JP2007108107A - Facility diagnostic device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To quickly specify an abnormal factor in the operation condition of a facility to conduct correct diagnosis, in a facility diagnostic device for collecting data by providing a sensor in a driving facility for driving a belt by a plurality of pulleys, and for conducting the diagnosis using principal component analysis. <P>SOLUTION: The time-serial data is collected by a time-serial data collecting part 4, a multi-dimensional data is generated by a multi-dimensional data generation part 5, the principal component analysis is carried out for a large amount of multi-dimensional data collected and generated under a prescribed condition, by a pattern generation part 6, to determine a feature space, and pattern information is stored in a storage part 7. A Mahalanobis distance from each stored pattern is computed in a new multi-dimensional data, by a distance computing part 8, and the degree of nearness to the each of the patterns is determined to diagnose the operation condition. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a diagnostic apparatus for diagnosing drive equipment that drives a belt with a plurality of pulleys, and more particularly to a diagnostic apparatus that collects waveform data such as vibration and sound and performs diagnosis using principal component analysis.

  A conventional method for diagnosing degradation of a target facility acquires waveform data of the facility in a plurality of time zones including a point in time when the facility is normal, and analyzes a part of the waveform data in each of the plurality of time zones. The analysis time zone is divided into divided time zones, the divided time zones are divided into T time zones, and the waveform data is divided by Fourier transform for each of the T time zones. Find the strength of each frequency band. Then, the strength for each frequency band is standardized, and a principal component score for each divided time zone is obtained using the standardized strength for each frequency band. When obtaining the principal component score for each of the divided time zones, the eigenvector used when obtaining the principal component score is an eigenvector when the equipment is normal (see, for example, Patent Document 1).

JP-A-2004-20193

  In such conventional equipment diagnosis, a map according to the distribution of normal data is determined using principal component analysis, and new acquired data is evaluated on this map. In such a diagnosis, it is easy to diagnose whether the facility is normal or abnormal, but when the facility is complicated, it is difficult to identify the cause of the abnormality in a short time.

  The present invention has been made in order to solve the above-described problems. A drive facility that drives a belt with a plurality of pulleys is used to determine whether the operation state of the facility is normal / abnormal, and the cause of the abnormality quickly. It is an object of the present invention to provide a highly reliable equipment diagnosis apparatus that performs specific diagnosis.

The facility diagnosis apparatus according to the present invention collects waveform data such as vibration and sound by providing sensors at predetermined locations of a facility that drives a belt with a plurality of pulleys, and diagnoses the operating state of the facility using principal component analysis. Do. A time-series data collecting means for collecting the waveform data in time series, including a time point when the equipment is normally operated; a multi-dimensional data generating means for generating multi-dimensional data from each time-series data; A pattern that performs principal component analysis on the multidimensional data in each operation state of the equipment including the state, generates a feature space for each operation state including the normal state, and stores pattern information indicating each feature space Information generation and storage means;
Distance calculation means for calculating the distance to each feature space from the newly collected and generated new multidimensional data based on the pattern information of each feature space, and the new multidimensional data based on the calculated distance. Determination means for determining an operation state in the dimension data.

  In such an equipment diagnosis apparatus, a feature space for each operation state including a normal state is generated by principal component analysis, each pattern information is stored, and new multidimensional data indicating the operation state at the time is acquired. In order to calculate the distance to each feature space, that is, how close to each operation state indicated by each stored pattern, it is possible to determine what the abnormal state is at the same time as determining whether it is normal. Can be identified quickly.

Embodiment 1 FIG.
A facility diagnosis apparatus according to Embodiment 1 of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a schematic configuration of a facility diagnosis apparatus according to Embodiment 1 of the present invention.
The equipment diagnosis apparatus 1 collects vibration data obtained from the acceleration sensor 3 provided at a predetermined location of the equipment 2 and diagnoses the operation state of the equipment 2. As shown in the figure, the facility diagnostic apparatus 1 generates time-series data collection unit 4 that collects vibration data obtained from the acceleration sensor 3 as time-series data in time series, and generates multidimensional data from each time-series data. A multi-dimensional data generation unit 5; a pattern generation unit 6 that performs principal component analysis on multi-dimensional data in each operation state of the facility 2 to generate a feature space for each operation state; A pattern storage unit 7 for storing information. In addition, a distance calculation unit 8 that calculates the distance from the new multidimensional data to each of the feature spaces, a determination unit 9 that determines a driving state based on the calculated distance, and an input / output unit 10 are provided.
The pattern generation unit 6 and the pattern storage unit 7 constitute a pattern information generation storage unit, and the input / output unit 10 constitutes an output unit and an input unit together.

The equipment 2 diagnosed by the equipment diagnostic apparatus 1 configured as described above is shown in FIG. As shown in FIG. 2, it is the installation 2 which drives the conveyor belt 12 with the some pulley 11, and transfers the sludge 13a, 13b, and is used for the treatment facility etc. of a sewage. A predetermined amount of sludge 13a is dropped onto the conveyor belt 12 from above and transferred to a predetermined processing position.
In order to diagnose such an operation state of the facility 2, a sensor 3 is installed at a predetermined location of the facility 2, and waveform data such as sound, vibration, and temperature are collected from the facility 2 during operation. In this case, vibration data is collected using the acceleration sensor 3.

Next, operation | movement of each part of the equipment diagnostic apparatus 1 is demonstrated below.
The time series data collection unit 4 collects vibration data detected by the acceleration sensor 3 as time series data in time series every predetermined unit time.
The multidimensional data generation unit 5 generates multidimensional data from each time series data. For example, as shown in FIG. 3, time-series data as one waveform data is converted into a frequency spectrum, and P-dimensional data X = [x ₁ , x ₂ ,. _P ] is generated.
The pattern generation unit 6 performs principal component analysis on a large number of multidimensional data generated from time-series data collected under a predetermined condition, and determines a feature space formed by these multidimensional data. That is, the average value, rotation matrix, and eigenvalue of the multidimensional data under this condition are calculated, and these are stored in the pattern storage unit 7 as pattern information indicating the feature space under this condition.

The principal component analysis in the pattern generation unit 6 will be described.
Assume that there are N samples for P-dimensional data X = [x ₁ , x ₂ ,..., X _P ] with P variables. An average value of each variable is calculated to calculate an average value M = [m ₁ , m ₂ ,..., M _P ] of the multidimensional data, and deviation data Y = [y ₁ , y ₂ from the average value is calculated. ,..., Y _P ] (= X−M) are prepared, and a feature space centered on M is determined.
Principal component analysis is independent of the information held by P variables y _p (p = 1, 2,..., P) given as a linear combination of y _p while minimizing information loss. This is a technique of expressing using Q (Q ≦ P) principal components (overall indices) z _q (q = 1, 2,..., Q). In this case, assuming that the number of principal components is P at the maximum, and R is a rotation matrix for conversion to the constituent elements Z = [z ₁ , z ₂ ,..., Z _P ] of the feature space, Z = R × Y. That is, the deviation data Y is axis-converted using the rotation matrix R. Here, the first variances of the principal components z ₁ and Z is the largest of the dispersion with the any linear expression of y _p, variance of the q main ingredient z _q is first to (q-1) It is the largest of the variances of the linear expression uncorrelated with all of the principal components.
The average value M of the multidimensional data, the rotation matrix R, and the eigenvalues λ = [λ ₁ , λ ₂ ,..., Λ _P ] obtained by performing the principal component analysis in this way, The pattern storage unit 7 holds the pattern information indicating the feature space based on the N pieces of P-dimensional data X.

  In the distance calculation unit 8, when new multidimensional data generated by the multidimensional data generation unit 5 is acquired, the acquired new multidimensional data and each of the acquired multidimensional data and each of the pieces of pattern information stored in the pattern storage unit 7 are obtained. The distance from the pattern (feature space) is calculated. Here, the Mahalanobis distance considering the variance is used as the distance, and the distance becomes smaller as the variance becomes larger. It is assumed that the pattern storage unit 7 holds pattern information from the pattern I to the pattern X including at least one normal state pattern, and the distance calculation unit 8 calculates the Mahalanobis distance from the pattern α, for example, Calculate as follows.

The pattern information of pattern α
Average value of pattern α: M _α = [m _α1 , m _α2 ,..., M _αP ]
Rotation matrix of pattern α: R _α
Eigenvalue: λ = [λ _α1 , λ _α2 ,..., Λ _αP ]
The new multidimensional data X = [x ₁ , x ₂ ,... X _P ] is converted as follows.
Z _α = R _α × (X−M _α ) = [z _α1 , z _α2 ,..., Z _αP ]
Further, each element of Z _α is multiplied by (λ _α1 ) ^(−1/2) , (λ _α2 ) ^(−1/2) ,..., (Λ _αP ) ^(−1/2) to obtain the following: Compress in the axial direction.
D _α = [z _α1 · (λ _α1 ) ^(−1/2) , z _α2 · (λ _α2 ) ^(−1/2) ,..., Z _αP · (λ _αP ) ^(−1/2) ] = [D _α1 , d _α2 ,..., D _αP ]
Calculating the distance D from the point to the origin of the P-dimensional space shown the P-dimensional data D _alpha is, the distance D is the Mahalanobis distance from the pattern alpha new multidimensional data X.

As described above, the Mahalanobis distance from each pattern of the new multidimensional data X is calculated for all the patterns from the pattern I to the pattern X held in the pattern storage unit 7.
Assuming that the feature space described above is generated by compressing so that the variance is made uniform by multiplying the axial direction of each principal component by the (−1/2) power of the eigenvalue, it is converted from the acquired new multidimensional data X The obtained P-dimensional data _Dα is data obtained by converting the new multidimensional data X into coordinates on the feature space indicated by the pattern α.

  Based on the Mahalanobis distance from each pattern of the new multidimensional data X, which is the calculation result of the distance calculation unit 8, the determination unit 9 determines how close the driving state indicated by each pattern is. FIG. 4 shows a case where the Mahalanobis distance from each pattern of the new multidimensional data X is calculated for the feature space of the pattern I and the feature space of the pattern II, and the degree of proximity from each pattern is determined. 14a is new multidimensional data shown on the feature space of the pattern I, 14b is new multidimensional data shown on the feature space of the pattern II, and the new multidimensional data is based on the Mahalanobis distances 15a and 15b from each pattern. It can be seen that X is relatively far from pattern I and relatively close to pattern II.

  In the input / output unit 10, the determination result from the determination unit 9 can be displayed as shown in FIG. This is a graphical representation of the reciprocal of the Mahalanobis distance from each pattern I-X of the new multidimensional data, representing the proximity to each pattern, and it can be seen that the acquired new multidimensional data is closest to the pattern II. .

Next, the flow of diagnosing the facility 2 by the facility diagnostic apparatus 1 will be described below.
The facility diagnosis apparatus 1 has a diagnosis mode for performing facility diagnosis and an initial learning mode for learning a normal state prior to the facility diagnosis, and operates by switching between the two modes.
In order to carry out equipment diagnosis of the equipment diagnostic apparatus 1, it is necessary to store at least pattern information in a normal state in the pattern storage unit 7, and when the equipment diagnostic apparatus 1 is first installed in the equipment 2, etc. A certain period prior to facility diagnosis is set as a normal data collection period, and the normal state is learned in the initial learning mode. In this normal data collection period, the equipment 2 is operated while being kept in a predetermined normal state.

  That is, in the initial learning mode, time-series data in which the equipment 2 is in a normal state is collected by the time-series data collection unit 4, and multidimensional data generation unit 5 generates multidimensional data. Then, for the multidimensional data collected and generated in the normal data collection period, the pattern generation unit 6 performs the principal component analysis as described above, determines the feature space formed by these multidimensional data, The average value, rotation matrix, and eigenvalue of these multidimensional data are calculated and stored in the pattern storage unit 7 as pattern information in the initial normal state.

Next, the facility diagnosis apparatus 1 is switched to the diagnosis mode to perform facility diagnosis. Note that the multi-dimensional data collected by the time-series data collection unit 4 in time series every predetermined unit time and generated by the multi-dimensional data generation unit 5 is retained for a certain period.
In the diagnosis mode, the Mahalanobis distance to each feature space corresponding to each pattern information stored in the pattern storage unit 7 is calculated by the distance calculation unit 8 as described above with respect to the newly acquired new multidimensional data. To do. Then, the degree of proximity to each feature space is determined by the determination unit 8 and displayed by the input / output unit 10. When only the pattern information in the initial normal state, for example, the pattern I is stored in the pattern storage unit 7, the result of whether or not the initial normal state is continued is obtained by how close to the initial normal state. When a plurality of pieces of pattern information are stored in the pattern storage unit 7, the driving state indicated by the new multidimensional data can be determined from the degree of proximity to each feature space.

  In the diagnostic mode, the determination unit 8 has a function of determining whether the new multidimensional data is an unknown case that is separated from any feature space by a predetermined Mahalanobis distance that is greater than or equal to a predetermined value. Then, the input / output unit 10 notifies the operator by outputting an alarm or the like. The operator checks this state and confirms whether it is a new normal state due to a change in operation mode such as maintenance of the equipment 2 or sludge transport amount, or an abnormal state due to a failure or the like. In this case, the pattern generation unit 6 generates a new feature space by performing principal component analysis on the multidimensional data collected in a predetermined period including the new multidimensional data determined to be an unknown case. Then, the pattern information is stored in the pattern storage unit 7. In addition, an identification symbol for identifying the operation state, such as a normal type and an abnormal type, can be input from the input / output unit 10, and when the operator inputs the identification symbol, newly stored pattern information is the above-described identification symbol. Add and store.

The input / output unit 10 displays the determination result from the determination unit 9. As shown in FIG. 5, the reciprocal of the Mahalanobis distance from each pattern I to X of the new multidimensional data is displayed in a graph to visually represent the proximity to each pattern. In this case, I to X are displayed as identification symbols of the pattern information.
The alarm output from the input / output unit 10 is not only when the new multidimensional data is an unknown case, but also when the degree of proximity from the pattern (feature space) indicating an abnormal state is equal to or greater than a predetermined value. An alarm is output as the abnormal state indicated by. In this case, whether or not a pattern having a degree close to a predetermined value is an abnormal pattern can be determined by the added identification symbol. For example, if an identification symbol is set and used with a pattern starting from I as an abnormal state and a pattern starting from A as a normal state, it is easy to understand even when displayed.
In the case of an abnormal state, the diagnosis is continued after the operator performs necessary measures on the equipment 2.

In the facility diagnosis apparatus 1 according to this embodiment, the principal component analysis is performed on a large number of multidimensional data generated from time-series data collected under a predetermined condition, and these multidimensional data constitute the feature. A space is determined and pattern information is stored. And, to determine the driving state by calculating the distance from the feature space indicated by each retained pattern for the new multidimensional data, determine how close the new multidimensional data is to which pattern, and the driving state In addition to normality or abnormality, diagnosis of the cause of abnormality can be performed in a short time. In addition, since multi-dimensional data is generated from time-series data by converting each time-series data into frequency intensity for each band, multi-dimensional data can be easily generated from one waveform data. Component analysis can be performed.
Further, since the Mahalanobis distance taking into account statistical dispersion is used as the distance used for the determination, it is possible to determine to what pattern the driving state is close to how much, and to improve the reliability of diagnosis.

  In addition, since an initial normal state pattern is generated and stored in an initial learning mode that learns a normal state prior to equipment diagnosis, and then the diagnosis mode can be started to start diagnosis. The initial normal state that is the basis of diagnosis can be learned efficiently without depending on the above, and the diagnosis can be promptly shifted to after diagnosis. Furthermore, in the diagnosis mode, when new multidimensional data is more than a predetermined distance from any pattern, a new feature space is generated as an unknown case and pattern information is stored, so a new driving state is always learned. The pattern information that is the basis of diagnosis can be easily updated, and pattern information can be stored for various abnormal states and multiple normal states. Is further improved.

  When the new multidimensional data is an unknown case, the pattern information is calculated by performing principal component analysis on the multidimensional data collected including the new multidimensional data in a predetermined period before the new multidimensional data. Since it memorize | stores, a new feature space can be produced | generated quickly and reliably with respect to an unknown case, and pattern information can be memorize | stored. Further, in the case of an unknown case, an alarm is output to notify the operator that the abnormality is different from the stored operation state, and a necessary treatment can be promoted. In addition, since an identification symbol to be added to newly generated pattern information can be input, information for identifying an abnormal type and a normal type can be easily added to the pattern information and used. Furthermore, the determination result by diagnosis, that is, how close the new multidimensional data is to which pattern is displayed by adding the abnormal type and normal type added to the pattern information, so that the diagnosis and determination result can be displayed accurately. .

  Note that when generating pattern information in the initial normal state, the distribution of multidimensional data collected during the normal data collection period may be nonlinearly spread due to maintenance adjustments or the like. For this reason, it is possible to generate a pattern showing multiple initial normal states by calculating a distance representing the similarity between each data during the normal data collection period and performing cluster analysis that combines similar data together It is also good. As a result, excessive spread of the feature space indicating the initial normal state is prevented, and the reliability of the equipment diagnosis is improved.

Embodiment 2. FIG.
In the first embodiment, vibration data is collected using the acceleration sensor 3, but a plurality of sensors may be used. For example, when accelerometers are installed at two locations of equipment 2 and data is collected synchronously, changes in the transfer function of vibration between them can be known, so the amount of information when an abnormality occurs increases and diagnosis is performed. Reliability is improved. In this case, multi-dimensional data that is one vector data is generated from a plurality of time-series data from the plurality of sensors, but the data collected from the plurality of sensors is converted into the frequency intensity for each band, respectively. Multidimensional data is generated by expressing it as one vector data. Alternatively, a transfer function between data from a plurality of sensors may be obtained and used for multidimensional data. Further, instead of converting the time series data into the frequency spectrum, a cross spectrum may be used.
The plurality of sensors may be different types of sensors such as an acceleration sensor and a voice microphone.

  In addition, multi-dimensional data can be generated by calculating multiple dimensional and non-dimensional parameters (effective value, kurtosis, etc.) for each frequency from time series data, or basic statistics such as execution value, variance and moment of time series data Multi-dimensional data may be generated by calculating the quantity.

It is a block diagram which shows schematic structure of the equipment diagnostic apparatus by Embodiment 1 of this invention. It is a figure which shows the installation by Embodiment 1 of this invention. It is a figure explaining the production | generation of the multidimensional data in Embodiment 1 of this invention. It is a figure explaining the nearness from each pattern of the novel multidimensional data in Embodiment 1 of this invention. It is a figure which shows the output display of the determination result in Embodiment 1 of this invention.

Explanation of symbols

1 equipment diagnostic equipment, 2 equipment, 3 sensors, 4 time-series data collection unit,
5 multi-dimensional data generation unit, 6 pattern generation unit, 7 pattern turn storage unit,
8 distance calculation unit, 9 determination unit, 10 input / output unit, 11 pulley, 12 conveyor belt,
14a, 14b New multidimensional data, 15a, 15b Mahalanobis distance.

Claims (9)

In a facility diagnostic apparatus that collects waveform data such as vibration and sound by installing sensors at predetermined locations of a facility that drives a belt with a plurality of pulleys, and diagnoses the operating state of the facility using principal component analysis,
Time series data collection means for collecting the waveform data in time series, including the time when the equipment is normally operated;
Multidimensional data generating means for generating multidimensional data from each of the time series data,
Principal component analysis is performed on the multidimensional data in each operation state of the equipment including the normal state, a feature space for each operation state including the normal state is generated, and pattern information indicating each feature space is stored. Pattern information generation and storage means;
Distance calculating means for calculating a distance to each feature space based on the pattern information of each feature space from newly collected and generated multidimensional data;
A facility diagnosis apparatus comprising: a determination unit that determines an operation state in the new multidimensional data based on the calculated distance. The equipment diagnosis apparatus according to claim 1, wherein the distance to each feature space calculated by the distance calculation means is a Mahalanobis distance in consideration of dispersion of each feature space. A diagnostic mode for performing equipment diagnosis, and an initial learning mode for learning a normal state prior to the equipment diagnosis,
In the initial learning mode, time-series data in which the equipment is in a normal state is collected by the time-series data collection unit, multi-dimensional data is generated in the multi-dimensional data generation unit, and the multi-dimensional data in the normal state On the other hand, the pattern information generation storage means performs principal component analysis to generate and store pattern information in an initial normal state,
In the diagnostic mode, the new multidimensional data is collected, and the distance calculation means calculates the distance to each feature space corresponding to each pattern information stored in the pattern information generation storage means, and the determination The degree of proximity to each feature space is determined by means, and it is determined whether or not the case is an unknown case that is separated from any feature space by a predetermined distance or more. The equipment diagnosis apparatus according to claim 1 or 2, wherein the information generation storage means generates a new feature space by principal component analysis based on the new multidimensional data and stores the pattern information. When the determination unit determines that the new multidimensional data is the unknown case, the pattern information generation and storage unit includes the new multidimensional data and the multidimensional data collected in a predetermined period before the new multidimensional data. 4. The equipment diagnosis apparatus according to claim 3, wherein the principal component analysis is performed on the equipment. An output means for informing when the new multidimensional data is determined as the unknown case by the determination means; and an input means for inputting the abnormal type or normal type of the unknown case from the outside, and generating the pattern information 5. The equipment diagnosis apparatus according to claim 3, wherein the abnormality type or the normal type is added to pattern information generated and stored by the storage unit based on the new multidimensional data. The output means displays the degree of proximity of the new multidimensional data to the feature spaces together with normal types and abnormal types that are added to the stored pattern information and identify the feature spaces. The facility diagnosis apparatus according to claim 5. 7. In the initial learning mode, when the pattern information generation storage unit generates and stores pattern information in an initial normal state, a plurality of pattern information can be generated by cluster analysis. The facility diagnostic apparatus according to any one of the above. The equipment diagnosis apparatus according to claim 1, wherein the multidimensional data generation unit generates the multidimensional data by converting the time series data into a frequency intensity for each band. The multidimensional data generation means generates the multidimensional data which is one vector data based on a plurality of time-series data obtained from a plurality of the sensors. The equipment diagnostic apparatus as described.

Images