JP2015174256A - Fault prediction system, fault prediction device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate a prediction model capable of accurately predicting generation of a fault in a device to be monitored.SOLUTION: Learning data adjustment means 32 sets a fault symptom period which is the period dated back based on a generation point of time of a fault and in which a symptom of the fault is predicted to appear for every case of the fault, on the basis of data on a use situation of an image forming apparatus and a collection point of time of the data, classifies operation data on the image forming apparatus into fault symptom period data or non-fault symptom period data by whether or not the collection point of time of the data belongs to the fault symptom period, fault predictor learning means 33 performs learning by using the operation data classified into the fault symptom period data or the non-fault symptom period data by the learning data adjustment means 32, and generates a fault predictor (prediction model of the fault).

Description

  The present invention relates to a failure prediction system, a failure prediction device, and a program.

As an image forming apparatus having a function of forming an image on a recording material such as paper, a copying machine, a printer apparatus, a facsimile apparatus, and a multifunction machine having these functions are known.
In such an image forming apparatus, when a failure (including failure or malfunction) that hinders its operation occurs, it causes inconvenience to the user of the image forming apparatus. Therefore, the occurrence of a failure in the image forming apparatus is predicted, and the use of the image forming apparatus can be performed immediately before the occurrence of the failure or after the failure has occurred so that necessary measures such as parts replacement and repair can be performed promptly. There is a desire to reduce the time during which is limited.

Up to now, various inventions have been proposed regarding failure prediction for an apparatus such as an image forming apparatus.
For example, in Patent Document 1, a failure occurrence prediction probability is calculated based on a data distribution of a period before failure made by acquiring ΔT1 units before each failure and a data distribution of a period in which no failure occurs. The invention is disclosed.
For example, Patent Document 2 discloses an invention in which a failure prediction period and a normal period are visually estimated, learning data is created, and learning is performed using a boosting method.

JP 2013-109483 A JP 2009-128636 A

  An object of the present invention is to propose a technique for generating a prediction model capable of accurately predicting the occurrence of a failure in a monitored apparatus.

  The present invention (1) includes a first storage means for storing operation parameter data used for operation control by the monitored device and a collection time point thereof, a second storage means for storing a failure occurrence time in the monitored device, Third storage means for storing monitoring device usage status data and its collection time point, and usage status data stored in the third storage means and its collection time point, stored in the second storage means A failure sign period that is assumed to be a sign of a failure and that is traced back to the time of occurrence of the failure is set for each case of the failure, and the operation parameter data stored in the first storage means is collected. Based on the classification means for classifying according to whether the time point belongs to the failure sign period and the classification result of the operation parameter data by the classification means, the prediction model used for the failure prediction of the monitored device is generated And forming means, a fault prediction system, characterized by comprising a.

  According to the present invention (2), in the present invention (1), a correspondence relationship between the trend of the usage status of the monitored apparatus and the length of the failure sign period is predetermined, and the classification means uses the usage status at the time of occurrence of the failure. The length of the failure sign period is adjusted for each case of the failure in light of the correspondence relationship, and the failure sign period is set by applying the adjusted length of the failure sign period to each case of the failure. It is a failure prediction system characterized by this.

  According to the present invention (3), in the present invention (1) or (2), the prediction model calculates a failure occurrence probability in the monitored device, and the failure prediction system is configured to detect a failure in the monitored device. Predicting means for calculating a failure occurrence probability using the prediction model for a plurality of time points before the occurrence time point, and the classification means includes a time-series change of failure occurrence probability calculated by the prediction means and failure occurrence Based on the relationship with the usage data at the time of collection corresponding to the time of occurrence, the length of the failure symptom period is adjusted for each case of failure, and the setting of the failure symptom period and the classification of the operation parameter data are performed again. The generation unit is a failure prediction system, wherein the prediction model is generated again based on data of operation parameters after the classification is performed again by the classification unit.

  According to the present invention (4), in the present invention (1) to (3), the third storage means stores usage status data of the monitored device and its collection time point for a plurality of types of usage status, and the classification means Sets a failure symptom period for each case of failure based on the data of the plurality of types of usage status stored in the third storage means and the time of collection, and data of operation parameters stored in the first storage means Is classified according to whether or not the collection point belongs to a failure sign period.

  The present invention (5) includes a first storage means for storing operation parameter data used for operation control by the monitored device and a time point for collecting the data, a second storage means for storing a failure occurrence time in the monitored device, Third storage means for storing monitoring device usage status data and its collection time point, and usage status data stored in the third storage means and its collection time point, stored in the second storage means A failure sign period that is assumed to be a sign of a failure and that is traced back to the time of occurrence of the failure is set for each case of the failure, and the operation parameter data stored in the first storage means is collected. Based on the classification means for classifying according to whether the time point belongs to the failure sign period and the classification result of the operation parameter data by the classification means, the prediction model used for the failure prediction of the monitored device is generated And forming means, a predictive failure device characterized by comprising a.

  According to the present invention (6), a first storage function for storing operation parameter data used by the monitored device for operation control and a collection time point in the computer, and a second storage function for storing a failure occurrence time in the monitored device. And a third storage function for storing the usage status data of the monitored device and the time of collection thereof, and the second storage function based on the usage status data stored by the third storage function and the time of collection thereof. Operation parameter data stored by the first storage function by setting a failure sign period for each failure case, which is a period that is traced back to the stored occurrence point of the failure and is assumed to have a failure sign. A function of classifying the data according to whether or not the collection point belongs to a failure indication period, and a failure prediction of a monitored device based on a result of classification of operation parameter data by the classification function Is a program for realizing a generation function of generating a prediction model used.

  According to the present invention (1), (5), and (6), it is possible to perform prediction with higher accuracy in the generation of a prediction model for predicting the occurrence of a failure in a monitored device than when the present invention is not applied. A predictive model can be generated.

  According to the present invention (2), the process of setting the failure sign period for each failure case in the generation of the prediction model can be easily performed.

  According to the present invention (3), the failure sign period set for each failure case in generating the prediction model can be made more appropriate.

  According to the present invention (4), it is possible to set a failure symptom period in consideration of a plurality of types of usage conditions in the monitored apparatus.

It is a figure explaining the basic operation | movement of the failure prediction by the failure prediction system which concerns on one Embodiment of this invention. It is a figure explaining the production | generation of a discriminator. It is a figure which shows the example of learning data. FIG. 6 is a diagram for explaining a difference in failure sign period due to a difference in usage status of an image forming apparatus. It is a figure explaining the conventional learning method. It is a figure explaining the learning method in the 1st example of composition. It is a figure explaining the learning method in a 2nd structural example and a 3rd structural example. It is a figure which shows the example of the functional block of the failure prediction system which concerns on a 1st structural example. It is a figure which shows the structural example of an apparatus operation condition memory | storage means. It is a figure which shows the example of the table | surface which classifies the tendency of each usage condition into a level value on the basis of design specifications. It is a figure which shows the example of the table | surface which replaced the design specification in the table | surface of FIG. 10 with the specific numerical value. It is a figure which shows the example of the table | surface which defined the corresponding | compatible relationship between the combination of each level value, and the length of a failure symptom period. It is a figure which shows the example of the functional block of the failure prediction system which concerns on a 2nd structural example. It is a figure which shows the example of transition of the failure occurrence probability with progress of time. It is a figure which shows the example of the relationship between the number of days after a failure occurrence probability exceeds a threshold value until it leads to a failure, and the tendency of a single kind of usage condition. It is a figure which shows the example of the processing flow by the failure prediction system which concerns on a 2nd structural example. It is a figure which shows the example of the detailed process flow which concerns on the production | generation of a failure symptom period length estimator. It is a figure which shows the example of the functional block of the failure prediction system which concerns on a 3rd structural example. It is a figure which shows the example of the relationship between the days after a failure occurrence probability exceeds a threshold value until it reaches a failure, and the tendency of multiple types of usage. It is a figure which shows the example of the table | surface used for estimating the length of a failure symptom period from the number of printed sheets of the last 10 days. It is a figure which shows the example of the graph showing the relationship between the number of printed sheets of the last 10 days, and the number of days until a failure occurs. It is a figure which shows the example of the graph showing transition of the failure occurrence probability with progress of time. It is a figure which shows the example which compared the failure occurrence probability calculated with the method of the prior art example, and the failure occurrence probability calculated in Example 2.

An embodiment of the present invention will be described with reference to the drawings.
In the following description, a failure prediction system in which an image forming apparatus (an example of a monitored device) is a target of failure prediction will be described as an example. Here, the image forming apparatus is an apparatus that performs an image forming process for forming an image on a recording material such as paper, and includes a combination of functions of these apparatuses in addition to apparatuses such as a printer, a copier, and a facsimile apparatus. Multi-function machines are also included.

First, the basic operation of failure prediction by the failure prediction system of this example will be described with reference to FIG.
As shown in FIG. 1, the failure prediction system of the present example includes a learning stage for generating a discriminator 13 used for fault prediction and a test stage for performing fault prediction using the generated discriminator 13.
The discriminator 13 discriminates whether the image forming apparatus targeted for failure prediction is in a normal state or a failure sign state (including a state where a failure has occurred) in a stage before the failure (a stage where a failure sign appears). To do. Here, a probability of failure occurrence in the image forming apparatus (hereinafter referred to as failure occurrence probability) P is calculated and compared with a threshold value to determine whether the image forming apparatus is in a normal state or a failure sign state. And

In the learning stage, the learning data 11 is input to the learning device 12, and the learning device 12 generates the discriminator 13 based on the learning data 11. The learning data 11 is data used for generation of the discriminator 13, and mainly data collected at a past time point from the image forming apparatus is used.
In the test stage, test data 14 is input to the discriminator 13, and the discriminator 13 performs failure prediction based on the test data 14. The test data 14 is data used for calculating the failure occurrence probability P of the image forming apparatus and calculating the accuracy thereof, and mainly data collected from the image forming apparatus at the current time point is used. .

The generation of the discriminator in the learning stage will be described with reference to FIG.
For simplicity of explanation, a discriminator that discriminates two types of similar flowers using a logistic regression model will be described instead of a discriminator related to failure prediction.
In the example of FIG. 2, based on a plurality of learning data that associates the length and width of the flower that is the condition data with the type of the flower that is the response data, the probability that the flower type is B according to the following formula: A discriminator for calculating P is generated.
P = 1 / {1 + exp (363-132 * length + 110 * width)}
If condition data consisting of flower length = 5.5 and width = 3.7 is input to this discriminator as test data, the probability that the flower type is B is 2 * 10 ⁻²⁰ , and the flower type is It is determined that it is not B. Further, when condition data having a flower length = 5.1 and a width = 2.4 is input as test data, the probability that the flower type is B is 1, and it is determined that the flower type is B. .
If there is answer data for the condition data (test data), the accuracy of the discriminator (probability calculation accuracy) can be calculated by matching the discrimination results against the answer data.

Here, in the failure prediction system of this example, as shown in FIG. 3, the data of various operation parameters collected from the image forming apparatus at the past time is used as the condition data, and the presence / absence of a failure sign at that time (normal state) A discriminator for predicting a failure of the image forming apparatus is generated using the learning data in which the response data is associated with each other.
The learning data in the example of FIG. 3 includes the date, values of the operation parameters X, Y, and Z used by the image forming apparatus for operation control of the image forming process, and the presence / absence of a sign of failure. In the example shown in the figure, it is unclear whether or not a sign of failure appears in 8/20 to 23, a sign of failure appears in 8/24, and a trouble actually occurs in 8/25. Yes.

Note that it is unclear whether or not a sign of failure appears in 8/20 to 23, depending on the use condition of the image forming apparatus, and the sign of failure (abnormality) in the operation parameters of the image forming apparatus. This is because the length of the period in which () appears (failure sign period) is different.
FIG. 4 shows a graph (a) representing a time-series change in operation parameters in an image forming apparatus, and a graph (b) representing a time-series change in operation parameters in another image forming apparatus having a different usage status. Is shown. The graph (a) has a relatively long period in which a sign of failure appears in the operating parameter, whereas the graph (b) has a relatively short period in which the sign of failure appears in the operating parameter.
In this way, the length of the period (failure sign period) in which a sign of failure appears in the operating parameter varies depending on the usage status of the device. Will be created.

In the conventional learning method, as shown in FIG. 5, when operation parameter data collected at a plurality of past points in time is received as learning data 11 from the image forming apparatus, the learning data 11 is not used using a fixed failure sign period. It classifies into failure sign period data 11a and failure sign period data 11b, and inputs to the learning device 12 to generate the discriminator 13.
In the configuration using the fixed failure sign period, it is not considered that the length of the trouble sign period varies depending on the usage status of the apparatus. Therefore, the operation parameter data of the image forming apparatus in a normal state (non-failure sign period data) 11a) is classified as failure sign period data 11b, or operation parameter data (data to be classified as failure sign period data 11b) when the image forming apparatus is abnormal is non-failure sign period data. In some cases, the accuracy of the discriminator 13 may be lowered.
Therefore, in the present invention, the length of the failure sign period is adjusted for each failure case, and the non-failure sign period data 11a and the failure sign period data 11b are appropriately separated and used for generation of the discriminator 13. The accuracy of 13 was improved.

  In the first configuration example to be described later, as shown in FIG. 6, the learning data adjusting means 21 is provided in the previous stage of the learning device 12. The learning data adjustment unit 21 adjusts the length of the failure sign period based on the use state of the image forming apparatus for each case of the failure, and the learning data 11 is converted into the non-failure sign period data according to the adjustment result. The data is classified into 11a and failure sign period data 11b and input to the learning device 12. Thereby, the learning device 12 can perform learning using appropriately classified data, and can generate the discriminator 13 with high accuracy.

  Further, in the second configuration example and the third configuration example, which will be described later, as shown in FIG. 7, a failure symptom period length estimator generating means 22 and a failure symptom period length estimator 23 are further provided. Then, the learning data 11 is applied to the discriminator 13 generated in the same manner as in the first configuration example, the failure occurrence probability P of the image forming apparatus is calculated, and the failure indication is based on the calculated failure occurrence probability P. The period estimator generation means 22 generates a failure symptom period length estimator 23. Thereafter, the length of the trouble sign period estimated by the trouble sign period length estimator 23 is given to the learning data adjustment means 21, so that the learning data 11 is more appropriately based on the length of the trouble sign period. The data is reclassified into non-failure sign period data 11 a ′ and fault sign period data 11 b ′ and input to the learning device 12. As a result, the learning device 12 can perform learning using data classified more appropriately, and can generate the discriminator 13 with high accuracy.

Next, the prediction model used for the discriminator 13 will be described using the Naive-Bayes model as an example.
In the Naive-Bayes model, a conditional probability P (T = yes | Data = d) that a failure occurs when a data set Data = d of a plurality of feature quantities (X1, X2,..., Xn) is obtained. ) Is given by (Equation 1) below.

In (Formula 1), T = yes means that a failure occurs, and T = no means that no failure occurs.
P (T = yes) is a probability (prior probability) that a failure will occur, P (T = no) is a probability (prior probability) that a failure will not occur, and P (T = yes) + P ( T = no) = 1.
P (Data = d | (T = yes)) is a probability that Data = d when a failure occurs, and P (Data = d | (T = no)) is a failure. If not, the probability is Data = d.
Here, when the denominator and the numerator are divided by P (T = yes) · P (Data = d | T = yes) of the numerator of (Expression 1), (Expression 1) can be transformed into the following (Expression 2).

  In the Naive-Bayes model, since each feature amount is assumed to be independent, P (Data = d | T = no) / P (Data = d | T = yes) in the denominator of (Equation 2) further includes a plurality of features. Since it can be expressed as the product of the distribution ratio for each quantity, (Equation 2) can be transformed into (Equation 3) below.

Therefore, the discriminator 13 performs calculation by applying the values of the operation parameters collected from the image forming apparatus (for example, the values of the parameters X, Y, and Z in FIG. 3) to (Equation 3), so that Data = d. The failure occurrence probability P (T = yes | Data = d) is calculated for the image forming apparatus that has obtained the above.
In the above description, the Naive-Bayes model is used. However, another prediction model may be used.

<First configuration example>
FIG. 8 shows an example of functional blocks of the failure prediction system according to the first configuration example.
The failure prediction system according to the first configuration example includes a device operation state storage unit 31, a learning data adjustment unit 32, a failure prediction, and a failure prediction server that is communicably connected to one or more image forming apparatuses that are targets of failure prediction. This is a system configuration provided with a device learning means 33.
In the first configuration example, the learning data adjustment unit 32 corresponds to the learning data adjustment unit 21 in FIG. 6, and the failure predictor learning unit 33 corresponds to the learning unit 12 in FIG. The failure predictor to be learned corresponds to the discriminator 13 in FIG.
In the following, only the description of the learning stage will be given, and the description of the test stage will be omitted.

As shown in FIG. 9, the apparatus operation status storage means 31 includes an operation data storage means 41, a failure occurrence information storage means 42, and an apparatus usage status storage means 43, and other function units can access them as necessary. Thus, data can be written or read.
Although not specified below, these storage units 41 to 43 store each data in association with the identification information of the image forming apparatus so that each image can be identified with respect to which image forming apparatus. It is.

The operation data storage unit 41 collects data of operation parameters (so-called image formation parameters) used for operation control of image forming processing from an image forming apparatus (an example of a monitored apparatus), and operates in association with the collection date and time. Store as data.
In this example, as the operation parameter data of the image forming apparatus, measured values and target values of various image forming parameters such as charging potential, developing bias, exposure power, and toner density are collected. In addition, statistical values such as standard deviations and correlation coefficients calculated based on the values at the plurality of points may be collected as operation parameter data. In short, there are various types that can contribute to failure prediction of the image forming apparatus. It is only necessary to collect data.
The collection of data of these operation parameters is preferably performed every day while the image forming apparatus is operating, and may be collected a plurality of times a day.

The failure occurrence information storage means 42 collects failure data (information indicating the type of failure) from the image forming apparatus in which the failure has occurred, and stores it as failure occurrence data in association with the failure occurrence date and time.
Examples of the trigger for collecting failure data include the generation of a fault code notifying an abnormal operation of the image forming apparatus or the input of service visit maintenance information requested by a user of the image forming apparatus. .

The apparatus usage status storage unit 43 collects data on the usage status of the image forming apparatus when the fault occurs from the image forming apparatus in which the fault has occurred, and stores the data as apparatus usage status data in association with the collection date and time. .
The usage status of the image forming apparatus includes an index value (for example, the number of printed sheets) indicating the operation rate of the image forming apparatus, an elapsed usage time of the image forming apparatus, a maintenance status of the image forming apparatus, and environmental information such as temperature and humidity. Various information can be used to indicate how the forming apparatus is being used. These usage conditions affect the length of the failure sign period from when a failure sign appears in the image forming apparatus until the actual failure occurs.
The device usage status data is collected at the same timing as the operation data collection.

The learning data adjustment unit 32 sets a failure sign period for each failure case based on the failure occurrence data and the device usage status data stored in the device operation status storage unit 31 and is stored in the device operation status storage unit 31. The operation data is classified into failure sign period data and non-failure sign period data.
That is, considering the fact that the difference in the usage status of the image forming apparatus affects the length of the fault indication period, the length of the fault indication period is determined based on the tendency of the apparatus usage status data at the time of occurrence of the fault. Adjustment is made for each failure case, and a period that goes back the adjusted length from the occurrence date of the failure is specified for each failure case, and the specified period is set as a failure sign period. Then, the operation data collected during the failure sign period set for each failure case is classified as failure sign period data, and the operation data collected during other periods is classified as non-failure sign period data.
In this example, a label for distinguishing between the failure sign period data and the non-failure sign period data is assigned to each action data, so that it is possible to grasp which action data is classified.
For operation data collected from an image forming apparatus that is in a special state of use of the apparatus, such as an image forming apparatus with extremely low use frequency of the apparatus, the failure sign period data and the non-failure sign period data are also included. You may make it not classify.

The failure predictor learning unit 33 learns using the operation data classified (labeled in this example) into the failure sign period data or the non-failure sign period data by the learning data adjustment unit 32, and the failure predictor (failure prediction). Model).
Learning methods include methods such as discriminant analysis, support vector machines, and Bayesian networks, but are not limited thereto.
Note that the processing of the learning data adjustment unit 32 and the failure predictor learning unit 33 is performed for each type of failure.

The adjustment of the length of the failure sign period according to the usage status of the image forming apparatus will be described with reference to FIGS.
The image forming apparatus has been developed with a design specification that assumes a certain standard usage situation. The design specification (standard usage situation) is compared with the actual usage situation at the time of occurrence of the fault, and the fault is determined. Adjust the length of the sign period.

FIG. 10 is a table that divides numerical data representing a tendency of an actual usage situation at the time of failure occurrence into five levels of levels “1” to “5” with respect to various usage situations based on the design specifications. An example is shown. (A) is an example of a table that divides the total PV (number of printed sheets) for the most recent 10 days into level values, and (b) is the level of the number of maintenance (for example, the number of color adjustment calibrations) for the last 100 days. (C) is an example of a table that divides the number of years of use of the device into standard values, and (d) is a table that classifies the average environmental temperature over the last 10 days into standard values. It is an example of a table | surface, (e) is an example of the table | surface which classifies the average environmental humidity for the last 10 days into a level value.
FIG. 11 shows an example in which the design specifications in the table of FIG. 10 are replaced with specific numerical values.

FIG. 12 shows an example of a table in which the correspondence relationship between the combination of each level value obtained in light of the usage status of the image forming apparatus and the length of the failure symptom period is shown in the tables of FIGS.
The learning data adjusting means 32 is numerical data (for each case of failure) that represents the tendency of various usage statuses based on the device usage status data at the time of occurrence of the failure or the device usage status data at a plurality of time points before that. The total PV for the most recent 10 days) is calculated, the calculated numerical data is compared with the tables of FIG. 10 and FIG. 11 to obtain the level value for each type of usage status, and the combinations of the acquired level values are shown in FIG. Determine the length of the symptom period in light of the table.

  As described above, in the failure prediction system according to the first configuration example, the length of the failure sign period can be adjusted based on the actual usage situation at the time of failure occurrence. By classifying operation data by setting a sign period and performing learning based on the result, a failure predictor capable of accurately performing failure prediction can be generated.

<Second configuration example>
FIG. 13 shows an example of functional blocks of the failure prediction system according to the second configuration example.
The failure prediction system according to the second configuration example includes a device operation status storage unit 31, a learning data adjustment unit 32, and a failure prediction in a failure prediction server that is communicably connected to one or more image forming apparatuses that are targets of failure prediction. In this system configuration, a failure learning means 33, failure occurrence probability calculation means 34, failure symptom period length estimator generation means 35, and failure symptom period length estimator 36 are provided.
In the second configuration example, the learning data adjustment unit 32 corresponds to the learning data adjustment unit 21 in FIG. 7, the failure predictor learning unit 33 corresponds to the learning unit 12 in FIG. 7, and the failure occurrence probability calculation unit 34 7, the failure symptom period length estimator generation means 35 corresponds to the failure symptom period length estimator generation means 22 in FIG. 7, and the failure symptom period length estimator 36 corresponds to FIG. 7 corresponds to the length estimator 23 of the failure sign period.
In the following, description of functional units similar to those in the first configuration example (see FIG. 8) is omitted, and only differences are described.

The failure occurrence probability calculating unit 34 uses the failure predictor (failure prediction model) generated by the failure predictor learning unit 33 to calculate the failure occurrence probability of the image forming apparatus. That is, by inputting the operation data stored in the apparatus operation status storage unit 31 for the image forming apparatus targeted for failure prediction to the failure predictor, the failure occurrence probability of the image forming apparatus is calculated.
The failure occurrence probability calculation means 34 is used in the learning stage and the test stage.
At the learning stage, the operation data collected in the past (operation data as learning data 11) is repeatedly calculated while changing the collection time, so that the transition of failure occurrence probability (time-series change) with time elapses. It becomes possible to capture. The failure occurrence probability calculated in the learning stage is given to the failure symptom period length estimator generation means 35.
In the test stage, the operation data collected at the present time (operation data as the test data 14) is input to the failure predictor, so that the failure occurrence probability at the present time can be calculated. The failure occurrence probability calculated in the test stage is notified to a preset notification destination (for example, a user or administrator of the image forming apparatus, a service person in charge of maintenance).

The failure symptom period length estimator generation unit 35 is based on the failure occurrence probability calculated by the failure occurrence probability calculation unit 34 and the device usage status data stored in the device operation status storage unit 31 in the learning stage. A failure symptom period length estimator 36 is generated that estimates the length of the failure symptom period from one type (single type) usage trend.
Specifically, as shown in FIG. 14, for each case of failure, the length of the failure sign period from when the failure occurrence probability exceeds a certain threshold value V until reaching the failure while maintaining the exceeded state ( Count). Then, a failure sign period length estimator 36 is generated based on the relationship between the tendency of a single type of usage situation in each case of a failure and the length of the failure sign period.

FIG. 15 shows an example of the relationship between the number of days until a failure occurs after the failure occurrence probability exceeds a certain threshold value V, and the number of printed sheets (PV) in the last 10 days of the failure occurrence. According to the figure, the number of days until the failure occurs after the failure occurrence probability exceeds a certain threshold V is shorter as the number of printed sheets in the last 10 days is larger. It can be seen that it fluctuates accordingly.
Therefore, the failure sign period length estimator generating means 35 of this example performs a regression analysis on the relationship between these two variables, thereby determining the period from the failure occurrence probability exceeding the threshold V to the occurrence of the failure as the failure sign period. As described above, the failure symptom period length estimator 36 that estimates the length of the failure symptom period from the number of printed sheets in the latest 10 days is generated.
The generation of the failure symptom period length estimator 36 is not limited to the regression analysis as described above, and may be performed by other methods.

  The failure symptom period length estimator 36 estimates the length of the failure symptom period from the number of printed sheets for 10 days immediately before the occurrence of the failure for each case of the failure in the learning stage, and the result is sent to the learning data adjustment unit 32. input.

When the learning data adjustment unit 32 receives an input of the length of the failure symptom period from the learning data adjustment unit 32 in the learning stage, the learning data adjustment unit 32 resets the failure symptom period according to this input and stores the failure symptom period in the apparatus operation status storage unit 31. Redo the classification of motion data.
Further, the failure predictor learning unit 33 is based on the operation data reclassified into the failure symptom period data or the non-failure symptom period data by the learning data adjustment unit 32 in the learning stage, and the failure predictor (failure prediction model). Is regenerated.
Here, in the second configuration example, the length of the failure symptom period is adjusted twice or more. Therefore, in the first adjustment, the length of the failure symptom period may be uniform for each failure.

  A processing flow of the failure prediction system according to the second configuration example will be described with reference to FIGS. 16 and 17. FIG. 16 shows an example of the processing flow in the learning stage by the failure prediction system according to the second configuration example.

In the failure prediction system according to the second configuration example, first, an input of the number N of repetitions of generation of the failure symptom period length estimator 36 based on the usage status of the image forming apparatus is received (step S11).
Further, an input as to whether or not the failure symptom period is made uniform in initial learning (first learning) is accepted, and the input value is determined (step S12).
These inputs may be appropriately input by the user when starting the process, or those used in the previous process may be stored and input.

  In step S12, when the input indicating that the failure sign period is made uniform in the initial learning is made (Yes), the learning data adjustment unit 32 sets the length of the uniform failure sign period for all the failure cases. Is applied to classify the operation data stored in the apparatus operation status storage unit 31 according to the setting, and then the failure predictor learning unit 33 performs learning based on the classified operation data. Then, a failure predictor used by the failure occurrence probability calculating means 34 is generated (step S13).

  On the other hand, when an input indicating that the failure sign period is not made uniform in the initial learning is made in step S12 (No), the learning data adjustment unit 32 uses the method of the first configuration example for each failure case. The failure indication period is set by adjusting the length of the period, and the operation data stored in the apparatus operation status storage unit 31 is classified according to the setting, and then the failure predictor learning unit 33 performs the classified operation data The failure predictor used by the failure occurrence probability calculation means 34 is generated by performing learning based on (Step S14).

Next, the failure occurrence probability calculation means 34 calculates a failure occurrence probability for each case of failure using the failure predictor generated by learning in step S13 or S14 (step S15).
Next, based on the failure occurrence probability calculated by the failure occurrence probability calculation unit 34 and the device usage status data stored in the device operation status storage unit 31, the failure sign period length estimator generation unit 35 A failure symptom period length estimator 36 that estimates the length of the failure symptom period with respect to the usage status trend is generated (step S16).
Next, the learning data adjustment unit 32 resets the failure symptom period according to the length of the failure symptom period estimated by the failure symptom period length estimator 36, and classifies the operation data stored in the apparatus operation status storage unit 31. The failure predictor learning unit 33 regenerates a failure predictor based on the operation data after reclassification by the learning data adjustment unit 32 (step S17).

FIG. 17 shows an example of a detailed processing flow related to generation of the failure symptom period length estimator 36 (step S16).
First, a variable i representing the current number of repetitions is initialized to 1 (step S21).
Next, the length (number of days) of the failure sign period in which the failure occurrence probability exceeds the threshold before the failure occurrence time is calculated (step S22).
Next, the relationship between the tendency of the usage status of the image forming apparatus and the length of the failure sign period is learned, and the trouble sign period length estimator 36 is generated (step S23).

Next, it is determined whether or not the condition of i ≦ N−1 is satisfied (step S24).
If it is determined in step S24 that the condition of i ≦ N−1 is satisfied (Yes), the length of the fault sign period is estimated using the generated fault sign period length estimator 36, and based on the result. Then, the failure sign period setting, the operation data classification, and the failure predictor generation are performed again (step S25). Then, using the generated failure predictor, a failure occurrence probability is calculated for each failure case (step S26), and the process returns to step S22.
On the other hand, when it is determined that the condition of i ≦ N−1 is not satisfied (No), the process of step S16 is finished.

  As described above, in the failure prediction system according to the second configuration example, by repeating a series of processes such as adjustment of the length of the failure symptom period, classification of operation data, and generation of a failure predictor, the failure symptom period Adjustment of length and classification of operation data can be made more appropriate, and as a result, a failure predictor with higher accuracy can be generated.

<Third configuration example>
FIG. 18 illustrates an example of functional blocks of the failure prediction system according to the third configuration example.
The failure prediction system according to the third configuration example includes a device operation status storage unit 31, a learning data adjustment unit 32, and a failure prediction in a failure prediction server that is communicably connected to one or more image forming apparatuses that are targets of failure prediction. In this system configuration, a failure learning means 33, failure occurrence probability calculation means 34, failure symptom period length estimator generation means 35, failure symptom period length estimator 36, and prediction result notification means 37 are provided.
In the third configuration example, similarly to the second configuration example, the learning data adjustment unit 32 corresponds to the learning data adjustment unit 21 in FIG. 7, and the failure predictor learning unit 33 corresponds to the learning unit 12 in FIG. 7. The failure occurrence probability calculating means 34 corresponds to the discriminator 13 in FIG. 7, and the failure symptom period length estimator generating means 35 corresponds to the failure symptom period length estimator generating means 22 in FIG. The period length estimator 36 corresponds to the failure symptom period length estimator 23 of FIG.
In the following, description of functional units similar to those in the first configuration example (see FIG. 8) and the second configuration example (see FIG. 13) will be omitted, and only differences will be described.

The failure symptom period length estimator generation unit 35 is configured to select a plurality of types based on the failure occurrence probability calculated by the failure occurrence probability calculation unit 34 and the device usage status data stored in the device operation status storage unit 31 in the learning stage. A failure symptom period length estimator 36 is generated that estimates the length of the failure symptom period from the trend of the usage status. That is, in the second configuration example, the length of the failure sign period is estimated from the tendency of the single type of usage situation, whereas in the third configuration example, the length of the failure sign period is calculated from the tendency of the multiple types of usage situation. Estimate.
Specifically, as shown in FIG. 14, for each case of failure, the length of the failure sign period from when the failure occurrence probability exceeds a certain threshold value V until reaching the failure while maintaining the exceeded state ( Count). Then, the failure sign period length estimator 36 is generated based on the relationship between the tendency of the plurality of types of usage situations in each case of the failure and the length of the failure sign period.

FIG. 19 shows the number of days until the failure occurs after the failure occurrence probability exceeds a certain threshold value V, the number of printed sheets (PV) for the most recent 10 days of the failure occurrence, and the number of maintenance operations for the most recent 100 days before the failure occurrence ( For example, an example of the relationship between the number of color adjustment calibrations), the average environmental temperature for the last 10 days before the occurrence of the failure, and the average environmental humidity for the last 10 days before the occurrence of the failure is shown. According to the figure, the number of days until the failure occurs after the failure occurrence probability exceeds a certain threshold V is not limited to the number of printed sheets in the most recent 10 days, but also the number of maintenance operations in the most recent 100 days and the average environment over the most recent 10 days. It can be seen that it fluctuates under the influence of temperature and humidity.
Therefore, the failure symptom period length estimator generating means 35 of this example performs a multiple regression analysis on the relationship between these many variables, thereby determining the period from the failure occurrence probability exceeding the threshold V to the occurrence of the failure. As a period, a failure sign period length estimator 36 that estimates the length of the trouble sign period from the number of printed sheets for the last 10 days before the occurrence of the trouble, the number of maintenance operations for the last 100 days, the average environmental temperature and the average environmental humidity for the last 10 days Generate.
The generation of the failure symptom period length estimator 36 is not limited to the multiple regression analysis as described above, and may be performed by other methods.

  In the learning stage, the failure symptom period length estimator 36 determines a failure from the number of printed sheets for the last 10 days before the occurrence of the failure, the number of maintenance operations for the last 100 days, the average environmental temperature and the average environmental humidity for the last 10 days. The length of the sign period is estimated, and the result is input to the learning data adjustment unit 32.

The prediction result notifying means 37 indicates the failure occurrence probability calculated by the failure occurrence probability calculating means 34 at the test stage, in accordance with a preset notification destination (for example, a user or administrator of the image forming apparatus, or a service person in charge of maintenance). Etc.).
In this example, notification is performed when the failure occurrence probability exceeds a predetermined notification threshold, but conditions for notification can be arbitrarily determined according to various requirements such as the system operation mode. That is, for example, notification may be performed when a state exceeding the notification threshold continues for a certain period, notification may be performed unconditionally, or conditions for performing notification may be changed for each image forming apparatus. .

  As described above, in the failure prediction system according to the third configuration example, the adjustment of the length of the failure symptom period and the classification of operation data can be made more appropriate in consideration of combinations of various usage situations. As a result, it is possible to generate a failure predictor with higher accuracy.

In the first embodiment, the failure prediction system (failure prediction model) for predicting the failure of the conveyance belt in the image forming apparatus is learned as follows by the failure prediction system having the system configuration shown in the first configuration example. .
That is, a drive error rate reflecting the operation state of the conveyor belt is acquired from the image forming apparatus, and a tendency of increase / decrease in the drive error rate is calculated. Further, the number of printed sheets by the image forming apparatus was acquired as the usage status of the image forming apparatus. These pieces of information were acquired every 2 hours while the image forming apparatus was operating. Also, the point of failure is defined as the point at which a fault code that indicates abnormal operation of the conveyor belt is generated. Here, the longer the number of printed sheets of the conveyor belt, the shorter the period from when the failure sign appears until the failure occurs, so the length of the failure sign period is changed depending on the number of printed sheets.

FIG. 20 shows an example of a table used to estimate the length of the failure symptom period from the number of printed sheets for the last 10 days. In the first embodiment, by using a table as shown in the figure, the length of the failure sign period is estimated for each failure case to set the failure sign period, and the failure sign period data and the non-failure sign period are set in the operation data. The label which distinguishes data was given, and the failure predictor was learned by the Bayesian network using the action data (the classification result of action data) after the label was given.
As described above, in this embodiment, for each failure example of the conveyance belt, the failure sign period is set by adjusting the length of the failure sign period according to the number of printed sheets, and classification and learning are performed based on the result. As a result, a failure predictor capable of accurately predicting the failure of the conveyor belt can be generated.

In the second embodiment, the failure prediction system (failure prediction model) for predicting the failure of the conveyance belt in the image forming apparatus is learned as follows using the failure prediction system having the system configuration as shown in the second configuration example. .
That is, as in the first embodiment, the drive error rate is acquired, the tendency of the increase / decrease is calculated, and the number of printed sheets as the usage status of the image forming apparatus is acquired. In addition, the time when a fault code for notifying the operation abnormality of the conveyor belt is set as the failure point, and the length of the failure sign period is changed depending on the number of printed sheets.

Then, for each failure case, the failure sign period is adjusted by adjusting the length of the failure sign period in the same manner as in the first embodiment, and the failure is detected by the Bayesian network using the operation data classified based on the result. The predictor was learned. Next, with the learned failure predictor, the failure occurrence probability before the failure occurrence is calculated for each failure case, the number of days from when the failure occurrence probability exceeds the threshold until the failure occurs is counted, and the result is As illustrated in FIG. 21, the number of printed sheets for the last 10 days is plotted on a horizontal axis, and the number of days from when the failure occurrence probability exceeds a threshold to when a failure occurs is plotted on a graph. Next, a regression analysis is performed on the number of printed sheets for the last 10 days and the number of days until the failure occurs after the failure occurrence probability exceeds the threshold, and the calculation formula for calculating the length of the failure indication period of the conveyor belt from the number of printed sheets Asked.
Then, using the calculation formula obtained above, estimate the length of the failure sign period for each failure case and set the failure sign period, and distinguish the failure sign period data from the non-failure sign period data in the operation data A label was assigned, and the failure predictor was learned by the Bayesian network using the action data after the label was attached (the action data classification result).

FIG. 22 shows an example in which the failure occurrence probability calculated for each day using the created failure predictor is plotted on a graph with the date as the horizontal axis and the failure occurrence probability as the vertical axis.
In the example of the figure, the period when the failure occurrence probability exceeds the threshold before the failure is 8 days from 7/30 to 8/6. Here, when the failure occurrence probability is expressed as P (x) as a function of the day x, P (a + 1−n) ≧ 0.5, ∀n = 1, 2,... The maximum b that satisfies the condition is the number of days in the period when the failure occurrence probability exceeds the threshold before the failure.
Then, the calculated number of days (the number of days in which the failure occurrence probability exceeds the threshold before the failure) is regression-analyzed based on the number of printed sheets to derive a calculation formula for calculating the failure sign period. As an example, an equation for calculating the length Y of the failure sign period by the number of printed sheets X is Y = 1.973 × 10 ⁹ × X− ^2.043 .

FIG. 23 shows an example in which the failure occurrence probability calculated by the conventional method is compared with the failure occurrence probability calculated in the present embodiment.
According to the example of FIG. 23A, in the conventional example, after the failure occurrence probability exceeds the threshold value (0.5 in this example), the failure occurrence probability once falls below the threshold value before the failure occurrence date. In the present embodiment, the drop in the failure occurrence probability at that time is small and does not fall below the threshold value. Further, according to the example of FIG. 23B, noise occurs after the failure occurrence probability once exceeds the threshold value in the conventional example, but in this embodiment, the increase in failure occurrence probability at that time is small. The threshold is not exceeded.
Thus, according to the comparative example of FIG. 23, it can be seen that the calculation accuracy of the failure occurrence probability is improved as compared with the conventional example, and the occurrence of the failure can be accurately predicted.
In addition, the false detection rate is calculated from the number of days that the failure occurrence probability exceeded 0.8 before the failure occurrence date and the number of days that the failure occurrence probability exceeded 0.8 days other than the failure occurrence date (number of false detection days). As a result, the false detection rate was 30% in the conventional example, whereas the false detection rate was 7% in the present example. This example has a lower false detection rate than the conventional example. The accuracy of failure prediction has been improved.

  As described above, in this embodiment, for each failure example of the conveyor belt, the failure sign period is set by adjusting the length of the failure sign period according to the number of printed sheets, and classification and learning are performed based on the result. By doing so, it was possible to generate a failure predictor capable of accurately predicting the failure of the conveyor belt.

In the third embodiment, a failure predictor (failure failure in this example) predicts an image quality abnormality (in this example, an image granularity abnormality) with respect to a print image by the image forming apparatus using the failure prediction system having the system configuration as shown in the third configuration example. The prediction model was learned as follows.
That is, the measured value and target value are acquired from the image forming apparatus for each parameter of toner density, patch density, developing bias, charging potential, transfer potential, and fixing temperature, and the difference between the measured value and the target value is calculated for each parameter. did. Further, as the usage status of the image forming apparatus, the temperature and humidity of the installation environment of the image forming apparatus, the number of days elapsed since the developer was replaced, the image density of the printed image, and the number of printed sheets by the image forming apparatus were acquired. These pieces of information were acquired every 2 hours while the image forming apparatus was operating. Further, a failure occurred when the image quality abnormality occurred in the printed image by the image forming apparatus and the user of the image forming apparatus requested the service staff to visit and maintain.

For each case of failure, the length of the failure sign period is adjusted in the same manner as in Example 1 using the number of printed sheets for the last 10 days before the failure occurs, and the failure sign period is set based on the result. The failure predictor was trained by Bayesian network using the classified motion data. Next, with the learned failure predictor, the failure occurrence probability before the failure occurrence is calculated for each failure case, and the number of days until the failure occurs after the failure occurrence probability exceeds the threshold is counted. Based on this, a table is created that correlates the number of days until failure occurs with various usage conditions, and a multiple regression analysis is performed to calculate the length of the failure sign period from various usage conditions. Asked.
Then, using the calculation formula obtained above, estimate the length of the failure sign period for each failure case and set the failure sign period, and distinguish the failure sign period data from the non-failure sign period data in the operation data A label was assigned, and the failure predictor was learned by the Bayesian network using the action data after the label was attached (the action data classification result).
Then, using the learned failure predictor, the failure occurrence probability of image quality abnormality is calculated based on the data newly acquired from the image forming device that is the target of failure prediction, and is notified to the administrator of the image forming device I tried to do it.

  As described above, in this embodiment, for each failure example of image quality abnormality, the failure sign period is set by adjusting the length of the failure sign period according to the use status of a plurality of types such as the number of printed sheets. By classifying and learning based on the above, it was possible to generate a failure predictor capable of accurately predicting failure of image quality abnormality.

  Here, the failure prediction server in the failure prediction system of each of the above configuration examples and embodiments includes a CPU (Central Processing Unit) that performs various arithmetic processes, a RAM (Random Access Memory) that is a work area of the CPU, and basic control. A main storage device such as a ROM (Read Only Memory) that records programs, an auxiliary storage device such as an HDD (Hard Disk Drive) that stores various programs and data, a display device and an operation for displaying and outputting various information Hardware such as an input / output I / F that is an interface with an input device such as an operation button or a touch panel used by a user, and a communication I / F that is an interface for performing wired or wireless communication with other devices Compile with hardware resources It is realized by the computer.

Then, the program according to the present invention is read from the auxiliary storage device or the like, expanded in the RAM, and executed by the CPU, so that each function of the failure prediction apparatus according to the present invention is realized by the computer of the failure prediction server. .
In this example, the first storage means according to the present invention is realized by the operation data storage means 41, the second storage means according to the present invention is realized by the failure occurrence information storage means 42, and the third storage according to the present invention is realized. The means is realized by the apparatus usage status storage means 43, and the classification means according to the present invention is realized by the learning data adjustment means 32 (and the fault symptom period length estimator generation means 35, the fault symptom period length estimator 36). The generation means according to the invention is realized by the failure predictor learning 33, and the prediction means according to the present invention is realized by the failure occurrence probability calculation means 34.

Here, the program according to the present invention is set in the computer of the failure prediction server by, for example, a format read from an external storage medium such as a CD-ROM storing the program or a format received via a communication network. The
Note that the present invention is not limited to a mode in which each functional unit is realized by a software configuration as in this example, and each functional unit may be realized by a dedicated hardware module.

In this example, each function of the failure prediction device according to the present invention is provided in one device (failure prediction server), but each function is distributed to a plurality of devices connected to be communicable with each other. It is good also as a structure provided.
Also, a failure predictor (failure prediction model) generated by the failure prediction apparatus according to the present invention is distributed to each image forming apparatus, and each image forming apparatus performs self-diagnosis (performs failure prediction for its own apparatus). It is good.

  Further, in the description so far, the processing for generating the failure predictor based on the result of adjusting the failure symptom period for each failure case has been described using the image forming apparatus as an example. Other devices that affect the length of the failure symptom period may be monitored devices, and there is a mechanism that can collect data necessary for adjusting the failure symptom period and generating a failure predictor from each device. That's fine.

  The present invention can be used in various systems and devices that perform failure prediction using a device whose difference in use status affects the length of a failure sign period as a monitored device, and programs and methods thereof.

11: Learning data, 12: Learning device, 13: Discriminator, 14: Test data,
21: Learning data adjustment means, 22: Fault symptom period length estimator generation means, 23: Fault symptom period length estimator,
31: Device operation status storage means 32: Learning data adjustment means 33: Failure predictor learning means 34: Failure occurrence probability calculation means 35: Failure sign period length estimator generation means 36: Failure sign period length estimator 37: Prediction result notification means,
41: Operation data storage means, 42: Failure occurrence information storage means, 43: Device usage status storage means

Claims (6)

First storage means for storing operation parameter data used for operation control by the monitored device and the time of collection thereof;
Second storage means for storing the time of occurrence of a failure in the monitored device;
Third storage means for storing usage status data of the monitored device and the time of collection thereof;
Based on the usage status data stored in the third storage means and the time of collection thereof, it is assumed that a failure sign appears in a period retroactive to the time of occurrence of the failure stored in the second storage means Classifying means for setting the failure sign period to be set for each case of failure, and classifying the operation parameter data stored in the first storage means according to whether or not the collection time point belongs to the failure sign period;
Generating means for generating a prediction model for use in predicting a failure of the monitored device, based on the classification result of the operation parameter data by the classification means;
A failure prediction system comprising: Correspondence between the trend of the usage status of monitored devices and the length of the failure sign period is predetermined,
The classification means adjusts the length of the failure sign period for each case of the failure in light of the correspondence relation at the time of occurrence of the failure, and sets the adjusted length of the failure sign period for each case of the failure. Set the failure symptom period by applying to
The failure prediction system according to claim 1. The prediction model is for calculating the occurrence probability of a failure in the monitored device,
The failure prediction system includes a prediction unit that calculates a failure occurrence probability using the prediction model for a plurality of times before a failure occurrence time in the monitored device.
The classification means determines the length of the failure indication period based on the relationship between the time-series change in the failure occurrence probability calculated by the prediction means and the usage data at the collection time corresponding to the failure occurrence time. Adjusting for each case, re-set the failure sign period and classify the operating parameter data,
The generation unit regenerates the prediction model based on data of operation parameters after the classification is performed again by the classification unit.
The failure prediction system according to claim 1 or claim 2, characterized by that. The third storage means stores the usage status data of the monitored device and its collection time point for a plurality of types of usage status,
The classifying unit sets a failure symptom period for each failure case based on the data of the plurality of types of use states stored in the third storage unit and the time of collection thereof, and the operation stored in the first storage unit Classify the parameter data according to whether the time of collection belongs to the failure indication period,
The failure prediction system according to any one of claims 1 to 3, wherein the failure prediction system is characterized in that: First storage means for storing operation parameter data used for operation control by the monitored device and the time of collection thereof;
Second storage means for storing the time of occurrence of a failure in the monitored device;
Third storage means for storing usage status data of the monitored device and the time of collection thereof;
Based on the usage status data stored in the third storage means and the time of collection thereof, it is assumed that a failure sign appears in a period retroactive to the time of occurrence of the failure stored in the second storage means Classifying means for setting the failure sign period to be set for each case of failure, and classifying the operation parameter data stored in the first storage means according to whether or not the collection time point belongs to the failure sign period;
Generating means for generating a prediction model for use in predicting a failure of the monitored device, based on the classification result of the operation parameter data by the classification means;
A failure prediction apparatus comprising: On the computer,
A first storage function for storing data of operation parameters used for operation control by the monitored device and the time of collection thereof;
A second storage function for storing the time of occurrence of a failure in the monitored device;
A third storage function for storing usage status data of the monitored device and the time of collection thereof;
Based on the usage status data stored by the third storage function and the time of collection thereof, it is assumed that a failure sign appears in a period retroactive to the occurrence time of the failure stored by the second storage function. A failure indication period to be set for each failure case, and a classification function for classifying the operation parameter data stored by the first storage function according to whether or not the collection time point belongs to the failure indication period;
Based on the classification result of the operation parameter data by the classification function, a generation function for generating a prediction model used for failure prediction of the monitored device;
A program to realize

Images