next generation through a seed solution generated by crossing the remaining chromosomes except for the superior chromosome And an initial solution to which a value of 1 is assigned is selected to indicate that the initial solution to which 0 is substituted is not selected, and the initial solution to which a value of 1 is assigned is selected.

The fitness evaluation sector may further include a selection characteristic value having a value of 1 in an initial solution or a descendent solution according to the class, wherein the fitness of the selection characteristic value is obtained through the density of the selection characteristic value in the class and the distance between the class It is preferable that the measured fitness is compared with each selected characteristic value and the fitness of the selected characteristic value is evaluated according to the comparison result.

According to the machine learning type predictive diagnosis apparatus of the nuclear power plant of the present invention, defect signal data is processed so as to predict and diagnose defects that may occur in a nuclear plant, in particular, a fan. Therefore, reliability and performance It is possible to greatly improve the utility of the predictive diagnostic device by introducing machine learning.

In addition, it is possible to prevent the occurrence of minute defects of the bearing due to normal wear, which is difficult to detect early among the defects that may occur in the nuclear power generation facilities, especially, the fan, early damage due to excessive fatigue damage of the metal parts due to axial thrust, It is possible to dramatically improve the predictive diagnostic performance of the fan.

1 is a block diagram of a machine learning type predictive diagnosis apparatus of a nuclear power plant according to an embodiment of the present invention;
FIG. 2 is a diagram showing a process of generating a plurality of subbands by dividing a vibration signal in the vibration signal dividing portion a plurality of times. FIG.
3 is a block diagram showing a sub-configuration of the characteristic selection module of FIG.
FIG. 4 is an objective function graph plotting three characteristic values as x, y, and z axes, respectively, in order to evaluate the fitness of the characteristic value solution.
5 is a flowchart showing a process of predictive diagnosis of a fan by the predictive diagnostic apparatus of FIG.
FIG. 6 is a flowchart illustrating a defect signal extracting process of FIG. 5 in detail.
7 is a flowchart illustrating the defect signal extraction method of FIG. 6 in more detail.
FIG. 8 is a flowchart illustrating details of a characteristic value selection process of FIG. 5;

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, a machine learning type predictive diagnosis apparatus of a nuclear power plant according to an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

The predictive diagnosis apparatus of the present invention includes a vibration signal receiving module 3, a signal preprocessing module 5, a characteristic extracting module 7, a characteristic selecting module 9, and a diagnosis And a module (11).

First, as shown in FIG. 1, the vibration signal receiving module 3 receives a vibration signal generated from a fan, which is a target of a predictive diagnosis of a nuclear power generation facility. The vibration signal receiving module 3 includes components such as fan blades, That is, the vibration signal sent from the vibration sensor (displacement and acceleration sensor) provided in the vibration source is received for two seconds, for example, for a unit time, and classified by the type of the corresponding component and stored in the DB.

Although not shown in FIG. 1, the vibration signal data stored in the DB undergoes a pass filtering process such as a low frequency, a medium frequency, and a high frequency through a noise eliminator before proceeding to a preprocess, which will be described later. This process removes the noise component caused by the quantization error in the process of digitizing the analog signal collected through the vibration sensor such as displacement and acceleration sensor in the nuclear power generation facility. In particular, when the component is rotating equipment, Thereby preventing the convergence phenomenon. That is, it plays a role of eliminating and reducing noise that is inevitably generated in the digitization process, which is an essential element, in advance.

The signal preprocessing module 5 preprocesses the vibration signal received by the vibration signal receiving module 3 and includes a noise eliminator 12, an averaging unit 13, And an extracting unit (15).

Here, as shown in FIG. 1, the noise removing unit 12 is subjected to a pass filtering process such as a low frequency, a medium frequency, and a high frequency according to a frequency band in which noise may be included in the vibration signal data stored in the DB. This process removes the noise component caused by the quantization error in the process of digitizing the analog signal collected through the vibration sensor such as the displacement sensor and the acceleration sensor in the fan of the nuclear power plant, Thereby preventing a defective component of the vibration signal from converging. That is, it plays a role of eliminating and reducing noise that is inevitably generated in the digitization process, which is an essential element, in advance.

The averaging unit 13 is a part for reducing unstable components in the vibration signal, and averages the vibration signal by the predetermined number of times of the predetermined fan. That is, the averaging unit 13 performs averaging processing on the vibration signal data from which the noise has been removed by varying the number of times for each vibration source. At this time, since the fan is likely to cause a diagnosis error due to a disturbance factor (vibration component) due to operation other than the optimum efficiency point of the driving characteristic and vibration of other parts, the number of averaging is increased to increase the non- ). On the other hand, the averaging unit 13 reduces the data size while performing the averaging process, thereby enabling efficient data accumulation management.

In addition, the defect signal extracting unit 15 divides the averaged vibration signal into an arbitrary section in consideration of the characteristics of a fan having a high time dependency of the defect signal, that is, a short time dependency, And outputs it to the characteristic extraction module 7. At this time, high time dependence of the defect signal means that the defect signal varies greatly with the lapse of time. In other words, relatively small and light parts such as a fan mean that the vibration signal is influenced by neighboring parts having different vibration signals, and once the defect occurs, the defect signal occurs intermittently.

1, the defect signal extracting unit 15 is provided with a subband designation sector 23. As mentioned above, the subband designating sector 23 is a part of a time- The time waveform of the averaged vibration signal is divided into an arbitrary section, that is, a subband, and then divided into a time division area and a frequency division area.

At this time, as shown in FIG. 1, the subband designating sector 23 includes a defect signal extraction stage 31 and a vibration splitting signal output stage 33. The defect signal extraction stage 31 includes a subband The vibration signal transmitted to the designated sector 23 is subjected to a wavelet-based envelope analysis and divided into a plurality of subbands. The divided vibration signal, that is, the vibration divided signal data, is divided into a set of specific basis functions, and the set of basis functions used is divided into a set of basis functions in a time axis direction for one basic wavelet basic function And parallel movements of the image. In addition, the defect signal extracting unit 31 evaluates the vibration division signal level of each of the divided sub-bands and judges whether or not the corresponding vibration division signal is defective.

To this end, the defect signal extracting stage 31 is composed of the defect signal comparing portion 41, the vibration signal dividing portion 43, the defect determining portion 45 and the defect signal extracting portion 47, The comparator 41 checks the frequency bandwidth of the defect signal, which is already determined for each element constituting the part, from the DB, and compares the confirmed value with each other to find an element having the minimum frequency bandwidth of the defect signal. For example, if the fan part of a power generation facility is a blade, the frequency bandwidth of a defect signal of each element of the blade, that is, a blade or a bushing, is checked in DB to find an element having a minimum defect signal frequency bandwidth I will.

2, the vibration signal dividing portion 43 outputs the vibration signal until the minimum width including the defective signal frequency component of the element found in the defect signal comparing portion 41 as described above, It is divided into four times. At this time, the vibration signal generates n + 1 sub-bands at every division, that is, when two sub-bands are divided by one time division, three sub-bands are divided by two division, When four subbands are divided into four subbands, five subbands are generated respectively.

In addition, the defect determination part 45 includes a total of n (n + 1) / 2 + n, i.e., 14 subbands generated during n = 4 divisions of the vibration signal in the upper vibration signal dividing part 43 The level of the vibration division signal of each sub-band is evaluated as an object, and the possibility of defect growth of the corresponding element, that is, the vibration division signal of the blade is judged.

In addition, the defect signal extracting section 47 extracts the defect signal from the vibration signal of each sub-band evaluated in the upper defect determination section 45 when it is determined that there is a possibility of defect growth, The vibration dividing signal is extracted as a defective signal and is transmitted to the vibration dividing signal output stage 33. [

At the same time, the defect determination part 45 evaluates the levels of the respective vibration division signals for the n (n + 1) / 2 + n upper, i.e., all 14 subbands, For example, the possibility of defect growth of the vibration division signal generated in the bushing. As a result of the determination, if it is determined that there is a possibility of defect growth in another element, the defect signal extraction section 47 likewise extracts the corresponding vibration division signal as a defect signal and delivers it to the vibration division signal output terminal 33.

The vibration dividing signal output stage 33 includes a plurality of sub-band demultiplexing stages 31 for dividing the demultiplexing signal output stage 31 into a plurality of sub- As shown in FIG. 6, into a time division area and a frequency division area, and outputs the signal to the characteristic extraction module 7. However, at this time, similarly, in order to output the frequency division area data, FFT is performed on the vibration division signal.

Therefore, since the vibration signal transmitted to the subband designation sector 23 is divided into an arbitrary section, a defective component with high time dependency, which is difficult to detect with the existing signal processing technique, can be detected early.

As shown in FIG. 1, the characteristic extraction module 7 has a function of extracting calculation values having statistical and morphological characteristics with respect to the vibration signal data, which has been pre-processed by the signal preprocessing module 5 and outputted therefrom . The extraction is performed by extracting the number of characteristics set by the following equation. Each characteristic value extracted by Equations (1) to (13) is provided to the characteristic selection module 9 as data dedicated to a machine learning algorithm.

x (n): time domain signal used for fault feature extraction

N: the number of samples of x (n) (hereinafter the same)

RMS (Root-Mean-Square) is a value for the magnitude of a signal change. It means the magnitude and magnitude of a sinusoidal waveform such as a sinusoidal waveform.

: x(n)의 평균, σ : x(n)의 표준편차 (이하 동일)

: average of x (n), standard deviation of?: x (n)

In statistics, skewness is an indicator of the asymmetry of the probability distribution. The skewness also increases with the bias of the signal (the degree to which the distribution of signal values converges on the average of the signal).

The impulse factor is an index of the magnitude of the largest impulse in the signal (the point where the waveform spikes).

In statistics, kurtosis is a measure of the degree of sharpness of the probability distribution. The kurtosis increases as the distribution of signal values converges near a certain value.

The kurtosis factor is a modified value of the kurtosis, which is the value that the kurtosis is sensitive to the size of the whole signal.

SMR (Square-Mean-Root) is the same or the same amount of consecutive waveforms as RMS. It is more sensitive to the signal size than RMS.

Peak-to-peak is a measure of the overall width of a signal, the difference between the smallest and largest values in the signal.

The margin factor refers to the difference between the minimum and maximum values relative to the average size of the signal.

The crest factor means the difference between the minimum and maximum values compared to the average size of the signal in the same way as the margin, and RMS is used instead of SMR as the average size.

Shape factor is a measure of the ratio of DC and AC components in electronics. It is the ratio of the magnitude of the continuous waveform to the average of the signal.

The frequency center means the average of the frequency domain.

The frequency RMS means the RMS value in the frequency domain.

The root variance frequency is the variance of frequency values in the frequency domain.

The characteristic selection module 9 selects a characteristic value suitable for the pan among the characteristic values calculated through the above equation. The characteristic selection module 9 includes a classification unit 51, a normalization unit 53, And a selection unit 55.

The classifying unit 51 classifies the characteristic values extracted by the characteristic extracting module 7 according to classes and stores the characteristic values for each part. The classifying unit 51 classifies the characteristic values of the fan, that is, And the rotating shaft. At this time, before applying a genetic algorithm (GA) to the above characteristic values in the selector 55, the class value is classified according to a certain criterion in preparation for the fitness evaluation in the genetic algorithm (GA) And the characteristic values may be classified according to, for example, the date on which the vibration signal data value is acquired from the fan or the presence or absence of a defect in the vibration signal.

The normalization unit 53 matches the data ranges of the characteristic values stored for each class in the classifying unit 51 for each of the components as described above and makes the distributions similar to each other. And the data values of the calculated values calculated by the equations (1) to (13) are matched, and the respective characteristic values are vectorized in the process of normalization. Thus, when normalization is not performed, the property value is likely to be underestimated or overestimated during class classification, thereby reducing the reliability of the diagnostic results. This is because when the class is determined by the classifying unit 51, the state is classified by evaluating the distance between each class region and another class, and if the non-normalized data is utilized for analysis, Is statistically or morphologically different from 100 to 10, 1 to 10, and 1,000 to 10,000, etc., and thus has a large influence on the calculation of the distance measurement, and the correlation between the variables can be misjudged .

The selecting unit 55 is a part for selecting suitable characteristic values by placing a normalized characteristic value in the normalization unit 53 for each part as a candidate and performing a genetic algorithm (GA) on them As shown in Fig. 3, again includes an initialization setting sector 61, a fitness evaluation sector 63, and a descendant loss generating sector 65. [

Here, the initial solution setting sector 61 is the first part to start the genetic algorithm, and first, n solution candidate sets called populations should be initially set. Therefore, for example, the feature values accumulated for 30 days each day in the above equations (1) to (13) can be used, so that the size of the group is calculated every 30 days by Equations 1 to 13 It increases in proportion to the number of days to accumulate characteristic value. When the size of the group is determined, that is, the cumulative number of days of characteristic value is determined, an initial solution is made by combining the characteristic values (f1 to f30) in the group as shown in the table below, and a plurality of chromosomes Chromosomes (Chrom1 to Chrom6) are generated. At this time, random numbers of binary numbers are substituted into the characteristic values f1 to f30 to generate an initial solution. Therefore, each of the chromosomes Chrom1 to Chrom6 has the characteristic values f1 to f30 ) As the initial solution, which has a plurality of genes, 0 or 1 randomly assigned. Here, the initial solution to which 0 is substituted for the binary random number means that the characteristic values f1 to f30 are not selected, and the initial solution to which the value of 1 is assigned is the characteristic values f1 to f30 Means selected.

f1 f2 f3 f4 f5 f29 f30 chrom1 One 0 One One 0 One 0 chrom2 0 One One One One 0 One chrom3 0 0 One 0 0 ... 0 0 chrom4 One 0 0 0 One One 0 chrom5 0 One One One 0 0 One chrom6 0 One 0 One One 0 One

The goodness-of-fit evaluation sector 63 determines the respective genes of the plurality of chromosomes (Chrom1 to Chrom6) originally generated in the above initial solution setting sector 61, that is, the genes of the chromosomes of the second- And the fitness of the descendant solution is evaluated. The adaptation of the SVM (Support Vector Machine) algorithm is used to evaluate the fitness of each gene, that is, the characteristic values f1 to f30 defined by 0 or 1, respectively. For the fitness evaluation, the characteristic values having the initial solution of 1 are combined to create an objective function graph as shown in FIG. 4, for example.

According to this graph, for example, if f1 is the frequency RMS value, f3 is the frequency center value, and f4 is the standard deviation frequency value, f1 is set as the x axis, f3 is set as the y axis, and f4 is set as the z axis , And the values f1, f3, and f4 acquired for each day are displayed on the graph. At this time, in case of chromosome 1 (chrom1), all values of f1, f3, and f4 are displayed in the above graph. However, chromosome 2 (chrom2) does not show the value of f1, only f3 and f4 are shown in the graph above. Thus, if the initial solutions of all the chromosomes are displayed in the above graph, the marked points are classified according to specific criteria to form a class, as mentioned above. The objective function OF is generated on the graph according to the density D in each class and the distance L between the classes. Accordingly, all the points indicated on the graph, that is, the values of all genes (f1 to f30), are evaluated according to how well they match the objective function.

The progeny generation sector 65 is a part generating an offspring solution through a plurality of chromosomes as if a progeny is generated by crossing between chromosomes in an animal gene. In other words, a descendant solution is produced by intersecting the chromosomes except for the descendant harmful chromosome and the descendent harmful chromosome, which has a high proportion of the total good chromosomes with a high total score, to the next generation. To do this, the fitness of each of the genes (f1 to f30) evaluated above is summed to evaluate fitness for each chromosome (Chrom1 to Chrom6). Thus, for example, the total score of fitness is 1, chrome 2 is 3, chrome 3 is 2, chrome 4 is 5, chrome 5 is 6, chromosome 6 is chrome 1, (chrome 6) can be defined as 4, and the order of fitness is high in order of chromosome 5, chromosome 4, chromosome 6, chromosome 2, chromosome 3, and chromosome 1. At this time, if elitism is applied at 33%, the chromosome 5 and chromosome 4 selected as the optimal chromosome are transferred to the next generation, that is, the second generation descendant solution, as it is. The remaining four chromosome 6, chromosome 2, chromosome 3, and chromosome 1 are randomly crossed with two pairs to create four new chromosomes, a second generation progeny solution, resulting in a total of six progeny solutions . At this time, the intersection of two chromosomes is assumed to occur at a rate of 50% without mutation.

Thereafter, the fitness evaluation and the descent generation are repeatedly performed until the desired descendant solution is generated, and until the optimum solution or the family solution satisfying the predetermined number of generations or a predetermined error range is obtained, the selection unit 55 The fitness evaluation in the above fitness evaluating sector 63 and the generation of the descendent solution in the descent generating sector 65 are repeatedly performed. If the result of the iterative execution is that the optimum solution or the family solution is obtained, the selection unit 55 can select the optimal or family property value of the vibration signal received from the specific component of the fan. That is, it is possible to select the characteristic value that shows the vibration state of the specific part of the fan as the topmost or the bottom.

The diagnosis module 11 learns the state of the fan through the property value selected by the property selection module 9 and diagnoses the state of the fan through the learning result, ), And the normal graph is created on the target machine. Then, the machine recognizes the vibration signal newly received from a specific part of the fan, so that the current state of the fan can be predicted and diagnosed.

For this purpose, the SVM can be used to automatically classify and recognize new vibration signals received by the learned machine. The SVM can be used not only for linear but also for nonlinear data classification. Basically, binary classification is performed. However, In order to distinguish between different states, SVMs corresponding to the classifications of each class are studied. The discrimination of new data is performed by a multi-class SVM (Multi-Class SVM) SVM) may be applied. Therefore, the new vibration data input to the machine, that is, the characteristic value, is displayed on the graph through the SVM to visually indicate the vibration state of the fan in order to predict and diagnose the vibration state of the fan after completion of the machine learning. However, at this time, the diagnostic module 11 compares the defect graph formed by the characteristic selection module 9 with the normal graph as described above, so that the degree of defect of all the components can be grasped as compared with the normal case.

The operation of the machine learning type predictive diagnosis apparatus 1 of the nuclear power plant according to the preferred embodiment of the present invention will now be described.

According to the predictive diagnosis apparatus 1 of the present invention, as shown in FIG. 5, the vibration signal receiving module 3 shown in FIG. The vibration signal is received from the same part (S100).

Next, the vibration signal data received through the receiving module 3 is preprocessed through the signal preprocessing module 5 (S200).

Then, the statistical and morphological characteristic values are calculated from the vibration signal preprocessed in the characteristic extraction module 7 using the above Equations 1 to 13, and the calculated values are transmitted to the outside (S300)

Next, the characteristic selection module 9 selects an appropriate characteristic value from the characteristic value data received from the characteristic extraction module 7 (S400).

Finally, the diagnostic module 11 learns the characteristic values selected by the characteristic selection module 9 to the target machine, sends a test signal relating to the fan vibration to the target machine or a new vibration signal received from the fan, The current state of the fan is examined or predicted through a machine (S500), and a series of machine learning predictive diagnoses for the fan is completed.

At this time, as shown in FIG. 4, the noise signal data inputted to the signal preprocessing module 5 is primarily removed through the noise portion. In particular, the noise component due to the quantization error inevitably generated in the process of digitizing the analog signal collected from the vibration sensor is removed and reduced in advance (S210).

Next, the noise-removed vibration signal data is averaged by the averaging unit 13 as shown in FIG. 5 (S220). Since the target part is a fan having a high time dependency, the possibility of occurrence of a vibration error The average number of times is relatively increased, thereby eliminating the unstable property in the vibration signal.

Subsequently, the averaged vibration signal is transmitted to the defect signal extracting unit 15, and a defect signal is extracted (S230). To this end, the vibration signal data of the fan with high time dependency is sent to the subband designation sector 23 as shown in Fig. 6, so that the defect signal is extracted. At this time, the corresponding vibration signal data is sent to the defect signal extracting stage 31 to extract the defect signal from the defect signal extracting stage 31 (SS100), and the extracted defect signal is outputted through the vibration extracting signal output stage 33 (7). 6, the extracted defect signal is divided into time or frequency regions, and time division data is output in the time division region, FFT is first performed in the frequency division region, and then frequency division data (SS200).

On the other hand, in the defect signal extracting stage 31, as shown in FIGS. 6 and 7, the defect signal comparing section 41 firstly judges whether or not a defect The frequency bandwidths of the signals are compared with each other to find an element having the minimum width of the frequency band B of the defect signal (SS110). Then, the vibration signal is divided into n number of times until the minimum width including the defect signal frequency band B of the element found in the defect signal comparing part 41 is reached, and the total of n + 1 n + 1) / 2 + n subbands (SS120). Finally, the level of the vibration division signal of each sub-band is evaluated for all n (n + 1) / 2 + n sub-bands, and the possibility of defect growth of the vibration division signal of the corresponding element is determined (SS130) , And the defect signal is finally extracted (SS140).

In order to select the characteristic values in the characteristic selection module 9, as shown in FIG. 8, the characteristic values extracted from the characteristic extraction module 7 are classified and stored for each class in the classification unit 51 S410). At this time, for example, as mentioned above, the 30 kinds of characteristic values can be classified according to the date acquired from the fan or the presence or absence of a defect of the vibration signal, for example, as shown in FIG. 3,

Next, the plurality of characteristic values classified as above are normalized by the normalization unit 53 and expressed as a dimensionless vector value as described above (S420).

As described above, the normalized characteristic values are evaluated according to the genetic algorithm in the selecting unit 55, and the selecting unit 55 selects appropriate characteristic values according to the result (S430) . For this, the selector 55 firstly sets n candidate candidates in the initial solution setting sector 61 and starts the genetic algorithm (S431). Next, the selection unit 55 evaluates the fitness of the initial solution by the SVM algorithm in the fitness evaluation sector 63 (S432). Subsequently, a progeny solution of n = first order is generated in the seed generation sector 65 (S433), and a certain proportion of the excellent chromosome having a high degree of fitness is delivered as a first-order progeny solution according to the elitism. By crossing the genes with other chromosomes, another primary progeny is generated. Subsequently, it is checked whether the first-order progeny solution is a suitable solution (S434). Since the first-order progeny solution is not an appropriate solution, the second-order progeny solution is generated by repeating the above- When it is judged that the descendent solution is a suitable solution, the corresponding solution is selected as the optimum or property value of the family register (S435).

When the selection of the characteristic value by the characteristic selection module 9 is completed in this way, the diagnostic module 11 learns the selected characteristic value to the target machine, and diagnoses the new vibration signal received from the following fan by the learning- (S500), the machine is allowed to learn the current state of the specific component by the property value selected by the property selection module 9 (S510).

However, in order to learn the machine learning, it is necessary to include information on the current state of the component as well as the numerical value of the selected characteristic value. In other words, the current state of a specific part of a fan when a certain characteristic value is calculated as an arbitrary numerical value must be DB and learned to a machine. However, since the present state of the part is normal, the normal state is represented by one data. However, since the defect state can be increased indefinitely, the more the defect data, the higher the performance of the diagnosis by the machine in the future. In other words, if a part is in a defect state, defects may occur in any of the components constituting the component, and the type, number, and size of defects may be different for one component. Therefore, It can be done. For example, if the vibration source of the vibration signal is called a bearing, and the selected characteristic value is referred to as a frequency RMS and 200 as the frequency RMS value, the bearing is disassembled in advance and the defect generating the vibration signal having the frequency RMS value of 200 And if a 1-mm-long crack of the inner ring, for example, is the cause of the defect, they should all be DB. Then, in addition to the information that the frequency RMS value is 200 at the time of machine learning, it is necessary to also learn defect information that the bearing has a crack of 1 mm on the inner ring.

After the learning of the machine is completed in this way, the current state of a specific part can be diagnosed by the above graph after several tests (S520). At this time, if a vibration signal having a frequency RMS value of 200 is received from the fan, the machine predicts that the present fan predicts that a 1 mm crack has occurred in the inner ring of the bearing.

1: Predictive diagnostic device 3: Vibration signal receiving module
5: Signal preprocessing module 7: Characteristic extraction module
9: Characteristic selection module 11: Diagnostic module
13: averaging unit 15: defect signal extracting unit
23: Sub band designation sector 31: Defect signal extraction stage
33: vibration division signal output stage 41: defective signal comparison section
43: vibration signal division part 45: defect judgment part
47: defect signal extracting section 51:
53: normalization unit 55: selection unit
61: initial setting sector 63: fitness evaluation sector
65: Generation Sectors Sector

Claims (4)

delete A vibration signal receiving module 3 for receiving and storing the vibration signal from the fan;
A signal preprocessing module 5 for preprocessing the vibration signal received by the vibration signal receiving module 3;
A characteristic extraction module 7 for calculating statistical and morphological characteristic values from the vibration signal preprocessed in the signal preprocessing module 5;
A characteristic selection module (9) for selecting a plurality of characteristic values suitable for the pan among the characteristic values extracted from the characteristic extraction module (7); And
And a diagnosis module (11) for learning the state of the fan with the characteristic value selected by the characteristic selection module (9) and diagnosing the fan through the learning result,
The signal preprocessing module (5)
An averaging unit 13 for averaging the vibration signal received by the vibration signal receiving module 3; And
The vibration signal of the fan that has been averaged by the averaging unit 13 is divided into an arbitrary section and divided into a time division area and a frequency division area and then outputted to the characteristic extraction module 7, (15)
The characteristic selection module (9)
A classifying unit 51 for classifying and storing characteristic values extracted by the characteristic extracting module 7 for each class;
A normalization unit 53 for matching the data ranges of the characteristic values stored for each class to the classifying unit 51 and making the distributions similar; And
And a selection unit (55) for selecting an appropriate characteristic value by performing a genetic algorithm with the characteristic value normalized by the normalization unit (53) as a candidate. . The method of claim 2,
The selector (55)
An initial solution setting sector 61 for generating a plurality of chromosomes by randomly assigning binary random numbers to initial solutions of respective characteristic values;
A fitness evaluation sector (63) for evaluating a fitness of a progeny solution of chromosomes set in the initial solution setting sector (61) or an initial solution set in the initial solution setting sector (61); And
Generation of the next generation is generated through a descendant solution transmitted to the next generation as it is and a descendant solution generated by intersecting the remaining chromosomes except for the superior chromosome, And a descendant loss generating sector (65)
Wherein the initial solution to which 0 is substituted for the binary random number is not selected and the initial solution to which a value of 1 is assigned is selected. The method of claim 3,
The goodness-of-fit evaluation sector 63 is configured to classify a selection characteristic value having a value of 1 in an initial solution or a descendent solution according to the class and determine a fitness value of the selection characteristic value through the density of the intra- And the fitness of the selected characteristic value is evaluated according to the comparison result by comparing the measured fitness with each selected characteristic value to evaluate the fitness of the selected characteristic value.

Images