JP2020061109A - Method for integrally monitoring quality and facility in production line - Google Patents

Abstract

To provide an integral monitoring function capable of tracing a temporal state change of a machine tool and a metal mold, detecting even slight deterioration in an operation state of the machine tool that affects quality deterioration of a product to be produced, and stably maintaining the quality of the product to be produced in a favorable state.SOLUTION: Upon determination of a failure in a primary determination step, from among observation variable groups, a factor variable group with a large degree of involvement in feature amounts for time of detecting the failure and a most recent time zone is extracted, a factor variable group with a large degree of involvement in a feature amount of a failure state in a proceeding time zone where a failure detection time zone is defined as a start point is extracted, in a total of three time zone groups for the time of failure detection and the most recent time, the factor variable group is classified into a plurality of categories based on the magnitude relation of the degree of involvement of the observation variable with respect to the feature amount of each time zone, and a cause and effect model is formed where the category is defined as a cause group and the observation variable group as a factor candidate is defined as a result group to thereby perform fairness diagnosis of a target process and product quality evaluation and specify a cause of the failure.SELECTED DRAWING: Figure 1

Description

    The present invention relates to a method for integrated quality and equipment monitoring in a production line.

In a production line that manufactures and processes machine parts and products using machine tools, data and information related to changes in the operating status of machine tools are acquired to evaluate the quality of production objects and the soundness of machine tools. If a sign of abnormality is detected, the cause is searched for, the machine tool is adjusted / repaired / improved at an appropriate time, and further, the design is changed or the operation is improved if necessary. In the quality inspection of the above production object, data is collected and measured according to each inspection item, the acquired data and information are processed and converted as necessary, and the target value set in advance is set. The quality is judged by comparison. On the other hand, in the soundness diagnosis of a machine tool, data related to the operating state of the device is collected, processed / converted as needed, and compared with a preset threshold value to determine an abnormality. When a defect in quality or an abnormality in a machine tool is detected, the cause is estimated based on the know-how of a deterioration factor or a collection of troubleshooting, or based on the know-how and experience of the expert.
However, in the current quality inspection, suddenly defective products are detected, such as the problem of the accuracy of the target value set in advance to judge the quality of the quality, and it is not always possible to collect data that fits the purpose of the inspection. There was a problem in the performance of defect detection, such as detection and missed subtle defective products. Also, in the soundness diagnosis of machine tools, there is a large temporal variation of data in the machining process, and various disturbances such as changes in machining conditions are mixed.Therefore, at the time of issuing an alarm by comparison with a predetermined threshold value. Depending on the type of abnormality, there is a problem in the sensitivity and reliability of abnormality detection, in which the risk of failure occurrence is high or the normal state is erroneously determined to be abnormal. Further, even when the quality of the production target is not inferior, it is not possible to detect that the minute deterioration state of the equipment such as machine tools causes the deterioration of the quality. There was a problem that it was difficult.

Patent No. 5753301 Takayoshi Yamamoto has proposed "a comprehensive diagnostic method for plant equipment and a comprehensive diagnostic equipment for plant equipment". In this conventional technology, a principal component analysis method is applied based on the correlation between observation data, and a detection method in which the degree of divergence between the feature amount of normal data and the feature amount of the observed data is used as a determination index for the presence or absence of an abnormal sign Then, when the deviation degree exceeds the determination threshold value, it is determined that “there is an abnormal sign”, and the observation variable group candidates related to the abnormal sign are listed. However, when the observed data is a relatively slow change, it is effective for early detection of abnormal signs, but the observed data on the production line targeted by the present invention often has a sharp temporal change, and There was a problem in the performance of anomaly detection under various disturbances. In addition, it is stated that among the observed variable groups, the variable groups that are candidate factors are listed, and the causal relationships between the variables that are candidate factors are searched during trend monitoring for several weeks to several months after the abnormality is detected. In the defect detection of the production line, the time to continue the above trend monitoring is not allowed, and it is difficult to confirm the defect determination when the defect is detected, build a causal model, and quickly estimate the cause. There was a problem of being there.

JP 2001-113399 A The proposals of Toyota Motor Corporation and Kyoho Manufacturing Co., Ltd. are as follows.
The method of diagnosing a press machine according to the related art is to measure the load applied to the machine frame of the press machine and diagnose the press machine based on the measured load. It is specialized in diagnosing backlash, and lacks the viewpoint of evaluating quality by comprehensively taking into account changes in the state of the mold and the properties of the work material, and is a sign of deterioration in the quality of the production target. There was a problem that I could not grasp.

Japanese Patent Application Laid-Open No. 2005-0744545 Okuma Co., Ltd. proposes to perform sensing with an AE sensor, accurately detect abnormality or deterioration of the spindle, judge damage to the spindle, and display an alarm. In response to the spindle rotation command and feed axis drive command, the vibration data is measured using the AE sensor, and the number of times the data exceeds the threshold value obtained from the spindle rotation information is counted by the level count circuit, and the count The number is compared with the alarm level to determine the state of the spindle, and an alarm indicator is displayed if necessary. At this time, the event rate of the measured AE vibration waveform is counted, and when determining the abnormality of the bearing, the average event rate that is the sum of the measured event rate and the event rate for any number of times previously measured. Although there is a device to judge the state of the spindle by comparing the one-time event rate and the average event rate with a preset bearing abnormality event rate level, there is a device to monitor the spindle for abnormalities. However, no consideration is given to the fact that the machining process is related to the machine tool, the die, and the work piece, and the state change of the die and the property of the work material. However, there is a lack of a viewpoint of evaluating quality by comprehensively considering the above, and there is a problem that it is impossible to grasp a sign of quality deterioration of a production target.

Japanese Patent Laid-Open No. 2007-237667 Kikuchi Seisakusho Co., Ltd. has proposed a method for inspecting the quality of a die-molded product with high accuracy. This system is provided with CCD camera units X, Y and Z in the X-axis direction, Y-axis direction and Z-axis direction, a laser length measuring unit, various temperature sensors, and a discriminating means. By comparing and collating the measurement data of the three-dimensional shape with the reference image data of the non-defective product of the molded product in the three-dimensional direction and the reference measurement data of the three-dimensional shape, the quality of the molded product is discriminated from multiple eyes, and the press is also performed. According to the temperature data of the machine, the suitability of the temperature of each part of the heat generating part including the mold is determined, and the quality of all the molded products is evaluated in a multidimensional manner. Although the quality of the molded product is evaluated by directly measuring the object to be produced with an image / laser and judging whether the temperature data is good or bad by comparing it with the reference data, the quality between the mold and the work piece is evaluated. The lack of consideration for the relationship over time (e.g. wear progress of each part, change in the clearance between the punch and die) and aging of the press machine (for example, change in the amount of slide position deviation). Is not solved.

JP, 2017-170578, A Nagasaki Prefecture has proposed a machine tool monitoring and predictive control device having a tool recognition function. In the machine tool system, the current data and vibration data measured by the measuring unit that measures the state of the spindle are used to compare and analyze the standard current data and vibration data for each tool, thereby operating the machine tool system. To monitor.
Then, the time series of the current data or the vibration data measured in association with the specified tool is compared and analyzed with the saved time series from the past to predict the occurrence of an abnormality in the machine tool system. However, in comparison with the standard values of current and vibration data, during normal machining, each actual data fluctuates greatly with time and the good range is wide, so false positive (normal is erroneously judged as abnormal. However, the accuracy of detection of minute abnormalities becomes poor and it is not practical. Further, it is obvious that the prediction accuracy is not practical because the prediction method of comparing these time series data is basically a time series comparison method based on the strength of current and vibration. In short, when the operation is very irregular and unsteady like a machine tool, there is a problem that the monitoring and prediction functions cannot be put to practical use with the method of using the judgment standard in comparison with the standard value. .

Detects minute abnormalities in machine tools, molds, and work materials that affect the quality deterioration of production objects, predicts the occurrence of machining defects, and also detects early signs of abnormalities in the machine tool itself. Furthermore, it is difficult to realize integrated monitoring that estimates the cause of the defect or abnormality when the defect or abnormality is detected.
This is because there are many deterioration modes in various machine tool mechanisms, and the three states of machine tool, die, and work material are the interactions during machining that affect the quality deterioration of the object to be produced. It is also possible to change the machining conditions and change over time such as wear of each part of the machine tool. This is because the variables that are latent factors in the observed data are mixed and change with time.
In addition, the observation data contains a mixture of a slow time change and a steep time change depending on the type of the target machine tool and the processing conditions.

  The present invention has been made in view of the above problems, and an object of the present invention is not only to detect a symptom of abnormality by tracking a change in state of a machine tool / die over time, but also to produce the object. Provides a comprehensive monitoring function that can detect even a slight deterioration of the operating condition of the machine tool that affects the deterioration of the quality of the product and stably maintain the quality of the production target in a good condition. Especially.

That is, the method is characterized by the following.
-Detects minute abnormalities in the machining process in which changes in the state of machine tools, dies, and workpieces are interrelated and prevents deterioration of the quality of the object to be produced.
-When there is no observation variable that can be a factor candidate at the time when the result of the primary determination is "poor", the primary determination result is judged to be suspended and continuously monitored, and the observation variable that becomes the above factor candidate. If there is, the reliability of the judgment result is secured by confirming the “defective” judgment.
・ The observation variable group that becomes the factor candidates at the time when the above-mentioned defect determination is confirmed and in the time zone preceding that time is extracted as a result group, and the observation variable group is classified into three categories, and each category is Since it is narrowed down to an accurate causal model by using the latent variable, an efficient causal analysis can be performed, and the cause group can be estimated without performing the trend monitoring after the defect determination.

The main invention is to diagnose the soundness of a target process of a product production line, and evaluate the quality of the product to identify the cause of the defect when the soundness diagnosis result or the product quality is defective. In order to do so, based on time series data of observation variables such as ultrasonic waves, vibration, current, size, surface roughness, etc., the process quality of the target process and the product is evaluated and tracked, and the cause of the defect is monitored. In the method
In the primary determination step of determining the quality of the target process or quality, and in the case of being determined to be defective in the primary determination step, in the observation variable group, in the time zone immediately before the defect detection time. Based on the most recent factor candidate extraction step of extracting a factor variable group having a large degree of involvement with the feature amount, and whether or not the factor candidate can be extracted in the most recent factor candidate extraction step or the magnitude relationship of the degree of involvement values, A determination determination step for determining whether the determination result of is determined to be a true defect, the determination result is determined to be good due to a disturbance or the like, or the determination result is to be suspended; Starting from the defect detection time zone confirmed in the confirmation step, the leading factor candidate extraction for extracting the factor variable group having a high degree of involvement with the feature amount of the defective state in the preceding time zone Observation that classifies the factor variable group into a plurality of categories based on the magnitude relationship between the step and the degree of involvement of the observation variable with respect to the feature amount of each time zone in a total of three time zone groups at the time of defect detection and the latest time A variable group classification step, and a causal model construction step of constructing a causal model using the plurality of categories classified in the observation variable group classification step as a cause group and the observation variable group that is a factor candidate as the result group. Is a monitoring method for performing the soundness diagnosis of the target process and the evaluation of the product quality to identify the cause of the defect.

  As described above, the integrated monitoring method of quality and equipment in the production line of the present invention is such that even if a wide variety of disturbances such as a change in processing conditions are mixed, a minute deterioration state on the equipment side such as a machine tool is concerned. Since it can be detected that the quality is deteriorated, stable quality control can be provided.

Overall flowchart relating to a method for diagnosing the integrity of a manufacturing process of a production line and identifying the cause of a defect according to the present invention Process from the start of integrated monitoring according to the present invention to the detection of the primary defect Concept of feature amount and deviation degree in D method according to the present invention Concept of feature amount and deviation degree in S method according to the present invention Time axis when extracting the factor variable group in the preceding time zone according to the present invention Example of constructing a causal model according to the present invention Outline of the mechanism of sheet metal processing by the press machine of the embodiment Description of the crank mechanism of the press machine of the embodiment Description of the slide mechanism of the press machine of the embodiment Description of the sensor attached to the die of the example Relationship between upper die punch and lower die in the shearing process of the embodiment Explanation of cut end by shearing of press machine of example Relationship between cost and clearance between upper die and lower die in the example Causal relationship in shearing by the press machine of the example Temporal composition of processing steps in shearing by the press machine of the example Vibration data on crank bearings of the press machine of the example AE data in the die part during shearing of the press machine of the example Motor current data of the press machine of the example Temperature data at the die part 1 during shearing of the example Temperature data in the die part 2 during shearing of the example Results of quality judgment in the integrated monitoring of shearing in the example Causal model at the time of failure in integrated monitoring of shearing in the example Relationship between causal model and causal relationship at the time of failure in integrated monitoring of shearing in the example

  At least the following matters will be made clear by the description in the present specification and the accompanying drawings.

To diagnose the soundness of the target process of the production line of the product, and evaluate the quality of the product, if the result of the soundness diagnosis or the product quality is defective, to identify the cause of the defect, Based on time series data of observation variables such as sound waves, vibrations, currents, dimensions, surface roughness, etc., in a monitoring method that evaluates and tracks the process quality of the target process and product and identifies the cause of the defect,
In the primary determination step of determining the quality of the target process or quality, and in the case of being determined to be defective in the primary determination step, in the observation variable group, in the time zone immediately before the defect detection time. Based on the most recent factor candidate extraction step of extracting a factor variable group having a large degree of involvement with the feature amount, and whether or not the factor candidate can be extracted in the most recent factor candidate extraction step or the magnitude relationship of the degree of involvement values, A determination determination step for determining whether the determination result of is determined to be a true defect, the determination result is determined to be good due to a disturbance or the like, or the determination result is to be suspended; Starting from the defect detection time zone confirmed in the confirmation step, the leading factor candidate extraction for extracting the factor variable group having a high degree of involvement with the feature amount of the defective state in the preceding time zone Observation that classifies the factor variable group into a plurality of categories based on the magnitude relationship between the step and the degree of involvement of the observation variable with respect to the feature amount of each time zone in a total of three time zone groups at the time of defect detection and the latest time A variable group classification step, and a causal model construction step of constructing a causal model using the plurality of categories classified in the observation variable group classification step as a cause group and the observation variable group that is a factor candidate as the result group. Is a monitoring method for performing the soundness diagnosis of the target process and the evaluation of the product quality to identify the cause of the defect.

    With the monitoring method that diagnoses the soundness of the target process of the production line and evaluates the product quality and identifies the cause of defects, it is possible to diagnose the condition considering the interaction among the machine tool, the mold, and the work material. Therefore, it is possible to understand the effect of deterioration signs on the processing equipment side on the quality of the production target object, not only to ensure the soundness of the machine tool side but also to maintain stable processing quality, and the processing time up to the point of failure occurrence. In, the observation variable group that is a factor candidate is extracted, and the latent variable that is the background is estimated, so that the cause of the defect can be specified without performing the trend monitoring after the defect occurs.

Further, in the above-described primary determination step, as a method of determining pass / fail by using the deviation degree of the characteristic amount hidden in the time series data in the three time zone groups, the initial time zone data is fixed as reference data, and 1 The present invention is characterized by having an initial reference method for performing feature extraction using the next determination time zone and the latest time zone as target data, and an adjacency comparison method for performing feature extraction between three consecutive time zone group data that are sequentially adjacent. It is a monitoring method that performs soundness diagnosis of the target process and evaluation of product quality, and identifies the cause of defects.
In such a case, the initial reference method considers the interaction within and between data for two time zones apart from the initial time zone, and the adjacency comparison method considers the data within and for three consecutive time zones. Since feature extraction is performed in consideration of the interaction between them, even when a deterioration event progresses slowly or when a deterioration event progresses rapidly, minute deterioration signs can be detected under various disturbances. It is possible to obtain information for avoiding the quality deterioration of the target product affected by the occurrence of minute deterioration signs of the machine and the mold.

In addition, in the adjacent comparison method described above, the time series data in a total of three time zone groups, that is, the primary determination time zone and the two most recent time zones, are converted into two front time zone data. The soundness of the target process and the evaluation of product quality are characterized by using the degree of deviation between the hidden front side feature amount and the back side feature amount hidden in the two time zone data on the rear side as an index for quality judgment. This is a monitoring method for identifying the cause of failure.
In such a case, the feature amount in each data in the two front time zones and the data amount in each data in the two rear time zones of the three time zone data groups The degree of divergence from the feature amount is used as a judgment index, so even if the signs of deterioration during processing of the target machine tool, mold, and work material under various external disturbances are sharp, minute deterioration can be detected. Is possible.

Further, in the above-described initial reference method, the front side feature quantity hidden in the two time zone data of the initial time zone and the time zone closest to the primary determination, and the initial time zone and the time zone of the primary determination are 2 A monitoring method for diagnosing the soundness of the target process and evaluating the product quality, and identifying the cause of the defect, which is characterized by using the degree of deviation from the backside feature amount hidden in the individual time zone data as an index for the quality judgment. Is.
In such a case, the initial time zone data is fixed as a reference, and within the reference time data and within the two time zone data that are separated by the progress of the machining process, and between the reference data and the two time zone data. Since the degree of divergence of the feature amount obtained by the feature extraction is used as a judgment index, even if the signs of deterioration during processing of the target machine tool, mold, and work material under various disturbances are gradual, they are small. It becomes possible to detect deterioration

In addition, when calculating the divergence degree by calculating the characteristic amount hidden in the time series data in a total of three time zone groups in the adjacent comparison method described above, it is obtained for the two front time zone data. The two front-side weighting factors are used to obtain two front-side feature amounts for the head time zone data and the intermediate time zone data, and then the same weighting coefficient as the front side weighting coefficient described above is used to determine the intermediate time zone data and the end time zone data. Soundness of the target process characterized in that two rear-side feature amounts are obtained for the time zone data, and two types of deviation degrees are calculated for the above-mentioned two front-side feature amounts and the two rear-side feature amounts. It is a monitoring method that performs sex diagnosis and product quality evaluation to identify the cause of defects.
In such a case, the front side weighting coefficient is commonly used for the two front side feature quantities and the two rear side feature quantities in the time-series data in the total of three time zone groups to calculate the two front side feature quantities. After fixing the latent features in the data set of, the degree of deviation between the above-mentioned front-side feature amount and the back-side feature amount is obtained, so that the sensitivity of defect detection can be maintained high, and the sharpness can be maintained under various disturbances. Stable detection of minute deterioration signs buried in state changes is possible.

Further, when calculating the feature amount hidden in the time-series data in a total of three time zone groups in the initial reference method to obtain the degree of deviation, the initial time zone and the two most recent primary determinations are calculated. Using the two front-side weighting factors obtained for the time-zone data, two front-side feature quantities are obtained for the initial time-zone data and the primary determination latest time-zone data, and then the same weight as the above-mentioned front-side weighting factor is obtained. Using the coefficient, two rear side feature quantities are obtained for the initial time zone data and the primary determination time zone data, and two types of divergence degrees with respect to the above two front side feature quantities and the two rear side feature quantities. Is a monitoring method for performing the soundness diagnosis of the target process and the evaluation of the product quality to identify the cause of the defect.
In such a case, in the time series data in a total of three time zone groups of the initial time zone data and the two time zone data separated from each other, the initial time zone data and the front side time zone data of the two time zones separated from each other. When calculating the front side feature amount and the rear side feature amount of the rear side time zone data of the two pieces apart from the initial time zone data, the front side weight coefficient is commonly used to calculate the front side two Since the latent feature in the dataset of is fixed and the deviation degree between the above-mentioned front-side feature amount and the back-side feature amount is obtained, the sensitivity of defect detection can be kept high, and it gradually increases under various disturbances. It is possible to stably detect signs of deterioration that are progressing to.

  In addition, when the result of the primary determination is determined to be defective based on the adjacent comparison method described above in the latest factor candidate extraction step described above, the total of the time zone at the time of detection and the two latest time zones For the three time-series data groups, two front side feature quantities are calculated using the two front side weighting factors obtained from the two front side time zone data, and the front time zone data and the corresponding time are calculated. Each observation obtained by calculating the product of the first involvement value of each observation variable obtained by calculating the product with the front side feature quantity corresponding to the zone, and the intermediate time zone data and the front side feature quantity corresponding to the time zone On the other hand, two rear side feature values are calculated using the second involvement value of the variable and the two rear side weighting factors obtained from the two rear side time zone data, and the intermediate time zone data is calculated. And the third involvement value of each observation variable obtained by calculating the product of the rear side feature quantity corresponding to the time zone , The fourth involvement value of each observation variable obtained by calculating the product of the terminal time zone data and the back-side feature quantity corresponding to the time zone, and the magnitude of the four kinds of involvement degree values for each observation variable In the case where at least two types of involvement value exceed a preset involvement threshold value, the observation variable corresponding to the involvement value is set as the latest factor candidate at the time of the defect detection and the latest time. It is a monitoring method for performing the soundness diagnosis of the target process and evaluating the product quality to identify the cause of the defect.

  In such a case, when the result of the primary determination is determined to be defective based on the above-mentioned adjacent comparison method, the observation variable whose degree of involvement with the feature amount exceeds a preset degree of involvement threshold is targeted. It is possible to extract not the factors of the change of the operation conditions and various disturbances but the latest factor candidates in the three time period groups at the time of the defect detection.

  In addition, in the latest factor candidate extraction step described above, when the result of the primary determination is determined to be defective based on the initial reference method described above, a total of three in the initial time period, the primary determination, and the latest For the time-series data group of, the two front side feature quantities are calculated using the two front side weighting factors obtained from the initial time zone data and the primary determination latest time zone data, and The first involvement value of each observation variable obtained by calculating the product of any one of the two front side feature quantities corresponding to the time zone, the initial time zone data, and the time The second involvement value of each observation variable obtained by calculating the product of the above two front side feature quantities corresponding to the zone and the other front side feature quantity, and on the other hand, the initial time zone data and the time at the time of primary determination By using the two rear side weighting factors obtained from the band data, two rear side characteristics are Each observation obtained by calculating the amount and calculating the product of the first determination latest time zone data and the back side feature quantity of any one of the above two back side feature quantities corresponding to the time zone Each observation obtained by calculating the product of the third involvement value of the variable, the time zone data at the time of the primary determination, and the other rear side feature quantity of the above two rear side feature quantities corresponding to the time zone If at least two types of involvement values exceed a preset involvement level threshold among the four types of involvement value values for each observed variable with the fourth involvement value of the variable, It is a monitoring method that performs the soundness diagnosis of the target process and the evaluation of the product quality by making the observed variable corresponding to the value the candidate of the latest factor at the time of the defect detection and the latest time, and identifies the cause of the defect. .

  In such a case, when the result of the primary determination is determined to be defective based on the above-mentioned initial reference method, the observation variable whose degree of involvement with the feature amount exceeds a preset degree of involvement threshold is targeted. It is possible to extract not as a factor of a change in operating conditions or various disturbances but as a latest factor candidate at the time of the defect detection and the latest time.

  Further, in the determination confirming step, when at least two engagement level values exceed a preset engagement level threshold among the four engagement level values described above, or If the magnitudes of the involvement levels are all less than the involvement threshold and there are no observation variables that should be factor candidates, and the number of determinations continues four or more times, the primary determination result must be determined to be a true failure. On the other hand, among the above-mentioned four types of involvement degree values, when only one type of involvement degree value exceeds a preset involvement degree threshold value, the observation variable relates to the operating condition of the target machine tool. If it is a variable, it is determined that the cause of the failure in the primary determination result is due to a change in the operating condition, and similarly, when the engagement degree value of only one type exceeds the preset engagement degree threshold value, the above Not related to machine tool operating conditions In this case, it is determined that the cause of the failure of the primary determination result is due to disturbance or the like, and on the other hand, the magnitudes of the four types of involvement values are all smaller than the involvement threshold, and the observed variables that should be candidate factors are If the number of nonexistent determinations is three or less, the determination is withheld because the cause is unknown, and the process proceeds to quality determination in the next time zone group. This is a monitoring method that identifies the cause.

  In such a case, when it is determined that the primary determination result is poor, the four types of involvement degree values calculated for each observation variable exceed the preset degree of involvement degree threshold, or the degree of involvement value is the threshold value. Whether the primary determination result is a true defect or the operating condition of the target machine tool is changed, depending on the number of determinations that is less than and there is no observation variable extracted as a factor candidate, or By classifying whether it is due to a disturbance or the like, or whether it is suspended due to unknown factors, it is possible to perform highly reliable integrated monitoring even if the machining conditions are changed or various disturbances exist.

  In the preceding factor candidate extraction step, in the case of the adjacency comparison method described above, from the time data of the defect detection and the last two time zones, or in the case of the initial reference method described above, the initial time zone and Using the rear-side weighting coefficient obtained from the two time zone data at the time of defect detection, each feature amount for each preceding time zone data three pieces before the detection time zone is calculated, and the time zone data and the feature are calculated. The value of the degree of involvement of each observation variable obtained by calculating the product with the quantity is sequentially obtained by tracing back the time zone, and when the preset degree of involvement threshold is exceeded, the corresponding observation variable is This is a monitoring method for performing soundness diagnosis of a target process and evaluating product quality, and identifying a cause of failure, which is characterized by being extracted as a candidate for a preceding factor in a preceding time period.

  In such a case, when the participation degree value of each observation variable with respect to each preceding time zone three times before the failure detection time zone exceeds a preset involvement degree threshold value, the corresponding observation variable is preceded in the preceding time zone. As a causal model can be effectively narrowed down by including all the factor candidate groups in addition to the observed variables that are the most recent factor candidates obtained in the most recent factor candidate extraction step as described above, it is possible to perform efficient causal analysis. It is possible to identify the cause.

  In the observation variable group classification step and the causal model building step described above, the first involvement value or the third involvement value among the fourth involvement values from the first involvement value of each observed variable obtained above. The observation variable group having the maximum value is the first classification, and the observation variable group having the maximum second involvement value or the fourth involvement value is the second classification, and the preceding factors in the preceding time period extracted above The characteristic is that a causal model is constructed by using a candidate observation variable group as a third classification, regarding each of the three classifications as one latent variable, and using the observation variable group corresponding to each latent variable as a result group. This is a monitoring method that performs soundness diagnosis of the target process and evaluation of product quality, and identifies the cause of the defect.

  In such a case, the order of the maximum value in the fourth degree of participation value from the first degree of participation value of each observation variable is related to the background cause of the defective event, and the latest factor candidate is If it is clear at the time of detection, it is classified as the first classification.If the latest candidate is revealed at the latest time, not at the time of defect detection, it is classified as the second classification. By using each of them as a latent variable, the causal model can be narrowed down to an appropriate causal model in consideration of the time of occurrence of a bad event, so that an efficient causal analysis can be performed.

In the construction of the causal model described above, the latent variable corresponding to the observation variable group of the third classification is a common cause of the latent variable group corresponding to the first classification and the second classification This is a monitoring method that performs the soundness diagnosis and product quality evaluation to identify the cause of defects.
In such a case, the latent variable corresponding to the third classification in the preceding time zone is likely to be a cause for the latent variable at the time of detection and the most recent time, and the initial cause when the deterioration event occurs and progresses. Since it is possible to perform a multi-layered causal analysis that causes other new deterioration events, it is possible to identify complicated causes.

(Overall flow regarding soundness diagnosis of the manufacturing process of the production line and identification of the cause of defects)
FIG. 1 shows a flowchart of a method of diagnosing the soundness of a manufacturing process of a production line and identifying a cause of a defect.
The entire flow chart in this embodiment mainly includes step A (1 in FIG. 1) and step B (4 in FIG. 1).
In step A, the degree of divergence of the feature amount in the three time period group data is determined by the initial reference method (hereinafter abbreviated as S method) and the adjacent comparison method (hereinafter abbreviated as D method) described in claim 2. DI value) is calculated, and when the maximum DI value exceeds the determination threshold value, it is determined to be “primary defect” and the process proceeds to step B (2 in FIG. 1). On the other hand, when the maximum DI value is less than the determination threshold value, it is determined to be a non-defective product, the process shifts to the next time zone, and monitoring is continued (3 in FIG. 1).

In the above step B, when the determination of the “primary defect” depends on the D method, a total of three time zone group data at the time of defect detection and the two time zones most recent thereto, or the “primary defect”. If the determination is based on the S method, the product of the correlation matrix of each time zone data and the corresponding weighting coefficient vector is calculated for a total of three time zone group data of the initial reference time, the defect detection time, and the latest time zone. The value of the degree of involvement of each observation variable in the feature quantity is obtained by calculating (4 in FIG. 1).
Here, with respect to one observation variable, there are four kinds of involvement level values regarding three time zone group data. A distribution indicating which kind of the participation level value exceeds the participation level threshold is called an appearance pattern of the feature type. In the appearance pattern of the feature type, if there is only one type of the degree of involvement value of the observation variable that may be the candidate factor, it is an observation variable that is directly related to the change of the operating conditions of the target machine tool. If so, it is determined that the observed variable is not due to the primary failure factor but due to a change in the operating condition, and the monitoring is shifted to the next time zone (5, 6, 7, 8, 10 in FIG. 1). ). Alternatively, if there is only one kind of the degree of involvement value, and if the observed variable is irrelevant to the operation condition, it is determined that the primary failure is due to another disturbance, and the next time zone The process shifts to monitoring (5, 6, 7, 9, 10 in FIG. 1). Further, in the appearance pattern of the feature type, if the time zone in which the involvement level threshold is exceeded is only at the time of the defect detection, it is determined that it is due to normal disturbance and the next time zone is entered and monitoring is continued.

  On the other hand, the observation variables that may be the above factor candidates are not directly related to the change of the operating conditions of the machine tool, and in the appearance pattern of the above feature type, two or more of the four involvement values are When the value exceeds the participation threshold, the determination of the primary failure is confirmed, and the observation variable is set as one of the most recent factor variable groups (11 and 12 in FIG. 1).

  On the other hand, when all of the above-mentioned four types of involvement level values are less than the involvement level threshold value, it is determined that the cause is unknown and the determination of the primary defect is suspended and the next monitoring is continued (13 in FIG. 1). 14). Then, when the holding process continues four times or more, the determination result of the primary defect is determined (13 and 11 in FIG. 1). However, in this case, the observed variables that are candidate factors are unknown.

After the determination result of the primary failure is determined to be defective, the factor variable group in the preceding time period is searched. When the defect determination depends on the D method, the weighting coefficient vector obtained from the defect detection and the latest time zone data is used, or when the defect determination depends on the S method, the initial reference time and the defect are determined. By using the weighting factor vector obtained from the detection time period data, the product of the correlation matrix of the data in each preceding time period and the weighting factor vector is calculated from the time period three times before the defect detection time. Find the degree of involvement of the observed variable in the feature. Then, an observation variable whose participation level value exceeds a preset participation level threshold is set as a factor variable in the preceding time zone (15 in FIG. 1).
Further, the observation variable group that is the factor candidate is classified into the third classification, and it can be estimated that the corresponding latent variable is one of the causes behind the defective event of the defect.

  After the failure of the primary judgment result described above is confirmed, the observation variable group extracted as the factor candidate group at the time of detection and the observation variable group extracted as the factor candidate group in the preceding time zone are combined to form a result group. Then, a causal model is constructed by using latent variables corresponding to the first classification to the third classification as a cause group, and causal analysis is performed to specify the cause of the defective event (16, 17 in FIG. 1).

  The basic processing in the overall flowchart shown in FIG. 1 will be described in detail.

In Fig. 2, the integrated monitoring is continued from the start of processing (initial standard ▲ 1) to each processing time slot ▲ 2 ▼, ▲ 3 ▼ .. Indicates.
Here, the time period at the time of defect detection is n time period, and the analysis result of the three time period group data including the latest two time periods (n-1) and (n-2) ( This is the case where the primary defect is determined from the comparison between the maximum DI value) and the primary determination threshold.
The data matrix X1 time zone ▲ 1 in _{▼ (x 1 (1) ···} x i (1) ··· x q (1)), a data matrix of a time zone ▲ 2 in ▼ X2 _(x 1 ₍ 2) ... x _i (2) ... x _q (2)), and the data matrix in the time zone (3) is X3 (x ₁ (3) ... x _i (3) ... x _When expressed as _q (3), it becomes as shown in [Table 1], [Table 2], and [Table 3]. Here, the subscripts indicate the order of observation variables, and the numerical values in parentheses indicate the order of three time period group data. Also, <n> pieces of data are sampled for each observation variable in one time zone.
The observation variable names are displayed as f ₁ , ... F _i , ... F _q (q types).

With the above, data necessary for the integrated monitoring of the present invention was prepared.
Next, the definition and calculation method of the deviation degree, which is the index of the quality judgment in step A of FIG. 1, will be described.

( Difference of feature amount in D method; 1 in Fig. 1 )
The concept of the D method is shown in FIG. In the D method, the data matrix X1 is set for the two front time zones (n-2) and (n-1) in the three time zone groups (n-2), (n-1), and n. , The three types of correlation coefficients R _x1x1 , R _x2x2 , and R _x1x2 in the data matrix X2 and between the data matrices X1 and X2 are obtained, and the weighting matrices A and B are obtained by solving the eigen equation. Note that R _x1x1 and R _x2x2 include two different types of information, that is, interdependence within a variable group, and R _{x1x2 includes} a correlation between variable groups.

By replacing each variable in the data matrix X1 of the time zone (n−2) and the data matrix X2 of the time zone (n−1) with a set of some linear combination of those variables, Pattern analysis can be performed. Among the above linear combinations, the weighting factors A and B are determined so that the linear combination of the variable group X1 has the maximum correlation with the linear combination of the variable group X2. Here, the weighting coefficient vector A; a ₁ , a ₂ , ... A _q , B; b ₁ , b ₂ , ... B _q is represented by the data matrix X 1 (in the two front time zones). For x ₁ (1) ... x _i (1) ... x _q (1)) and X2 (x ₁ (2) ... x _i (2) ... x _q (2)) , And the two linear couplings F _x1A and F _x2B are [ _Equation 1] and [ _Equation 2], respectively.
The above two linear combinations F _x1A and F _x2B are defined as two feature quantities of the data matrices X1 and X2 in the time zones (n-2) and (n-1).

Then, the rear two hours, similar to the front as described above, the weighting coefficients in the data matrix X3 of the data matrix X2 and time zone n of time slot (n-1) vector <A>;_<a1> , _<a 2>, ··· and _{<a q>, <B>;} <b 1>, <b 2>, seeking ··· _{<b q>,} the data matrix X2 _(x 1 ₍₂₎ · .. Two linear couplings F _{x2 <A for} x _i (2) ... x _q (2)) and X3 (x ₁ (3) ... x _i (3) ... x _q (3)) _> And F _{x3 <B>} are [ _Equation 3] and [Equation 4].
The above F _{x2 <A>} and F _{x3 <B>} are defined as two feature quantities of the data matrices X2 and X3 in the time zone (n−1), n.

Next, the divergence degree of the feature amount in the three time zone group data is defined. This degree of divergence is used as an index for determining the quality of the target state, and in order to capture the state change with high sensitivity, the feature amount is calculated from the two time zone data on the rear side in the three time zone groups. It is effective that the weighting coefficient vector for calculating is the same as the weighting coefficient vector obtained from the two front time zone data. This is modeled on the fact that humans have a sharp sense of “something different from usual” that remembers the characteristics inherent in external signals in the synaptic weights in the brain, and The weight matrix obtained from the time zone data is fixed.
The two types of deviation in the D method are calculated by the following [Equation 5] and [Equation 6]. Here, the time zones (n-2), (n-1), and n are expressed as (1), (2), and (3), respectively.
here,
F _x1A : Using the weighting matrix A obtained in the two time zones (1) and (2), the average value F _x2A of the feature quantity F _x1A for the n data in Table 1 F _x2A : Two time zones (1) Using the weighting matrix A obtained in ▼ and ▲ 2, the average value F _x2B of the feature quantity F _x2A for the n data in Table 2 is the weight obtained in the two time zones ▲ 1 and ▼ 2. Using the matrix B, the average value F _x3B of the feature values F _x2B for the n pieces of data in Table 2 is shown in Table 3 using the weight matrix B obtained in the two time zones (1) and (2). n pieces of the mean value of the feature amount _{F X3b} for data σ _Fx1A: Additional feature _F standard deviation _x1A sigma _FX2A: standard deviation of the feature amount _{F x2A} ABS: shows the absolute value processing.
σ _Fx2B : Standard deviation of the above feature amount F _x2B SQRT: Square root processing.
σ _Fx3B : standard deviation of the above feature amount F _x3B

( Difference degree of feature amount in S method; 1 in FIG. 1 )
The concept of the S method is shown in FIG. In the S method, in the data matrix X1 for the initial reference time zone (1) and the two time zones (n-1), the two front time zones (1) and (n-1) in n. , In the data matrix X2, and between the data matrices X1 and X2, three types of correlation coefficients R _x1x1 , R _x2x2 , and R _x1x2 are obtained, and the weighting matrices A and B are obtained by solving the eigen equation.
By exchanging each variable in the data matrix X1 of the time zone (1) and the data matrix X2 of the time zone (n-1) with some linear combination sets of these variables, the pattern analysis between the variable groups is performed. It can be performed. Among the above linear combinations, the weighting factors A and B are determined so that the linear combination of the variable group X1 has the maximum correlation with the linear combination of the variable group X2. The weighting coefficient vector A; a ₁ , a ₂ , ... A _q , B; b ₁ , b ₂ , ... B _q can be expressed as a data matrix X1 (x ₁ (1) ... x _{i (1) ··· x q (1} )) and _{X2 (x 1 (2) ···} x i (2) with respect ··· _x q (2)), two linear combinations _{F x1A} and _{F X2B} Are [Equation 7] and [Equation 8], respectively.
The above two linear combinations F _x1A and F _x2B are defined as two feature quantities of the data matrices X1 and X2 in the time zone (1), (n-1).
Next, in the two time zones on the rear side, the weighting coefficient vectors <A>;<a1>,<a _{1 in} the data matrix X1 in the time zone ( ₁ ) and the data matrix X3 in the time zone n are the same as in the front side. a ₂ >, ... <a _q> and <B>;<b ₁ >, <b ₂ >, ... <b _q >, and a data matrix X1 (x ₁ (1) ... x _{i (1) ··· x q (} 1)) and _{X3 (x 1 (3) ···} x i (3) two linear combinations against _{··· x q (3)) F} x1 <A> and F _{x3 <B>} becomes [ _Equation 9] and [Equation 10].
The above two linear combinations F _{x1 <A>} and F _{x3 <B>} are defined as two feature quantities of the data matrices X1 and X3 in the time zone (1), n.

Next, the deviation degree of the feature amount in the three time zone group data in the S method is defined. As in the case of the D method, this degree of deviation is used as an index for determining the quality of the target state, and in order to capture the state change with high sensitivity, the degree of deviation of 2 in the rear of the three time zone groups is used. The weighting coefficient vector when calculating the feature amount from the data of n time zones (1) and n is the same as the weighting coefficient vector obtained from the data of the two previous time zones (1) and (n-1). Is effective. Here, the time zones (1), (n-1), and n are displayed as (1), (2), and (3), respectively.
The two types of deviation degree in the S method are calculated by the following [Equation 11] and [Equation 12].
here,
F _x1A : Using the weighting matrix A obtained in the two time zones (1) and (2), the average value F _x2A of the feature quantity F _x1A for the n data in Table 1 F _x2A : Two time zones (1) Using the weighting matrix A obtained in ▼ and ▲ 2, the average value F _x1B of the feature quantity F _x2A for the n data in Table 2 is the weight obtained in the two time zones ▲ 1 and ▼ 2. Using the matrix B, the average value F _x3B of the feature values F _x1B for the n pieces of data in Table 1 is shown in Table 3 using the weight matrix B obtained in the two time zones (1) and (2). n pieces of the mean value of the feature amount _{F X3b} for data σ _Fx1A: Additional feature _F standard deviation _x1A sigma _FX2A: standard deviation of the feature amount _{F x2A} ABS: shows the absolute value processing.
σ _Fx1B : Standard deviation of the above feature amount F _x1B SQRT: Square root processing.
σ _Fx3B : standard deviation of the above feature amount F _x3B

If it exceeds, it is determined as a primary failure. (2, 4 in Figure 1)
Alternatively, if the maximum DI value among the four DI values is less than the above-mentioned determination threshold value, it is determined to be good, and the process proceeds to monitoring in the next time zone. (2, 3 in Figure 1)
In the above, the definition and the calculation method of the deviation degree which is the index of the quality judgment in 1; Step A of FIG. 1 have been described.

  ◎ Next, a method of extracting an observation variable group that may be a factor of the defect in step B when it is determined to be the primary defect in step A will be described. (4 in Fig. 1)

( Regarding the method of extracting observation variables that may be a factor; 4 in Figure 1 )
The information for interpreting the content of the feature quantity by examining the correlation between the feature quantity and the original observation variable group will be called a structural matrix.
Generally, the structural matrices S _X1 and S _X2 for the data X1 and X2 in the two time zones are defined by the product of the respective feature quantities F _X1A and F _X2B and the corresponding data matrices X1 and X2. It is calculated by the following [Equation 13] and [Equation 14].
When the above structural matrices S _X1 and S _X2 are written down, the following [Equation 15] is obtained.
S _x1 (1), S _x1 (2) ... S _x1 (i) ... S _x1 (q) displayed at the right end of S _x1 and S _x2 of [Equation 15] are observed variable names f, respectively. _{1, f 2, ··, f} i, of · · _{f q,} represents the relevance value to the feature amount _{F x1A.} Similarly, S _x2 (1), S _x2 (2) ... S _x2 (i) ... S _x2 (q) are observed variable names f ₁ , f ₂ , ..., F _i , ... F _{q, respectively. Represents} the degree of involvement value of the other feature amount F _x2B . Then, it can be construed that the observation variable having the large participation degree value strongly contributes to the configuration of the feature amount.
The structural matrices S _X1 and S _X2 for the data X1 and X2 in the two time zones indicate which observation variables are related to the corresponding feature amount and to what extent each observation variable has It will be referred to as a characteristic type S value.

  In the present invention, since three time zone groups are analyzed, there are four types of the above-mentioned degree-of-involvement values. Which of these degree-of-interest values exceeds the degree-of-involvement threshold, and Whether or not the type of involvement level value exceeds the involvement level threshold value is used to determine whether the factor is a true factor candidate, a disturbance, or the like. (5-9, 11 in FIG. 1)

( About change of operation conditions or exclusion of disturbance due to disturbance in the observed variable group of factor candidates; 6 to 9 in FIG. 1 )
In the observation variable group that exceeds the above-mentioned participation threshold and is extracted as a candidate for a factor, if it corresponds to an observed variable related to a change in the operating conditions of the target machine tool, etc., the relevant observed variable is excluded from the candidate for the factor. . In this case, the object to be produced or the machine tool is in a good state, so the process shifts to monitoring in the next time zone. (7, 8, 10 in Figure 1)

  If any of the four types of involvement values for each of the above observed variables is less than the involvement level threshold value, the primary failure determination result is temporarily suspended and the process proceeds to monitoring in the next time zone. To do. This is because even if the maximum DI value exceeds the determination threshold by the D method or the S method and it is determined that the defect is a primary defect, if there is no observation variable that may be the cause, the defect determination has determinism. In order to improve the reliability of the pass / fail judgment, it is temporarily suspended, and the monitoring is performed by moving to the next time zone. (5, 13, 14, 3 in FIG. 1)

  In addition, even if there is no possible observation variable that may be a factor, if the monitoring is continued in the next time zone and the number of pending monitoring continues to be 4, the primary judgment result is Determined as defective. This is based on the analysis / evaluation in which the pass / fail judgment is made on the set of three time zone group data, only one time zone data is determined to be defective, and the other two time zone data are good. This is because even if there is, if the number of times of determination that the result of the primary determination is defective becomes four times or more in succession, it can be estimated that the defect is not a one-off event and a defect event is continuing. (13, 11 in FIG. 1)

( Regarding the latest extraction of the factor candidate variable group after the defect is determined; 12 in FIG. 1)
When the defect is confirmed in the n time zone by the D method, the feature amount F shown in FIG. 3 is obtained in the three time zone data including the most recent (n-2) and (n-1) time zones. _{x1 (a), F x2 (} B), F x2 <A>, F x3 <B> with the corresponding data X1, X2, wherein type S value was determined correlation between _{X3 S X1, S X2, S} ' _X2 and S _X3 can be obtained by the following [Equation 16], [Equation 17], [Equation 18], and [Equation 19].
4 types of feature type S value described above, when the failure detected by the D method, each observed variable names _{f 1, f 2, ··,} f i, ··, the _{f q,} relevance value to the feature amount And is defined in order as a first involvement degree value, a second involvement degree value, a third involvement degree value, and a fourth involvement degree value.
Then, when any of the above-mentioned participation degree values corresponding to each observation variable exceeds the participation degree value threshold value, the observation variable is set as the defective factor candidate variable.

On the other hand, when the defect is confirmed in the n time zone by the S method, the characteristics shown in FIG. 4 are shown in the three time zone data including the initial reference (1) and the latest (n-1) time zone. Feature type S values S _X1 , S ′ _X1 for which the correlation between the quantities F _x1 (A), F _x1 (B), F _x2 <A>, F _x3 <B> and the corresponding data X1, X2, X3 is obtained. , S _X2 , and S _X3 can be obtained by the following [Equation 20], [Equation 21], [Equation 22], and [Equation 23].
During defect detection by S method, 4 kinds of feature type S value described above, each of the observed variable names _{f 1, f 2, ··,} f i, ··, the _{f q,} relevance value to the feature amount And is defined in order as a first involvement degree value, a second involvement degree value, a third involvement degree value, and a fourth involvement degree value.
Then, when any of the above-mentioned participation degree values corresponding to each observation variable exceeds the participation degree value threshold value, the observation variable is set as the defective factor candidate variable.

( About extraction of the factor candidate variable group in the preceding time zone after the defect is determined ; 15 in FIG. 1)
FIG. 5 shows a time axis when extracting the factor variable group in the preceding time period after the defect is determined. In the n time period in which the defect is detected and confirmed and the latest (n-1) time period, the latest factor candidate variable group is extracted. Then, the extraction of the variable group that may be a factor candidate in the preceding preceding time zone j from the (n−2) time zone is performed by the following [Equation 24], and the characteristics in the time zone j are calculated. Whether or not the type S value S _Xj exceeds the participation threshold value is determined as a factor candidate variable. Note that the weighting coefficient vector <A> obtained in the above-described defect detection n time zone and the latest (n-1) time zone data is fixed, and the correlation matrix R _XjXj of the j time zone data X _j is sequentially set. The feature type S value S _Xj is calculated by _{obtaining the} product.
Here, R _XjXj indicates the correlation matrix of the data X _j .
Sequentially going back to j = n−2 (initial time zone), the feature type S value S _Xj in each j time zone is obtained, and the observation variable group as a factor candidate is extracted in the first time zone.

( Regarding the classification of the observation variables of all factor candidates and the setting of latent variables : 16 in Fig. 1)
The observation variable group that is a total factor candidate including the latest factor candidate and the factor candidate in the preceding time zone described above can be classified into three. The reason is that, for an observation variable group in which the degree of involvement value in the defect is large in the three time zone group data, the time zone in which the maximum or significant degree of involvement value appears is the detection time, This is because the observed variables are likely to be derived from different latent factors when the time zone in which a significant degree of involvement value appears is the latest time.
In addition, it is highly probable that the observation variable extracted as a factor candidate in the preceding time zone is derived from a latent factor different from that at the time of the defect detection and the latest time.
In the above classification, the observed variables are classified into classification numbers I, II or III as shown in [Table 4] from the appearance patterns of the feature types in which the first involvement value to the fourth involvement value appear. . The observation variable that matches the appearance pattern of the feature type corresponding to [Table 4] is a confirmation rule for determining that it is a candidate for a failure factor. Note that in [Table 4], the defect is confirmed by the D method, and the symbol displayed with the [] mark corresponds to the case of the S method.

Further, in [Table 5], as the appearance pattern of the feature type at the time of defect detection and at the latest time, when the maximum participation degree value appears only in the n time period at the time of detection, the factor is a change in the operation condition or the like. It is supposed to be a disturbance caused by.

( About construction of causal model : 17 in Fig. 1)
The causal model is constructed by using the classification numbers I, II, and III, which are classifications of the observation variable groups that are all the factor candidates described above, as latent variables and the corresponding observation variable groups as a result group.
FIG. 6 shows an example of building a causal model. The latent variables are indicated by circles and the observed variables are indicated by squares, and are connected by arrows from the cause side to the result side. Among the latent variables, arrows are drawn from the third classification to the first and second classifications, and the classification III corresponds to the latent variables causing the classifications I and II. An observation variable group that should be a factor candidate described above is connected to each latent variable by an arrow as a result group.
Here, since it is unclear which of the classification I and the classification II is the cause, there are two types of causal models as a case 1 where the classification I is the cause and a case 2 where the classification II is the cause. A causal analysis will be performed.

(Embodiment of the invention When integrated monitoring is performed on sheet metal processing by a press machine)
As an example of the present invention, details of integrated monitoring in sheet metal processing by a press will be described.

FIG. 7 shows an outline of the mechanism of sheet metal working by the press machine.
In order for sheet metal working by a press machine to be successful, it is essential that the press machine, the die (upper die punch and lower die), and the three elements of the work material interact and move. The balance of these three elements largely depends on the person, and the press machine processing is performed while affecting the processing accuracy and productivity of the product.
That is,
○ Press machine: Elastic deformation only (ideally, ideally no deformation occurs)
○ Mold ・ ・ ・ Elastic deformation only (similar to press machine)
○ Workpiece: Elastic deformation + plastic deformation (ideally, there should be no elastic deformation)
Therefore, in the pressing process, not only the mold is important but the work material and the elements of the pressing machine must be considered. If the material of the material to be processed changes, the load at the time of shearing during press working will change, and there will be differences in characteristics such as the length of time required for shearing. Further, in the press machine, pressure is generated by converting the rotational movement of the crankshaft into the vertical movement of the slide by the crank mechanism, so the press machine's capability is not that it can be generated at any slide position, It is important to understand and confirm the relationship between the load and slide position in all processes. The punch 1 of the upper die is attached to the slider 1 in FIG. 7, while the die 4 of the lower die is installed on the bolster 3 on the lower side of the press machine. Then, the workpiece 5 is inserted and set at the center position of the press machine by the feeding device, the slider 1 is lowered, and the workpiece 5 is sheared between the punch 2 and the die 4. When the processing is completed, the slider 1 rises and the product 6 is discharged from the press machine. The target product is mass-produced by continuously repeating this processing step.

  The above crank mechanism is shown in FIG. Rotational energy from a motor (not shown) is transmitted to the flywheel 1 to rotate the crankshaft 2 at a speed of a stroke number n. The slide 4 and the crankshaft 2 are connected by a connecting rod 3, and the rotation of the crankshaft 2 causes the slide 4 to move up and down. The slide 4 moves up and down within the range of the stroke length S. Further, vibration data obtained by installing the acceleration sensors 7 and 8 on the bearings 1 and 2 at both ends of the crankshaft 2 are analyzed to diagnose the soundness of the rotational performance of the crankshaft 2 and estimate the deterioration factor. And so on.

  The range of vertical movement of the slide in the above crank mechanism will be described with reference to FIG. The bottom dead center position 5 of the slide 4 is determined by turning the slide adjusting screw 2. The space between the bottom dead center position 5 and the bolster plate 7 is the die height 6. Of course, the adjustable range of the slide adjustment screw 2 is defined by the adjustment upper limit 3.

FIG. 10 shows a sensor group attached to the lower die 1 of the mold. Two AE sensors CH1 and CH2 and two temperature sensors T1 and T2 are attached. This AE sensor group monitors whether "elongation" has occurred due to insufficient cracks or plastic deformation occurring in the work material during shearing in the pressing process.
Further, the temperature of the die 1 during shearing is measured by the temperature sensor. The press machine itself is a heat source due to motor heat generation, compressed air heat generation, friction heat at the slide part, etc. Also, the work material to be processed and the mold become heat sources, so if each part continues to operate. Grows. From this, the bottom dead center position 5 shown in FIG. 9 is gradually lowered. This becomes more serious as the shape of the target product becomes complicated and the required accuracy of the product increases.

FIG. 11 shows the relationship between the punch 1 of the upper die and the lower die 2 during the shearing process. At the time of the shearing process, the clearance 3 which is a gap between the punch size and the die size is properly adjusted, but the burr 6 is generated along with the product 4.
The properties of the shear surface differ depending on the size of the clearance, and also vary depending on the uneven shape of the product.
FIG. 12 shows the configuration of the cut surface. In order from the upper part of FIG. 12, it comprises a droop 1, a shear surface 2, a fracture surface 3, and a burr (also called burr) 4. Then, as shown in the lower part of FIG. 12, when the clearance (3 in FIG. 11) increases, the area of the shear surface decreases, the burr (burrs) increases, and conversely, when the clearance decreases, a secondary shear surface occurs. . This burr requires a work of "deburring" in a later process, and shear energy increases with a small clearance such as a secondary shear surface. Therefore, maintaining an appropriate clearance is an important factor in maintaining product quality and optimizing input energy, but even with the initial settings, it will gradually deviate from the optimum range when machining is repeated. .

  In FIG. 13, the horizontal axis represents the size of the clearance and the vertical axis represents the processing cost, and the optimum point of the size of the clearance is shown. The shearing cost 1 is the smallest in the clearance 4 shown by the circle in the figure, while the deburring cost 2 increases as the clearance increases, so the total cost 3 including the shearing and deburring costs is It can be seen that the clearance 5 indicated by the triangle mark has the lowest cost.

  When the press machine is operated and the machining process is repeated, the state of the press machine, the mold, and the work material change while being mutually related. FIG. 14 shows a causal relationship between various groups of factors resulting from the interaction or propagation of deterioration events and groups of results of failure signs, functional deterioration, or deterioration items. Since the slide position of the press machine deviates from the heat generation 1 at each part, the bottom dead center is lowered, which causes load eccentricity 2 and press die expansion 3 during pressing. On the other hand, as the machining process continues, die wear 5 progresses, and the clearance change 4 occurs together with the die expansion 3. Further, as the die wear 5 progresses, fine scratches 6 may occur on the upper die punch, and the fine scratches 6 cause damage (defective) 13 in the product after processing and enlargement of burr (burrs) 14 and This leads to a shortening 17 of the die life.

The clearance change 4 described above results in a dimensional change (defect) 10 of the processed product, a change (defective) 11 of the shear surface property, an increase in processing energy 12, and a damage (defect) 13 of the processed product.
Further, as the wear 7 of the crank bearing progresses, a large vibration 8 is generated, which causes an eccentric load 9 due to the displacement of the slide position. Machine deterioration 16 and die life shortening 17 progress. The contents described above correspond to the relation of the item group which changes with the passage of time.
On the other hand, the case where there is a change in the material to be processed, the die, or the state of die assembly will be described. Similarly, as shown in B of FIG. 14, due to the difference or change in the surface roughness 18 of the work material, the surface roughness 19 of the die, and the assembly accuracy 20 of the die, the clearance difference 21 and the wear or loss of the upper die punch are caused. 21, resulting in damage 13 of the processed product, enlargement 14 of burrs (burrs), and shortening 17 of the die life.

(Embodiment of the invention; Case example of defect detection by integrated monitoring of sheet metal processing by a press machine)
As an example of the present invention, a detailed description will be given of a case in which good monitoring, change of operating conditions, and detection of failure are performed when integrated monitoring is performed in sheet metal working by a press.

(Collecting various data when processing sheet metal with a press machine)
In this example, data collection and information acquisition were performed for 12 types of observation variable groups f1, ..., F12.

  FIG. 15 shows a temporal structure of a processing step when performing integrated monitoring during sheet metal processing using a press machine. The composition of the processing unit consists of three areas, which are loading and setting of the work material to the press machine, operation of the press machine, and unloading of the processed product (product) from the press machine. Various data will be collected at.

  FIG. 16 shows vibration data f6 and f7 in the acceleration sensors 1 and 2 in the crankshaft bearing shown in FIG. After 500 msec from the start of the press machine, the slide descends and the upper punch contacts the material to be processed (a in the figure), and through the process of shearing (b in the figure), the bottom dead center is reached. The waveforms up to the end of processing (section c in the figure) are shown.

  FIG. 17 shows waveform data f4 and f5 obtained by the AE sensors installed in the die parts of CH1 and CH2 shown in FIG. Various ultrasonic waveforms were observed from the start of the press machine, and the shearing process was completed from about 300 msec described as a bottom dead center of 10 mmUP in the figure to about 500 msec to the bottom dead center. It is observed that the corrugated portion A in FIG. 17 corresponds to the actual shearing between the die and the workpiece, and that the CH1 waveform data f4 has two small peaks. However, no information about the quality of the shearing process is available.

  FIG. 18 shows waveform data f8 of the motor current of the press machine. The waveform data of the motor current greatly changes corresponding to the three areas a, b, and c shown in FIG. 16 described above. It can be seen that there is a significant change over time.

  FIG. 19 shows temperature data in the die unit 1, and FIG. 20 shows temperature data in the die unit 2. Within each processing unit, 15 times are sampled during the operation time of the press machine of about 500 msec and taken on the horizontal axis. From time zone (1) at the start of production to time zone (10) at the time of primary defect detection. Show. When the press machine is in operation, heat is generated by shearing, so the temperature of the mold increases cumulatively as the machining process progresses (the arrow in the figure), and the mold expands. Become.

  [Table 6] shows stroke length f1, number of strokes per minute f2, bottom dead center position (DH) f3, AE data f4 in die CH1, AE data f5 in die CH2, vibration data of crankshaft bearing 1. f6, vibration data f7 of the bearing 2, motor current f8, data f9 by the temperature sensor T1 in the die unit 1, data f10 by the temperature sensor T1 in the die unit 2, surface roughness (Ra) of the work material f11 and the misalignment amount f12 between the bolster and die are changed from the start of production (1) to the material to be processed in time zone (6), and the primary defect is detected in time zone (10). The whole data of the observation variable group in the processing unit of is shown. In this embodiment, the stroke length f1 and the number of strokes per minute f2 in all the machining steps remained substantially constant.

Since the temperature of the press machine, the mold, etc. rises as the working process progresses, the bottom dead center position (DH) f3 was 300 mm at the start of production, but 296 in the time zone (10) when the primary defect was detected. You can see it goes down to 1mm. Moreover, as long as the changes in the waveforms of the AE data f4 and f5 in the lower die of the mold, the vibration data f6 and f7 in the crankshaft bearing, and the motor current f8 are observed, the size and strength of the waveform lead to abnormalities or defects. There is no such point.
Further, the surface roughness (Ra) f11 of the material to be processed varies within the range of 0.2 μ to 0.4 μ due to the variation in the material when it is inserted and set in the press machine. Even when the material to be processed was changed in, the surface roughness was within the above range.

  Similarly, the misalignment f12 between the bolster and the die shown in [Table 6] is based on the position set in (1) at the start of production as a reference 0, and the lower die is moved forward, backward, leftward and rightward relative to the bolster as the machining process progresses. The position shift occurs. It can be seen that the absolute value of the amount of positional deviation fluctuates considerably from 0.05 mm to 0.25 mm.

(Result of primary pass / fail judgment during sheet metal working by press machine)
For the above 12 types of observation variables f1, ..., f12, in the time zone of each processing unit
It shows in [Table 7]. The maximum DI values of these D method and S method are shown in FIG. When it is judged by setting the pass / fail judgment threshold value to 2.0, it is good in the time zones (1) to (5), and in the time zone (6) when the workpiece is changed, the D method and the S method are used. Both of them are determined to be defective, and since the D method is based on the feature amount in three consecutive time zones, the influence of the time zone (6) remains, and therefore the defect is also determined in the next time zone (7). It's gone. On the other hand, in the S method, the residual amount is not seen because it is based on the feature amount based on (1) at the start of production, and the judgment returns to good in the time zone (7). Then, the machining process proceeded satisfactorily in the time zones (7), (8), and (9), and finally, in the time zone (10), both the D system and the S system were judged to be defective.

(Extraction of observed variable group that may be the factor at the latest time after it is judged as primary failure)
As a result of the above-described quality determination, when it is determined that the defect is the primary defect, the observation variable group that may be a factor of the defect is extracted.
In the above-mentioned [Table 7] and in the good working time zones (1), (2) and (3) shown in FIG. 21, and in the working time zones (6) and (6) when the workpiece is changed. Table 8 shows the characteristic type S value of each observation variable in the cases of 5), 4), and the processing time zones of 10), 9), and 8 which were determined to be primary defects. It should be noted that here, it corresponds to the result of the quality determination by the D method.

  At the time of good shown in the four columns on the right side of [Table 8], the feature type S values of all the observation variables are less than the participation threshold value 0.6, and the degree of participation with the corresponding feature amount is small, Obviously, there are no observed variables that could be a factor. In the central part of [Table 8], the feature type S of each observation variable in three time zones (4), (5), and (6) related to the time zone (6) when the work material is changed Indicates a value. The observation variable having the characteristic type S value exceeding the involvement degree threshold value 0.6 exceeds the involvement degree threshold value only in the time zone (6) with respect to the stroke length f1 and the stroke number f2. When comparing with the confirmation rules of the determination results of Table 4] and [Table 5], the appearance patterns of the feature types in which the first involvement value to the fourth involvement value appear correspond to [Table 5]. It can be seen that the defects are due to disturbances such as changes in operating conditions.

Next, the observation time of each observation variable in the time zone (10) judged as the primary failure shown in the four columns on the left side of [Table 8] and the two time zones (9) and (8) before the time zone The feature type S value is the same as the confirmation rule shown in [Table 4] because the appearance patterns of the feature types that have exceeded the involvement degree threshold value 0.6 in the observation variables f4 and f5 are the first involvement degree and the third involvement degree. It is an observation variable group that becomes the true factor of the defect from the collation of.
Further, since the observation variables f6 and f12 exceed the involvement degree threshold value 0.6, and the appearance patterns of the feature type are the second involvement degree and the fourth involvement degree, matching with the confirmation rule shown in [Table 4] is performed. Therefore, it is an observed variable that may be the true cause of the defect.
The observation variables f4 and f5 correspond to the classification number I in [Table 4], the observation variables f6 and f12 correspond to the classification number II in [Table 4], and the primary defect is determined.

  Further, in [Table 8], f11 is listed as an observation variable whose feature type S value exceeds the involvement degree threshold value of 0.6, but since the appearance pattern of the feature type is only the second degree of involvement, [Table 4 ] Confirmation rule cannot be applied, and the observation variable f11 is unknown as a factor candidate for the defect.

(Observation variable extraction that may be a factor in the preceding time zone after the primary failure is determined)
[Table 9] shows the characteristic type S value of each observation variable in the preceding time zone when the defect determination is confirmed based on the extraction result of the observation variable group that may be the latest factor. ing. Observation variables that may be a factor are f9 and f10 in which the feature type S value exceeds the involvement level threshold of 0.6. Then, the observation variables f9 and f10 correspond to the classification number III by collating with [Table 4].

(Construction of causal model)
The observation variable groups f4, f5, f6, and f12 that may be the factor candidates in the latest time and the observation variable groups f9 and f10 that may be the factor candidates in the preceding time period are classified into three groups. Based on the basic causal model shown in FIG. 6 in which the respective classification numbers I, II, and III are latent variables and the corresponding observation variable groups are result groups, a causal model for the embodiment is constructed, and the result is It shows in FIG.
Temperatures 1 and 2 at the die part of observation variables f9 and f10 that may be candidate factors extracted in the preceding time zone are presumed to be influenced by the third latent variable III, and the other two Influence the latent variables I and II of. The latent variable I affects the AE waveforms 1 and 2 of the observed variables f4 and f5, and the latent variable II is the vibration data of the observed variables f6 and f12 at the crankshaft bearing 1 and the amount of positional deviation between the bolster and the die. It is presumed to have an influence.
In addition, between the latent variables I and II, it is considered that one of the latent variables influences the other latent variable, but since information for determining any one is not obtained in advance, A causal analysis is executed for the causal directions of cases 1 and 2 shown in FIG.

  The various groups of factors resulting from the interaction of the press machine, the mold, and the work material in the shearing process by the above-mentioned press machine or the propagation of the deterioration event, and the result group of the signs of failure, the deterioration of function, or the deterioration item. Referring to FIG. 14 showing the causal relationship, the causal model shown in FIG. 22 is interpreted and the validity of the causal model is evaluated.

  Since the data of the AE waveforms 1 and 2 which are the observation variables f4 and f5 are greatly related to the feature amount at the time of the defect detection and the latest time, the damage occurrence 13 of the workpiece during the shearing process in FIG. It is presumed that the variation 15 occurred in the ratio of the sheared surface / fractured surface of the cut.

FIG. 23 shows the above-mentioned causal relationship that causes damage to the processed product or deterioration of the quality, which is cited from the overall image of the causal relationship shown in FIG. In addition, in the figure, items corresponding to latent variables are displayed with a mesh, and related items and flow lines are indicated by thick lines.
With the heat generation 1 of each part and the wear 5 of the mold as latent variables III, the temperatures 1 and 2 of the mold rise from the heat generation 1 and the expansion 3 of the mold is caused, and the upper die punch and the lower die die are separated. It affects the phenomenon that the change 4 in the size of the clearance occurs, and it can be interpreted that the change in the clearance causes the damage 13 of the workpiece as the latent variable I.
On the other hand, the eccentric load 9 due to the slide position deviation, which is supposed to have resulted in the above-mentioned variation 15 of the processed product, is the abrasion 7 of the crankshaft bearing which causes the large vibration 8 in the crankshaft bearing, and the latent variable II. Can be interpreted as
Further, the heat generation 1 of each part, which is one of the latent variables III, causes a phenomenon that the bottom dead center of the press machine is lowered, which results in an eccentric load 2, and the vibration starts from the wear 7 of the crankshaft bearing of the above latent variable II. As a result of the deviation 15 of the work piece together with the eccentric load 9 due to the slide position shift generated through the large 8.

From the detailed description and interpretation of the causal relationship shown in FIG. 23, it can be seen that the causal model shown in FIG. 22 constructed in the present example is appropriate.
The integrated monitoring method of quality and equipment in the production line has been described above, but the embodiment of the invention described above is for the purpose of facilitating the understanding of the present invention and does not limit the present invention. The present invention can be modified and improved without departing from the spirit of the present invention, and it goes without saying that the present invention includes equivalents thereof.

1 step A (1st pass / fail judgment)
2 Comparison with the pass / fail judgment threshold value 3 When the primary judgment result is non-defective, move to the next time zone 4 When the primary judgment result is bad Step B (extraction of factor possibility variable group)
5 Comparison of 4 types of involvement values and factor possibility thresholds 6 6 Judgment as to whether the defect is a true defect or a disturbance such as an operating condition based on the number of 4 types of involvement values exceeding the threshold 7 Is it an operation variable? Judgment as to whether or not it is a disturbance 8 Operational conditions (good)
9 Disturbance (good product)
10 Shift to monitoring in the next time zone 11 Defect determination 12 Latest factor candidate group determination 13 4 types of involvement level values are all below the threshold Judgment whether the number of monitoring times is 4 or more 14 If the number of times is less than 4 in step 13, the cause is unknown, and the primary defect determination result is suspended and the process shifts to monitoring in the next time period 15 Search for factor candidate group in preceding time period 16 Classification of all factor candidate groups And setting of latent variables 17 Construction of causal model

Patent Document 3: Japanese Patent Laid-Open No. 2005-074545
Okuma Co., Ltd. proposes to perform sensing with an AE sensor, accurately detect abnormality or deterioration of the spindle, judge damage to the spindle, and display an alarm. In response to the spindle rotation command and feed axis drive command, the vibration data is measured using the AE sensor, and the number of times the data exceeds the threshold value obtained from the spindle rotation information is counted by the level count circuit, and the count is made. The number of alarms is compared with the alarm level to determine the state of the spindle, and an alarm is displayed on the alarm display if necessary. At this time, the event rate of the measured AE vibration waveform is counted, and when determining the abnormality of the bearing, the average event rate that is the sum of the measured event rate and the event rate for any number of times previously measured. Although there is a device to judge the state of the spindle by comparing the one-time event rate and the average event rate with a preset bearing abnormality event rate level, there is a device to monitor the spindle for abnormalities. However, no consideration is given to the fact that the machining process is related to the machine tool, the die, and the work piece, and the state change of the die and the property of the work material. However, there is a lack of a viewpoint of evaluating quality by comprehensively considering the above, and there is a problem that it is impossible to grasp a sign of quality deterioration of a production target.

(Overall flow regarding soundness diagnosis of the manufacturing process of the production line and identification of the cause of defects)
FIG. 1 shows a flowchart of a method of diagnosing the soundness of a manufacturing process of a production line and identifying a cause of a defect.
The entire flow chart in this embodiment mainly includes step A (1 in FIG. 1) and step B (4 in FIG. 1).
In step A, the degree of divergence of the feature amount in the three time period group data is determined by the initial reference method (hereinafter abbreviated as S method) and the adjacent comparison method (hereinafter abbreviated as D method) described in claim 2. (Referred to as DI value) is calculated, and when the maximum DI value is equal to or larger than the determination threshold value, it is determined as "primary defect" and the process proceeds to step B (2 in FIG. 1). On the other hand, when the maximum DI value is less than the determination threshold value, it is determined to be a non-defective product, the process shifts to the next time zone, and monitoring is continued (3 in FIG. 1).

Or alienation by D scheme obtained in _{DI ▲ 1 ▼ / ▲ 2 ▼} / ▲ 3 ▼ (A), DI ▲ 1 ▼ / ▲ 2 ▼ / ▲ 3 ▼ (B) and S system by deviance DI _{▲ 1 If} the maximum DI value of _{▼ / ▲ 2 ▼} (A) and DI _{▲ 1 ▼ / ▲ 3 ▼} (B) is greater than or equal to a preset determination threshold value , it is determined as a primary failure. (2, 4 in FIG. 1) Alternatively, when the maximum DI value among the four DI values is less than the above-mentioned determination threshold value, it is determined to be good, and the process shifts to monitoring in the next time zone. (2 and 3 in FIG. 1) The definition and the calculation method of the deviation degree, which is the index of the quality determination in 1 of FIG. 1; step A, have been described above.

On the other hand, the case where there is a change in the material to be processed, the die, or the state of die assembly will be described. Similarly, as shown in FIG. 14B, due to the difference or change in the surface roughness 18 of the work material, the surface roughness 19 of the mold, and the assembly accuracy 20 of the mold, the clearance difference 21 and the wear or loss of the upper die punch. produce 2 2, thus Kitashi shortening 17 large 14 and die life workpiece damage 13 Yakaeri Bari.

In addition, when calculating the divergence degree by calculating the characteristic amount hidden in the time series data in a total of three time zone groups in the adjacent comparison method described above, the two time zone data in the old time zone are compared. Using the two old time zone weighting factors obtained as described above, two old time zone feature amounts are obtained for the leading time zone data and the intermediate time zone data, and then the same weighting factor as the old time zone weighting factor described above is obtained. Two new time zone feature quantities are obtained for the intermediate time zone data and the end time zone data by using the above, and two types of deviation degrees are calculated for the above two feature quantities and the two new time zone feature quantities. It is a monitoring method for performing the soundness diagnosis of the target process and the evaluation of the product quality to identify the cause of the defect.
In such a case, the old time zone weighting coefficient is calculated in common for the two old time zone feature amounts and the two new time zone feature amounts in the time series data of the total of three time zone groups. , fix the characteristics potentially to the old time zone two data sets by, since obtaining the degree of deviation between the old time zone characteristic quantity and the new time zone feature amount of the, possible to maintain a high sensitivity of the defect detection Therefore, it is possible to stably detect a minute deterioration sign buried in a steep state change under various disturbances.

Further, when calculating the feature amount hidden in the time-series data in a total of three time zone groups in the initial reference method to obtain the degree of deviation, the initial time zone and the two most recent primary determinations are calculated. Using the two old time zone weighting factors obtained for the time zone data, two old time zone feature amounts are obtained for the initial time zone data and the primary determination latest time zone data, and then the above-mentioned old time obtains two new time zone characteristic quantity for the initial time zone data and primary determining time zone data with the same weighting coefficient as band weighting factor, two old time zone characteristic quantity of the and the two new time This is a monitoring method for performing the soundness diagnosis of the target process and the evaluation of the product quality, and identifying the cause of the defect, which is characterized by calculating two types of deviations with respect to the band feature amount.
In such a case, in the time-series data in the total of three time zone groups of the initial time zone data and the two distant time zone data, the old time zone data out of the two distant from the initial time zone data Of the old time zone feature quantity and the new time zone feature quantity of the two new time zone time zone data apart from the initial time zone data, the old time zone weighting coefficient is commonly calculated. , fix the characteristics potentially to the old time zone two data sets by, since obtaining the degree of deviation between the old time zone characteristic quantity and the new time zone feature amount of the to maintain high sensitivity of defect detection Therefore, it is possible to stably detect signs of deterioration that gradually progress under various disturbances.

Degree of deviation DI by the D method obtained above: (1) / (2) / (3) (A), DI (1) / (2) / (3) (B) and deviation by the S method DI (1) If the maximum DI value of ▼ / ▲ 2 ▼ (A) or DI ▲ 1 ▼ / ▲ 3 ▼ (B) exceeds a preset judgment threshold value, it is judged as a primary failure. (2, 4 in FIG. 1) or when the maximum DI value among the four DI values is less than the above-mentioned determination threshold value, it is determined to be good, and monitoring is performed in the next time zone. (2 and 3 in FIG. 1) The definition and the calculation method of the deviation degree, which is an index of the quality judgment in 1 of FIG. 1; Step A, have been described above.

On the other hand, the case where there is a change in the material to be processed, the die, or the state of die assembly will be described. Similarly, as shown in FIG. 14B, due to the difference or change in the surface roughness 18 of the work material, the surface roughness 19 of the mold, and the assembly accuracy 20 of the mold, the clearance difference 21 and the wear or loss of the upper die punch. 22 results in damage 13 to the processed product, enlargement 14 of burrs (burrs), and shortening 17 of die life. The contents described above correspond to the relation of the item group which changes with the passage of time.

Claims (12)

To diagnose the soundness of the target process of the production line of the product, and evaluate the quality of the product, if the result of the soundness diagnosis or the product quality is defective, to identify the cause of the defect, Based on time series data of observation variables such as sound waves, vibrations, currents, dimensions, surface roughness, etc., in a monitoring method that evaluates and tracks the process quality of the target process and product and identifies the cause of the defect,
In the primary determination step of determining the quality of the target process or quality, and in the case of being determined to be defective in the primary determination step, in the observation variable group, in the time zone immediately before the defect detection time. Based on the most recent factor candidate extraction step of extracting a factor variable group having a large degree of involvement with the feature amount, and whether or not the factor candidate can be extracted in the most recent factor candidate extraction step or the magnitude relationship of the degree of involvement values, A determination determination step for determining whether the determination result of is determined to be a true defect, the determination result is determined to be good due to a disturbance or the like, or the determination result is to be suspended; Starting from the defect detection time zone confirmed in the confirmation step, the leading factor candidate extraction for extracting the factor variable group having a high degree of involvement with the feature amount of the defective state in the preceding time zone Observation that classifies the factor variable group into a plurality of categories based on the magnitude relationship between the step and the degree of involvement of the observation variable with respect to the feature amount of each time zone in a total of three time zone groups at the time of defect detection and the latest time A variable group classification step, and a causal model construction step of constructing a causal model using the plurality of categories classified in the observation variable group classification step as a cause group and the observation variable group that is a factor candidate as the result group. The monitoring method for diagnosing the soundness of the target process, evaluating the product quality, and identifying the cause of the defect.   In the primary determination step according to claim 1, as a method of determining pass / fail by using the deviation degree of the characteristic amount hidden in the time series data in the three time zone groups, the initial time zone data is fixed as reference data, It is characterized by having an initial reference method for performing feature extraction with the primary determination time zone and the latest time zone as target data, and an adjacency comparison method for performing feature extraction between three consecutive time zone group data that are sequentially adjacent. A monitoring method that diagnoses the soundness of the target process and evaluates the product quality, and identifies the cause of defects.   In the adjacency comparison method according to claim 2, for the time series data in a total of three time zone groups, that is, the primary determination time zone and the two most recent time zones, two front time zone data Diagnosis of the target process and evaluation of product quality, characterized in that the degree of deviation between the front side feature amount hidden in the rear side and the back side feature amount hidden in the two time zone data on the rear side is used as an index for quality judgment. A monitoring method to identify the cause of defects.   3. The initial reference method according to claim 2, wherein the front side feature quantity hidden in the two time zone data of the initial time zone and the time zone immediately preceding the primary determination, and the two of the initial time zone and the primary determination time zone. A monitoring method for diagnosing the soundness of the target process and evaluating the product quality, and identifying the cause of the defect, which is characterized by using the degree of deviation from the backside feature amount hidden in the individual time zone data as an index for the quality judgment. .   In the adjacency comparison method according to claim 3, when calculating the divergence degree by calculating the characteristic amount hidden in the time-series data in a total of three time zone groups, for the two time zone data on the front side Using the two front-side weighting factors thus obtained, two front-side feature quantities are determined for the head time zone data and the middle-time zone data, and then the same front-side weighting factors as the above-mentioned front-side weighting factors are used to generate intermediate-time zone data. In the target process, two rear side feature quantities are obtained for the terminal time zone data, and two types of deviation degrees are calculated for the two front side feature quantities and the two rear side feature quantities. A monitoring method that performs soundness diagnosis and product quality evaluation to identify the cause of defects.   In the initial reference method according to claim 4, when calculating the feature amount hidden in the time-series data in a total of three time zone groups to obtain the deviation degree, the initial time zone and the two most recent primary determinations are calculated. Using the two front-side weighting coefficients obtained for the time zone data, two front-side feature quantities are obtained for the initial time zone data and the primary determination latest time zone data, and then the same as the above-mentioned front-side weighting coefficient. Using the weighting factors, two rear side feature quantities are obtained for the initial time zone data and the primary determination time zone data, and there are two types of deviations with respect to the above two front side feature quantities and the two rear side feature quantities. A monitoring method for assessing the soundness of the target process and evaluating the product quality, and identifying the cause of defects   When the result of the primary determination is determined to be defective based on the adjacency comparison method according to claim 2, in the latest factor candidate extraction step of claim 1, the time zone at the time of detection and the last two time zones For three total time series data groups, the two front side feature quantities are calculated by using the two front side weighting factors obtained from the two front side time zone data, and the front time zone data is calculated. Obtained by calculating the product of the first involvement value of each observation variable obtained by calculating the product with the front side feature quantity corresponding to the time zone, and the intermediate time zone data and the front side feature quantity corresponding to the time zone On the other hand, two rear-side feature values are calculated using the second involvement value of each observation variable and the two rear-side weighting factors obtained from the rear-side two time zone data, and the intermediate time is calculated. Third contribution of each observation variable obtained by calculating the product of the band data and the backside feature value corresponding to the time zone Value and the fourth involvement value of each observation variable obtained by calculating the product of the terminal time zone data and the back-side feature amount corresponding to that time zone, and the magnitude of the four types of involvement degree values for each observation variable If at least two types of involvement values exceed a preset involvement level threshold value, the observation variable corresponding to the involvement level value is set as the factor candidate at the time of the defect detection and the latest time. A monitoring method that diagnoses the soundness of the target process and evaluates product quality, and identifies the cause of defects.   In the most recent factor candidate extraction step of claim 1, when the result of the primary determination is determined to be defective based on the initial reference method of claim 2, a total of 3 in the initial time period, the primary determination, and the latest For two time series data groups, two front side feature quantities are calculated using the two front side weighting factors obtained from the initial time zone data and the primary determination latest time zone data, and the initial time zone data is calculated. And the first involvement value of each observation variable obtained by calculating the product of any one of the two front side feature quantities corresponding to the time zone, the initial time zone data, and The second involvement value of each observation variable obtained by calculating the product of the above two front side feature quantities corresponding to the time zone and the other front side feature quantity, on the other hand, the initial time zone data and the primary determination time Two rear sides using the two rear side weighting factors obtained from the time zone data Each obtained by calculating a characteristic amount and calculating a product of the primary determination latest time zone data and any one of the above-mentioned two rear side characteristic quantities corresponding to the relevant time zone. Each obtained by calculating the product of the third involvement value of the observation variable, the time zone data at the time of the primary determination, and the other rear side feature quantity of the above two rear side feature quantities corresponding to the time zone. If at least two types of involvement values exceed a preset involvement level threshold among the four types of involvement value values for each observed variable with the fourth involvement level value of the observed variable, A monitoring method for performing soundness diagnosis of a target process and evaluating product quality, and identifying a cause of a defect, wherein an observation variable corresponding to a frequency value is used as a factor candidate at the time of the defect detection and the latest time.   In the determination determining step of claim 1, at least two types of involvement values out of the four types of involvement level values according to claim 7 or 8 exceed a preset involvement level threshold value. If all of the four types of involvement values are smaller than the involvement level threshold and there is no observation variable that should be a factor candidate, the number of determinations continues for four or more consecutive times. It is decided that the defect is a true defect, and on the other hand, if the degree of involvement value of only one of the above four types of degree of involvement exceeds a preset degree of involvement threshold, It is judged that the cause of the failure in the next judgment result is disturbance or the like, and the size of all of the four types of involvement values is smaller than the involvement threshold and there are no observation variables that should be candidate factors. Factors are unknown if Monitoring method pending determination, we evaluate the soundness diagnosis and product quality of the subject process, wherein a proceed to quality determination in the next time slot group, performs a specific failure cause.   In the preceding factor candidate extraction step of claim 1, in the case of the adjacency comparison method of claim 2, from the time data of the defect detection and the latest two time zone data, or of the initial reference method of claim 2. In this case, using the rear-side weighting coefficient obtained from the initial time zone and the two time zone data at the time of defect detection, each feature amount for each preceding time zone three before the detection time zone is calculated, When the degree of involvement value of each observation variable obtained by calculating the product of the time zone data and the feature amount is sequentially obtained by tracing back the time zone, and when the preset degree of involvement threshold is exceeded, , Monitoring method to identify failure cause by performing soundness diagnosis of target process and evaluating product quality, characterized by selecting corresponding observation variables as factor candidates   In the observation variable group classification step and the causal model building step of claim 1, the first involvement value out of the first involvement degree value of each observation variable obtained in claim 7 or 8 The observation variable group having the maximum degree value or the third involvement degree value is the first classification, and the observation variable group having the maximum second involvement value or the fourth involvement degree value is the second classification, The observation variable group that is a factor candidate in the preceding time period selected in 10 is set as the third classification, each of the three classifications is regarded as one latent variable, and the observation variable group corresponding to each latent variable is set as the result group. A monitoring method for identifying a cause of failure by performing soundness diagnosis of a target process and evaluating product quality, which is characterized by constructing a causal model.   The causal model construction according to claim 11, wherein the latent variable corresponding to the observation variable group of the third classification is set as a common cause of the latent variable groups corresponding to the first classification and the second classification. A monitoring method that diagnoses the soundness of the target process and evaluates the product quality, and identifies the cause of defects.