Patent Application
20150276557
This application is based upon and claims the benefit of priority of the prior Japanese Patent Application No. 2014-069166, filed on Mar. 28, 2014, the entire contents of which are incorporated herein by reference.
The present invention relates to a state monitoring system, a state monitoring method and a state monitoring program.
A variety of state monitoring technologies are applied to plants and other various types of equipment. Further, over the recent years, with developments of a sensing technology and an information processing technology, high-level state monitoring for predicting abnormality has been conducted by making use of a tremendous quantity of data of sensors (refer to, e.g., Patent documents 1-2).
[Patent document 1] International Publication WO2013/030984
[Patent document 2] International Publication WO2013/111397
The state monitoring technology includes a technology for informing of abnormality when a sensor output deviates from, e.g., a default range and a technology for informing of the abnormality when an indication value of the sensor output varies. However, for example, in such equipment that a variety of manipulations to fluctuate the sensor output are performed during a processing step as by process control of a chemical plant, normally the sensor output momentarily varies during the processing step. Hence, it is not an easy task to detect a sign of the abnormality from the data of the sensor in the equipment such as this.
Under such circumstances, it is an object of the present application to provide a state monitoring system, a state monitoring method and a state monitoring program, which are capable of detecting a sign of abnormality of such equipment that a variety of manipulations to fluctuate a sensor output are performed during a processing step.
In order to solve the problem described above, the present invention is devised to include a normal model based on data of a sensor output of equipment when normally finishing processing a raw material and to monitor, upon an input of time-series evaluation data of the sensor output on the occasion of finishing processing the specified raw material through a series of manipulations according to a default sequence, a state of the processing equipment on the basis of a comparison between the inputted evaluation data and the normal model.
Specifically, the present invention is a state monitoring system including: a storage unit to be stored with normal models obtained by analyzing, per series of manipulations, time-series learning data of sensor outputs indicated by respective units of processing equipment when normally finishing processing a raw material through the series of manipulations according to a default sequence; and a processing unit to diagnose a state of the processing equipment on the occasion of processing a specified raw material, upon an input of time-series evaluation data of the sensor output indicated by each of the units of the processing equipment on the occasion of finishing processing the specified raw material through the series of manipulations, on the basis of a comparison between the inputted evaluation data and the normal model.
The state monitoring system described above is configured such that the storage unit is stored with a model based on the data given when performing the series of manipulations according to a default sequence in order to monitor the equipment in which to perform a variety of manipulations to fluctuate the sensor output during the processing step as by process control of, e.g., a chemical plant. This model is obtained in a way that handles, as learning data, the data given when normally finishing processing the raw material through the series of manipulations and analyzes the learning data. A reason why the data given when normally finishing processing the raw material are handled as the learning, lies in that a probability of the processing equipment being normal is considered high if the processing of the raw material is normally finished. The state monitoring system handles the model obtained by analyzing the learning data as a normal model.
In the state monitoring system, the storage unit is stored with such a normal model, and, when the time-series data of the sensor outputs indicated by the respective units of the processing equipment are inputted as the evaluation data, the state of the processing equipment is diagnosed based on the comparison between the inputted evaluation data and the normal model. Further, the normal model is based on the sensor outputs indicated by the respective units of the processing equipment in the series of manipulations being performed for the raw material when normally finishing processing the raw material. It is therefore feasible to detect a sign of the abnormality of the processing equipment even in such a type of processing equipment that the variety of manipulations to fluctuate the sensor output are performed during the processing step and the sensor output exhibits a complicated behavior. The normal model is obtained based on the data given when normally finishing processing the raw material, resulting therefore in such determination that the state of the processing equipment when collecting the evaluation data is normal if the evaluation data fall within a range of the normal model.
The storage unit may be stored with, per type of product, the normal model obtained by analyzing the learning data per type of the product generated by processing the raw material, and the processing unit may diagnose, per type of the product, the state of the processing equipment, upon the input of the evaluation data, on the basis of the comparison between the inputted evaluation data and the normal model.
Normally, when a type of the product is different, contents of the series of manipulations being performed for the raw material in order to produce the product and the sensor outputs indicated by the respective units of the processing equipment, become different per type of the product. In this point, as described above, the normal model is prepared per type of the product, and the diagnosis based on the comparison between the evaluation data and the normal model is conducted per type of the product, in which case the sign of the abnormality of the processing equipment can be detected with high accuracy even in such a configuration that the processing equipment produces the variety of products.
Further, the storage unit may be stored with the normal model obtained by analyzing, per system, the time-series learning data of the sensor output indicated by each of the units of the processing equipment including a plurality of systems to process the raw material in parallel, and the processing unit may diagnose, per system, the state of the processing equipment, upon the input of the evaluation data, on the basis of the comparison between the inputted evaluation data and the normal model.
For example, in the processing equipment including the plurality of same systems in parallel, the sensor outputs indicated by the respective units of the processing equipment are different per system as the case may be irrespective of whether mutually the same type of products are produced in the systems or whether mutually different types of products are produced in the respective systems. In this point, as described above, the normal model is prepared per system of the processing equipment, and the diagnosis based on the comparison between the evaluation data and the normal model is conducted per system of the processing equipment, in which case the sign of the abnormality of the processing equipment can be detected with the high accuracy even when the data differs between the systems of the processing equipment.
Still further, the storage unit may be stored with the normal model obtained by analyzing the time-series learning data of the sensor output indicated by each of the units of a specified system of the processing equipment including a plurality of systems to process the raw material in parallel, and the processing unit may diagnose the state of the processing equipment, upon the input of the evaluation data of each system, on the basis of a comparison between the inputted evaluation data of each system and the normal model obtained from the learning data of the specified system.
In the processing equipment including the plurality of same systems in parallel, the sensor outputs indicated by the respective units of the processing equipment become approximately the same between the systems as the case may be. Hence, as described above, the normal model is prepared based on the data of the specified system of the processing equipment, and the diagnosis based on the comparison between the evaluation data and the normal model is made with respect to the data of each system of the processing equipment, in which case it is feasible to detect the sign of the abnormality not only in the system with the normal model being generated but also in the system with the normal model not being generated in the processing equipment.
The processing unit may diagnose, based on an anomaly measure, the state of the processing equipment when processing the specified raw material by calculating the anomaly measure, upon the input of the evaluation data, on the basis of a comparison between the inputted evaluation data and the normal model.
Herein, the “anomaly measure” is an index that represents a degree of the abnormality of the processing equipment, the degree being calculated from the learning data and the evaluation data. The “anomaly measure” is also an index being incremented and decremented in proportion to, e.g., a degree of divergence of the evaluation data from the normal model generated based on the learning data. When the abnormality of the processing equipment is diagnosed based on the index described above, the state of the processing equipment can be quantitatively grasped, and hence the index is effective in detecting the sign of the abnormality.
It is to be noted that the present invention can be grasped also in terms of an aspect of a method. For example, the present invention is a state monitoring method including: storing normal models obtained by analyzing, per series of manipulations, time-series learning data of sensor outputs indicated by respective units of processing equipment when normally finishing processing a raw material through the series of manipulations according to a default sequence; and diagnosing a state of the processing equipment on the occasion of processing a specified raw material, upon an input of time-series evaluation data of the sensor output indicated by each of the units of the processing equipment on the occasion of finishing processing the specified raw material through the series of manipulations, on the basis of a comparison between the inputted evaluation data and the normal model.
Yet further, the present invention can be grasped also in terms of an aspect of a program. For example, the present invention is a state monitoring program to make a computer execute: a process of storing normal models obtained by analyzing, per series of manipulations, time-series learning data of sensor outputs indicated by respective units of processing equipment when normally finishing processing a raw material through the series of manipulations according to a default sequence; and a process of diagnosing a state of the processing equipment on the occasion of processing a specified raw material, upon an input of time-series evaluation data of the sensor output indicated by each of the units of the processing equipment on the occasion of finishing processing the specified raw material through the series of manipulations, on the basis of a comparison between the inputted evaluation data and the normal model.
The state monitoring system, the state monitoring method and the state monitoring program are capable of detecting the sign of the abnormality of the equipment in which to perform the variety of manipulations to fluctuate the sensor output during the processing step.
An embodiment of the invention of the present application will hereinafter be described. The embodiment, which will be illustrated as below, is one aspect of the invention of the present application but does not limit the technical scope of the invention of the present application.
<Architecture of State Monitoring System>
The data accepting unit 11 is a function unit to handle a process of accepting a data input of plant equipment 101. Further, the data pre-processing unit 12 is a function unit to handle pre-processing that is performed on the data accepted by the data accepting unit 11. Further, the data segmenting unit 13 is a function unit to handle a process of segmenting the data accepted by the data accepting unit 11. Still further, the data selecting unit 14 is a function unit to handle a process of selecting the data being segmented on a batch-by-batch basis by the data segmenting unit 13. Moreover, the normal model generating unit 15 is a function unit to handle a process of generating a normal model from learning data accepted by the data accepting unit 11. Furthermore, the anomaly measure calculating unit 16 is a function unit to handle a process of calculating a degree of abnormality (which will hereinafter be referred to as an “anomaly measure”) of the plant equipment 101 from the learning data accepted by the data accepting unit 11 and from evaluation data. Additionally, the abnormality distinguishing unit 17 is a function unit to handle a process of distinguishing whether or not the abnormality exists in the evaluation data accepted by the data accepting unit 11.
<One Example of Equipment Being Monitorable by State Monitoring System>
The controller to perform feedback control is frequently used in the plant equipment 101. The controller to perform the feedback control adjusts, when the set value (SV) is changed, the manipulate value (MV) so that the process value (PV) becomes coincident with the set value (SV).
<Content of Process Executed by State Monitoring System>
A content of a state monitoring process executed by the state monitoring system 1 will hereinafter be described by exemplifying a case of monitoring the state based on the data of the plant equipment 101 as described above.
When the operator of the state monitoring system 1 requests the data accepting unit 11 to generate the normal model, the data accepting unit 11 executes a process of accepting an input of the data of the sensor signal etc of the plant equipment 101 (S101). The data accepting unit 11 stores inputted learning data in the storage 5.
After inputting the learning data, the data pre-processing unit 12 pre-processes the learning data into a format suited to generating the normal model (S102).
For example, the data of the plant equipment 101 contain a sensor signal with a relatively small dispersion, a monotonously increasing sensor signal and an invalid sensor signal not exhibiting a valid value as the case may be. Such being the case, the data pre-processing unit 12 executes, e.g., a process of removing the sensor signal with the relatively small dispersion, the monotonously increasing sensor signal and the invalid sensor signal (this process is a process of fetching feature signals and will hereinafter be therefore termed a “feature selection”).
A feature selection method is exemplified by a method of performing, e.g., a correlation analysis about the data of the multidimensional time-series sensor signals, determining that the plurality of signals exhibiting a relatively high similarity because of a correlation value being approximate to “1” are redundant signal, and deleting the redundant signals from the plurality of signals while setting non-redundant signals to remain. The “feature extraction” may involve, though it is considered to use the sensor signal as it is, extracting a feature representing a time variation of the data from a feature vector given by (window width (3, 5, . . . ))×(sensor count) in away that provides a window of “±1, ±2, . . . ” against a certain point of timing. Further, decomposition into a frequency component may also be attained by applying frequency demultiplexing such as DWT (Discrete Wavelet Transform).
Note that each feature may preferably involve canonicalization to perform transformation so that an average becomes “0” and a dispersion becomes “1” by use of the average and a standard deviation. The average and the standard deviation of each feature are stored so that the same transformation can be done when making an evaluation. Alternatively, normalization may also be performed by use of a maximum value and a minimum value or an upper limit value and a lower limit value, which are preset. These processes serve to simultaneously handle the sensor signals being different in terms of a unit and a scale.
The feature transformation involve using multiple algorithms such as a PCA (Principal Component Analysis) algorithm, an ICA (Independent Component Analysis) algorithm, an NMF (Non-negative Matrix Factorization) algorithm, a PLS (Projection to Latent Structure) algorithm and a CCA (Canonical Correlation Analysis) algorithm. However, any one of these algorithms may be used, or combinations thereof may also be used, or the transformation may not be performed. The principal component analysis algorithm, the independent component analysis algorithm and the non-negative matrix factorization algorithm do not require setting of response values and are therefore easy to be utilized. Parameters such as a transformation matrix necessary for the transformation are stored beforehand so that the same transformation as when generating the normal model is conducted when making the evaluation.
Moreover, the data pre-processing unit 12, if the acquired multidimensional time-series signals have deficits, such items of data are deleted. For example, when a majority of sensor signals are “0”, all of the signal data at the corresponding time are deleted.
After pre-processing the learning data, the data segmenting unit 13 executes the process of segmenting on a batch-by-batch basis the learning data stored in the storage 5 (S103). The data segmenting unit 13 segments the learning data, being continuous in time-series, of the plant equipment 101, into the data per series of manipulations according to a default sequence implemented when carrying out the processing for the raw material.
A variety of methods can be applied to the data segmenting process executed by the data segmenting unit 13. For instance, when a series of manipulations according to the default sequence are started and if the item of data indicating the variation first and the item of data indicating the variation lastly are previously known in terms of design, the data segmenting unit 13 may segment the data at a point of date/time when these items of data begin to vary and a point of date/time when the variations are converged. The date/time when these items of data begin to vary and the date/time when the variations are converged can be specified based on, e.g., a magnitude of a variation ratio of the data.
Further, the data segmenting unit 13 may, when the operator of the state monitoring system 1 designates an arbitrary point from within the time base of the graph displayed on the display device 6 of the state monitoring system 1, segment the data at the point of date/time designated by the operator. Furthermore, for instance, when the data of the plant equipment 101 contain starting date/time information and finishing date/time information of the batch process, the data segmenting unit 13 may segment the data at a point of the starting date/time and at a point of the finishing date/time. Moreover, e.g., when the data of the plant equipment 101 contain information of contents of the manipulations of the respective appliances of the plant equipment 101, the data segmenting unit 13 may segment the data at a point of manipulation date/time of the appliance being manipulated first and a point of manipulation date/time of the appliance being manipulated lastly on the occasion of performing the series of manipulations according to the default sequence.
After segmenting the data, the data selecting unit 14 executes a process of selecting the respective pieces of data segmented by the data segmenting unit 13 (S104). The data selecting unit 14 selects the respective pieces of data segmented by the data segmenting unit 13 on the basis of whether the quality of the product produced by the batch processes corresponding to the respective items of data is acceptable or not.
A variety of methods can be applied to the data selecting process executed by the data selecting unit 14. For example, when the data related to whether the quality of the product is acceptable or not takes such a format as to be readable by the computer 7 that attains the state monitoring system 1, the data selecting unit 14 selects the data with the product's quality being determined to be “acceptable” from within the respective pieces of data segmented by the data segmenting unit 13. Further, when the data related to whether the quality of the product is acceptable or not takes such a format as to be unreadable by the computer 7 that attains the state monitoring system 1, the data selecting unit 14 may accept a manipulation for selecting the data with the product's quality being determined to be “acceptable” from within the respective pieces of data segmented by the data segmenting unit 13 by displaying, e.g., a screen for accepting the data selecting manipulation by the operator of the state monitoring system 1.
After selecting the data, the normal model generating unit 15 executes a normal model generating process described as below. To be specific, the normal model generating unit 15 performs learning by using the data obtaining by removing one-batch data from the data being determined to be “acceptable” through the selection, thereby generating the normal model (S105). Next, the normal model generating unit 15 calculates an anomaly measure in a way that employs the generated normal model by inputting the one-batch data removed in step S105 (S106). The normal model generating unit 15 checks whether the calculations of the anomaly measures with respect to the entire batchwise learning data are finished or not (S107). If not yet finished, the normal model generating unit 15 repeats step (S105) of generating the normal model and step (S106) of calculating the anomaly measure with respect to other batchwise learning data about which the anomaly measure is not yet calculated (S108). The normal model generating unit 15, when finishing the calculations of the anomaly measures with respect to the entire batchwise learning data (S107), sets a threshold value for distinguishing the abnormality based on the calculated anomaly measure (S109). The normal model generating unit 15 finally generates the normal model by use of all of the learning data (S110).
An in-depth description of respective steps from S105 to step S110 will hereinafter be made.
C=(Q−X)T(Q−X) [Mathematical Expression 1]
After obtaining the correlation matrix C, the point “b” is calculated in the following mathematical expression.
In this method, an affine subspace cannot be created unless inputting the evaluation data, and hence the normal model generating unit 15 builds up a “kd” tree for efficiently searching for the point of the normal model in the vicinity of the data “q” in step S105 and step S110. The “kd” tree is defined as a space-division data structure to classify points existing in a k-dimensional Euclidean space. In step S106, the point “b” of the normal model in the vicinity of the data “q” is obtained by making use of the “kd” tree, and a distance between the point “b” and the data “q” is calculated and is set as the anomaly measure.
Next, in step S109, a threshold value is set based on the anomaly measure. The data used for the learning are structured of the normal data being obtained through the selection in step S104, and therefore a maximum value of the anomaly measure is set as the threshold value.
Through the procedure described above, the normal model generating process is completed. The normal model generating unit 15 saves the generated normal model in the storage 5. After completing generating the normal model, the state monitoring system 1 executes the following evaluation process.
The state monitoring system 1 has completed the process of generating the normal model, the data accepting unit 11 executes a process of accepting an input of evaluation target data (S201). The evaluation data, of which the input is accepted by the data accepting unit 11, are, e.g., data to be newly provided to the state monitoring system 1 from the plant equipment 101 after completing generating the normal model. The evaluation data may be inputted off-line by use of the storage medium stored with an aggregation of the evaluation data for one or plural batches, and may also be inputted on-line by use of the communication line that interconnects the plant equipment 101 and the state monitoring system 1. The data accepting unit 11 stores the inputted evaluation data in the storage 5.
After inputting the evaluation data, the data pre-processing unit 12 executes the pre-processing to convert the evaluation data into a format suited to the evaluation based on the normal model (S202). The data pre-processing unit 12 performs the pre-processing such as the feature selection and the feature conversion in the same way as in step S102 described above.
After pre-processing the evaluation data, the data segmenting unit 13 executes a process of segmenting the evaluation data stored in the storage 5 on the batch-by-batch basis (S203). The data segmenting unit 13 segments the evaluation data, being continuous in time-series, of the plant equipment 101 into the data per series of manipulations according to the default sequence implemented when processing the raw material.
After segmenting the evaluation data, the anomaly measure calculating unit 16 calculates the anomaly measure from the evaluation data (S204). The anomaly measure calculating unit 16 calculates the anomaly measure by the same method as in step S106 on the basis of the normal model generated by the normal model generating unit 15 in step S110.
After calculating the anomaly measure, the abnormality distinguishing unit 17 distinguishes whether the abnormality exists in the learning data or not (S205). The abnormality distinguishing unit 17 compares the anomaly measure with the threshold value set by the normal model generating unit 15 in step S109 and, if equal to or larger than the threshold value, detects that the abnormality exists in the learning data. The abnormality distinguishing unit 17 checks whether the calculations of the anomaly measures and the determinations about the abnormality with respect to the entire batches of evaluation data are finished or not (S206). If not yet finished, the abnormality distinguishing unit 17 repeats step (S204) of calculating the anomaly measure and step (S206) of determining as to the abnormality with respect to the batches of evaluation data about which the calculation of the anomaly measure and the abnormality determination are not yet conducted (S207). The abnormality distinguishing unit 17, when finishing calculating the anomaly measures and making the abnormality determinations about the entire batches of evaluation data (S206), finishes evaluating the evaluation data. Through the procedure described above, even in plant equipment 101 where the variety of manipulations to fluctuate the sensor outputs are performed during the processing step, a sigh of the abnormality can be detected.
Note that the abnormality distinguishing unit 17 may display a result of evaluating the evaluation data as follows.
Moreover, the abnormality distinguishing unit 17 may also display the evaluation result of the evaluation data in the way of being associated with graphs of the evaluation data and the anomaly measure.
Further, the abnormality distinguishing unit 17 may also present items of data having a large degree of influence on the abnormality with respect to the evaluation result of the evaluation data.
A first modified example of the state monitoring system 1 according to the embodiment will hereinafter be described.
Then, when the single plant equipment 101 produces the multiple types of products, the process of segmenting the data of the plant equipment 101 in step S103 described above may also be executed per series of manipulations according to the default sequence being determined per type of the product. In this case, in the data selecting process in step S104 described above, the process of selecting the data in accordance with whether the quality of the product is acceptable or not may be executed per type of the product. Moreover, a series of processes from step S105 to step S110 described above may be executed a plural number of times per type of the product, and a plurality of normal models may also be generated per type of the product.
Further, when the plurality of normal models is generated per type of the product, the process of segmenting the evaluation data of the plant equipment 101 in step S203 described above may be executed per series of manipulations according to the default sequence being determined per type of the product. In this case, the series of processes from step S204 to step S207 described above may be executed a plural number of times per type of the product, and the abnormality determination of the evaluation data may also be made by use of the normal models corresponding to the types of the products.
According to the first modified example, the plurality of normal models is generated per type of the product, and the determination as to the abnormality of the evaluation data can be made according to the type of the product. Hence, even when the single plant equipment 101 generates the variety of products such as the product X and the product Y, according to the first modified example, the abnormality determination using the normal model prepared per type of the product enables the sign of the abnormality to be detected with high accuracy.
A second modified example of the state monitoring system 1 according to the embodiment will hereinafter be described.
When all the same type of products are generated in the A-system through the C-system of the plant equipment 101′, it follows that the manipulation procedures and the sensor outputs are substantially common among the respective systems. Such being the case, when all the same type of products are generated in the plurality of systems, a series of normal model generating processes from step S101 through S110 described above may be executed with respect to all sets of system data from the A-system through the C-system and may also be executed with respect to only the system data of any one of the systems. Furthermore, a series of evaluation processes from step S201 through S207 may be executed to evaluate the evaluation data of all of the A-system through the C-system by use of the normal model generated per system and may also be executed to evaluate the evaluation data by use of the normal model generated bases on the system data of any one of the systems.
Note that when the plant equipment 101′ generates the mutually different types of products in the respective systems, it follows that the manipulation procedures and the sensor outputs are differentiated in the respective systems. Then, when generating the mutually different types of products in the respective systems, the series of processes from step S101 through S110 described above are executed with respect to all of the system data of the A-system through the C-system. Subsequently, the series of evaluation processes from step S201 through S207 are executed to evaluate the evaluation data of the whole systems from the A-system through the C-system by use of the normal model generated per system data of each system.
By the way, as illustrated in
When the abnormal sensor signals derived from the fault in the utility 108 are recorded in the evaluation data of the plant equipment 101′, in the series of evaluation processes from step S201 through S207, it follows that an increase of the anomaly measure is observed in the evaluation data of all of the systems from the A-system through the C-system in which to generate the products when the fault occurs in the utility 108. The increase of the anomaly measure is observed in both of the case that the respective systems generating so far the products when the fault occurs in the utility 108 generate mutually the same type of products and the case that the respective systems generate the mutually different types of products.
Thus, the plant equipment 101′ equipped with the plurality of same systems in parallel and further the system being common among the respective systems, evaluates the evaluation data of the respective systems by using the normal model generated based on the data per system or the data of any one of the systems, and is thereby enabled to observe the abnormality of the system such as the utility 108 being common among the respective systems.
By the way, in the case of desiring to monitor an inside of a whole site including a multiplicity of various plant equipment such as the plant equipment 101′ illustrated in
<<Readable-by-Computer Recording Medium>>
A program for making a computer, other machines and devices (which will hereinafter be referred to as the computer etc) realize any one of the functions can be recorded on a recording medium readable by the computer etc. Then, the computer etc is made to read and execute the program on this recording medium, whereby the function thereof can be provided.
Herein, the recording medium readable by the computer etc connotes a recording medium capable of storing information such as data and programs electrically, magnetically, optically, mechanically or by chemical action, which can be read from the computer etc. Among these recording mediums, for example, a flexible disc, a magneto-optic disc, a CD-ROM, a CD-R/W, a DVD, a Blu-ray Disc, a DAT, an 8 mm tape, a memory card such as a flash memory, etc are given as those removable from the computer. Further, a hard disc, a ROM (Read-Only Memory), etc are given as the recording mediums fixed within the computer etc.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2014-069166 | Mar 2014 | JP | national |
