WO2020194534A1 - 異常判定支援装置 - Google Patents
異常判定支援装置 Download PDFInfo
- Publication number
- WO2020194534A1 WO2020194534A1 PCT/JP2019/012973 JP2019012973W WO2020194534A1 WO 2020194534 A1 WO2020194534 A1 WO 2020194534A1 JP 2019012973 W JP2019012973 W JP 2019012973W WO 2020194534 A1 WO2020194534 A1 WO 2020194534A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- abnormality
- determination
- primary
- target data
- analysis target
- Prior art date
Links
- 230000005856 abnormality Effects 0.000 title claims abstract description 181
- 238000004458 analytical method Methods 0.000 claims abstract description 139
- 238000004519 manufacturing process Methods 0.000 claims abstract description 88
- 238000010801 machine learning Methods 0.000 claims abstract description 69
- 239000000284 extract Substances 0.000 claims abstract description 11
- 230000002159 abnormal effect Effects 0.000 claims description 67
- 238000000034 method Methods 0.000 claims description 34
- 239000000463 material Substances 0.000 claims description 24
- 238000005259 measurement Methods 0.000 claims description 24
- 238000003860 storage Methods 0.000 claims description 13
- 238000013480 data collection Methods 0.000 claims description 10
- 238000013526 transfer learning Methods 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 7
- 238000004364 calculation method Methods 0.000 claims description 5
- 230000008569 process Effects 0.000 claims description 5
- 230000008054 signal transmission Effects 0.000 claims description 5
- 238000013135 deep learning Methods 0.000 claims description 3
- 238000013075 data extraction Methods 0.000 abstract 1
- 238000009826 distribution Methods 0.000 description 48
- 238000005096 rolling process Methods 0.000 description 41
- 239000000047 product Substances 0.000 description 24
- 238000012545 processing Methods 0.000 description 11
- 238000010586 diagram Methods 0.000 description 9
- 230000009471 action Effects 0.000 description 7
- 230000006870 function Effects 0.000 description 6
- 230000001419 dependent effect Effects 0.000 description 5
- 238000007726 management method Methods 0.000 description 4
- 239000002436 steel type Substances 0.000 description 3
- 229910000831 Steel Inorganic materials 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000010438 heat treatment Methods 0.000 description 2
- 230000033001 locomotion Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 239000010959 steel Substances 0.000 description 2
- 230000007704 transition Effects 0.000 description 2
- RYGMFSIKBFXOCR-UHFFFAOYSA-N Copper Chemical compound [Cu] RYGMFSIKBFXOCR-UHFFFAOYSA-N 0.000 description 1
- CWYNVVGOOAEACU-UHFFFAOYSA-N Fe2+ Chemical compound [Fe+2] CWYNVVGOOAEACU-UHFFFAOYSA-N 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- XAGFODPZIPBFFR-UHFFFAOYSA-N aluminium Chemical compound [Al] XAGFODPZIPBFFR-UHFFFAOYSA-N 0.000 description 1
- 229910052782 aluminium Inorganic materials 0.000 description 1
- 230000002547 anomalous effect Effects 0.000 description 1
- 230000001174 ascending effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 229910052802 copper Inorganic materials 0.000 description 1
- 239000010949 copper Substances 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000003745 diagnosis Methods 0.000 description 1
- 238000002405 diagnostic procedure Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 239000012467 final product Substances 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000007257 malfunction Effects 0.000 description 1
- 238000003908 quality control method Methods 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 238000005728 strengthening Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
- G05B23/0267—Fault communication, e.g. human machine interface [HMI]
- G05B23/0272—Presentation of monitored results, e.g. selection of status reports to be displayed; Filtering information to the user
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
- G05B23/0224—Process history based detection method, e.g. whereby history implies the availability of large amounts of data
- G05B23/024—Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
- G05B23/0221—Preprocessing measurements, e.g. data collection rate adjustment; Standardization of measurements; Time series or signal analysis, e.g. frequency analysis or wavelets; Trustworthiness of measurements; Indexes therefor; Measurements using easily measured parameters to estimate parameters difficult to measure; Virtual sensor creation; De-noising; Sensor fusion; Unconventional preprocessing inherently present in specific fault detection methods like PCA-based methods
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
- G05B23/0243—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults model based detection method, e.g. first-principles knowledge model
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B23/00—Testing or monitoring of control systems or parts thereof
- G05B23/02—Electric testing or monitoring
- G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
- G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
- G05B23/0275—Fault isolation and identification, e.g. classify fault; estimate cause or root of failure
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Definitions
- the present invention relates to a device that assists in determining whether the manufacturing equipment itself or the product quality is normal or abnormal in the manufacturing equipment involved in manufacturing the product.
- a rolling mill that manufactures plate materials facilitates processing into automobiles and electric appliances by rolling and thinning lumps of steel materials and non-ferrous materials such as aluminum and copper.
- the rolling mills for example, there are two rough rolling mills and seven finishing rolling mills, and although the detailed specifications are different, such as a large-capacity electric motor for driving the upper and lower rolling rolls and a shaft connecting the rolls and the electric motor, the equipment configuration is different. Often similar.
- Patent Document 1 Patent Document 2
- Patent Document 3 Patent Document 3
- Patent Document 1 describes a method of identifying the presence or absence of an abnormality with two indexes using the amplitude of the current of the rotating machine. According to the method described in Patent Document 1, it is determined whether or not it is abnormal depending on whether or not it exceeds a preset determination standard. However, there is no mention of strengthening judgment through learning such as machine learning.
- Patent Document 2 describes a method of diagnosing an abnormality using the current of a rotating machine. However, as in Patent Document 1, the diagnostic method determines whether or not a certain index exceeds a threshold value. In addition, although there is a description that machine learning can also be applied, there is no mention of a specific method.
- Patent Document 3 a model to be diagnosed is prepared in advance by regression analysis using normal data, and the difference from the model created from the current data is evaluated to determine whether or not the model is abnormal. Is described.
- the method described in Patent Document 3 is also a method of determining whether or not an abnormality is made based on whether or not a preset threshold value is exceeded.
- normality may be determined to be abnormal, or abnormality may be determined to be normal.
- FFT fast Fourier transform
- the number and types of abnormal data are very small compared to the number and types of normal data, and it generally takes a lot of time to collect cases indicating abnormalities.
- the present invention has been made in view of such a problem, and an object of the present invention is to provide an abnormality determination support device that assists in accurately determining whether an abnormality has occurred in a manufacturing facility.
- the abnormality determination support device is an abnormality determination support device that provides a determination material for determining whether or not an abnormality has occurred in a manufacturing facility, and is an analysis target data creation unit, a primary determination unit, and a secondary determination unit.
- the analysis target data creation unit is configured to acquire a time-series signal representing at least one of the state of the manufacturing equipment or the product quality from the data collection device of the manufacturing equipment and extract the analysis target data from the time-series signal.
- the primary determination unit is configured to derive a plurality of primary determination results from common analysis target data by applying a plurality of different analysis methods to the analysis target data extracted by the analysis target data creation unit.
- the secondary determination unit has a machine learning device that has learned the pair of the primary determination result obtained by the primary determination unit, the abnormality determination result as the corresponding correct answer, and the cause of the abnormality as a teacher signal, and has a primary determination. It is configured to input a plurality of primary judgment results obtained from common analysis target data in the unit to the machine learning device, and output the secondary judgment result output from the machine learning device and the cause of the estimated abnormality as judgment materials. ..
- Each process of the analysis target data creation unit, the primary determination unit, and the secondary determination unit may be executed by the computer constituting the abnormality determination support device.
- the abnormality determination support device is composed of a computer including at least one processor and at least one memory for storing at least one program, and when the program read from the memory is executed by the processor, the processor The program may be configured to operate as an analysis target data creation unit, a primary determination unit, and a secondary determination unit.
- the primary determination unit converts the analysis target data into a plurality of numerical indexes by applying a plurality of different analysis methods to the analysis target data, and a plurality of numerical indexes. May be configured to output as a plurality of primary determination results.
- the machine learning device may be configured to input a numerical index obtained by the primary determination unit and to learn using a teacher signal having an actual abnormality determination result and an abnormality cause as a correct answer.
- the secondary judgment unit inputs a plurality of numerical indexes obtained for each analysis method by the primary judgment unit into the machine learning device, and outputs the abnormality judgment result and the estimated abnormality cause output from the machine learning device as judgment materials. It may be configured in.
- the primary determination unit converts the analysis target data into a plurality of numerical indexes by applying a plurality of different analysis methods to the analysis target data, and a plurality of numerical indexes.
- the presence or absence of abnormality and the calculation of the degree of abnormality may be performed based on each of the above, and a plurality of judgment results and the degree of abnormality obtained for each analysis method may be output as a plurality of primary judgment results.
- the machine learning device may be configured to input the determination result and the degree of abnormality obtained by the primary determination unit and to learn using the teacher signal with the actual abnormality determination result and the cause of the abnormality as the correct answer.
- the secondary judgment unit inputs a plurality of judgment results and the degree of abnormality obtained for each analysis method by the primary judgment unit into the machine learning device, and uses the abnormality judgment result and the estimated abnormality cause output from the machine learning device as judgment materials. It may be configured to output.
- the primary determination unit converts the analysis target data into a plurality of numerical indexes by applying a plurality of different analysis methods to the analysis target data, and obtains a plurality of numerical values.
- the index may be configured to be output as a plurality of primary determination results.
- the machine learning device uses the numerical index obtained by the primary judgment unit as an input, and learns using a teacher signal whose correct answer is the presence or absence of an abnormality determined from the numerical index and the degree of abnormality calculated from the numerical index. It may be configured in.
- the secondary judgment unit inputs a plurality of numerical indexes obtained for each analysis method by the primary judgment unit into the machine learning device, and outputs the abnormality judgment result and the estimated abnormality cause output from the machine learning device as judgment materials. It may be configured in.
- the analysis target data creation unit extracts data in two states, a load state and a non-load state of the manufacturing equipment while the manufacturing equipment is in operation, and further extracts low frequency components from the extracted data. It may be configured to calculate the high frequency component excluding the data as the analysis target data.
- the primary determination unit converts the analysis target data into a plurality of numerical indexes by applying a plurality of different analysis methods to the analysis target data in each of the load state and the non-load state, and each of the plurality of numerical indexes.
- the analysis target data creation unit extracts data in two states, a measurement state and a non-measurement state, of the sensor for measuring product quality while the manufacturing equipment is in operation. Further, the high frequency component obtained by removing the low frequency component from the extracted data may be calculated as the analysis target data.
- the primary determination unit converts the analysis target data into a plurality of numerical indexes by applying a plurality of different analysis methods to the analysis target data in the measurement state and the non-measurement state, respectively, and converts the analysis target data into a plurality of numerical indexes.
- the presence or absence of abnormality is determined based on each of the above, and if it is abnormal in the measurement state and normal in the non-measurement state, it is judged as an abnormality in product quality, and if it is abnormal in the measurement state and abnormal in the non-measurement state. It may be configured to determine that there is an abnormality in the sensor system that measures product quality or an abnormality in the signal transmission system.
- the machine learning device is any one of learning by a neural network having one intermediate layer, deep learning by a neural network having a plurality of intermediate layers, and rule-based learning. It may be configured to perform learning by one method.
- the machine learning device targets a dimensionless variable having no physical unit among the variables indicating the primary determination result obtained by the primary determination unit as a learning target, and the dimensionless variable is manufactured by another manufacturer. It may be configured to be applied to transfer learning from an abnormality determination support device of equipment or transfer learning from an abnormality determination support device of other manufacturing equipment.
- the abnormality determination support device may further include a data and result storage unit and a display unit.
- the data and result storage unit records at least one of the analysis target data created by the analysis target data creation unit, the judgment progress and result of the primary judgment unit, and the judgment progress and result of the secondary judgment unit, for example. Configured to save.
- the display unit is at least one of the time-series signal obtained from the data collection device, the analysis target data created by the analysis target data creation unit, the judgment progress and result of the primary judgment unit, and the judgment progress and result of the secondary judgment unit. It is configured to display one visually.
- the abnormality determination support device in addition to the primary determination result obtained by the primary determination unit, the secondary determination result by the secondary determination unit and the estimated cause of abnormality can be obtained.
- the secondary judgment unit the secondary judgment result and the cause of the estimated abnormality are obtained by inputting the plurality of primary judgment results obtained by the primary judgment unit into the machine learning device, so that a highly accurate judgment that does not depend on the analysis method or the threshold value can be performed. It is possible. Therefore, according to the abnormality determination support device according to the present invention, it is possible to assist in accurately determining whether or not an abnormality has occurred in the manufacturing equipment.
- FIG. 1 It is a figure which shows the system example of the manufacturing equipment to which the abnormality determination support apparatus of embodiment of this invention is applied. It is a block diagram which shows the structure of the abnormality determination support apparatus of embodiment of this invention. It is a figure explaining an example of the processing flow of the analysis target data creation part of embodiment of this invention. It is a table explaining the correspondence relationship between the state of the signal of the numerical index and the estimated equipment abnormality part. It is a figure explaining an example of the processing flow of the primary determination part of the Embodiment of this invention. It is a table explaining the example of the past data accumulation table in embodiment of this invention. It is a figure explaining the control chart in embodiment of this invention. It is a figure explaining the probability density distribution in embodiment of this invention.
- FIG. 1 is a diagram showing a system example of manufacturing equipment to which the abnormality determination support device according to the embodiment of the present invention is applied.
- the manufacturing facility 20 to which the abnormality determination support device 2 is applied in the present embodiment is a hot thin sheet rolling line.
- the hot sheet rolling line is a manufacturing facility 20 including various devices such as a heating furnace 21, rough rolling mills 22 and 23, a bar heater 24, a finishing rolling mill 25, a runout table 26, and a winder 27.
- the rolled material 100 heated in the heating furnace 21 is rolled by the rough rolling mills 22 and 23.
- the rolled material 100 rolled by the rough rolling mills 22 and 23 is conveyed to the finishing rolling mill 25 via the bar heater 24.
- the finish rolling mill 25 has seven rolling stands F1 to F7 arranged in series, and rolls the rolled material 100 to a desired plate thickness.
- the rolled material 100 rolled by the finish rolling mill 25 is cooled by the runout table 26 and then wound into a coil by the winder 27.
- a coiled thin plate formed by thinly rolling a rolled material 100 is a final product manufactured by the manufacturing equipment 20.
- thermometer 30 for measuring the temperature on the inlet side of the finish rolling mill 25
- sensor 31 for measuring the plate thickness and the plate width
- thermometer 32 for measuring the temperature on the outlet side of the finish rolling mill 25
- thermometer 33 for measuring the temperature on the inlet side of the winder 27 and the like are arranged.
- the manufacturing facility 20 is provided with a data collection device 1.
- the data collection device 1 includes set values and actual values for each device constituting the manufacturing facility 20, measured values by sensors 30 to 33 arranged in the manufacturing facility 20, and each device.
- Various data such as the amount of operation for proper operation are continuously or intermittently collected and recorded in a recording device 1a such as a hard disk.
- the data collection device 1 may be composed of a single computer or a plurality of computers connected to the network.
- the abnormality determination support device 2 is connected to the data collection device 1 by, for example, a LAN.
- the abnormality determination support device 2 is a device that assists the user in determining an abnormality in the manufacturing equipment 20. More specifically, the abnormality determination support device 2 is a device that provides the user with determination material for determining whether or not an abnormality has occurred in the manufacturing equipment 20, and collects analysis target data used for determining the abnormality of the manufacturing equipment 20. By extracting from the time-series signal recorded in the sampling device 1, analyzing the data, and providing the analysis result to the user, the abnormality determination performed by the user is supported.
- the abnormality determination support device 2 is a computer having at least one memory and at least one processor. Various programs and various data used for abnormality determination are stored in the memory.
- FIG. 2 is a diagram showing the configuration of the abnormality determination support device 2, and the functions of the abnormality determination support device 2 are represented by blocks.
- the abnormality determination support device 2 includes an analysis target data creation unit 3, a primary determination unit 4, a secondary determination unit 5, an information input unit 6, a data and result storage unit 7, and a display unit 8.
- the analysis target data creation unit 3, the primary determination unit 4, and the secondary determination unit 5 are realized by the processor as software by executing the program read from the memory by the processor.
- the information input unit 6, the data and result storage unit 7, and the display unit 8 can be provided separately from the abnormality determination support device 2.
- the information input unit 6 is, for example, a keyboard
- the data and result storage unit 7 is, for example, a recording device such as a hard disk
- the display unit 8 is, for example, a display device.
- the analysis target data creation unit 3 acquires a time-series signal indicating the state of the manufacturing equipment 20 such as vibration, current, and load from the data collection device 1 and a time-series signal indicating the product quality, and these time-series. Data necessary for analysis and determination performed by the primary determination unit 4 is extracted from the signal. However, since information on whether the manufacturing equipment 20 is abnormal or normal cannot be obtained unless the manufacturing equipment 20 is operating, the analysis target data creation unit 3 extracts the data in which the manufacturing equipment 20 is operating from the time series signal.
- the analysis target data creation unit 3 extracts data in two states, a loaded state and a non-loaded state. Further, in the sensors 30 to 33 for measuring product quality, data in two states, a measurement state and a non-measurement state, of the product quality measurement sensors 30 to 33 during the operation of the manufacturing equipment 20 are extracted.
- the analysis target data creation unit 3 transmits all the data extracted from the time series signal to the primary determination unit 4. At that time, the data can be processed so as to be a signal suitable for analysis and determination by the primary determination unit 4. For example, the analysis target data creation unit 3 can calculate the deviation from the high frequency component excluding the low frequency component, that is, the low frequency component from the value of the extracted data itself.
- the load state signal generally has a large value, while the non-load state signal has a small value, and when the two are compared, the magnitude of the load state signal becomes significant, and the equipment and quality in the non-load state. It becomes difficult to extract the state of. In order to compare the signal in the loaded state and the signal in the unloaded state on the same basis, it is preferable to extract the high frequency component.
- the processing for the extracted data in addition to the processing for extracting the high frequency component as described above, for example, a processing for reducing the noise of the data by applying a low-pass filter can be applied.
- FIG. 3 is a diagram illustrating an example of the processing flow of the analysis target data creation unit 3.
- step S101 when the rolling of the rolled material to be analyzed is completed in the manufacturing facility 20, a time series signal including before and after rolling is acquired from the data collecting device 1.
- This time-series signal includes data representing the state of the manufacturing equipment 20 and sensor data representing the product quality.
- step S102 data such as rolling load, rolling torque, electric machine current, and speed of rotating equipment are being rolled (load) for each rolling facility (two rolling mills and seven rolling stands constituting the finishing rolling mill). It is classified into data of (state) and data during non-rolling (non-load state).
- step S103 sensor data such as plate thickness and plate width indicating product quality are classified into measurement state data and non-measurement state data.
- step S104 high frequency components are extracted for each of the original data of steps S102 and S103.
- high-frequency components can be extracted by directly applying a high-pass filter to the original data.
- the high frequency component can be extracted by applying a low-pass filter to the original data and subtracting the output result of the low-pass filter from the original data.
- the high frequency component may be referred to as deviation data
- the original data before extracting the high frequency component may be referred to as absolute value data with respect to the deviation data.
- the original data means direct data collected from manufacturing equipment such as electric current and rolling load of an electric motor.
- the data to be analyzed includes both the original data and the deviation data.
- the original data also includes data converted into deviation data in a sensor or the like.
- the analysis target data creation unit 3 passes deviation data, which is a high-frequency component, to the primary determination unit 4, and also performs primary determination such as absolute value data, which is the original data, and data in which noise is reduced by applying a low-pass filter to the original data. All of the requested data can be passed to the primary determination unit 4.
- the primary judgment unit 4 uses a plurality of different analysis methods for the analysis target data classified into a load state and a non-load state, or a measurement state and a non-measurement state in the analysis target data creation unit 3. Apply. Specifically, the primary determination unit 4 converts the analysis target data into a numerical index suitable for determining the number of analysis methods by applying a plurality of different analysis methods to the common analysis target data. Further, the primary determination unit 4 determines the abnormality of the manufacturing equipment 20 and calculates the degree of abnormality based on each of the plurality of numerical indexes, and also determines the abnormality of the product quality and calculates the degree of abnormality.
- the primary determination unit 4 determines an abnormality based on the characteristics of the numerical index signal.
- the vibration state of the signal of the numerical index can be mentioned as an example.
- the vibration of the signal of the numerical index is large, it can be determined to be abnormal, and when the vibration of the signal of the numerical index is small, it can be determined to be normal. Further, by combining the determination based on the signal characteristics in the load state and the determination based on the signal characteristics in the non-load state, it is possible to estimate the abnormal part of the equipment.
- FIG. 4 is a table explaining the correspondence between the state of the signal of the numerical index and the estimated equipment abnormality location.
- the state of the signal has the patterns of a1, b1 and c1 shown in the table. If the state of the signal is a1, that is, if the vibration is large (that is, abnormal) in the loaded state and the vibration is small (that is, normal) in the unloaded state, it can be determined that the mechanical system is abnormal.
- the equipment can be determined to be normal.
- a2 that is, if the vibration is large (that is, abnormal) in the loaded state and the vibration is small (that is, normal) in the unloaded state
- the quality is abnormal.
- the state of the signal is b2, that is, if the vibration is large in the load state and the vibration is large even in the non-load state, it can be determined that the sensor system is abnormal or the signal transmission system is abnormal.
- the state of the signal is c2, that is, if the vibration is small in the loaded state and the vibration is small even in the unloaded state, the equipment can be determined to be normal.
- the primary determination unit 4 outputs the numerical index obtained by the above processing and / or the abnormality determination result and the degree of abnormality to the secondary determination unit 5 as the primary determination result.
- FIG. 5 is a diagram illustrating an example of the processing flow of the primary determination unit 4.
- the analysis target data is provided by the analysis target data creation unit 3.
- the data to be analyzed includes types of rolling equipment (including loaded and unloaded states), product quality (including measured and non-measured states), and absolute value data or deviation data.
- the primary determination unit 4 sorts the data to be analyzed so as to cover all cases such as the 1st stand, rolling load, load, deviation data, 7th stand, motor current, no load, absolute value data, etc. ..
- step S112 one of a plurality of different analysis methods is selected as the analysis method to be applied to the analysis target data.
- the details of the analysis method will be described later. For example, a method of obtaining the standard deviation of the data to be analyzed is selected, and the method is applied to the first stand, rolling load, load, and deviation data.
- a numerical index is generated from the original data (absolute value data) to be analyzed.
- the standard deviation is calculated from the original data to be analyzed and used as a numerical index.
- the probability density distribution is calculated from the original data to be analyzed, the difference from the normal distribution based on the data is evaluated by the Kullback-Leibler distance, etc., and this is used as a numerical index.
- step S114 the average A of m numerical indexes obtained from the past normal data is compared with the numerical index B based on the newly collected analysis target data. Then, the difference between A and B is calculated, or the theory of Hotelling is further applied to calculate the higher numerical index. The details of how to obtain the numerical index will also be described later.
- step S115 the normality / abnormality of the manufacturing equipment and product quality is determined by the higher numerical index calculated in step S114.
- the numerical index based on Hotelling's theory is applied to the chi-square distribution to obtain the degree of anomaly (possibility of anomaly).
- the degree of abnormality is 99% or more, a red alarm is displayed, if it is 95% or more, a yellow alarm is displayed, and if it is in between, a gradation color from red to yellow is displayed.
- step S116 if all the analysis methods to be applied are covered, the process proceeds to step S118, and if not, the analysis method is changed in step S117.
- step S118 if the analysis target data is covered, the process is terminated, and if not covered, the analysis target data is changed in step S119.
- variables that are dimensionless that is, variables that do not have physical units such as mm and kg, are (6) waveform rate, (7) wave height rate, (8) impact index, and (10) skewness. And (11) sharpness.
- This table has a table for each equipment and product quality item, and also for each absolute value and deviation of the original data, and is classified into steel type classification (TS pieces), plate thickness classification (TT pieces), and plate width classification (T pieces).
- these divisions may be made finer, and unnecessary divisions may be eliminated.
- Each cell of the table has m storage areas.
- the standard deviation calculated from the normal data is stored there as a normal numerical index.
- the numerical indexes for the past m pieces are extracted.
- the standard deviation calculated by the data string including the newly collected data to be analyzed is compared with the numerical index extracted from the cell, and the difference is evaluated. As a result of the evaluation, if it is determined that the newly collected analysis target data is normal, the oldest numerical index of the cell is deleted, and the numerical index calculated from the newly collected analysis target data is newly added to this cell. To do.
- Equation 1 the numerical index of the newly collected data to be analyzed
- H the index based on Hotelling's theory
- m past standard deviations are stored as normal numerical indexes in one cell stored in the table of FIG.
- the mean x_ave and standard deviation ⁇ of the past m numerical indexes (standard deviations) in Equation 2 can be calculated.
- the standard deviation of the newly collected data to be analyzed is calculated as x in Equation 2.
- H is a dimensionless value.
- the value of the chi-square distribution is generally a mathematical table or can be calculated by the following equation 3.
- ⁇ is a gamma function.
- FIG. 7 shows an example of a control chart.
- the control upper limit and the control lower limit are generally set to 3 ⁇ ( ⁇ : standard deviation), and if they exceed them, it is judged to be abnormal.
- ⁇ standard deviation
- the ⁇ of the numerical indexes can be calculated.
- the degree of abnormality is 99.73%.
- the slightly lower control standard is 2.5 ⁇
- the degree of abnormality is 97.5%
- that of 2 ⁇ is 95.4%.
- the above 2 ⁇ 2, 3 ⁇ 3, etc. are dimensionless values.
- Equation 4 is the equation of the control upper limit UCL
- equation 5 is the equation of the control lower limit LCL
- ⁇ 1 in the equation 6 is the skewness.
- the past m numerical indexes accumulated in the cells of the same steel type, plate thickness, and plate width as the newly collected data to be analyzed are taken out, and the average value is calculated. .. Calculate the difference between the average value of the past m numerical indicators and the newly collected data to be analyzed, and if the difference is, for example, 3 times the standard deviation, a yellow alarm, 4 times, a red alarm, etc. It is also possible to do. However, it may be necessary to make trial and error in the field to determine how many times to increase. It should be noted that the above three times, four times, 3, 4 and the like are dimensionless values.
- the probability density distribution represents the probability that the data x will be a certain value in the range when it changes in a certain range, and when all the probabilities in the range are added, it becomes 1 (100%).
- FIG. 8 shows an example of the probability density distribution of normal data and an example of the probability density distribution of data including abnormal data.
- the graph (a) of FIG. 8 exemplifies the probability density distribution of only normal data
- the graph (b) of FIG. 8 exemplifies the probability density distribution of data including abnormal data.
- the probability density distribution shown in the graph (b) has a larger spread on the horizontal axis than the probability density distribution shown in the graph (a), but the degree of deviation from the normal distribution is also large.
- the magnitude of the spread on the horizontal axis also appears in the magnitude of the standard deviation shown in the explanation of the above statistics, and therefore appears in the numerical index called the standard deviation.
- the degree of deviation from the normal distribution is considered.
- Equation 7 is an equation of the Kullback-Leibler distance (Kullback-Leibler Divergence) D KL
- Formula 8 is a formula for the error square sum D SQ
- Equation 9 in calculations of error absolute value sum D ABS is there.
- P A (x) is actual probability density taking the original data x
- P N (x) is a normal distribution.
- the target data x is not absolute value data but deviation data. Since the deviation data has a strong high frequency component, it can be regarded as almost noise. Generally, the noise is mostly white noise, and its distribution is normally distributed. However, if the original data contains a noise signal due to some abnormality, the deviation data is likely to have a distribution different from the normal distribution, and an attempt is made to detect it.
- the Kullback-Leibler distance D KL is used as a numerical index
- a table similar to the table shown in FIG. 6 is prepared, and the D KL calculated for normal data is used. Is stored. When new data comes in there, the D KL is calculated and compared with the past normal m D KLs to determine normal / abnormal.
- the Hotelling theory and the control chart determination method described above can be used. The same applies when the error squared sum D SQ and the error absolute value sum D ABS are used.
- the numerical indexes that can be used in the second example of the analysis method are not limited to these numerical indexes D KL , DSQ , and D ABS . Further, as described as yet another determination method in the first example of the analysis method, the degree of abnormality can be manually set and determined.
- Third Example of Analysis Method As a third example of the analysis method, a method of calculating the probability density distribution for each of the maximum value and the minimum value of the deviation data and using the difference from the Rayleigh distribution as a numerical index will be described.
- the distribution of the data at the normal time is not a normal distribution but a Rayleigh distribution as shown in FIG.
- the calculation of the numerical index and the normal / abnormal determination method are the same as described above.
- the above probability density value is a dimensionless value.
- the regression model expresses the relationship between the dependent variable and the independent variable in the form of a first-order polynomial, for example, and the dependent variable and the independent variable may be different variables.
- the dependent variable is the rolling load
- a regression model can be created with the independent variables as the deformation resistance, rolling speed, and material temperature.
- the dependent variable and the independent variable have the same data type, but the dependent variable is the current value, and the independent variable uses the past value. For example, it corresponds to identifying the rolling load by its own value in the past (rolling load).
- the autoregressive model is represented by, for example, Equation 10 below.
- ⁇ is white noise
- ⁇ K-1 is an autoregressive coefficient.
- the value of the autoregressive coefficient is a dimensionless value.
- FIG. 10 shows an example of changes in the autoregressive coefficient of the autoregressive model.
- the horizontal axis 0 means the value of the constant term ⁇ 0
- the horizontal axis k (k is a natural number) means the coefficient ⁇ k of the value k before.
- the vertical axis is the coefficient value.
- the line shown in FIG. 10 also includes the result of identification by the abnormal data. If the time to be identified is constant (fixed value of 12 in this case) and normal data is targeted, the autoregressive coefficient is likely to continue to be almost constant, but some in FIG. The line behaves differently than the other lines, which is due to anomalous data. Therefore, when the value of the coefficient identified by the autoregressive model is different from the value of the coefficient identified by the past normal data, it can be determined that something is wrong.
- a numerical index based on the past value of the normal original data is calculated, and a new numerical index based on the newly added data is calculated based on the numerical index.
- it is possible to determine normal / abnormal by calculating a numerical index based on the data obtained from the similar facility and comparing it with the numerical index based on the data obtained from the target facility. it can.
- the axial direction of the rolled material represents the transition of time. This axial comparison is a method of comparing with the past m values described above.
- comparison can also be made in the axial direction of the equipment. If the numerical index shows different behavior from other equipment, it can be judged as abnormal.
- the selection of the analysis method, the calculation method of the numerical index, and the judgment of normal / abnormal are the same as those described above.
- the primary determination unit 4 calculates a numerical index for each manufacturing facility and each product quality for each analysis method, determines normality / abnormality based on the numerical index, and calculates the degree of abnormality. Will be done.
- FIG. 12 is a diagram showing a first example of a machine learning device included in the secondary determination unit 5.
- the numerical index of the primary determination unit is the input 121 to the machine learning device 122
- the normal or abnormal determination result (secondary determination result) and the estimated abnormality cause are the output 123.
- the input 121 and the resulting output 123 are given to the machine learning device 122 as a pair of teacher signals.
- the output 123 is used as the determination result.
- the primary determination unit 4 there are a plurality of types of analysis methods, and there are also a plurality of numerical indexes calculated from them. Therefore, even if the same target data is used, the judgment of normality / abnormality of equipment and quality may differ depending on the case. This means that the events that we are good at may differ depending on the analysis method, and the judgment result may differ depending on the boundary between normal and abnormal, that is, how to set the threshold value of whether it is abnormal or not. is there.
- the learning function in the secondary determination unit 5 uses the input 121 as a plurality of numerical indexes output by the primary determination unit 4 at the learning stage.
- the machine learning device 122 has a causal relationship such as learning by a neural network having one intermediate layer, deep learning by a neural network having a plurality of intermediate layers, or event A occurring by cause C with a probability of B. It has a rule-based learning method that describes.
- the output 123 indicates the determination of normality or abnormality and the estimation result of the cause of the abnormality.
- FIG. 13 shows an example of a learning mode in the machine learning device 122.
- the information input unit 6 lists hierarchical cause candidates such as the location of the equipment as shown in the table of FIG. 13, the abnormality cause-1 indicating the outline cause, and the abnormality cause-2 indicating the detailed cause.
- the number of layers is not limited to two. It also has an editing function so that a person skilled in the art can newly input (13 or later in FIG. 13) or correct the cause.
- the input by the information input unit 6 shall be performed by a person skilled in the art related to the target manufacturing equipment such as a rolling mill, that is, an operator or an engineer who has sufficient knowledge about the target manufacturing equipment.
- a person skilled in the art can use the input numerical index while referring to the table of FIG. 13 as an abnormality cause-1 or a further abnormality cause.
- the pair of the input numerical index and the abnormality cause-1 or the abnormality cause-2 is used as a teacher signal for learning the machine learning device 122.
- FIG. 14 shows how the data of a certain target manufacturing facility changes in a management diagram, and the relationship with the artificial action is shown in the diagram. It is assumed that some kind of target manufacturing equipment is approached at t1 and the data begins to move toward outliers in the positive direction. It is assumed that the primary determination unit 4 indicates that the numerical index has exceeded the management upper limit three times, and a person skilled in the art has made some action at the time of t2. Furthermore, it is assumed that some kind of action is taken even at the time of t3. In this case, it is highly probable that the action performed at t1 was the cause of the abnormality, and it is considered that the action performed at t2 was a measure to eliminate the abnormality. The work done at t3 has not had much effect.
- the input 121 when the machine learning device 122 learns is the equipment and quality when the numerical index exceeds the management upper limit three times, and the value of the numerical index at that time, and the teacher signal works in FIG.
- the operation diary may be a paper or an electronic operation diary. In the case of an electronic operation diary, these actions can be incorporated into the system relatively easily. In a paper operation diary, it is necessary to convert the description into electronic information.
- FIG. 15 is a diagram showing a second example of the machine learning device included in the secondary determination unit 5.
- the input 131 from the primary determination unit 4 is not a numerical index but a result of determining normality or abnormality and the degree of abnormality.
- the input is different between the first example and the second example, but the others are the same.
- the input 131 from the primary determination unit 4 is not a numerical index but a result of determining normality / abnormality
- the threshold value used for the normal / abnormal determination in the primary determination unit 4 is changed, the normal / abnormal determination is made. Results can vary significantly. If that happens, you will have to start over from the beginning.
- the weight for the degree of abnormality is increased according to the normal / abnormal determination result of the primary determination unit 4 for learning.
- Equation 2 which is an index of Hotelling theory is used
- the value of the chi-square distribution corresponds to the degree of anomaly.
- H (x) 3.0
- the value of the chi-square distribution is 0.051, so the probability of being normal is 0.051, that is, the probability of being abnormal is 0.949, and 0.949.
- the degree of abnormality is not affected by the change of the threshold value set by the primary determination unit 4.
- the number and types of abnormal data are very small compared to the number and types of normal data, and it generally takes a lot of time to collect cases showing abnormalities. is there. That is, in order for the machine learning device 122 of the first example shown in FIG. 12 and the machine learning device 132 of the second example shown in FIG. 15 to learn, the frequency of obtaining a teacher signal in which an input including an abnormal state and a correct answer are paired. Is small, and it takes time to have sufficient learning ability.
- the machine learning device 137 of the third example shown in FIG. 16 is used.
- a plurality of numerical indexes output by the primary determination unit 4 are input, and the normal / abnormal determination of the manufacturing equipment and product quality output from the primary determination unit 4 is performed.
- a teacher signal is used with the result and the degree of abnormality as the correct answer.
- a plurality of numerical indexes by the primary determination unit 4 are input 136, and the normal / abnormal determination result and the degree of abnormality of the manufacturing equipment and the product quality are extracted as the output 138.
- the machine learning device 137 learned at the stage where there are few abnormal cases mainly inputs normal data the feature of determining normality is larger than that of determining abnormalities. If abnormal data comes in there, it is judged to be abnormal because it is different from normal. As the number of abnormal cases increases, machine learning becomes more sophisticated by inserting and learning the input / output relationships as described with reference to FIGS. 12 and 15 in the teacher signal.
- Transfer learning of machine learning equipment can be applied as a countermeasure when there are few abnormal data.
- transfer learning is a method used when sufficient learning data is not yet accumulated in machine learning, and the result of learning at another place or at another opportunity (machine learning, for example, the connection weight inside the neural network) is applied.
- machine learning for example, the connection weight inside the neural network
- This is a method that can be used for machine learning.
- the original data string X represented by the formula 1 is obtained in the manufacturing facility A, the characteristics of the manufacturing facility A are strongly reflected in the original data string X.
- the manufacturing facility A is a factory that mass-produces more than the manufacturing facility B and rolls more hard steel grades, the rated capacity of the electric motor of the manufacturing facility A is usually larger than that of the manufacturing facility B.
- the motor current and the like are all obtained as large values, and when machine learning is performed using them directly, even if the learning results can be used in the manufacturing facility A, they are often not suitable for the manufacturing facility B. ..
- the normalization method is to calculate the average and standard deviation of the original data and correct the input and output so that the average is 0 and the standard deviation is 1, or to find the maximum and minimum values of the original data and set the range.
- the data of the manufacturing facility B with few abnormal data may not be able to sufficiently express the characteristics of the manufacturing facility B due to the normalization because the distribution range of the original data is narrow.
- the original data obtained from the manufacturing equipment is not used as it is for learning, but the dimensionless variables are used for learning.
- the numerical index is calculated using the past normal data.
- dimensionless variables such as waveform rate, crest factor, impact index, skewness, sharpness, probability density distribution and normal distribution difference evaluated by equations 7 to 9, and basic statistics. Since the values obtained by calculating the difference between all or the probability density distribution and the normal distribution using the index of Hoteling theory are dimensionless, learning is performed using these. Then, the result learned in the manufacturing facility A can be directly diverted to the manufacturing facility B or another manufacturing facility without being corrected, so that the transfer learning can be easily performed. That is, a small amount of abnormal data can be effectively and easily used in each manufacturing facility.
- the data and result storage unit 7 stores the analysis target data created by the analysis target data creation unit 3, the judgment progress and result of the primary judgment unit 4, and the judgment progress and result of the secondary judgment unit 5 in a storage device such as a hard disk. save. This is so that the reason and grounds for the judgment can be extracted later.
- the display unit 8 is a time-series signal obtained from the manufacturing equipment, a time-series signal indicating product quality, analysis target data created by the analysis target data creation unit 3, determination progress and result of the primary determination unit 4, and secondary determination unit.
- the judgment progress and result of 5 are visually displayed.
- the graph shown in FIG. 17 is an example in which 14 Kullback-Leibler indexes are calculated for the data of four rolled materials and their transitions are plotted three-dimensionally.
- the eighth index of the fourth rolled material is by far the largest, and it can be seen that it is necessary to pay attention to this numerical index.
- Data collection device 2 Abnormality judgment support device 3: Analysis target data creation unit 4: Primary judgment unit 5: Secondary judgment unit 6: Information input unit 7: Data and result storage unit 8: Display unit 20: Manufacturing equipment 122 , 132, 137: Machine learning device
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Human Computer Interaction (AREA)
- Testing And Monitoring For Control Systems (AREA)
- General Factory Administration (AREA)
Abstract
Description
図1は、本発明の実施の形態の異常判定支援装置が適用される製造設備のシステム例を示す図である。本実施の形態において異常判定支援装置2が適用されている製造設備20は、熱間薄板圧延ラインである。熱間薄板圧延ラインは、加熱炉21、粗圧延機22,23、バーヒータ24、仕上圧延機25、ランアウトテーブル26、巻き取り機27などの各種の装置からなる製造設備20である。
異常判定支援装置2は、例えばLANによってデータ採取装置1に接続されている。異常判定支援装置2は、ユーザによる製造設備20の異常判定を支援する装置である。より詳しくは、異常判定支援装置2は、製造設備20において異常が生じているか判定するための判断材料をユーザに対して提供する装置であり、製造設備20の異常判定に用いる解析対象データをデータ採取装置1に記録された時系列信号から抽出し、解析し、その解析結果をユーザに対して提供することにより、ユーザが行う異常判定を支援する。異常判定支援装置2は、少なくとも1つのメモリと少なくとも1つのプロセッサとを有するコンピュータである。メモリには、異常判定に用いる各種のプログラムや各種のデータが記憶されている。
3-1.解析対象データ作成部
解析対象データ作成部3は、データ採取装置1から振動、電流、荷重などの製造設備20の状態を表す時系列信号及び製品品質を表す時系列信号を取得し、それら時系列信号から一次判定部4で行う解析や判定に必要なデータを抽出する。ただし、製造設備20が稼働していないと異常か正常かの情報は得られないので、解析対象データ作成部3は製造設備20が稼働中のデータを時系列信号から抽出する。
一次判定部4は、解析対象データ作成部3において負荷状態と非負荷状態、或いは測定時状態と非測定時状態とに分類された解析対象データに対し、複数種類の異なる解析方法を適用する。詳しくは、一次判定部4は、複数種類の異なる解析方法を共通の解析対象データに適用することで、解析対象データを解析方法の数の判定に適する数値指標に変換する。また、一次判定部4は、複数の数値指標のそれぞれに基づき製造設備20の異常の判定と異常度合の計算とを行い、また、製品品質の異常の判定と異常度合の計算とを行う。
(1)平均値
(4)実効値
(5)Peak値
例えば正の最大値から大きい順に10個の平均値、又は負の最小値から小さい順に10個の平均値
(6)波形率
(7)波高率
(8)衝撃指数
(9)間隙率
(10)歪度
(11)尖り度
解析方法の第2例として、偏差データの確率密度分布と正規分布との差を用いることについて説明する。確率密度分布とは、データxがある範囲を変化するとき、その範囲の中のある値となる確率を表したもので、その範囲内の確率を全て加算すると1(100%)となる。
解析方法の第3例として、偏差データの極大値及び極小値のそれぞれを対象として確率密度分布を計算し、レイリー分布との差を数値指標とする方法について説明する。極大値、極小値の場合は、正常時のデータの分布は正規分布ではなく、図9に示すようなレイリー分布となる。数値指標の計算、正常/異常の判定方法は、前記と同様である。なお、上記の確率密度の値は無次元の値である。
解析方法の第4例として、自己回帰モデルを用いる方法について説明する。一般に、回帰モデルとは、従属変数と独立変数との関係を例えば一次多項式の形で表したものであり、従属変数と独立変数は異なる変数でもよい。例えば、従属変数を圧延荷重とした場合、独立変数を変形抵抗、圧延速度、材料温度として回帰モデルを作ることができる。自己回帰モデルとは、従属変数と独立変数が同じデータ種別であるが、従属変数は現在の値であり、独立変数はその過去の値を用いるものである。例えば、圧延荷重を過去の自身の値(圧延荷重)で同定することに相当する。自己回帰モデルは、例えば以下の式10で表される。ここで、εは白色ノイズ、α0、α1、…、αK-1は自己回帰係数である。なお、自己回帰係数の値は無次元の値である。
解析方法としては、上記の例の他にもFFT(高速フーリエ変換)、ウェーブレット変換などの解析方法もあり、上記の例には限定されない。
ケース1:標準偏差や歪度などの統計量そのもの
ケース2:ケース1に基づくHotelling理論の指標
ケース3:原データの確率密度分布と正規分布との差を表す式7~9の値
ケース4:ケース3に基づくHotelling理論の指標
ケース5:原データの極大値・極小値の確率密度分布とレイリー分布との差を表す式7~9の値
ケース6:ケース5に基づくHotelling理論の指標
ケース7:自己回帰モデルで同定された自己回帰係数そのものの値
ケース8:ケース7に基づくHotelling理論の指標
次に、二次判定部5について説明する。
図12は、二次判定部5が備える機械学習装置の第1例を示す図である。図12では、一次判定部の数値指標が機械学習装置122に対する入力121であり、正常又は異常の判定結果(二次判定結果)と推定した異常原因とが出力123である。学習段階では、教師信号として、入力121とその結果としての出力123を一対として機械学習装置122に与える。二次判定を行う段階では、入力121のみを機械学習装置122に与え、その結果である出力123を判定結果とする。
図15は、二次判定部5が備える機械学習装置の第2例を示す図である。図12に示す第1例とは異なり、第2例の機械学習装置132では、一次判定部4からの入力131は、数値指標ではなく、正常か異常を判定した結果と異常度合である。第1例と第2例とは入力が異なっているが、その他は同じである。しかしながら、一次判定部4からの入力131を、数値指標ではなく正常か異常を判定した結果とした場合、一次判定部4における正常/異常の判定に用いた閾値を変更すると、正常/異常の判定結果が大きく変わることがある。そうなると学習を最初からやり直さないといけない。
一般に、正常データの数・種類に比べて異常データの数・種類は非常に少なく、異常であることを示す事例を集めるのに多くの時間がかかることが一般的である。すなわち図12に示す第1例の機械学習装置122、図15に示す第2例の機械学習装置132が学習するためには、異常状態を含む入力と正解が対になった教師信号を得る頻度が小さく、十分な学習能力を持つまでに時間がかかるという課題がある。
異常データが少ない場合の対策として、機械学習における転移学習を適用することができる。一般に転移学習とは、機械学習でまだ十分学習データが貯まっていないときに用いられる手法であり、別の場所や別の機会に学習した結果(機械学習、例えばニューラルネットワーク内部の結合重み)を当該機械学習に転用する方法である。別の製造設備で多くの異常データが得られている場合、そこで機械学習した結果、異常データが少ない当該製造設備に移すことができる。ただし、製造設備の特性が異なる場合には、特性の違いを適切に評価して差を減らさないといけない。
再び図2に戻り、データ及び結果保存部7と表示部8とについて説明する。データ及び結果保存部7は、解析対象データ作成部3により作成された解析対象データ、一次判定部4の判定経過及び結果、二次判定部5の判定経過及び結果をハードディスク等の記憶装置などに保存する。後々、判定した理由や根拠を取り出せるようにするためである。
2:異常判定支援装置
3:解析対象データ作成部
4:一次判定部
5:二次判定部
6:情報入力部
7:データ及び結果保存部
8:表示部
20:製造設備
122,132,137:機械学習装置
Claims (12)
- 製造設備において異常が生じているか判定するための判断材料を提供する異常判定支援装置であって、
前記製造設備のデータ採取装置から前記製造設備の状態又は製品品質の少なくとも一方を表す時系列信号を取得し、前記時系列信号から解析対象データを抽出する解析対象データ作成部と、
前記解析対象データ作成部で抽出された前記解析対象データに複数種類の異なる解析方法を適用することによって、共通の前記解析対象データから複数の一次判定結果を導出する一次判定部と、
前記一次判定部で得られた一次判定結果とそれに対応する正解としての異常判定結果及び異常原因との対を教師信号として学習された機械学習装置を有し、前記一次判定部において共通の前記解析対象データから得られた前記複数の一次判定結果を前記機械学習装置に入力し、前記機械学習装置から出力される二次判定結果及び推定異常原因を前記判断材料として出力する二次判定部と、
を備えることを特徴とする異常判定支援装置。 - 前記一次判定部は、前記解析対象データに前記複数種類の異なる解析方法を適用することによって前記解析対象データを複数の数値指標に変換し、前記複数の数値指標を前記複数の一次判定結果として出力し、
前記機械学習装置は、前記一次判定部で得られた数値指標を入力とし、実際の異常判定結果及び異常原因を正解とする教師信号を用いて学習され、
前記二次判定部は、前記一次判定部で前記解析方法ごとに得られた前記複数の数値指標を前記機械学習装置に入力し、前記機械学習装置から出力される異常判定結果及び推定異常原因を前記判断材料として出力する
ことを特徴とする請求項1に記載の異常判定支援装置。 - 前記一次判定部は、前記解析対象データに前記複数種類の異なる解析方法を適用することによって前記解析対象データを複数の数値指標に変換し、前記複数の数値指標のそれぞれに基づき異常の有無の判定と異常度合いの計算とを行い、前記解析方法ごとに得られた複数の判定結果及び異常度合を前記複数の一次判定結果として出力し、
前記機械学習装置は、前記一次判定部で得られた判定結果及び異常度合いを入力とし、実際の異常判定結果及び異常原因を正解とする教師信号を用いて学習され、
前記二次判定部は、前記一次判定部で前記解析方法ごとに得られた前記複数の判定結果及び異常度合いを前記機械学習装置に入力し、前記機械学習装置から出力される異常判定結果及び推定異常原因を前記判断材料として出力する
ことを特徴とする請求項1に記載の異常判定支援装置。 - 前記一次判定部は、前記解析対象データに前記複数種類の異なる解析方法を適用することによって前記解析対象データを複数の数値指標に変換し、前記複数の数値指標を前記複数の一次判定結果として出力し、
前記機械学習装置は、前記一次判定部で得られた数値指標を入力とし、前記数値指標から判定される異常の有無と前記数値指標から計算される異常度合いとを正解とする教師信号を用いて学習され、
前記二次判定部は、前記一次判定部で前記解析方法ごとに得られた前記複数の数値指標を前記機械学習装置に入力し、前記機械学習装置から出力される異常判定結果及び推定異常原因を前記判断材料として出力する
ことを特徴とする請求項1に記載の異常判定支援装置。 - 前記解析対象データ作成部は、前記製造設備が稼働中における前記製造設備の負荷状態と非負荷状態の2つの状態におけるデータを抽出し、さらに抽出したデータから低周波数成分を除いた高周波数成分を前記解析対象データとして算出する
ことを特徴とする請求項1乃至4の何れか1項に記載の異常判定支援装置。 - 前記解析対象データ作成部は、前記製造設備が稼働中における製品品質測定用のセンサの測定時状態と非測定時状態の2つの状態におけるデータを抽出し、さらに抽出したデータから低周波数成分を除いた高周波数成分を前記解析対象データとして算出する
ことを特徴とする請求項1乃至4の何れか1項に記載の異常判定支援装置。 - 前記一次判定部は、前記負荷状態と前記非負荷状態のそれぞれにおける前記解析対象データに前記複数種類の異なる解析方法を適用することによって前記解析対象データを複数の数値指標に変換し、前記複数の数値指標のそれぞれに基づき異常の有無の判定を行い、前記負荷状態で異常かつ前記非負荷状態で正常であれば機械系の異常と判定し、前記負荷状態で異常かつ前記非負荷状態で異常であれば電気系の異常、信号伝達系の異常、又は制御系の異常と判定する
ことを特徴とする請求項5に記載の異常判定支援装置。 - 前記一次判定部は、前記測定時状態と前記非測定時状態のそれぞれにおける前記解析対象データに前記複数種類の異なる解析方法を適用することによって前記解析対象データを複数の数値指標に変換し、前記複数の数値指標のそれぞれに基づき異常の有無の判定を行い、前記測定時状態で異常かつ前記非測定時状態で正常であれば製品品質の異常と判定し、前記測定時状態で異常かつ前記非測定時状態で異常であれば製品品質を測定するセンサ系の異常、又は信号伝達系の異常と判定する
ことを特徴とする請求項6に記載の異常判定支援装置。 - 前記機械学習装置は,中間層が1つであるニューラルネットワークによる学習、中間層が複数から成るニューラルネットワークによる深層学習、及びルールベースの学習のうちの何れか一つの方法により学習を行う
ことを特徴とする請求項1乃至8の何れか1項に記載の異常判定支援装置。 - 前記機械学習装置は、前記一次判定部で得られた一次判定結果を示す変数のうち物理単位を持たない無次元変数を学習対象とし、前記無次元変数は他の製造設備の異常判定支援装置への転移学習、又は他の製造設備の異常判定支援装置からの転移学習に適用される
ことを特徴とする請求項1乃至9の何れか1項に記載の異常判定支援装置。 - 前記解析対象データ作成部により作成された解析対象データ、前記一次判定部の判定経過及び結果、前記二次判定部の判定経過及び結果のうちの少なくとも一つを記録装置に保存するデータ及び結果保存部、
を備えることを特徴とする請求項1乃至10の何れか1項に記載の異常判定支援装置。 - 前記データ採取装置から得られる時系列信号、前記解析対象データ作成部により作成された解析対象データ、前記一次判定部の判定経過及び結果、前記二次判定部の判定経過及び結果のうちの少なくとも一つを視覚的に表示する表示部、
を備えることを特徴とする請求項1乃至11の何れか1項に記載の異常判定支援装置。
Priority Applications (7)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
KR1020207030237A KR102398307B1 (ko) | 2019-03-26 | 2019-03-26 | 이상 판정 지원 장치 |
US16/977,476 US11392114B2 (en) | 2019-03-26 | 2019-03-26 | Abnormality determination support apparatus |
JP2020558554A JP7044175B2 (ja) | 2019-03-26 | 2019-03-26 | 異常判定支援装置 |
PCT/JP2019/012973 WO2020194534A1 (ja) | 2019-03-26 | 2019-03-26 | 異常判定支援装置 |
CN201980027213.4A CN112041771A (zh) | 2019-03-26 | 2019-03-26 | 异常判定辅助装置 |
EP19921842.1A EP3764184B1 (en) | 2019-03-26 | 2019-03-26 | Abnormality determination assistance device |
TW108128299A TWI728422B (zh) | 2019-03-26 | 2019-08-08 | 異常判定支援裝置 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2019/012973 WO2020194534A1 (ja) | 2019-03-26 | 2019-03-26 | 異常判定支援装置 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2020194534A1 true WO2020194534A1 (ja) | 2020-10-01 |
Family
ID=72609660
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2019/012973 WO2020194534A1 (ja) | 2019-03-26 | 2019-03-26 | 異常判定支援装置 |
Country Status (7)
Country | Link |
---|---|
US (1) | US11392114B2 (ja) |
EP (1) | EP3764184B1 (ja) |
JP (1) | JP7044175B2 (ja) |
KR (1) | KR102398307B1 (ja) |
CN (1) | CN112041771A (ja) |
TW (1) | TWI728422B (ja) |
WO (1) | WO2020194534A1 (ja) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11300952B2 (en) * | 2019-03-29 | 2022-04-12 | Festo Se & Co. Kg | Anomaly detection in a pneumatic system |
WO2022113501A1 (ja) * | 2020-11-26 | 2022-06-02 | 三菱重工業株式会社 | 異常検知システム、異常検知方法およびプログラム |
WO2022158037A1 (ja) * | 2021-01-25 | 2022-07-28 | オムロン株式会社 | 品質予測システム、モデル生成装置、品質予測方法、及び品質予測プログラム |
WO2022249315A1 (ja) * | 2021-05-26 | 2022-12-01 | 三菱電機株式会社 | 異常判定装置 |
WO2023286795A1 (ja) * | 2021-07-13 | 2023-01-19 | 荏原環境プラント株式会社 | 情報処理方法、情報処理装置、および情報処理プログラム |
JP2023094659A (ja) * | 2021-12-24 | 2023-07-06 | 三菱重工パワーインダストリー株式会社 | 伝熱管損傷原因推論装置及び伝熱管損傷原因推論方法 |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114730180A (zh) * | 2019-11-13 | 2022-07-08 | 杰富意钢铁株式会社 | 生产设备的运行方法以及运行系统 |
TWI792086B (zh) * | 2020-10-30 | 2023-02-11 | 友達光電股份有限公司 | 行動式設備診斷裝置及設備診斷資訊顯示方法 |
JP2022107463A (ja) * | 2021-01-08 | 2022-07-21 | 株式会社日立製作所 | プラント制御装置、プラント制御方法及びプログラム |
US11734013B2 (en) * | 2021-06-17 | 2023-08-22 | International Business Machines Corporation | Exception summary for invalid values detected during instruction execution |
TWI784718B (zh) * | 2021-09-17 | 2022-11-21 | 和碩聯合科技股份有限公司 | 廠區告警事件處理方法與系統 |
CN114488950B (zh) * | 2022-02-09 | 2023-02-28 | 无锡微茗智能科技有限公司 | 机床机械部件动态保护方法及数控机床设备 |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS5828948B2 (ja) | 1975-08-19 | 1983-06-18 | 松下電器産業株式会社 | デンゲンデンアツキヨウキユウソウチ |
JP6252675B2 (ja) * | 2014-05-20 | 2017-12-27 | 東芝三菱電機産業システム株式会社 | 製造設備診断支援装置 |
JP2018106562A (ja) * | 2016-12-27 | 2018-07-05 | 株式会社ジェイテクト | 解析装置および解析システム |
JP2019016039A (ja) | 2017-07-04 | 2019-01-31 | Jfeスチール株式会社 | プロセスの異常状態診断方法および異常状態診断装置 |
JP2019020278A (ja) | 2017-07-19 | 2019-02-07 | 株式会社日立製作所 | 回転機システムの診断装置、電力変換装置、回転機システム、および回転機システムの診断方法 |
Family Cites Families (23)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP1292723A2 (en) * | 2000-04-11 | 2003-03-19 | Recherche 2000 Inc. | Method and apparatus for acquisition, monitoring, display and diagnosis of operational parameters of electrolysers |
JP4605132B2 (ja) * | 2006-09-29 | 2011-01-05 | パナソニック電工株式会社 | 異常検出装置、異常検出方法 |
JP5519472B2 (ja) * | 2010-10-27 | 2014-06-11 | 株式会社日立製作所 | 被圧延材張力制御装置、被圧延材張力制御方法および熱間タンデム圧延機 |
JP5647917B2 (ja) * | 2011-03-04 | 2015-01-07 | 東芝三菱電機産業システム株式会社 | 制御装置及び制御方法 |
CN103384572B (zh) * | 2011-03-18 | 2015-05-27 | 株式会社日立制作所 | 轧制控制装置、轧制控制方法以及轧制控制程序 |
CN102496028B (zh) * | 2011-11-14 | 2013-03-20 | 华中科技大学 | 一种复杂装备的事后维修故障分析方法 |
JP2015075821A (ja) * | 2013-10-07 | 2015-04-20 | 横河電機株式会社 | 状態診断方法および状態診断装置 |
CN104102773B (zh) * | 2014-07-05 | 2017-06-06 | 山东鲁能软件技术有限公司 | 一种设备故障预警及状态监测方法 |
WO2016033247A2 (en) | 2014-08-26 | 2016-03-03 | Mtelligence Corporation | Population-based learning with deep belief networks |
JP5828948B2 (ja) | 2014-10-09 | 2015-12-09 | 株式会社高田工業所 | 回転機械系の異常診断方法 |
JP6156355B2 (ja) * | 2014-12-24 | 2017-07-05 | コニカミノルタ株式会社 | 画像形成システム、記録材の搬送停止方法、および記録材の搬送停止プログラム |
JP6148316B2 (ja) * | 2015-07-31 | 2017-06-14 | ファナック株式会社 | 故障条件を学習する機械学習方法及び機械学習装置、並びに該機械学習装置を備えた故障予知装置及び故障予知システム |
JPWO2017109903A1 (ja) * | 2015-12-24 | 2018-03-22 | 株式会社東芝 | 異常原因推定装置及び異常原因推定方法 |
JP6140331B1 (ja) * | 2016-04-08 | 2017-05-31 | ファナック株式会社 | 主軸または主軸を駆動するモータの故障予知を学習する機械学習装置および機械学習方法、並びに、機械学習装置を備えた故障予知装置および故障予知システム |
JP6661559B2 (ja) * | 2017-02-03 | 2020-03-11 | 株式会社東芝 | 異常検出装置、異常検出方法およびプログラム |
JP6841558B2 (ja) * | 2017-02-24 | 2021-03-10 | Thk株式会社 | 転がり案内装置の状態診断システム及び状態診断方法 |
US10928814B2 (en) | 2017-02-24 | 2021-02-23 | General Electric Technology Gmbh | Autonomous procedure for monitoring and diagnostics of machine based on electrical signature analysis |
CN107132410B (zh) * | 2017-03-13 | 2018-09-25 | 广东电网有限责任公司信息中心 | 10kV线路线损率异常原因检测方法和系统 |
EP3379357B1 (en) | 2017-03-24 | 2019-07-10 | ABB Schweiz AG | Computer system and method for monitoring the technical state of industrial process systems |
JP6860406B2 (ja) * | 2017-04-05 | 2021-04-14 | 株式会社荏原製作所 | 半導体製造装置、半導体製造装置の故障予知方法、および半導体製造装置の故障予知プログラム |
JP6380628B1 (ja) | 2017-07-31 | 2018-08-29 | 株式会社安川電機 | 電力変換装置、サーバ、及びデータ生成方法 |
CN107817404B (zh) * | 2017-11-18 | 2023-06-20 | 广西电网有限责任公司电力科学研究院 | 一种便携式计量自动化终端故障诊断装置及其诊断方法 |
CN109218114B (zh) * | 2018-11-12 | 2021-06-08 | 西安微电子技术研究所 | 一种基于决策树的服务器故障自动检测系统及检测方法 |
-
2019
- 2019-03-26 CN CN201980027213.4A patent/CN112041771A/zh active Pending
- 2019-03-26 KR KR1020207030237A patent/KR102398307B1/ko active IP Right Grant
- 2019-03-26 WO PCT/JP2019/012973 patent/WO2020194534A1/ja unknown
- 2019-03-26 JP JP2020558554A patent/JP7044175B2/ja active Active
- 2019-03-26 US US16/977,476 patent/US11392114B2/en active Active
- 2019-03-26 EP EP19921842.1A patent/EP3764184B1/en active Active
- 2019-08-08 TW TW108128299A patent/TWI728422B/zh active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPS5828948B2 (ja) | 1975-08-19 | 1983-06-18 | 松下電器産業株式会社 | デンゲンデンアツキヨウキユウソウチ |
JP6252675B2 (ja) * | 2014-05-20 | 2017-12-27 | 東芝三菱電機産業システム株式会社 | 製造設備診断支援装置 |
JP2018106562A (ja) * | 2016-12-27 | 2018-07-05 | 株式会社ジェイテクト | 解析装置および解析システム |
JP2019016039A (ja) | 2017-07-04 | 2019-01-31 | Jfeスチール株式会社 | プロセスの異常状態診断方法および異常状態診断装置 |
JP2019020278A (ja) | 2017-07-19 | 2019-02-07 | 株式会社日立製作所 | 回転機システムの診断装置、電力変換装置、回転機システム、および回転機システムの診断方法 |
Non-Patent Citations (2)
Title |
---|
BETUL KANBERNA YAZICI: "The Individuals Control Chart in Case of Non-Normality", JOURNAL OF MODERN APPLIED STATISTICAL METHODS, 2005 |
See also references of EP3764184A4 |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11300952B2 (en) * | 2019-03-29 | 2022-04-12 | Festo Se & Co. Kg | Anomaly detection in a pneumatic system |
WO2022113501A1 (ja) * | 2020-11-26 | 2022-06-02 | 三菱重工業株式会社 | 異常検知システム、異常検知方法およびプログラム |
WO2022158037A1 (ja) * | 2021-01-25 | 2022-07-28 | オムロン株式会社 | 品質予測システム、モデル生成装置、品質予測方法、及び品質予測プログラム |
JP7567504B2 (ja) | 2021-01-25 | 2024-10-16 | オムロン株式会社 | 品質予測システム、モデル生成装置、品質予測方法、及び品質予測プログラム |
WO2022249315A1 (ja) * | 2021-05-26 | 2022-12-01 | 三菱電機株式会社 | 異常判定装置 |
WO2023286795A1 (ja) * | 2021-07-13 | 2023-01-19 | 荏原環境プラント株式会社 | 情報処理方法、情報処理装置、および情報処理プログラム |
JP2023094659A (ja) * | 2021-12-24 | 2023-07-06 | 三菱重工パワーインダストリー株式会社 | 伝熱管損傷原因推論装置及び伝熱管損傷原因推論方法 |
Also Published As
Publication number | Publication date |
---|---|
JP7044175B2 (ja) | 2022-03-30 |
KR20200135453A (ko) | 2020-12-02 |
US20210232131A1 (en) | 2021-07-29 |
CN112041771A (zh) | 2020-12-04 |
EP3764184A1 (en) | 2021-01-13 |
TW202036016A (zh) | 2020-10-01 |
EP3764184B1 (en) | 2024-09-18 |
EP3764184A4 (en) | 2021-04-21 |
TWI728422B (zh) | 2021-05-21 |
US11392114B2 (en) | 2022-07-19 |
JPWO2020194534A1 (ja) | 2021-04-30 |
KR102398307B1 (ko) | 2022-05-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP7044175B2 (ja) | 異常判定支援装置 | |
CN107949813B (zh) | 制造设备诊断辅助装置及制造设备诊断辅助方法 | |
US10996662B2 (en) | Manufacturing equipment diagnosis support system | |
CN109996615B (zh) | 轧制设备的异常诊断方法及装置 | |
JP5158018B2 (ja) | 生産システムの設備診断装置および設備診断方法、並びに設備診断プログラムおよびこれを記録したコンピュータ読み取り可能な記録媒体 | |
CN108267312A (zh) | 一种基于快速搜索算法的地铁列车轴承智能诊断方法 | |
JP2010175446A (ja) | 状態診断装置 | |
JP6885321B2 (ja) | プロセスの状態診断方法及び状態診断装置 | |
US20230393113A1 (en) | Construction method of abnormality diagnosis model, abnormality diagnosis method, construction device of abnormality diagnosis model, and abnormality diagnosis device | |
JP6760503B2 (ja) | 製造プロセス監視装置 | |
CN118643384A (zh) | 一种多权重指标融合的滚动轴承健康状态退化失效评估方法 | |
Sharma et al. | Estimation of bearing remaining useful life using exponential degradation model and random forest algorithm | |
CN117019889A (zh) | 一种三辊旋轧设备智能检测故障的方法及系统 | |
CN117168849A (zh) | 汽车转向器装配质量的噪声检测方法 | |
JPH03258410A (ja) | タンデム圧延機における板厚異常の診断方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
ENP | Entry into the national phase |
Ref document number: 2019921842 Country of ref document: EP Effective date: 20201006 |
|
ENP | Entry into the national phase |
Ref document number: 2020558554 Country of ref document: JP Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 20207030237 Country of ref document: KR Kind code of ref document: A |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 19921842 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |