JP3484905B2 - State determination device - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a state discriminating apparatus, and more specifically, by executing a learning mode, knowledge (especially effective feature amount) used in discriminating a state is automatically obtained. A device capable of generating.

[0002]

2. Description of the Related Art As described in, for example, Japanese Patent Application Laid-Open No. 5-340799, the following state determination device is known. This device detects vibrations of various machines that are objects to be measured with a sensor, analyzes the sensor output with an information processing device such as a computer, and performs earthquake discrimination, or vibration of the object to be measured. It is determined as follows whether it is normal or abnormal.

First, the information processing apparatus is operated in the learning mode while the object to be measured is vibrating normally. In this learning mode, the vibration waveform from the vibration sensor is sampled for an appropriate period, the vibration waveform is analyzed according to a plurality of predetermined analysis items, the analysis data of each item is statistically processed, and the processing result (that is, , The characteristics of the waveform of normal vibration) are used to determine an algorithm for determining normal / abnormal vibration.

Here, the analysis items are the frequency, amplitude, maximum value, minimum value, peak-to-peak value, number of times the high level threshold is exceeded, and number of times the low level threshold is exceeded, of the vibration waveform, The number of occurrences of the minimum value, etc., is an item suitable for extracting the characteristics of the vibration waveform.

In the actual operation mode, the vibration waveform from the vibration sensor is sampled at any time, the vibration waveform is analyzed according to each analysis item as in the learning mode, and the analysis data of each item is processed according to the determination algorithm. , It is determined whether the vibration of the measured object is normal or abnormal.

[0006]

In the conventional device described above, a normal / abnormal determination algorithm is automatically generated based on the normal vibration waveform of the measured object input in the learning mode. Therefore, when a general user applies this state determination device to a specific device, it is not necessary to analyze the characteristics of normal vibration and abnormal vibration of the applied device and to consider an appropriate determination algorithm based on the analysis result. No need at all. That is, the user only needs to attach the vibration sensor to the applied device and operate the device in the learning mode, and in the subsequent actual operation mode, normal vibration / abnormal vibration is appropriately determined. In this respect, it is an extremely convenient and excellent state determination device.

However, in the conventional device, the analysis items when analyzing the vibration waveform from the vibration sensor in the learning mode and the actual determination mode are uniquely determined at the device designing stage, and therefore are applied to devices such as motors. However, when applied to equipment such as a hydraulic cylinder, the reliability of the determination is lowered, and there is a problem that the compatibility is good or bad depending on the application target. Therefore, it is necessary to set an appropriate analysis item in advance according to the inspection target, and such setting is actually complicated and difficult, and there is a possibility that an incorrect setting may cause an erroneous determination.

Furthermore, even when the object to be measured is known in advance, the design should be performed by relying on the know-how and intuition of a skilled engineer when determining analysis items at the design stage of the monitoring device. As it is done, a great deal of effort is required to set the optimum knowledge. If there is no skilled engineer, sufficient knowledge cannot be incorporated, and even a skilled engineer may have variations due to individual differences. As a result, there is a risk of misjudgment.

In addition, it may be necessary to reconstruct the optimum knowledge necessary for the state determination based on the secular change of the object to be measured or the change of the environment. It will be set, and the above problem will occur again.

The present invention has been made in view of the above background, and an object of the present invention is to solve the above problems and automatically execute a learning mode by executing a learning mode without any skilled technique. It is possible to select a feature amount (effective feature amount) suitable for determining the state of the measured object,
An object of the present invention is to provide a state determination device capable of performing accurate state determination by making a determination using it.

[0011]

In order to achieve the above-mentioned object, in a state discriminating apparatus according to the present invention, a sensing section for detecting a physical quantity of an object to be measured and a sensor output waveform output from the sensing section. From the measured object for learning, in a state determination device having a determination unit that determines a plurality of states including normality / abnormality of the measured object based on the calculated characteristic amount from the obtained characteristic amount Obtaining data in which information of a plurality of feature amounts calculated from the sensor output waveform output by the sensing unit that detects the generated physical quantity and information regarding the state of the measurement target object for learning are associated, and the learning is performed. The state of the measured object is calculated by calculating the degree of separation between the respective states of the measured object for each feature amount and selecting one or a plurality of the high-separation characteristic amounts as effective feature amounts. A feature extraction unit that separately extracts effective features and a knowledge generation unit that generates knowledge for making a determination from the extracted effective feature amounts are provided. The feature amount extraction unit and the knowledge generation unit The determination unit operates in the mode, and is configured to determine the state of the measured object based on the effective feature amount extracted by the execution of the learning mode and the generated knowledge (claim 1). ).

In order to execute the learning mode in order to generate the knowledge for determining the state, the object to be measured in a predetermined state (the kind of the state is known) is sensed by the sensing unit and the output waveform Get the signal. The output waveform signal is given to the feature extraction unit, and various feature amounts set in advance are obtained there. It learns iteratively by changing the state and collects various feature quantities in each state. Then, based on the obtained feature amount data, the degree of separation between the states of the measured object is calculated for each feature amount.

Here, the degree of separation means the degree to which each state can be separated by using a certain feature amount, and for example, it is 2
There are two states α and β, and it can be said that the separation is higher as the existence / appearance region of the feature amount in the state α and the existence / appearance region of the feature amount in the state β are farther apart. Then, specifically, it can be performed by various calculations as exemplified in the embodiment, but in the embodiment, the calculation can be easily performed, and it is not necessary to limit to the exemplified calculation formula, Various methods can be adopted. Further, if there is a feature amount that does not appear in the other party's region without overlapping even if the boundary of the region where the feature amount of each state exists is close, it can be said that the degree of separation is high.

Therefore, by extracting the one having a high degree of separation as the effective feature quantity and using it as the knowledge to be used in the actual state discrimination, the feature quantity suitable for the object to be measured can be discriminated. Can effectively and accurately determine the state. Moreover, since the extraction and selection of the effective feature amount is automatically performed, even a person without a skilled technique can extract the desired one. Also, there is no variation. Further, even when it is necessary to correct the discrimination knowledge of the object to be measured due to a change with time or the like, it is possible to easily reconstruct the desired knowledge at that time by performing the learning described above. Then, even when the knowledge is reconstructed, it can be performed without a skilled engineer.

Further, a multidimensional separation degree by a combination of a plurality of feature quantities is calculated from the calculated separation degree for each feature quantity, and a combination of the feature quantities having a high multidimensional separation degree is selected. It is also possible (claim 2). By doing so, the most effective combination of feature quantities for distinguishing a plurality of states is selected, and thus the accuracy of judgment is further improved.

Further, when selecting a combination of a plurality of feature quantities, it is possible not to select a combination of feature quantities having a high degree of correlation with each other due to the correlation coefficient. A feature amount having a high degree of correlation (correlation coefficient) has a similar change in the feature amount with respect to a change in state. Therefore, a case where a plurality of feature amounts having a high degree of correlation are used and a case where one of them is used is determined. The discrimination results and erroneous discrimination rates are almost the same. Therefore, by preventing multiple feature quantities with high correlation from being selected, it is possible to select feature quantities that have no correlation with each other and that are effective for determination, and it is possible to omit useless calculations and reduce the amount of calculation. Therefore, it is possible to efficiently and accurately perform the discrimination.

Then, for the selected effective feature quantity, the existence range of the effective feature quantity for each state of the object to be measured is determined from the average and deviation of the effective feature value data measured at the time of learning, and the object to be measured is determined in the judging section. The state may be determined by comparing with the value of the effective feature amount calculated from the sensor output waveform of the object (claim 4). In this way, knowledge corresponding to each state is constructed, so that the accuracy is further improved.

[0018]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 shows the overall configuration of a preferred embodiment of a state determination device according to the present invention. As shown in the figure, the output of the sensing unit 1 that detects physical quantities such as vibration, acceleration, pressure, sound, angular velocity, and temperature generated from the object to be measured is given to the A / D converter 2, where the sensing unit is provided. The output waveform (analog signal) output from 1 is converted into a digital signal at a predetermined sampling frequency and time length. The output of the A / D converter 2 is output to the feature extraction unit 4 for learning via the changeover switch 3.
It is adapted to be selectively switched and connected to the determination unit 5 which makes an actual determination.

Further, effective feature amount data suitable for determining the state extracted by the feature extracting unit 4 is given to the determining unit 5 and the knowledge generating unit 6. Further, the knowledge generation unit 6 generates knowledge necessary for the determination from the effective feature amount data, and is configured to give the generated knowledge to the determination unit 5. Furthermore, a state selection switch 7 is connected to the feature extraction unit 4, and it becomes possible to obtain information as to which state (abnormal / normal, etc.) the data currently given from the switch 7 relates to. ing. Then, the learning mode is executed by the feature extraction unit 4, the knowledge generation unit 6, and the state selection switch 7.

Next, each part will be described. The sensing unit 1 includes various semiconductor physical quantity sensors (capacitance type, piezoresistive type, etc.), a sensor for detecting a predetermined physical quantity such as a thermometer and a microphone, an amplifier for amplifying an output signal of the sensor, and It is configured to include a filter for removing an offset component and a high frequency noise component in the amplified signal. Further, when the frequency band to be detected is known, a filter (bandpass filter or the like) that passes only the frequency band concerned may be provided.

Then, the sensing unit 1 is brought into direct contact with the object to be measured or is arranged in proximity thereto. As a result, a waveform signal (analog) according to the state of the measured object
Is output from the sensing unit 1.

Here, considering the case of diagnosing the state of the motor, in the case of the motor, normality, unbalance, gear failure,
It is necessary to judge various states such as bearing failure and output the result. In this case, the object to be measured described above serves as a motor, and the sensing unit 1 uses an acceleration pickup / acceleration sensor, a thermistor, or the like. Then, the output of the acceleration sensor is 10 to 10 which is the frequency range including the vibration component of the motor after the amplification process is performed.
A bandpass filter process per kHz is performed.

On the other hand, the feature extracting section 4 is configured as shown in FIG. That is, the input side is provided with a feature amount calculation unit 4a, which receives an output waveform signal (digital signal) given through the A / D converter 2 and calculates a predetermined number of types of feature amounts. The feature quantities calculated / extracted at this time include the following (see FIG. 3). Note that the feature amount used in the present invention is not limited to the feature amount illustrated below, and other possible feature amounts may be added, or conversely, a feature amount determined to be completely unnecessary may be obtained. It may not be used (not used).
The point is that the value of the feature amount that may be effective for the determination may be calculated.

[0024]

[Table 1] (1) Maximum value (2) Minimum value (3) Range (maximum value-minimum value) (4) Number of times exceeding a predetermined threshold (5) Time when exceeding a predetermined threshold ( 6) Peak TO peak (maximum amplitude in one cycle) (7) Number of occurrences of minimum value (corresponding to number of peaks) (8) Single amplitude value (9) Average value (SUM (| data |) / number of data ) (10) RMS value (SQRT (SUM (| data |) ² / number of data)) (11) Crest value (RMS value / single amplitude value × 100) (12) Average slope (SUM (| slope |) / data (13) Integrated value (trapezoidal integration (data)) (14) Absolute integrated value (trapezoidal integration (| data |)) Then, the feature quantity calculated as described above is compared with the category classification unit 4b in the next stage. It is adapted to be applied to the relation number calculation unit 4c. Since the category selection unit 4b is also provided with information regarding the current state of the object to be measured from the state selection switch 7, the given feature amount is stored separately for each state (category). As a result, the values of the respective characteristic amounts obtained each time are sorted and stored in a table format as shown in FIG. 4, for example. Then, the separation data thus obtained is supplied to the separation degree calculating unit 4d in the next stage.

In this example, the state selection switch 7 allows the operator to manually change the state of the object to be measured, such as normal or abnormal state, such as console input by the operator, to give the state to the category classification section 4b. It has become. Of course, it is also possible to automatically give the kind of state of the measured object at that time by signals from various sensors and other external devices.

The state to be applied is not limited to the one for simply discriminating between two types, normal and abnormal, but the type of abnormality is given and the measured object itself is normal, but the earthquake / installation surface itself It can be appropriately set according to the state to be detected, such as when there is a disturbance such as vibration.

On the other hand, the correlation coefficient calculation unit 4c compares the feature amount data given each time to obtain the correlation coefficient between any two feature amounts, and the correlation coefficient (correlation degree) is high. To extract. That is, the correlation coefficient obtained exceeds a certain reference value. Then, the combination of the feature amounts having the high correlation coefficient is given to the effective feature amount selecting unit 4e.

The separation degree calculating unit 4d calculates the degree of separation indicating the distance between each category (state) for each given feature amount, and specifically, performs the following processing. (1) Calculation of Average Value and Standard Deviation of Feature Value That is, the average value μ and standard deviation σ of each feature value are obtained for each state. Thereby, for example, a table as shown in FIG. 5 is created. (2) Calculation of Distance Between Each State Referring to the table obtained in the process of (1), the distance d between arbitrary states for each feature amount is obtained. As an example, the average value of the state α for a certain feature value is μ _α , the standard deviation is σ _α, and similarly, the average value of the state β is μ _β , and the standard deviation is σ _β . The distance d _αβ can be obtained by d _αβ = (μ _α −μ _β ) ² / (σ _α × σ _β ). Then, the distances between the states for all the combinations are obtained based on the equation.

The distance is not limited to the one obtained by the above equation, but may be obtained based on the following equation, for example.

D _αβ = | μ _α −μ _β | / σ _α or d _{α β} = | μ _α −μ _β | / (σ _α ² + σ _β ² ) ^1/2 (3) Calculation of separation degree (2) Using the distance d obtained by the processing, the average of the distances between all the states for each feature is calculated, and the average value is taken as the degree of separation. Thereby, the degree of separation for each feature amount is obtained. As an example, there are three states α, β, γ for the feature amount A, and the mutual distances are d _αβ (distance between states α and β), d _βγ (distance between states β and γ),
_Letting d _γα (distance between states γ and α), the degree of separation (A) = (d _αβ + d _βγ + d _γα ) / 3 can be obtained. Then, the degree-of-separation data for each obtained feature amount is used as the effective feature amount selection unit 4e in the next stage.
Give to.

The effective feature amount selection unit 4e determines the effective feature amount (group) suitable for performing the state determination based on the given separation data, and determines the selected effective feature amount in the next stage. It is provided to the section 5 and the knowledge generation section 6.
Specifically, the one having a high degree of separation is selected as the effective feature amount. That is, a predetermined number is selected from the top in the descending order of the degree of separation, a feature amount having a degree of separation equal to or higher than a certain reference value is selected, or a combination thereof
Various methods such as D) can be adopted.

Further, in this example, based on the correlation degree data provided from the correlation coefficient calculation unit 4c, both feature quantities having a high correlation coefficient are not selected as effective feature quantities. That is, even if the state determination is performed using a plurality of feature amounts having a high correlation coefficient, or the state determination is performed using one of the feature amounts, the accuracy of the determination result does not change much. Therefore, in this example, a combination of feature quantities having a correlation coefficient equal to or larger than the setting is not selected, and when there are a plurality of feature quantities having such a high correlation coefficient, the feature quantity having the highest degree of separation among them is selected. Was selected as the effective feature amount. As a result, highly accurate discrimination can be efficiently performed with a small amount of calculation processing.

Further, in the present embodiment, the criterion for selecting effective feature amounts is based on the degree of separation for each feature amount, but the present invention is not limited to this, and multidimensional separation is performed. Degrees may be used. That is, the multidimensional separation S2 _ABC based on the combination of the feature quantities A, B, and C can be obtained, for example, by the following formula.

S2 _ABC = (S _A ² + S _B ² + S _C ² )
^1/2 where S _A is the degree of separation of the feature amount A, S _B is the degree of separation of the feature amount B, and S _C is the degree of separation of the feature amount C. The combination of the obtained feature amounts may be selected as the effective feature amount. Needless to say, the method for obtaining the multidimensional separation is not limited to the above.

The knowledge generation unit 6 automatically generates knowledge for performing state determination from the selected effective feature amount and its value (average value, standard deviation), and supplies it to the determination unit 5. There is.

Specifically, the membership function is obtained, and an example thereof is shown. For example, in the case of making a crisp inference, as shown in FIG. It is possible to generate a function in which the range of 1 is the goodness of fit and the other goodnesses are 0. When fuzzy inference is used for the determination, as shown in FIG. 7B, the goodness of fit is set to 1 at the average value μ, and the goodness of fit is set to 0 at a position separated by ± 3σ from the average value. A triangular function may be generated. Furthermore, as shown in FIG. 6C, the goodness of fit within the range of ± σ with respect to the average value μ is 1
Then, as it deviates from the average value further,
It is also possible to generate a trapezoidal function in which the goodness of fit decreases and the goodness of fit becomes 0 at a position apart from the average value by ± 3σ. Then, a membership function is generated for each state.

The rule used when performing fuzzy inference is, for example, "if all effective feature quantities are in state A, the object to be measured is state A. If all effective feature quantities are in state B, The object to be measured is in the state B. For all effective feature amounts, if the object to be measured is the state C, it is the state C ....

It should be noted that which shape of the above three functions is to be generated may be preset or all of them may be generated. Further, the shape to be generated is not limited to the one described above, and any other shape may be generated. Then, in this example, the trapezoidal membership function shown in FIG. 7C is generated in consideration of the variation in the data of the state of the measured object.

On the other hand, the determination unit 5 is composed of a feature amount calculation unit 5a and a feature amount comparison unit 5b as shown in FIG. 7 when performing crisp inference. That is, when the changeover switch 3 is switched to the determination unit 5 side, the sensing data (output waveform signal) output from the sensing unit 1 is converted to digital data and then provided to the feature amount calculation unit 5a. Here, the value of the effective feature amount is obtained.
The specific calculation method is the same as that of the feature amount calculation unit 4a in the feature amount extraction unit 4 for executing the learning mode,
The difference is that the feature amount calculation unit 4a obtains all the preset feature amounts, whereas the feature amount calculation unit 5a in the determination unit 5 obtains only the selected effective feature amount. Then, the value of the obtained characteristic amount is given to the characteristic amount comparison unit 5b.

The feature quantity comparison unit 5b compares the value of the obtained feature quantity with the membership function of each state, and the goodness of fit 1
It is determined whether or not it exists within the range of. Then, the state of the membership function belonging to the goodness of fit 1 is determined to be the state of the measured object. Since there are usually a plurality of effective feature amounts, the comparison results of all feature amounts do not necessarily become the same state.Therefore, when different results are obtained, the state can be determined by, for example, a majority method. it can.

When the state discrimination is performed by using the fuzzy inference, the discriminator 5 can be constructed as shown in FIG. 8, for example. That is, as in the case of FIG. 7, the feature amount calculation unit 5a calculates the effective feature amount and sends it to the fuzzy inference unit 5c. In the fuzzy inference 5c, the goodness of fit, which is the degree of belonging to each state, is calculated based on the membership function or the like given from the knowledge generation unit 6. Then, the degree of matching is determined by the definite calculation unit 5 at the next stage.
given to d, the state having the highest degree of conformity is determined therein, and the determined state is output. In addition, when the conformity of any of the states does not satisfy a certain criterion, it may be output that the state is not any (unknown). Furthermore, each degree can be output separately.

In any of the above cases, the determination result is output to an output device (not shown) to provide information to the operator or external equipment.

[0043]

As described above, in the state discriminating apparatus according to the present invention, each feature quantity appearing in each state is obtained in the learning mode, and the feature quantity is actually discriminated based on the degree of separation between the states. Since the effective feature quantity used for the above is selected, the feature quantity suitable for automatically determining the state of the measured object (the effective feature Quantity),
By using that to make a determination, it is possible to perform an accurate state determination.

[Brief description of drawings]

FIG. 1 is a diagram showing a preferred embodiment of a state determination device according to the present invention.

FIG. 2 is a diagram showing an example of an internal structure of a feature quantity extraction unit.

FIG. 3 is a diagram illustrating an example of a feature amount to be extracted.

FIG. 4 is a diagram illustrating a function of a category classification unit.

FIG. 5 is a diagram illustrating a function of a separation degree calculation unit.

FIG. 6 is a diagram illustrating a function of a knowledge generation unit.

FIG. 7 is a diagram illustrating an example of an internal structure of a determination unit.

FIG. 8 is a diagram showing another example of the internal structure of the determination unit.

[Explanation of symbols]

1 Sensing section 4 Feature extraction unit 5 Judgment section 6 Knowledge generator

─────────────────────────────────────────────────── ─── Continuation of front page (72) Inventor Atsushi Nagata 10 Odoron-cho, Hanazono Todo-cho, Ukyo-ku, Kyoto City, Kyoto Prefecture (56) References JP-A-6-213704 (JP, A) JP-A-5- 340799 (JP, A) JP 59-2187 (JP, A) JP 60-142788 (JP, A) JP 63-55677 (JP, A) JP 7-324976 (JP, A) JP-A-5-72026 (JP, A) JP-A-2-272326 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G01D 1/18 G01H G06F G06K 9/00

Claims (4)

(57) [Claims] 1. A sensing unit that detects a physical quantity of an object to be measured, a predetermined characteristic amount is calculated from a sensor output waveform output from the sensing unit, and the object to be measured is calculated based on the obtained characteristic amount. In a state determination device having a determination unit that determines a plurality of states including normality / abnormality, a plurality of feature amounts calculated from sensor output waveforms output by the sensing unit that detects a physical quantity generated from a measurement object for learning Of the information and the information relating to the state of the measured object for learning is obtained, and the degree of separation between the states of the measured object for learning is calculated for each feature amount, and A feature extraction unit that extracts a feature effective for determining the state of the object to be measured by selecting one or a plurality of feature values with high degree of separation as the effective feature amount, and distinguishes from the extracted effective feature amount. And a knowledge generator for generating knowledge for performing the feature extraction unit and the knowledge generating unit operates in the learning mode, the effective feature value extracted by the execution of the learning mode,
The state determination device, wherein the determination section is configured to determine the state of the object to be measured based on the generated knowledge. 2. A multidimensional separation degree based on a combination of a plurality of feature quantities is calculated from the calculated separation degrees for each feature quantity, and a combination of the feature quantities having a high multidimensional separation degree is selected. The state determination device according to claim 1, wherein: 3. When selecting a plurality of feature quantities, a combination of feature quantities having a high degree of correlation with each other is not selected due to the mutual correlation coefficient.
State determination device described in. 4. For the selected effective feature quantity, the existence range of the effective feature quantity for each state of the object to be measured is determined from the average and deviation of the effective feature quantity data measured at the time of learning, and the determination unit measures the measured value. The state determination device according to claim 1, wherein the state determination is performed by comparing with a value of an effective feature amount calculated from a sensor output waveform of the object.