JP2006169661A - Analyzer and mechanical equipment mounted therewith and analysis program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect appearance of new abnormal mode(s) in mechanical equipment or the like to be analyzed. <P>SOLUTION: An SOM(self-organizing map) is made to learn using abnormal tension waveform in a draw false-twister to label abnormality source(s) on SOM units. The SOM is made to learn again with combining abnormal data obtained thereafter and the original abnormal data to label the SOM with the original abnormal data, and the cluster of unlabeled units is detected. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to the analysis of the state of textile machinery such as drawing false twisting machine, spinning winder, spinning machine, rewinding machine, other mechanical equipment, power plant, building, etc. It relates to detecting the appearance of different types of abnormal modes.

  Inventors propose to categorize data such as yarn tension waveforms from sensors using a neural network such as a self-organizing map (SOM) in Patent Document 1 and analyze the cause of abnormalities in textile machinery. did. For neural networks, learning vector quantization (LVQ) and the like are preferable in addition to SOM, all of which use a set of codebook vectors, modify the value of the codebook vector using the learning data, and codebook Learning is performed by assigning the analysis target state to the vector as a label. When actually analyzing abnormal data, the codebook vector close to the abnormal data is extracted, and the label attached to the codebook vector or the label attached to the adjacent codebook vector is the cause of the abnormal data (abnormal Mode).

  Here, it is desirable that all abnormal data that may occur in actual use can be included in the initial set of learning data, but this is not possible. Therefore, it is sufficient that at least representative abnormality data is acquired in advance and included in the learning data for all abnormality modes that actually cause problems, but only after the machine has been installed and started operation. There is abnormal data that cannot be obtained. For example, this is data generated due to a change in the machine over time, a machine installation environment, a difference in operation state, or the like.

A new abnormal mode may appear after installation of machinery and equipment. This is important for machine maintenance. This means that the machine state, environment, and operating conditions have changed in some way from before, and it is necessary to clarify the cause of the abnormal mode. Note that Patent Document 1 discloses that relearning is performed using data obtained after the start of operation of the analysis apparatus, but does not consider detecting the appearance of a new abnormal mode.
JP 2004-183139 A

  An object of the present invention is to be able to detect that a new abnormal mode has appeared in a machine facility or the like to be analyzed.

  The analysis apparatus according to the present invention learns to correct values of a plurality of codebook vectors based on an initial learning data set including abnormal data, and attaches the analysis target state to the codebook vectors as a label. In the apparatus, the plurality of codebook vectors are relearned by a second set of learning data different from the set of learning data, and the known codebook vector is moved to the codebook vector close to the known state. Means for attaching a state as a label and means for extracting a codebook vector not labeled at the time of relearning are provided.

  Although extraction of clusters can be performed manually, it is preferable to provide means for detecting clusters of codebook vectors that have not been labeled during relearning. When an unlabeled codebook vector is expressed, it is preferable that a new category of abnormal mode is generated, particularly when it is detected that such a cluster of codebook vectors is expressed.

  The machine equipment according to the present invention includes an analysis device according to the present invention, a sensor for collecting data of the machine equipment including abnormal data, and means for storing the collected data as candidates for the second set of learning data. It is equipped with.

The analysis program according to the present invention includes an instruction for learning while correcting values of a plurality of codebook vectors based on a set of initial learning data including abnormal data, and a state to be analyzed in the codebook vector as a label. A codebook that causes the plurality of codebook vectors to be relearned by a second set of learning data different from the set of learning data and that is in proximity to data whose state is known The vector is provided with an instruction for attaching the known state as a label and an instruction for extracting a codebook vector that is not labeled at the time of relearning.
When an unlabeled codebook vector is expressed, it is preferable that a new category of abnormal mode is generated, particularly when it is detected that such a cluster of codebook vectors is expressed.

  In the present invention, a set of codebook vectors is relearned by the second set of learning data, and a codebook vector that is not labeled is extracted. When a new abnormal mode occurs, the codebook vector corresponding to this mode is not labeled. Therefore, when a codebook vector that is not labeled is extracted, the appearance of a new abnormal mode can be detected.

  If the codebook vectors that are not labeled form a cluster, there is a high probability that a new abnormal mode has occurred. If a means for detecting a cluster is provided, the appearance of an abnormal mode can be detected automatically and reliably.

  When the analyzer according to the present invention is incorporated in a mechanical equipment such as a textile machine together with a sensor for collecting data and a storage means for collected data, the state of the mechanical equipment is detected by detecting a cluster of codebook vectors without labels. It is easy to perform optimization, maintenance, maintenance, etc. of the operating conditions of mechanical equipment.

  In the following, an optimum embodiment for carrying out the present invention will be shown.

  The analysis apparatus and the program of the embodiment will be described by taking the analysis of the abnormal tension waveform of the yarn on the untwisting side of the drawing false twisting machine as an example. In the drawing false twisting machine 1 shown in FIG. 1, reference numeral 2 denotes a yarn supply package for supplying a yarn to be falsely twisted. A yarn 12 from the yarn supply package 2 is fed to a first heater H1 via a first feed roller FR1. And cooled by the cooling plate C1. The yarn supply package is sometimes called POY. Reference numeral 4 denotes a twister. As shown in FIG. 2, the yarn 12 passes between a pair of belts 14 and 16, for example, downward from the top of the figure and is twisted by the belts 14 and 16. Further, the yarn 12 is stretched between the first feed roller FR1 and the second feed roller FR2. The twist given by the twister 4 propagates to the first feed roller FR1 and is fixed via the heater H1 and the cooling plate C1.

  A tension sensor 6 is provided on the downstream side (untwisting side) of the twister 4 because the untwisting tension of the yarn is greatly related to the quality and properties of the processed yarn, and a tension signal of the yarn (in this case, the yarn tension) An untwisting tension signal (T) is input to the control unit 8. The control unit 8 applies feedback control to the twister 4 and the like based on the yarn tension waveform to control the yarn tension at the exit side of the twister 4 to be constant. Further, on the downstream side of the tension sensor 6, a second feed roller FR2 is provided to untwist up to the second feed roller FR2, and then a second heater H2 and a third feed roller FR3 are provided to give relaxation heat treatment to the yarn. In this way, the filament yarn having a smooth surface is given bulkiness and stretchability by the simultaneous simultaneous false twisting action. In this specification, upstream / downstream is determined along the traveling direction of the yarn, the side close to the yarn supply package 2 is the upstream side, and the side close to the final drawn false twisted yarn package 10 is the downstream side. The drawn false twisted yarn is wound up as a winding package 10. WD is a take-up drum that rotates in contact with the take-up package 10.

  FIG. 3 shows the configuration of the tension sensor. Reference numeral 18 denotes a rotating body, which rotates around a fulcrum 20, 21 to 23 are rollers, and 24 is a spring that pulls the rotating body 18 in a clockwise direction, for example. When the tension of the yarn is applied to the roller 22, for example, a counterclockwise moment W is applied to the rotating body 18, and the rotating body 18 rotates to a position where the moment W balances with the force F from the spring 24. Therefore, the tension of the yarn 12 can be monitored by monitoring the rotation angle of the rotating body 18 and its differential value. The type of yarn tension sensor is arbitrary.

  FIG. 4 shows an analysis apparatus. Abnormal data is extracted by the control unit 8 from the tension signal T of the tension sensor 6, input to the wavelet transform unit 30, and converted into a feature vector. The feature vector is input to the SOM 32, and a unit (SOM node) having a reference vector closest to the input feature vector is extracted and displayed on the monitor 40 or the printer 41. In this display, the unit with the closest reference vector is marked, and all or at least some of the units are labeled in the SOM, so the cause of the abnormality can be identified from the label of the marked unit and its surrounding units. Can be estimated. In SOM, the distance between the reference vectors of each unit is displayed by color coding, light and dark, etc., so if the marked unit is not labeled, determine which unit of the surrounding unit is the closest label it can.

  SOM initial learning is performed, for example, before or when the analysis apparatus is installed, and the learning data used for this may be referred to as a set of initial learning data. It is difficult to accurately analyze all abnormal data experienced after the installation of the analysis device by learning with a set of initial learning data. For example, it is difficult to cope with a change in abnormal data from the initial learning data due to an abnormal mode that appears after installation, a change over time of the drawing false twisting machine 1, an installation environment, an operating state, or the like. Therefore, the waveform of the abnormal data and the feature vector obtained by the wavelet transform are set as a set, stored in the storage unit 34, and accumulated as subsequent learning data.

  In re-learning, a set of abnormal data for re-learning is added to the set of post-learning data by adding a set of abnormal data whose labels are known and used in previous learning, such as data used for initial learning. obtain. If re-learning is performed only with data not used for initial learning, SOM 32 is suitable for the classification of recent abnormal data, but tends to be unsuitable for classification of data used for initial learning. For example, when the type of abnormal data fluctuates with a trend that repeats over a long period of time, it is not preferable to learn the SOM 32 so as to correspond only to recent data. It is preferable to re-learn using both the abnormal data used and the short-term data, that is, data not used in the previous learning.

  The labeling unit 36 labels each unit of the re-learned SOM 32. For example, a reference vector once labeled at the time of initial learning or the like is input to the SOM after re-learning. The input vector label is transferred to the unit having the reference vector closest to the input vector. Alternatively, the input vector is not limited to the range of the original reference vector, and the label may be a known vector, and the label is transferred to the closest unit. In this way, in labeling, an input vector whose label is known in some sense is input to the SOM 32, and the label is transferred to the unit having the closest reference vector. When multiple input vectors are close to one unit, even if these labels are written together, or the label with the shortest distance between the reference vector of the unit and the input vector is transferred. good.

  When labeling is performed after re-learning, a unit without a label may be generated. A cluster (unit group) of unlabeled units is detected, and when a cluster is formed, the number of units and the number of units belonging to each cluster are detected by the cluster detection unit 38. When the cluster is detected, the SOM is displayed as shown in FIG. 14, and the reference vector value of the unit belonging to the cluster and the tension waveform for the feature vector closest to the reference vector are output to the monitor 40 and the printer 42. The tension waveform here may be extracted from what is stored in the storage unit 32, for example.

  The occurrence of a new cluster often means the appearance of a new kind of abnormal mode. The cause of the abnormality is estimated by a technique such as a skilled operator checking the tension waveforms corresponding to the units included in the new cluster, and a new label is assigned to those that have been able to estimate the cause of the abnormality. Next, change the operating conditions of the drawing false twisting machine in response to the estimated cause of abnormality, or perform maintenance such as parts replacement, repair, inspection, etc. with reference to the estimated cause of abnormality. Responds to changes in equipment status. The re-learning of the SOM 32 is preferably repeated every time the subsequent learning data is accumulated in the storage unit 34.

FIG. 5 shows a feature vector extraction algorithm by wavelet transform. As an analyzing wavelet, for example,
Ψ (t) = t · ( 2π) -1/2 exp (-t 2/2) (1)
Is used. The analyzing wavelet is a wavelet that can be inversely transformed, and Ψ (t) corresponds to the inverted one of the first derivative of the Gaussian function. Using the analyzing wavelet Ψ (t) in equation (1), the input tension waveform x (t) is subjected to discrete wavelet transformation according to equation (2).
Xj (b) = 2 ^{-3j / 2} ∫Ψ ((tb) ・ 2 ^-j ) x (t) dt (2)
In equation (2), j is an integer, and the integration range is from −∞ to ∞. b is a time parameter, and when the start time of the abnormal waveform signal is, for example, time 0, a signal near time b is subjected to wavelet transform using a first-order derivative of Gaussian as a window function.

  Next, for j = 2, 4, the second to fourth order cumulants, the second order correlation (autocorrelation function) and the third order correlation are obtained. The second order cumulant represents the dispersion, the third order cumulant represents the distortion of the distribution, and the fourth order cumulant represents the sharpness of the distribution. The third-order correlation G (τ, σ) is expressed by <f (t) · f (t + τ) · f (t + σ)> when the signal is f (t). The operation to be obtained is shown, and τ and σ are time differences. These data were used as feature vectors. The type of feature vector is arbitrary, and the above feature vector is obtained by statistically processing data obtained by wavelet transform and converting it into a short vector. The feature vector may be a cumulant of each order such as a tension waveform or a differential value thereof without going through wavelet transform. Furthermore, the feature vector may be subjected to processing such as normalization of each data with a secondary cumulant.

FIG. 6 shows a learning algorithm for the self-organizing map SOM. First, the reference vector of each SOM unit (feature vector assigned to the unit) is initialized to an appropriate value, and the unit of the reference vector most similar to the learning input vector (feature vector) is obtained. Note that the distance between vectors and the distance between units are, for example, Euclidean distance (root of the sum of squares of differences of vector components) and Mahalanobis distance (R is a covariance matrix for a set of reference vectors in the SOM, and its inverse matrix. the ^{R -1,} the difference between the reference vector and the feature vector y, upon the transposed vector and ^{y T,} defined by ^{(y T R -1 y) 1/2} ). The reference vector between the obtained unit and its neighboring units is corrected so as to be close to the input vector. The above process is repeated for all the input vectors, and the reference vector correction amount in one learning is gradually decreased to converge the reference vector. This completes the learning, and thereafter fixes the reference vector of each unit. Next, the feature vector whose label is known is input to the SOM, for example, the feature vector used for learning is input, and the unit having the reference vector closest to it is labeled with the feature vector label (the cause of abnormal data or this). Post symbol).

FIG. 7 shows the SOM relearning algorithm. The feature vector obtained after the initial learning (the label may be known or unknown) and the feature vector used in the previous learning, for example, are combined to relearn the SOM. A feature vector with a known label is then input and the label is transcribed to the nearest reference vector unit. When a unit without a label is generated by re-learning, it is determined whether or not a cluster is formed. If a cluster is formed, the number n and size N are obtained and displayed on a monitor, a printer, etc. When the number or size is large, for example, (nN) ^1/2 is equal to or greater than a predetermined value, and a new abnormal mode is generated, and an alarm is given by a monitor or a printer.

  FIG. 8 shows an analysis program of the embodiment. The SOM initial learning instruction 51 initially learns the SOM, and the SOM labeling instruction 52 assigns a label to the SOM unit during initial learning. In the abnormal data accumulation command 53, abnormal data after the start of operation of the analyzer is stored together with the tension waveform data, the SOM is re-learned by the SOM re-learning command 54, and after the unit having no label is re-learned by the cluster extraction command 55 The cluster is extracted, and the occurrence of a new abnormal mode is warned according to the size and number of clusters.

  FIG. 9 shows the display of the SOM after the initial learning. The black small circle in the figure represents the unit position, and the brightness of the unit position indicates the average distance between the surrounding 6 units. The distance is a distance between reference vectors of units. The hexagonal contrast between the two black circles indicates the distance between these units. The darker the light and dark, the greater the distance, and the dark hexagonal series indicates the boundaries between clusters, and the same labels from 1 to 8 are attached, and the bright areas are clusters corresponding to the same types of abnormal causes. Even in FIG. 9, there is a unit without a label.

  Table 1 shows the contents of nine types of abnormal causes (abnormal modes) 0 to 8 used in the examples. In the examples, tension waveforms with different operating conditions were collected for each cause of abnormality, and a total of 400 waveforms were prepared. Among these, in the initial learning of FIG. 9, 200 tension waveforms were used so as to eliminate the cause of abnormality 0. In the subsequent relearning, each time, 100 are randomly selected from the 200 tension waveforms used in the immediately preceding learning, and 100 are extracted from the remaining 200 tension waveforms, for a total of 200 relearning data. It was. Here, the appearance probability of the abnormality cause 0 was gradually increased. Table 2 shows a breakdown of data used for the SOM learning of FIGS. Further, the labels in FIGS. 10 to 14 are obtained by inputting the reference vector of the SOM in FIG. 9 to the SOM after the relearning and transferring the label to the unit having the shortest distance.

Table 1
Label (number) Cause of error
0 Shifter guide dislodgement 1 Bad bearing guide 2 Displacement of the creel center 3 Scratch of the apron belt 4 Defective yarn disengagement 5 Nearly surging 6 Poor tension before false twist 7 Poor tension after false twist 8 End of raw yarn

Table 2
Cause of abnormality
Fig. 0 1 2 3 4 5 6 7 8
Fig. 10 0 23 26 26 23 27 27 23 25
Fig. 11 5 20 36 30 22 22 24 24 20 21
Fig. 12 16 16 31 24 21 26 29 21 16
Fig. 13 28 15 24 21 19 28 23 22 20
Fig. 14 39 23 22 21 16 22 18 20 19

  In FIG. 13 and FIG. 14, a cluster without a label grows in the upper left, and this indicates the appearance of a new abnormality cause (missing shifter guide) that did not exist in FIG.

  FIG. 15 and FIG. 16 show examples of cluster detection. In FIG. 15, starting from a unit without a label and marked with ◎ or Δ, the distance at the position of the unit on the SOM, or between reference vectors If there is a unit without a label within a predetermined distance on the basis of the distance of For the units connected in this way, those that satisfy the criteria such as large size, small average distance between reference vectors, and large average distance with surrounding labeled units are extracted as clusters. .

  In FIG. 16, starting from a unit without a label (marked with ◎ in the figure), the density of a unit without a label within a predetermined distance is determined by the distance at the position of the unit on the SOM and the distance between reference vectors. A cluster is detected by combining a group of units having a density equal to or higher than a predetermined value as a cluster candidate and combining the cluster candidates sharing a unit without a label.

  In the embodiment, the yarn tension is an analysis target, but the yarn thickness and the yarn temperature in the heaters H1, H2, the cooling plate C1, and the like are also suitable for the analysis target. For example, the temperature of the yarn may be detected as a differential signal with an infrared thermometer or the like, and the thickness of the yarn can be detected optically. These data are important data for yarn quality control. Classification of abnormal data by codebook vectors is performed by SOM, but learning vector quantization or the like may be used. In the embodiment, learning is performed using all abnormal data, but normal data may be added to initial learning or relearning.

  In the embodiment, the abnormality of the textile machine is analyzed by the SOM. However, it may be detected by the SOM only that a new abnormality mode is developed in the analyzed textile machine. For example, normal abnormal data is analyzed by another analysis device, and when abnormal data is accumulated after the start of use of the textile machine, the presence or absence of a new category of abnormal mode is analyzed by the SOM of the embodiment. May be. In this case, daily abnormal data is analyzed by an SVM (Support Vector Machine), a perceptron, a general ANN (Artificial Neural Network), or a discriminant analysis device to elucidate the cause of the abnormality. When a new abnormal mode is detected, a new mode can be detected by adding an SVM, re-learning the perceptron or ANN, or changing the discrimination conditions in discriminant analysis to detect this mode. Like that.

  The SVM is an entity such as a program or apparatus for classifying vector data, and one SVM is provided for each abnormal mode. In the SVM, it is determined that the abnormal data to be detected and other data can be separated, and it is known from the label whether the data is a detection target. For example, a vector representing each data is X, a vector indicating a boundary between the detection target and other data is B, a vector facing the detection target from the boundary is W, and the detection target data and other data are W -Determine the values of W and B so that the value of (X-B) can be sufficiently separated.

  In addition to drawing false twisting machines, other textile machines such as spinning winders, spinning machines, and rewinding machines, or cranes, robots, vehicles, machine tools, etc., plants, and structures other than textile machines. It may be targeted. As analysis data in that case, a tension waveform, a stress waveform, a vibration waveform, an acceleration waveform, a temperature waveform, a pressure waveform, or the like, or data obtained by combining signals from a plurality of sensors may be used. Preferably, the machine equipment consisting of one or more machines is combined with a sensor for collecting abnormal data and a storage unit for accumulating abnormal data, and an analysis device is combined to diagnose the state of the machine equipment. And the appearance of a new abnormal mode may be detected.

In the embodiment, the following effects can be obtained.
(1) The appearance of a new abnormal mode can be detected.
(2) This makes it possible to detect changes in the state of the machine, elucidate the cause of new abnormal modes, determine the degree of change over time, identify places where abnormalities are likely to occur, change operating conditions, replace parts, etc. Can be used as an opportunity to perform maintenance such as maintenance.
(3) The cause of the abnormality can be estimated by analyzing the abnormality data with SOM.
(4) By re-learning, the SOM can be modified to be suitable for analysis of new types of abnormal data.

Block diagram of a drawing false twisting machine that was analyzed for yarn tension in the examples Schematic diagram of the twister used in the drawing false twisting machine shown in FIG. Schematic diagram of the tension sensor used in the drawing false twisting machine shown in FIG. Block diagram of analysis apparatus of embodiment The flowchart which shows the extraction algorithm of the feature vector using the wavelet transform in an Example The flowchart which shows the learning algorithm of the self-organization map (SOM) in an Example The flowchart which shows the relearning algorithm of SOM in an Example Block diagram of the analysis program of the embodiment The figure which shows the example of the long-term memory | storage SOM used in the Example The figure which shows the example of the short-term memory | storage SOM which re-learned SOM using 100 feature vectors used for learning of the long-term memory | storage SOM of FIG. 9, and a new 100 feature vector The figure which shows the example of the short-term memory | storage SOM which re-learned SOM using 100 feature vectors used for the learning of SOM of FIG. 10, and 100 new feature vectors The figure which shows the example of the short-term memory | storage SOM which re-learned SOM using 100 feature vectors used for the learning of SOM of FIG. 11, and 100 new feature vectors The figure which shows the example of the short-term memory | storage SOM which re-learned SOM using 100 feature vectors used for the learning of SOM of FIG. 12, and 100 new feature vectors The figure which shows the example of the short-term memory | storage SOM which re-learned SOM using 100 feature vectors used for the learning of SOM of FIG. 13, and 100 new feature vectors Diagram showing a search algorithm for a cluster of unlabeled units Diagram showing another search algorithm for clusters of unlabeled units

Explanation of symbols

1 Drawing false twisting machine 2 Yarn supply package (POY)
4 Twister 6 Tension sensor 8 Control unit 10 Stretch false twisted yarn package 12 Filament yarns 14 and 16 Belt 18 Rotating body 20 Support points 21 to 23 Roller 24 Spring 30 Wavelet conversion unit 32 Self-organizing map (SOM)
34 Storage Unit 36 Labeling Unit 38 Cluster Detection Unit 40 Monitor 42 Printer 51 SOM Initial Learning Instruction 52 SOM Labeling Instruction 53 Abnormal Data Accumulation Instruction 54 SOM Relearning Instruction 55 Cluster Extraction Instruction
H1, H2 heater
C1 cooling plate
FR1 to FR3 Feed roller
WD take-up drum

Claims (6)

In the analysis apparatus, which learns to correct the values of a plurality of codebook vectors by using a set of initial learning data including abnormal data, and attaches the analysis target state to the codebook vector as a label.
The plurality of codebook vectors are relearned by a second learning data set different from the learning data set, and the known state is attached as a label to a codebook vector close to known state data. Means for
An analysis apparatus comprising: means for extracting a codebook vector that is not labeled during re-learning. 2. The analysis apparatus according to claim 1, further comprising means for detecting a cluster of codebook vectors not labeled during re-learning. 3. The analysis apparatus according to claim 1, wherein the occurrence of a new abnormal mode in the analysis target is detected based on the expression of a codebook vector that has not been labeled at the time of re-learning. The analysis device according to any one of claims 1 to 3, a sensor for collecting data on mechanical equipment including abnormal data, and means for storing the collected data as candidates for the second set of learning data; With mechanical equipment. An instruction for learning while correcting the values of a plurality of codebook vectors based on a set of initial learning data including abnormal data, and an instruction for attaching a state to be analyzed to the codebook vector as a label In the analysis program,
The plurality of codebook vectors are relearned by a second learning data set different from the learning data set, and the known state is attached as a label to a codebook vector close to known state data. Instructions for
An analysis program comprising: an instruction for extracting a codebook vector that is not labeled during re-learning. 6. The analysis program according to claim 5, wherein the occurrence of a new abnormal mode in the analysis target is detected by the expression of a codebook vector that has not been labeled at the time of re-learning.