JP2019028671A - Information processing device, information processing method, computer program and storage medium - Google Patents

Abstract

To provide an information processing device that precisely learns a predictive rule for a component replacement occurrence.SOLUTION: An information processing device 100 includes: a data storing unit 113 that stores component pre-replacement data which adds a component label indicating the component to an event log prior to the replacement of the components of devices 101 to 10N each including multiple components; a rule learning unit 114 that learns a predictive rule indicating a predictor of an occurrence of replacement of each component among the multiple components from the component pre-replacement data; a duplicated rule selecting unit 115 that selects, as a duplicated rule, the predictive rule duplicated among the multiple components; a data selecting unit 116 that selects, as duplicated rule learning data, the component pre-replacement data which contains an event pattern of the duplicated rule; an influence calculating unit 117 that calculates influence of multi-label data on the duplicated rule on the basis of the number of the multi-label data to which the labels of the multiple components are added from the duplicated rule learning data; and a rule adjusting unit 118 that adjusts the duplicated rule on the basis of the influence.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an information processing apparatus that learns predictive rules for predicting the occurrence of equipment failures and parts replacement.

The device stores an event log related to events such as processing, operations, errors, and failures for maintenance management. An event log is analyzed, and an event combination and an event occurrence order before the occurrence of a failure are extracted as an indication rule (an indication pattern, an indication pattern) indicating an indication of the occurrence of the failure. The predictive rule is used to predict a failure occurrence.
Patent Document 1 discloses a failure prediction system that extracts predictive rules by performing pattern mining on an event log before a failure occurs. This system applies the extracted predictive rule to the event log of the device that is the target of failure occurrence, and predicts the occurrence of a failure when the same pattern as the predictive rule is detected in the event log. In addition, this system accumulates an event log before the occurrence of a failure for each type of failure, and extracts a predictive rule for each type of failure from each event log.

JP 2007-172131 A

In an apparatus such as an image forming apparatus, maintenance work is performed to replace parts according to the occurrence of a failure or the life. When learning a predictive rule for occurrence of parts replacement from a history of parts replaced by maintenance work and an event log before parts replacement, there are the following problems.
For example, when replacing the failed part A, another part B may be replaced at the same time. This is because, in order to reduce the number of times a serviceman is dispatched for maintenance work, when the part B that does not need to be replaced is replaced, or the cause of the failure of the device cannot be specified, the true faulty part (part A ) As well as different parts (part B).

  For learning the sign rule of occurrence of parts replacement, an event log before replacement of each part is required. More specifically, data with a part label (teacher data) added to the event log is used as learning data. With respect to an event log in which a plurality of parts are simultaneously exchanged, learning data to which each label (multi-label) of each part (part A and part B in the above example) is added is obtained. However, the label of the part B is an erroneous label because the part B is not a true faulty part. For this reason, it is difficult to correctly learn the pre-replacement predictive rule of each part from learning data including such multi-label data.

  The present invention has been made in view of the above problems, and a main object of the present invention is to provide an information processing apparatus that accurately learns a predictive rule for occurrence of replacement of each component from a multi-label event log resulting from replacement of a plurality of components. And

  An information processing apparatus according to the present invention includes a data storage unit that stores pre-part replacement data in which a part label indicating the part is added to an event log before replacement of the part of a device including a plurality of parts; Rule learning means for learning a sign rule indicating a sign of occurrence of part replacement of each of the plurality of parts from data, duplication rule selecting means for selecting the sign rule that overlaps the plurality of parts as a duplication rule, and the parts Data selection means for selecting pre-part replacement data including the event pattern of the duplication rule as duplication rule learning data from the pre-replacement data, and multi-label data to which a plurality of part labels are added from the duplication rule learning data An influence degree calculating means for calculating the influence degree of the multi-label data on the duplication rule based on the number, and the influence degree And a rule adjusting means for adjusting the overlapping rule Zui characterized.

  According to the present invention, by reducing the influence of multi-label data, it is possible to accurately learn a predictive rule for occurrence of replacement of each of a plurality of parts.

FIG. 3 is a configuration explanatory diagram of an information processing apparatus. (A), (b) is explanatory drawing of maintenance work and an event log. Explanatory drawing of the data before components replacement and normal data. The modification example figure of component replacement data. The flowchart showing the learning process of a precursor rule. Explanatory drawing of the process which selects duplication rule learning data. (A)-(c) is explanatory drawing of adjustment of a duplication rule. FIG. 4 is an exemplary diagram of an event log of a device that is in operation. Explanatory drawing of the learning which associates the time until parts replacement. FIG. 3 is a configuration explanatory diagram of an information processing apparatus.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings.

(First embodiment)
The present embodiment is applied to a device composed of a plurality of parts. The equipment is subjected to maintenance work such as replacement according to the failure or life of each part by a service person. When an image forming apparatus such as a printer is taken as an example of the device, the photosensitive drum, the fixing device, the transfer unit, and the like are replaced parts. The information processing apparatus according to the present embodiment learns a sign rule indicating a sign of occurrence of parts replacement from an event log before parts replacement. The predictive rule is one event or a combination (event pattern) of a plurality of events. The information processing apparatus can perform a prediction process for predicting the occurrence of component replacement based on the learned sign rule.

(Constitution)
FIG. 1 is a diagram illustrating the configuration of the information processing apparatus according to the present embodiment. A device log storage unit 111 and a component replacement history storage unit 112 are connected to the information processing apparatus 100. The device log storage unit 111 stores an event log that is a history of events that occur in N devices 101 to 10N, each of which includes a plurality of components. The device log storage unit 111 is connected to each device 101 to 10N via a network or the like, and stores an event log collected from each device 101 to 10N.

  The devices 101 to 10N are maintained by the service person 110, and the parts that need to be replaced are replaced by the service person. The component replacement history storage unit 112 stores a component replacement history associated with maintenance work of the service person 110. In the present embodiment, the component replacement history is input by the service person 110 and stored in the component replacement history storage unit 112. Note that the device itself may be configured to detect component replacement and write the component replacement history in the component replacement history storage unit 112.

  Although not shown, the information processing apparatus 100 is a computer including a CPU (Central Processing Unit), a ROM (Read Only Memory), a RAM (Random Access Memory), and a storage. The CPU implements the processing and functions of the present embodiment by reading the computer program of the present embodiment from the ROM or storage and executing it using the RAM as a work area. The information processing apparatus 100 allows the data storage unit 113, the rule learning unit 114, the duplicate rule selection unit 115, the data selection unit 116, the influence degree calculation unit 117, the rule adjustment unit 118, and the predictor to be executed by the CPU executing a computer program. It functions as a rule storage unit 119.

  The data storage unit 113 is realized by storage, and stores learning data obtained by integrating a component replacement history and an event log. The learning data is generated by the CPU, for example. The learning data includes pre-part replacement data in which a label indicating the type of replacement part is added to the event log before part replacement, and normal data generated from an event log in a period other than before part replacement.

The rule learning unit 114 learns a sign rule indicating a sign of replacement of each part from the learning data. The duplication rule selection unit 115 selects a predictive rule (duplicate rule) composed of the same event pattern with a sign rule of different parts from the learned sign rule of each part. The data selection unit 116 selects data related to learning of duplication rules from the data storage unit 113 as duplication rule learning data.
The degree-of-influence calculation unit 117 determines the degree of influence that the multi-label data has given to the duplication rule learning based on the number of pieces of data before component replacement (multi-label data) to which multiple component labels included in the duplication rule learning data are added. calculate. Based on the degree of influence, the rule adjustment unit 118 adjusts which part rule is adopted as an indication rule indicating a replacement indication of which part.

The sign rule storage unit 119 stores the sign rule that is finally learned. The predictive rule is used to predict the occurrence of parts replacement. The predictive rule storage unit 119 is realized by storage. The data storage unit 113 and the predictive rule storage unit 119 are provided in different storage areas of the same storage or in different storages.
Details of processing of each functional block of the information processing apparatus 100 will be described later.

(Simultaneous replacement of parts and duplication rules)
The overlapping rules learned from the simultaneous replacement of parts and the event log at the time of simultaneous replacement will be described. FIG. 2 is an explanatory diagram of maintenance work of an apparatus including the parts A and B and an event log. And event log e _a shows a single event such as an error code. Details of the event will be described later. The service person 110 of the device performs maintenance work to deal with the failure by a report from the user or the like.

FIG. 2A shows a state where the failed part A is replaced at time t. The service person 110 performs the replacement 201 of the part A at time t. In this case, a set of event logs (event set 211) representing an event that occurred in a period just before time t becomes the pre-part replacement data for learning the replacement sign rule for part A. Specifically, an event set 211 event set _{E = {e d, e a} , e c} in the component A of the label is additional data is stored as part replacement before data in the data storage unit 113. The information processing apparatus 100 collects a large number of such pre-component replacement data and detects a unique event pattern before replacement of the component A by a pattern mining method described later. The rule learning unit 114 learns the detected event pattern as a predictive rule.

  FIG. 2B shows a state in which the parts A and B are exchanged at the same time. The service person 110 performs the replacement 201 of the part A and the replacement 202 of the part B at time t. Here, it is assumed that only the part A is actually broken and should be replaced, and the part B does not need to be replaced yet. In this way, the replacement of a plurality of parts at the same time, including parts that do not need to be replaced, can frequently occur for the following reasons.

  The service person 110 needs to dispatch to a place where the device is installed in order to perform maintenance of the device. Performing maintenance on the same device many times leads to an increase in maintenance costs. In particular, when the installation position of the device is far from the service base, maintenance work is costly. In order to reduce the number of dispatches as much as possible, the service person 110 may replace parts that are likely to fail after a while even if performing a normal operation at the time of dispatch for replacement of a failed part. In addition, the service person 110 cannot identify the true cause of the failure, and may attempt to restart the device by exchanging a plurality of replaceable parts. Even in such a case, parts other than the parts that originally need to be replaced are also replaced.

  It is difficult to collect information about a really failed part because it may be difficult to identify the cause of the failure (failed part) or the workload of the service person 110 increases. In addition, for the purpose of recovering expenses, the part replacement history “which part has been replaced” is usually collected information.

Against this background, in FIG. 2B, only the replacement history that the parts A and B were replaced at time t is obtained, and the true faulty part is not known. In order to use the replacement history and the event log as learning data for the predictive rule for component replacement, an event log for a period slightly before time t is used. That is, a label indicating that the parts A and B are exchanged is added to the event set 212 (event set E = {e _e , e _c , e _a }) included in a period slightly before the time t. As a result, the learning data learned by the rule learning unit 114 is added with the multi-labels of the parts A and B, and it is unclear which part needs to be replaced.

In learning of the predictor rule, this learning data is data for learning the replacement predictor rule for part A and data for learning the replacement predictor rule for part B. When a large number of such learning data exists, the event set 212 (event set E = {e _e , e _c , e _a }) is a predictive rule for predicting the replacement of the part A and at the same time the replacement of the part B Are learned as predictive rules for predicting In this way, the same event combination belonging to the predictive rule for replacement of a plurality of parts is called an “overlapping rule”.

  When parts replacement is predicted based on predictive rules including duplication rules, occurrence of replacement of parts A and B is predicted when duplication rules appear in the event log to be predicted. In this example, the original learning data (event set 212) is originally an event log in a period before the timing when only the part A should be replaced. Therefore, the duplication rule should predict the occurrence of replacement of only part A. Therefore, if the occurrence of replacement of the parts A and B is predicted, the prediction of the occurrence of replacement of the part B becomes “false information”. Such a false alarm causes an increase in maintenance costs such as unnecessary dispatch of the service person 110 and unnecessary part replacement.

  The information processing apparatus 100 reduces the influence of learning data (simultaneous exchange multi-label data) to which labels of a plurality of parts are added. Therefore, the information processing apparatus 100 deletes a predictive rule that may cause a false alarm by performing post-processing on the post-learned predictive rule. Post-processing examines the label of each part added to the pre-part replacement data related to the learning of the duplication rule, and determines the selection of the duplication rule based on the number of pre-part exchange data to which the multi-label is added. It is processing. Details of such processing will be described.

(Learning data)
The learning data stored in the data storage unit 113 will be described. As described above, the learning data is generated from the device event log stored in the device log storage unit 111 and the component replacement history stored in the component replacement history storage unit 112.

The device log storage unit 111 stores event logs output from the devices 101 to 10N. The “event log” is time-series data in which event codes such as jams, alarms, and errors are recorded together with the occurrence time when the device is an image forming apparatus, for example. One event is represented by _{e x.} Further, the event is not limited to data previously discretized in this way, and may include, for example, data obtained by discretizing sensor values indicating the internal state of the device. Specifically, each predetermined value range is set as an event, and an event in a value range corresponding to the acquired sensor value is generated and recorded in the log. For example, when the temperature data inside the device is discretized and used as an event, the temperature of 10 [° C.] to 20 [° C.] is set to event e _T1 , and the temperature of 20 [° C.] to 30 [° C.] is set to event e _T2 , And

  The part replacement history storage unit 112 stores a maintenance work history (maintenance history) of the service person 110. Specifically, the component replacement history storage unit 112 stores component replacement information related to the timing at which components of the device are replaced. Further, information related to the maintenance work of the service person 110 may be stored in the part replacement history storage unit 112 and handled as an event such as an error code. For example, the service person 110 observes abnormal noise or vibration of the device or abnormality of the print result at the time of maintenance and inspection, and records such information in the component replacement history storage unit 112. In this case, these pieces of information may be one event that constitutes the event log together with the recording time (or occurrence time). That is, the event indicating the device state may include events other than those automatically collected by the device.

  The data storage unit 113 stores pre-part replacement data and normal data for predictive rule learning in which the data of the device log storage unit 111 and the part replacement history storage unit 112 are integrated. FIG. 3 is an explanatory diagram of pre-part replacement data and normal data. FIG. 3 exemplifies event logs of the same model device α and device β (event log α and event log β, respectively) as event logs of a plurality of devices. The data storage unit 113 acquires the event logs α and β from the device log storage unit 111. In FIG. 3, component A replacement 301, component A and component B simultaneous replacement 302, and component B replacement 303 are shown as the component replacement timing. The CPU acquires these component replacement information from the component replacement history storage unit 112 and integrates the time according to the event log as shown in FIG. The CPU creates pre-part replacement data and normal data, which are learning data for the sign rule, from these data, and saves them in the data storage unit 113.

The pre-part replacement data includes an event set included in the event log for a period slightly before the replacement timing of each part, and a label (teacher data) indicating the replaced part type. For example, for the replacement 301 of the part A, an event included in the period 311 slightly before the time t1 indicating the replacement timing of the part A is cut out, and the pre-part replacement data P1 = {e _b , _ed ; A} Created. Here, “A” in the pre-part replacement data P1 is the label of the part A. In addition, since the sign rule described later does not consider the event occurrence order, the event occurrence order in the period 311 is not included in the pre-part replacement data. When creating a predictive rule considering the event occurrence order, it is necessary to include the event occurrence order in the pre-part replacement data.
Similarly, for the replacement 303 of the part B, pre-part replacement data P3 = {e _a , e _b , e _c ; B} is created from the event in the period 314. Regarding the simultaneous exchange 302 of the parts A and B, for one event set ({e _a , e _c }) as in the pre-parts replacement data P2 = {e _a , e _c ; A, B}, Two labels of part A and part B are added. Hereinafter, this is referred to as “multi-label data”.

  In the above description, an event in a period slightly before the time t1 to t3 representing the replacement timing of each component is used as the pre-component replacement data. However, a predetermined time margin may be provided. FIG. 4 is a diagram illustrating a modified example of component replacement data. In FIG. 4, an event occurring in a period 411 between a time (t4−tc) (tc is a predetermined time) and a predetermined time before the time t4 when the component replacement 401 is performed is included in the pre-part replacement data. It has become an event. When creating an indication rule from an event immediately before component replacement is required, the information processing apparatus 100 may learn an event pattern before component replacement as an indication rule. If the occurrence of parts replacement is predicted based on such a predictive rule, the parts need to be replaced immediately after the predictive rule is generated (a failure occurs), and the service person 110 cannot respond before the failure. For this purpose, a pre-part replacement rule for learning the predictor rule may be created except for the event before the part replacement.

The normal data is created from the event log for a period other than before the parts replacement. Basically, normal data is created from an event log for a period not related to parts replacement. For example, in FIG. 3, normal data is created from a period 313 and a period 315 that are not related to parts replacement. The normal data is an event set of normal data N1 = {e _a , e _b } and normal data N2 = {e _a , e _c , _ed }, respectively. In this embodiment, in order to learn the predictive rule of each part independently, when learning the predictive rule of a predetermined part, the data before replacement of another part is used as normal data for learning. be able to.

For example, when learning the replacement predictor rule for the part A, the example of FIG. 3 is used as follows. The pre-part replacement data when learning the predictive rule for part A is pre-part replacement data P1 = {e _b , _ed ; A} (event set of period 311) to which the part A label is added. Data P2 = {e _a , e _c ; A, B} (event set of period 312). The normal data includes normal data N1 = {e _a , e _b }, normal data N2 = {e _a , e _c , _ed }, and normal data P3 = {e _a , e _b , e _c ; B} (period 314 event set). In the predictive rule learning for the part A, the label “B” added to the normal data P3 is not used.
The learning data as described above is stored in the data storage unit 113.

(Learn predictive rules)
FIG. 5 is a flowchart showing the sign rule learning process by the rule learning unit 114. This process includes a process of extracting a frequent event pattern from data before parts replacement, and a process of selecting a specific event pattern before parts replacement by comparing the output limit of the extracted frequent event pattern with normal data. . In the present embodiment, for simplification of explanation, the predictive rule to be learned is a combination of events, and the order of event appearance and the number of appearances are not considered. However, the predictive rule may be considered in consideration of these. The rule learning unit 114 learns the predictive rule for each part. Here, an example of learning the predictive rule for the part A will be described.

  The rule learning unit 114 reads learning data (pre-component replacement data and normal data) for learning the predictive rule for the component A from the data storage unit 113 (S501). The data used for learning the part A is data with the label A added to the data before parts replacement, and the data other than before the replacement of the parts A about normal data. If the pre-part replacement data is data to which a label A is added, multi-label data to which a label of another part is added is also used for learning a predictive rule for the part A.

  The rule learning unit 114 extracts a combination of events that frequently appear in the pre-part replacement data (frequent event pattern) (S502). The rule learning unit 114 performs this process by using a so-called method called data mining or frequent pattern mining. These methods are methods for efficiently extracting frequent patterns from a huge number of event combination patterns. In this embodiment, various events such as an error code are used. There are cases where there are over 100 types of events, and it is difficult to examine all the appearances of the combinations. Therefore, the rule learning unit 114 searches for an event pattern that frequently appears before component replacement, using a frequent pattern mining technique.

In this embodiment, Priori is used to extract a frequent pattern before component replacement. The details of Apriori are disclosed in R. Agrawal and R. Srikant: Fast algorithms for mining association rules, In Proceedings of 20th Int. Conf. Very Large Data Bases, VLDB, pp. 487-499, 1994. In Priori, the appearance frequency of an event or a combination of events is referred to as support (support). Assume that a set of data before parts replacement is D _P = {P ₁ , P ₂ ,... P _N }. “P _n ” indicates, for example, pre-part replacement data P _n = {e _a , e _b , e _c ; A} for part A. The support of the event pattern (for example, E = {e _a , e _b }) in the data set D _p before parts replacement is expressed by the following expression.

Support (E, D _p ) = σ _Dp (E) / N (1)
“N” is the number of data before parts replacement, and “σ _Dp (E)” is the number of data before parts replacement including the event pattern E.

Priori, a threshold value called minimum support is given in advance, and an event pattern (or a single event) indicating support above the minimum support is extracted as a frequent pattern. Priori is a known algorithm. Therefore, although detailed description is omitted, Priori can efficiently find an event pattern that is equal to or greater than the minimum support by designating the minimum support in advance when performing a combination search of events. As described above, x event patterns that frequently appear before component replacement are obtained. This is represented by a set D _F = {F ₁ , F ₂ ,... F _x }. “F” is a frequent event pattern F = {e _a , e _b }.

The rule learning unit 114 selects a “unique” event pattern from _DF before component replacement (S503). The frequent event pattern set _DF extracted in the process of S502 is an event pattern that occurs frequently before a failure occurs, but may occur frequently before the failure occurs. That is, the frequent event pattern set _DF may be an event and an event pattern that are likely to appear regardless of the state of a device or a component (failure, whether it is before replacement). For this purpose, the rule learning unit 114 uses normal data to extract a unique event pattern from these event patterns before component replacement, and creates a predictive rule. Specifically, the rule learning unit 114 calculates the ratio of the support (occurrence frequency) before failure occurrence of each event pattern and the support with normal data, and uses this event pattern as a unique event pattern before failure occurrence. And This support ratio is called GrowthRate and is expressed by the following equation.

GrowthRate (F, D _y , D _N ) = Support (E, D _p ) / Support (E, D _n )
= {Σ _Dp (F) / N} / {σ _Dn (F) / N} (2)
“F” is one frequent event pattern that is a target for calculating Growth Rate. “Dn” is a set of M normal data, and D _n = {N ₁ , N ₂ ... N _M }. Hereinafter, Growth Rate is referred to as “GR”. Note that such a pattern that appears biased in one data set is called “Emerging Patterns”. Details of “Emerging Patterns” are disclosed in Guozhu Dong, Jinyan Li, “Efficient Mining of Emerging Patterns: Discovering Trends and Differences”, KDD 1999: pp. 43-52.

The higher the GR is, the more the event pattern is a unique event pattern before parts replacement. Therefore, the rule learning unit 114 selects an event pattern having a GR greater than or equal to a predetermined threshold from the frequent event pattern set _DF , so that the predictive rule set R _A = {R _A1 , R _{A2,. Ay} } is created. R _A1 or the like is one predictive rule, for example, an event pattern such as _a predictive rule R _A1 = {e _a , e _b }. In addition, since the predictive rule set _RA is a subset of the frequent pattern set DF, y ≦ x. Thus, in the present embodiment, the rule learning unit 114 creates a plurality of predictive rules for a predetermined part. Note that the event pattern selection based on the GR may select the event pattern of the top n GRs instead of the threshold value. For prediction processing to be described later, it is preferable that the support of each pre-replacement data and GR for each predictive rule R _A1 , R _A2 ,... Are stored together with the predictive rule.

  Through the above processing, the rule learning unit 114 creates a predictive rule for the part A. The rule learning unit 114 repeatedly executes the processing of S501 to S503 until it learns the predictive rules for all the parts that make up the device (S504: N). When the rule learning unit 114 learns the predictive rule of all the parts constituting the device, the rule learning unit 114 ends the predictive rule learning process (S504: Y).

  In the above processing, the predictive rule learning by two steps of frequent pattern extraction using Priori and event pattern selection peculiar to parts replacement by GR calculation has been described, but the predictive rule learning method is not limited to this. In other words, any other method may be used as long as it uses a pre-part replacement data and normal data to find a unique event pattern before part replacement. For example, as a frequent pattern extraction method, a method called FP-Growth is well known in addition to Priori. In addition, frequent pattern extraction may be performed using another algorithm with high calculation efficiency.

(Duplicate rule selection part)
The duplication rule selection unit 115 selects the same rule (duplication rule) from the predictive rules of a plurality of parts learned by the rule learning unit 114. When the rule learning unit 114 creates the predictive rule set R _A = {R _A1 , R _A2 ,... R _Ay } of the component A and the predictive rule set R _B = {R _B1 , R _B2 ,... R _Bz } of the component B Will be described. When the predictor rule R _A1 = {e _a , e _c , e _e }, and the predictor rule R _B2 = {e _a , e _c , e _e }, the duplicate rule selection unit 115 uses the same event pattern for these predictor rules. _Therefore , R _A1 and R _B2 are selected as overlapping rules for parts A and B. In the subsequent processing, the duplication rule selection unit 115 determines whether the duplication rule R _AB = {e _a , e _c , e _e } is used as a predictive rule for the part A or the part B, or is left as a predictive rule for both. , Is determined. Similarly, when there are three or more parts to be predicted, the duplication rule selection unit 115 determines that, if there is a predictive rule composed of the same event combination among the predictive rules of a plurality of parts, the duplication rule is selected. Choose as.

  When there is no overlapping rule, the overlapping rule selection unit 115 does not perform subsequent processing. In this case, the predictive rule learned by the rule learning unit 114 is stored in the predictive rule storage unit 119 as it is. Even if there is a duplicate rule, the duplicate rule selection unit 115 may store the sign rule other than the duplicate rule in the sign rule storage unit 119 at this point because the sign rule is not affected by future processing. . When there are a plurality of overlapping rules, the following processing of the data selection unit 116 and the influence calculation unit 117 is repeatedly executed for the number of overlapping rules.

(Data selection part)
The data selection unit 116 selects pre-part replacement data related to the learning of duplication rules from the data storage unit 113. Data before parts replacement related to learning of duplication rules is data before parts exchange including events of duplication rules. This pre-part replacement data is hereinafter referred to as “duplication rule learning data”. FIG. 6 is an explanatory diagram of a process for selecting duplication rule learning data. In FIG. 6, when the duplication rule R _AB = {e _a , e _c , e _e } is given, duplication rule learning data is selected from the pre-part replacement data 601 to 605 stored in the data storage unit 113. The A replacement part label is added to the pre-part replacement data. For example, a part A label is added to the pre-part replacement data 601, and a multi-label of part A and part B is added to the pre-part replacement data 603.

The data selection unit 116 selects data including the duplication rule R _AB = {e _a , e _c , e _e } from these pre-part replacement data 601 to 605. Since the duplication rule is a duplication rule for component A and component B, the data selection unit 116 selects data to which at least one label of component A or component B is added. In FIG. 6, pre-part replacement data 601, 602, 603 is data including an event pattern {e _a , e _c , e _e } and at least one label of part A and part B added. is there. For this purpose, the data selection unit 116 selects pre-part replacement data 601, 602, and 603 as duplicate rule learning data. Although FIG. 6 shows a state in which three overlapping rule learning data are selected for illustration, in practice, a certain amount of data is selected as the overlapping rule learning data.

Note that the pre-part replacement data 604 is not selected as duplication rule learning data because it does not include the event pattern {e _a , e _c , e _e }. The pre-part replacement data 605 includes an event pattern {e _a , e _c , e _e }, but since only the label of the part C is added, it is not selected as duplication rule learning data. If there is pre-part replacement data to which the labels of parts A and C are added, this is selected as duplication rule learning data. However, since we are interested in the labels of the parts A and B, the data to which the labels of the parts A and C are added are processed as single label data of the part A in the subsequent processing. This is a case in which the multi-label data of the parts A and C exists in the learning data, but the duplication rule of the parts A and C is not learned.

(Influence calculation section)
The influence degree calculation unit 117 calculates the influence degree of the multi-label data with respect to learning of the duplicate rule based on the number of multi-label data included in the duplicate rule learning data. Based on the degree of influence calculated by the degree-of-impact calculator 117, the rule adjustment unit 118 determines whether the duplication rule is a sign rule for a predetermined part or a sign rule for a plurality of parts. When the duplication rule is used as a predictive rule for a predetermined part, the rule adjustment unit 118 deletes the duplicate rule included in the predictive rule set of another part. As described above, the process of deleting the predictive rule corresponding to the duplicate rule from the predictive rule set of the predetermined part (or the weighting process for the predictive rule described later) is referred to as “duplicate rule adjustment”. In the present embodiment, the duplicate rule is adjusted by two steps, that is, the process of the influence calculation unit 117 and the process of the rule adjustment unit 118. Details of these processes will be described.

  FIG. 7 is an explanatory diagram of adjustment of duplication rules. As described above, here, the overlapping rule of the parts A and B is adjusted. The duplication rule learning data selected by the data selection unit 116 is attached with either the label of part A or part B or the multi-label when parts A and part B are exchanged simultaneously. As shown in FIG. 7, there are roughly three types of relations between the labels of these duplication rule learning data. FIG. 7 is a Venn diagram showing the relationship between the labels of a plurality of overlapping rule learning data.

  FIG. 7A shows a case where the multi-label data obtained by simultaneously exchanging the parts A and B is not included in the duplication rule learning data. In this case, although the same sign rule (duplicate rule) is learned in the parts A and B, the multi-label data is not included in the original learning data. Therefore, FIG. 7A shows a case where the same predictive rule is extracted by chance from independent data. The duplication rule in this case is considered to be appropriate as a rule for predicting replacement of both part A and part B.

  FIG. 7B shows a case where both the parts A and B have the same influence of the multi-label data. Specifically, the ratio of the multi-label data in the data to which the labels of all the parts A are added (including the multi-label data to which the labels of A and B are added) and the labels of all the parts B are added. This is a case where the ratio of the multi-label data in the processed data is approximately the same. In this case, although it is difficult to make an accurate determination, it is considered that the learning data includes simultaneous replacement data in which it is appropriate to replace parts A and B at the same time. It is a rule that predicts both exchanges. That is, the duplicate rule is not deleted from the duplicate rule set of any part.

  FIG. 7C shows a case where the influence of the multi-label data is biased between the parts A and B. Specifically, most of the data to which the label of the part B is added is multi-label data to which the label of the part A is also added, whereas there is a lot of single label data for the part A. Yes. In this case, the predictive rule for the part B is learned by being greatly influenced by the multi-label data as compared with the part A. That is, it is presumed that the predictive rule of the part B tends to be learned mainly by an event log in which the part B that has not yet failed is replaced during maintenance work for the purpose of replacing the part A. Therefore, in this case, the duplication rule is selected as the predictive rule for the part A only, and is deleted from the predictive rule for the part B.

  The influence degree calculation unit 117 calculates an index for judging these cases in order to remove the duplicate rule included in the part B in the case of FIG. This index is calculated as a multi-label ratio in the pre-part replacement data to which the label of each part is added as shown in the following equation. Hereinafter, this is referred to as “multi-label influence E”.

E _A = γ (A, B) / ρ (A) (3)
E _B = γ (A, B) / ρ (B) (4)
“E _A ” and “E _B ” are the multi-label influence degrees of the duplicate rule learning data of the parts A and B. “Ρ (A)” is the number of all data (including multi-label data) to which label A is added in the duplicate rule learning data. “Ρ (B)” is the number of all data (including multi-label data) to which the label B is added in the duplication rule learning data. In other words, ρ (A) and ρ (B) are the sum of the number of single label data of each part and the number of multi-label data. “Γ (A, B)” is the number of multi-label data to which the labels of both parts A and B are added.

FIGS. 7A to 7C are graphs showing the multi-label influence E in each case. For example, in the case of FIG. 7A, since there is no multi-label data, the multi-label influence E is “0” for both the parts A and B. In the case of FIG. 7C, since the ratio of the multi-label data to the total data to which the label of the part B is added is high, the multi-label influence degree E of the part B is larger than the multi-label influence degree E of the part A. Indicates a high value.
By comparing these multi-label influences E _A and E _B , it is possible to determine a case where the influence of the multi-label is biased (case in FIG. 7C). This determination is performed by the rule adjustment unit 118.

(Rule adjustment section)
The rule adjustment unit 118 determines whether or not the multi-label influence degree of a plurality of parts is biased, and a duplication rule indicating whether the duplication rule is a common sign rule for a plurality of parts or a sign rule for a specific part Make adjustments. Various methods can be used to determine the bias of multi-label influence. Simply, calculate the absolute value of the difference in multi-label influence of each component and perform threshold processing on this value. Is done.

^{_{E * = | E A -E B}} | ... (5)
"E ^*" is the absolute value of the difference between the multi-label impact E _B of the multi-label influence the degree of E _A and part B of part A. When E ^* is greater than or equal to the predetermined threshold θ _E , the rule adjustment unit 118 determines that the multi-label influence degree is biased (E ^* > θ _E ). The determination of the non-uniformity of the multi-label influence is not limited to this, and may be obtained from the ratio of the multi-label influence of each component.

When the rule adjustment unit 118 determines that the multi-label influence degree is biased by the determination process as described above, the rule adjustment unit 118 performs a process of deleting a duplicate rule of a part having a high multi-label influence degree. For example, the rule adjustment unit 118 determines to delete the duplication rule for the part B in the case of FIG.
More specifically, the above example will be described. In this example, the rule learning unit 114 performs the predictive rule set R _A = {R _A1 , R _A2 ,... R _Ay } for the component A, and the predictive rule set R _B = {R _B1 , R _{B2,. Bz} } is learned. As a result, it is determined that R _A1 = {e _a , e _c , e _e } and R _B2 = {e _a , e _c , e _e } are duplicate rules. In this case, the rule adjustment unit 118 deletes the duplicate rule _RB2 for the part B. Therefore, finally, the predictive rule set of the part B is R _B = {R _B1 , R _B3 ,... R _Bz−1 }.

  When the multi-label influence degree is not biased as shown in FIG. 7A or FIG. 7B, the overlapping rule is a common rule for predicting replacement of the parts A and B as described above. To that end, each predictor rule is left behind. The rule adjustment unit 118 does not make adjustments such as deletion.

(Variation of rule adjustment section: learning of weight of predictive rule)
In the above, the adjustment of the duplicate rule for removing the duplicate rule by the determination process based on the multi-label influence degree has been described. In addition to this, the rule adjustment unit 118 may calculate the weight for the prediction process based on the multi-label influence degree. In the prediction process using the predictor rule, as described later, a method of calculating a predictive score from the predictor rule, the support of the predictor rule, and the GR can be used. The rule adjustment unit 118 may calculate the prediction score by weighting each predictor rule when calculating the prediction score. The rule adjustment unit 118 calculates the weight based on the multi-label influence level for the predictive rule corresponding to the duplicate rule for calculating the weighted prediction score, and stores it as a learning result in the predictive rule storage unit 119 together with the predictive rule. Details of the method of calculating the weighted prediction score will be described later.

The weight will be specifically described based on the example of the overlapping rule for the parts A and B described above (case in FIG. 7C). In the previous explanation, in this case, it is determined that only the predictive rule of the part A is left from the calculation result of the multi-label influence degree of the overlapping rule R _AB = R _A1 = R _B2 {e _a , e _c , e _e }. . Here, the rule adjustment unit 118 does not delete the sign rule R _B2 completely, but learns a small weight for the sign rule R _A1 . Rule arrangement 118, sign rule set of the part _{A R A = {R A1,} R A2, ... R Ay} _{respect, W A = {w A1,} w A2, ... w Ay} as an initial value of the weight give. Further, the rule adjustment unit 118 sets W _B = {w _B1 , w _B2 ,... W _BZ } to an initial weight for the predictive rule set R _B = {R _B1 , R _{B2 ,.} Give as value.

Here, the initial values of the weights are all “1”. The weights related to the predictive rules R _A1 and R _B2 corresponding to the overlapping rules are w _A1 and w _B2 . In this example, the predictive rule R _{B2 for} the part B is a rule learned under the influence of multi-label data. It is presumed that there is a high possibility of generating false alarms when parts replacement is predicted using this predictive rule. Therefore, the rule adjustment unit 118 sets a small value for the weight w _B2 of the predictor rule R _B2 to reduce the influence of the predictor rule on the prediction score calculation during the prediction process. The value set for w _{B2 only} needs to be calculated based on the multi-label influence degree E _B , and is not set to the reciprocal of the multi-label influence degree, for example, w _{B2 ′} = 1 / E _B. By doing in this way, the influence to the prediction score calculation at the time of a prediction process can be made small, as an indication rule with a high multi-label influence degree. New weight _{w B2} calculated _'replaces the _{w B2} of _{W B, W B = {w} B1, w B2', ... w BZ} is stored as a weight of sign ruleset _{R B} as.

  The rule adjustment unit 118 stores the sign rule adjusted in this way in the sign rule storage unit 119. In the case of the method of calculating the weight of each predictive rule, the rule adjustment unit 118 stores the weight in the predictive rule storage unit 119 together with the weight. If another duplicate rule is selected by the duplicate rule selection unit 115, the processing by the data selection unit 116, the influence degree calculation unit 117, and the rule adjustment unit 118 is repeated, and the learning of the predictor rule is completed.

  Through the above processing, the information processing apparatus 100 reduces the influence of the multi-label from the learning data including the multi-label data to which a plurality of component labels are added by the simultaneous replacement of the components, and indicates the indication of the replacement of the components. You will be able to learn the rules. In particular, the information processing apparatus 100 can learn the predictor rule by deleting the predictor rule that is likely to cause a component replacement error report.

  In the above description, the processing for the overlapping rule that occurs between two components has been described. When overlapping rules of the same event pattern are learned with three or more parts, the processes of the influence degree calculation unit 117 and the rule adjustment unit 118 may be repeated as many times as the number of component combinations. For example, a case will be described in which the same event pattern is learned as a predictive rule for three components, component A, component B, and component C (when there is an overlapping rule). First, for the set of component A and component B, multi-label influence calculation and duplication rule adjustment (prediction rule deletion or weight calculation) are performed. Next, the calculation of the multi-label influence and the duplication rule adjustment are performed in the same manner for the set of the parts A and C. Further, for the combination of the parts B and C, the calculation of the multi-label influence degree and the duplication rule adjustment are performed in the same manner. Here, if it is determined that the duplication rule for part A is to be deleted in the process of the set of parts A and B, the process of the set of parts A and C can be omitted.

(Prediction of parts replacement using predictive rules)
As described below, the information processing apparatus 100 configured as described above executes processing for predicting the occurrence of component replacement using the learned predictor rule.

  FIG. 8 is an exemplary diagram of an event log of a device that is in operation, which is used in a process for predicting the occurrence of component replacement. In the prediction process, part replacement is predicted using an event set included in a predetermined period 801 a little before the current time tn. Note that the prediction process is executed under an arbitrary condition such as a predetermined interval such as once a day, or whenever a new event occurs.

In the prediction using the predictor rule, it is determined whether the event set E _tn (E _tn = {e _b , e _c , _ed } in FIG. 8) included in the period 801 matches the predictor rule of each component. . Thereafter, a predicted score is calculated using the matched predictive rule. This process is executed for each component when a plurality of components are predicted. Hereinafter, a case where the replacement prediction of the part A is performed will be described.

In the information processing apparatus 100, any one of the predictive rule sets R _A = {R _A1 , R _A2 ,... R _Ay } of the component A is changed to the event set E _tn = {e _b , e _c , e _d } to be predicted. Determine whether they match. For example, the information processing apparatus 100, if R _An = {e _b , e _c } and R _Am = {e _b , e _d } exist in the predictive rule set R _A , these predictive rules are included in the event set E _th. Is determined to match. Here, as a simple prediction processing method, the information processing apparatus 100 may predict the occurrence of component replacement without calculating a prediction score. Specifically, when even one predictive rule matches the event set _Etn to be predicted, the information processing apparatus 100 outputs that predicts component replacement, assuming that a predictor of component replacement has been found.

A process for calculating the prediction score from the matched predictor rule and performing threshold value processing on the prediction score to determine whether or not a prediction has occurred (whether there is a sign of component replacement) will be described. The predictive rule of this embodiment is “Emerging Patterns” created by extracting a unique event pattern before parts replacement. Details of the ensemble identification method using the emerging patterns are described in “G Dong, X Zhang, L Wong, J Li,“ CAEP: Classification by aggregating emerging patterns ”, Discovery Science, 1999”. In this embodiment, this method is applied to prediction of parts replacement. For this purpose, support of each predictor rule R _A1 , R _A2 ,... And GR are used. The predicted score S for occurrence of replacement of the part A is expressed by the following equation (6).

“R” is a predictive rule for the part A that matches the event set E _th to be predicted. For R, a prediction score is calculated by GR and support as shown in Expression (6). Further, when a plurality of predictive rules coincide with the event set _Eth , the sum of these prediction scores becomes the final prediction score. In the calculation of the prediction score of Equation (6), the probability of component replacement (in the learning data) when the sign rule R matches by GR is calculated and weighted by support. The calculation method of the prediction score is not limited to this. However, since the GR of the predictor rule indicates the degree that the predictor rule appears only before parts replacement, it is effective to calculate the prediction score using the GR.

  As described above, when the weight based on the multi-label influence degree is learned for each predictive rule, the prediction score may be calculated using this.

Equation (7) calculates the prediction score S (E _tn ) by weighting Equation (6) with the weight w _R of the predictor rule R. According to this method, it is possible to calculate a prediction score in which the influence of a predictive rule (a predictive rule that may cause a false alarm) in which a small weight is learned is reduced.

In the prediction process based on the score calculation, when the predicted score S (E _tn ) calculated as described above exceeds the threshold value θS, it is determined that a sign that component replacement will occur is detected. When a sign of occurrence of component replacement is detected, the information processing apparatus 100 performs output for predicting occurrence of component replacement.

  The information processing apparatus 100 may present information using the information on the duplicate rule during the prediction process. A case where the occurrence of part replacement is predicted based on the overlapping rule of part A and part B will be described. When the duplication rule at this time is a duplication rule learned with data including a large amount of data before parts exchange for simultaneous exchange of parts A and B as shown in FIG. The effect is presented to the user (serviceman 110). The information processing apparatus 100 presents a message to the service person 110, for example, “This prediction indicates the necessity of exchanging part A and part B simultaneously according to past data”. The rule adjustment unit 118 needs to determine the case of FIG. 7B and store in the predictive rule storage unit 119 in association with the predictive rule that the predictive rule requires simultaneous replacement. The rule adjustment unit 118 may perform the determination of the case of FIG. 7B based on conditions such that the multi-label influence is not biased and the multi-label influence is high for a plurality of components.

(Variation of predictive rule learning: time to exchange)
The parts replacement prediction based on the above sign rule was a binary output indicating whether or not parts replacement occurred. The information processing apparatus 100 may output the time until the parts need to be replaced. In this case, the rule learning unit 114 performs learning for associating the time until parts replacement for each predictive rule. FIG. 9 is an explanatory diagram of learning for associating the time until component replacement.

One of the predictive rule sets R _A of the part A learned by the rule learning unit 114 is assumed to be a predictive rule R _An = {e _b , e _c }. The rule learning unit 114 performs a process of associating a time until parts replacement with the predictive rule R _An . The rule learning unit 114 collects pre-part replacement data including the event of {e _b , e _c }, to which the component A label is added, from the data storage unit 113. FIG. 9 illustrates the collected n pre-part replacement data 901 to 90n.

Based on the collected pre-part replacement data 901 to 90n, the rule learning unit 114 calculates the time from the occurrence time of the event that matches the predictive rule to the part replacement. For example, the component replacement before data 901, from the time of the time and parts replacement event e _c occurs, the time m ₁ until parts replacement is calculated. In the predictive rule including a plurality of event patterns, the time m until the component replacement is calculated based on the time from the time when the last event of the plurality of events occurs until the component replacement. Similarly, the time m _{2 to mn} until the part replacement is calculated for the other part replacement data 902 to 90n. The time until component replacement associated with the predictive rule R _An is the average m _μ of these times m _{1 to mn} . In this way, the rule learning unit 114 collects pre-part replacement data including the predictor rule for each predictor rule, and learns the time between the predictor rule and the component replacement.

The information processing apparatus 100 determines the time (M = t _R + m) until the occurrence of the component replacement based on the time t _R when the predictor rule appears and the time m _μ until the component replacement associated with the predictor rule. _μ ) is output. In addition, when a plurality of predictive rules appear for a predetermined event log, the information processing apparatus 100 calculates a time M _μ by averaging the time M until occurrence of parts replacement calculated by each predictor rule, and replacing the parts. Output as predicted time until occurrence.

  In the above processing, the information processing apparatus 100 calculates the average time by learning the time until component replacement, or by processing when multiple predictive rules appear. Not limited to this, the information processing apparatus 100 may store and output these as time distributions until occurrence of parts replacement without averaging them. In this case, the user (service man 110) can confirm the prediction result of the time until component replacement as a distribution with respect to the time axis.

(Variation of forecast target)
The rule learning unit 114 learns an event pattern for detecting a sign before component replacement as a sign rule. The learning by the rule learning unit 114 can be used not only for learning targeting a component replacement sign but also for learning a failure sign. For example, it is assumed that an accurate failure type and a failed part can be determined when a device fails. When the failure type and the failure part label are obtained, the rule learning unit 114 changes the failure type and the failure part for each failure type and the failure part by changing the part label in the predictive rule learning method of the present embodiment example to the failure type and the failure part label. You can learn these predictive rules.

  Specifically, the rule learning unit 114 learns a failure predictor rule using the component label as a failure label and the pre-component replacement data as pre-failure data in the component replacement predictor rule learning described above. For example, the rule learning unit 114 sets the occurrence of the failure by the method described above by setting the replacement 301 of the component A in FIG. 3 as the occurrence of the failure A and the replacement 303 of the component B as the occurrence of the failure B. It is possible to learn a sign rule indicating a sign of The failure may be replaced by a part, or may be recovered by repair and treatment. Normal data in this case is created by an event log in a period other than before the failure.

  In addition, it may be difficult to specify a failure type or a failure part, and the service person 110 may add a plurality of failure types or failure component labels (multi-labels) to the event log when a failure occurs. Even in such a case, there is a possibility that a duplication rule is learned between failure types and parts, but by adjusting the duplication rule, a predictive rule that predicts the occurrence of a failure by reducing the effect of multi-label data Can learn.

(Effects of learning method)
The information processing apparatus 100 of the present embodiment as described above suppresses adverse effects of multi-label data by performing post-processing (adjustment of duplication rules) on the learned predictor rule. Further, the processing of the rule learning unit 114 also contributes to reducing the adverse effects of multi-label data. This is particularly due to the process of extracting a unique event pattern before parts replacement using the normal data in S503 of FIG. Hereinafter, this effect will be described.

  As described above, the pre-part replacement data when learning the predictive rule for the part A includes the multi-label data of the parts A and B, and the normal data is the pre-part replacement data with the label of only the part B added. including. Here, the rule learning unit 114 extracts a predetermined event pattern F from the pre-component replacement data including the multi-label in the process of extracting the frequent event pattern in S502. It is assumed that this event pattern F is included in a lot of multi-label data and may become a duplication rule for parts A and B.

  In the next processing of S503, the GR of the frequent event pattern F is calculated with normal data including the pre-part replacement data with the part B label added. When the event pattern indicates a high GR, the event pattern is determined to be a unique event pattern before replacement of the part A, and is selected as a predictive rule. Here, when the frequent event pattern F is a unique event pattern before the replacement of the component B, it is preferable that the predictive rule for only the component B is used. That is, it is not preferable that the rule is learned as a predictive rule for the part A and an overlapping rule is generated. In this case, it can be expected that the calculated value of the support of the event pattern F shows a relatively high value in the normal data including the pre-part replacement data of the part B. When the support of the event pattern F with normal data shows a high value, the GR becomes a low value (formula (2)). As a result, the event pattern F is not selected as a predictive rule for the part A, and no duplication rule is generated.

  In this way, by using the pre-part replacement data for labels other than the learning target as normal data, it is possible to suppress the generation of duplicate rules that may cause misprediction. That is, the information processing apparatus 100 can perform predictive rule learning that suppresses the influence of multi-label data. However, in order to perform such learning, it is premised that the number of parts pre-replacement data to which the part B label in the normal data is added is abundant to some extent. That is, it is necessary that the normal data includes abundant data before replacement with the part B label in the normal data. Therefore, since the duplication rule generation cannot be completely suppressed only by the predictive rule learning, when the duplication rule is learned by the rule learning unit, the duplication rule adjustment described in the present embodiment is performed. good.

(Second Embodiment)
The information processing apparatus according to the second embodiment performs duplication rule adjustment in consideration of prior information on the relationship between components. The configurations of the information processing apparatus 100, the device log storage unit 111, and the component replacement history storage unit 112 of the second embodiment are the same as those of the first embodiment, and thus the description thereof is omitted.

There is a high possibility that a common predictive rule (duplicate rule) is created for parts that are physically close and parts that are close to the processing process. Here, the “processing process” is, for example, the order in which sheets on which images are formed pass if the apparatus is an image forming apparatus. In other words, the closer the order in which the sheets pass, the closer the processing process. Parts that are physically close and parts that are close to the processing process are often relatively likely to fail at the same time. In other words, even if there are cases in which two parts are exchanged at the same time, there are relatively many cases in which the exchange of both parts is correct. In addition, since the physical and processing processes are close to each other, there is a high possibility that a combination of certain events becomes a predictive rule for a plurality of nearby parts.
Therefore, when a duplication rule is learned for a plurality of parts that are physically and processally close, it is preferable to leave the duplication rule more than usual (do not delete the predictive rule for a specific part).

Therefore, in the present embodiment, the threshold value of the rule adjustment unit 118 is changed according to the combination of components in which overlapping rules exist. The rule adjustment unit 118 of the first embodiment performs threshold processing on the difference absolute value E ^* (Equation 5) of the multi-label influence degree of two parts with a predetermined threshold θ _E and deletes the predictive rule of the predetermined part. Judging what should be done. In the second embodiment, by adjusting the threshold theta _E, it is possible to relax the criteria for determining to leave overlapping rules. Specifically, by than the normal threshold theta _E is set to a large value to the threshold, overlapping rules tend to remain. This new threshold value θ _E ′ can be calculated, for example, by giving a weight (Equation 8) or a bias (Equation 9) to the reference threshold value θ _E.

θ _E ′ = α _AB θ _E (w _AB ≧ 1) (8)
θ _E ′ = θ _E + β _AB (β _AB ≧ 0) (9)
“Α _AB ” and “β _AB ” are weights and biases when the parts A and B are combined, respectively, and are set in advance based on the combination of the parts and their physical and processing process proximity. .

  As described above, the information processing apparatus 100 can create a suitable duplication rule by introducing the prior information on the relationship between components to adjustment of duplication rule selection.

(Third embodiment)
The information processing apparatus according to the third embodiment, on the contrary to the information processing apparatus according to the second embodiment, preferentially deletes the overlapping rule for a predetermined part based on prior information. The configurations of the information processing apparatus 100, the device log storage unit 111, and the component replacement history storage unit 112 according to the third embodiment are the same as those in the first embodiment, and a description thereof will be omitted.

  As described above, in order to reduce the number of dispatches, the service person may replace other inexpensive parts at the time of replacement of the failed part regardless of whether or not there is a failure. Depending on the combination of parts to be simultaneously exchanged, as in the case of the simultaneous exchange history of parts having such a price difference, it may be possible to predict a part to be truly exchanged at the time of simultaneous exchange. In the following, an embodiment for weighting the multi-label influence degree will be described for a method of introducing such prior information.

When there are overlapping rules for component A and component B, the impact calculation unit 117 calculates the respective multi-label impacts E _A and E _B based on the number of multi-label data (formula (3), formula (4)). In the present embodiment, the influence degree calculation unit 117 calculates the multi-label influence degrees E _A ′ and E _B ′ weighted thereto.

E _A '= w _A E _A (10)
E _B '= w _B E _B (11)
“W _A ” and “w _B ” are weights based on the prices of the parts A and B, respectively. For example, when the part B is a cheaper part, w _A <w _B.

As a result, the multi-label influence level E _B ′ of the component B is higher than the multi-label influence level E _B before weighting. In addition, the difference between the multi-label influences E _A ′ and E _B ′ increases. For this reason, the rule of the part B is easily deleted by the process of the rule adjustment unit 118.

In the above description, the priority includes the price of the component, but other information may be used for this priority. For example, when there are a part with a long life and a part with a short life, it is highly likely that the true replacement part for simultaneous replacement is a part with a short life. If the true replacement part (part of failure cause) at the time of simultaneous replacement can be determined from the past maintenance record, this may be used. The weighting coefficient of each component, such as w _A and w _B in Expression (10) and Expression (11), is determined in advance in consideration of these effects. As described above, the information processing apparatus 100 can preferentially adjust a part from which the predictive rule is erased when a duplication rule occurs based on the prior information on the relationship between the parts.

(Fourth embodiment)
The information processing apparatus according to the fourth embodiment adjusts the duplication rule based on the event pattern of the duplication rule. The configurations of the information processing apparatus 100, the device log storage unit 111, and the component replacement history storage unit 112 according to the fourth embodiment are the same as those of the first embodiment, and a description thereof will be omitted.

  Among events, there are events associated with specific parts. For example, a jam code of the image forming apparatus is associated with a part of the part where the jam has occurred. As another example, an error code generated when a temperature measured by a temperature sensor provided in a specific component is equal to or higher than a predetermined threshold is associated with the specific component. On the other hand, examples of events that are not associated with a component are events in which the generation source cannot be specified as a device part, such as software error codes and events based on abnormal sounds.

The information processing apparatus 100 adjusts the deletion rule deletion priority based on which component the event included in the duplication rule is related to. Here, the event _exA is an event associated with the part A. For example, when the duplication rule for part A and part B is R _AB = {e _aA , e _cA , e _eA }, all the events constituting the duplication rule R _AB are events related to part A. Therefore, there is a high possibility that the overlapping rule R _AB is a predictive rule for component replacement only for the component A. Therefore, it is preferable that the duplication rule R _AB is a predictive rule for parts replacement of only the part A by deleting the duplication rule of the part B of R _AB .

More specifically, for example, as described below, the multi-label influence degree is weighted based on an event included in the duplication rule. The weights for the multi-label influences E _A and E _B of the parts A and B are expressed by the following equations.

E _A '= f (R _AB , A) E _A (12)
E _B '= f (R _AB , B) E _B (13)
“F” is a function for calculating a weight. For example, the formula _f in _{(12) (R AB, A} ) , among the events included in the overlaid rule _{R AB,} as the ratio of the event is high regarding part A, and calculates the weight to reduce the multilabel influence _{E A} .

  As described in the first embodiment, when the multi-label influence degree E of a predetermined part is small and the difference from the multi-label influence degree of other parts becomes large, a selection is made to leave only the duplication rule of the part in the predictive rule. It becomes easy. Accordingly, by weighting the multi-label influence degree E based on the relationship between the event and the part included in the duplication rule as described above, the duplication rule adjustment based on the relation between the event and the part can be performed. become.

(Fifth embodiment)
In the above embodiment, the rule adjustment unit 118 automatically adjusts the duplicate rule based on the multi-label influence degree. On the other hand, the user may make a final determination of duplication rule adjustment (whether or not to delete the predictive rule that is the duplication rule from a predetermined part). The user is a device administrator, a serviceman, a component replacement prediction system administrator, or the like, and is a person who confirms the learning result of the predictor rule.

  FIG. 10 is an explanatory diagram of a configuration of the information processing apparatus according to the present embodiment. The information processing apparatus 200 has a configuration in which a display unit 120 and an input unit 121 are added to the information processing apparatus 100 of FIG. The display unit 120 displays information related to the duplicate rule. The user can confirm the display result and perform input for adjusting duplicate rules by using the input unit 121. The display unit 120 is an output device such as a display or a printer. The input unit 121 is an input device such as a keyboard or a touch panel.

  The display unit 120 presents information related to duplication rules. “Information about duplication rule” is, for example, the name of a part in which the duplication rule has occurred, information on an event included in the duplication rule, and the like. “Event information” is information such as detailed contents indicated by the event, for example, an error code indicating what kind of state the device is, and an event associated with which part. A user who has specialized knowledge about the device can consider the adjustment of the duplication rule in consideration of these factors.

  The rule adjustment unit 118 according to the present embodiment performs determination for adjustment of duplication rules (deletes duplication rules for predetermined parts or leaves duplication rules for all parts) based on the multi-label influence degree. The display unit 120 displays the determination result by the rule adjustment unit 118. The user confirms the determination result displayed on the display unit 120 and determines whether or not to adopt the determination result. The user inputs the determination result to the information processing apparatus 200 via the input unit 121. When the user inputs to delete a duplication rule for a predetermined part, the rule adjustment unit 118 creates a sign rule set from which the duplication rule is deleted and stores it in the sign rule storage unit 119.

  The display unit 120 may present the multi-label influence level to the user together with the determination result of the rule adjustment unit 118 or alone. Or you may make it the display part 120 display the relationship of the number of labels contained in duplication rule learning data. In these cases, the display unit 120 displays, for example, the Venn diagram or graph of FIG. 7 or the numerical value itself. This is because a user who understands the predictive rule learning method may be able to make a detailed determination by confirming the data that is the basis of the determination rather than the determination result of the rule adjustment unit 118.

  When there are a large number of event logs and the evaluation data of the predictive rule can be prepared separately from the learning data, the display unit 120 may display how the prediction performance changes depending on the judgment of the user. In this case, first, the user predicts component replacement of the evaluation data with the predictive rule set before adjusting the duplication rule, and inputs the evaluation result (prediction accuracy) through the input unit 121. The input prediction accuracy is stored in the predictive rule storage unit 119. When the user inputs a determination to delete a predetermined duplication rule, the rule adjustment unit 118 predicts component replacement of the evaluation data with the sign rule set from which the duplication rule is deleted, and sends the evaluation result to the sign rule storage unit 119. Store. The information processing apparatus 200 can assist the user's determination by displaying the change in the evaluation result before and after the user's duplication rule adjustment on the display unit 120.

(Sixth embodiment)
The present invention supplies a program that realizes one or more functions of the above-described embodiments to a system or apparatus via a network or a storage medium, and one or more processors in a computer of the system or apparatus read and execute the program This process can be realized. It can also be realized by a circuit (for example, ASIC) that realizes one or more functions.

Claims (22)

Data storage means for storing pre-replacement data in which a part label indicating the part is added to an event log before replacement of the part of a device including a plurality of parts;
Rule learning means for learning a sign rule indicating a sign of occurrence of part replacement of each of the plurality of parts from the pre-part replacement data;
Duplicate rule selection means for selecting the predictive rule that is duplicated in the plurality of parts as a duplicate rule;
Data selection means for selecting, from the pre-part replacement data, pre-part replacement data including the event pattern of the duplication rule as duplication rule learning data;
From the duplication rule learning data, an influence degree calculating means for calculating an influence degree of the multi label data on the duplication rule based on the number of multi label data to which a plurality of component labels are added;
Rule adjustment means for adjusting the duplication rule based on the degree of influence, and
Information processing device. The degree of influence calculation means, for each of the plurality of parts in which the duplication rule appears, the degree of influence, the number of multi-label data included in the duplication rule learning data, and all the parts labels are added It is calculated as a ratio of the number of data of
The information processing apparatus according to claim 1. The rule adjustment means performs a process of deleting a duplication rule of a part having a higher influence degree when the influence degree is biased in the plurality of parts in which the duplication rule appears.
The information processing apparatus according to claim 1 or 2. The rule adjusting means calculates a weight to be used during a prediction process for each predictor rule of the plurality of parts in which the duplicate rule appears based on the degree of influence.
The information processing apparatus according to any one of claims 1 to 3. The rule adjustment means performs the determination of the bias of the influence degree based on a predetermined threshold value, and adjusts the threshold value based on the relationship between the parts where the duplication rule has occurred,
The information processing apparatus according to claim 3. The influence degree calculating means assigns a weight based on a type of component to the influence degree to calculate a weighting influence degree, and adjusts a duplication rule using the weighting influence degree,
The information processing apparatus according to any one of claims 1 to 5. The influence degree calculating means assigns a weight to the influence degree based on an event pattern included in a duplication rule to calculate a weighting influence degree, and adjusts the duplication rule using the weighting influence degree. To
The information processing apparatus according to any one of claims 1 to 5. It further comprises display means for displaying information to the user, and input means for accepting user input,
The display means displays the information calculated by the influence degree calculating means or the rule adjusting means, and presents information for determining the adjustment of the duplicate rule to the user,
The rule adjusting means adjusts a rule based on information input to the input means,
The information processing apparatus according to claim 1. An information processing apparatus for learning predictive rules for occurrence of multiple types of failures,
Data storage means for storing pre-failure data with a failure label added to the event log in the period before failure;
Rule learning means for learning a sign rule indicating a sign of occurrence of each of the plurality of types of failure from the pre-failure data;
Duplicate rule selection means for selecting the predictive rule that is duplicated by a plurality of failures as a duplicate rule;
Data selection means for selecting pre-replacement data including the event pattern of the duplication rule from the pre-failure data as duplication rule learning data;
An influence degree calculating means for calculating the influence degree of the multi-label data on the duplication rule based on the number of the multi-label data to which a plurality of failure labels are added from the duplication rule learning data;
Rule adjustment means for adjusting the duplication rule based on the degree of influence, and
Information processing device. Data storage means for storing the event log in a period other than before the parts replacement of the device as normal data, and storing the event log in the period before the parts replacement as data before the parts replacement,
Rule learning means for extracting event patterns that frequently appear only before parts replacement from the normal data and the data before parts replacement, and learning a plurality of predictive rules indicating signs of parts replacement of the device. Characterized by the
Information processing device. For the event log that is a target for finding a sign of component replacement, the event log further includes a prediction processing unit that finds a sign of component replacement by calculating a prediction score using the plurality of sign rules.
The information processing apparatus according to claim 10. The rule learning means associates the time before component replacement for the predictor rule from the pre-component replacement data and the learned predictor rule,
The prediction processing means estimates the time until component replacement using the time until component replacement associated with a plurality of predictive rules,
The information processing apparatus according to claim 11. Data storage means for storing the event log in a period other than before the equipment failure as normal data, and storing the event log in the period before the equipment failure as pre-failure data,
A rule learning means for extracting an event pattern that appears frequently only before a failure from the normal data and the pre-failure data, and creating a plurality of predictive rules indicating a sign of the failure of the device. And
Information processing device. A method executed by an information processing apparatus that acquires an event log from a device including a plurality of components,
Storing pre-part replacement data in which a part label indicating the part is added to the event log before replacement of the part in a predetermined storage unit;
Learning a sign rule indicating a sign of occurrence of part replacement of each of the plurality of parts from the pre-part replacement data;
Selecting the predictive rule that overlaps the plurality of parts as a duplicate rule;
Selecting pre-part replacement data including the event pattern of the duplication rule as duplication rule learning data from the pre-part exchange data;
Calculating the influence of the multi-label data on the duplication rule based on the number of multi-label data to which a plurality of component labels are added in the duplication rule learning data;
Adjusting the duplication rule based on the degree of influence, and
Information processing method. An information processing method for learning predictive rules of multiple types of failures,
Storing the pre-failure data in which the failure label is added to the event log of the period before the failure in a predetermined storage means;
Learning a sign rule indicating a sign of occurrence of each of the plurality of types of failure from the pre-failure data;
Selecting the predictive rule that overlaps with multiple failures as a duplicate rule;
Selecting pre-part replacement data including the event pattern of the duplication rule from the pre-failure data as duplication rule learning data;
Calculating the degree of influence of the multi-label data on the duplication rule based on the number of multi-label data to which a plurality of failure labels are added from the duplication rule learning data;
Adjusting the duplication rule based on the degree of influence, and
Information processing method. A method executed by an information processing apparatus that acquires an event log from a device including a plurality of components,
The event log of a period other than before the replacement of the part is set as normal data, and the event log of the period before the replacement is stored in a predetermined storage unit as data before the replacement of parts;
Extracting from the normal data and the pre-component replacement data an event pattern that appears frequently only before the component replacement, and creating a plurality of predictive rules indicating the pre-replacement of the components of the device. And
Information processing method. A method executed by an information processing apparatus that acquires an event log from a device including a plurality of components,
The event log of a period other than before the failure of the device is set as normal data, and the event log of the period before the failure of the device is stored as predetermined data in a predetermined storage unit;
Extracting from the normal data and the pre-failure data an event pattern that frequently appears only before the failure, and creating a plurality of predictive rules indicating a sign of the occurrence of the failure of the device. ,
Information processing method. To a computer that acquires event logs from devices with multiple parts,
Storing pre-part replacement data in which a part label indicating the part is added to the event log before replacement of the part in a predetermined storage unit;
Learning a sign rule indicating a sign of occurrence of part replacement of each of the plurality of parts from the pre-part replacement data;
Selecting the predictive rule that overlaps with the plurality of parts as a duplicate rule;
Selecting pre-part replacement data including the event pattern of the duplication rule from the pre-part exchange data as duplication rule learning data;
Calculating the degree of influence of the multi-label data on the duplication rule based on the number of multi-label data to which a plurality of component labels are added in the duplication rule learning data;
Adjusting the duplication rule based on the degree of influence;
A computer program for running. A computer that learns the predictive rules for multiple types of failures,
Storing the pre-failure data in which the failure label is added to the event log of the period before the failure in a predetermined storage means;
Learning a sign rule indicating a sign of a failure occurrence of each of the plurality of types of failures from the pre-failure data;
Selecting the predictive rule that overlaps with multiple failures as a duplicate rule;
Selecting pre-part replacement data including the event pattern of the duplication rule as duplication rule learning data from the pre-failure data;
A step of calculating the influence of the multi-label data on the duplication rule based on the number of multi-label data to which a plurality of failure labels are added from the duplication rule learning data;
Adjusting the duplication rule based on the degree of influence;
A computer program for running. To a computer that acquires event logs from devices with multiple parts,
A step of storing the event log in a period other than before replacement of the parts as normal data, and storing the event log in the period before replacement in a predetermined storage unit as data before replacement of parts;
Extracting from the normal data and the data before parts replacement an event pattern that appears frequently only before parts replacement, and creating a plurality of predictive rules indicating a sign of parts replacement of the device;
A computer program for running. To a computer that acquires event logs from devices with multiple parts,
The event log of a period other than before the failure of the device is set as normal data, and the event log of the period before the failure of the device is stored as pre-failure data in a predetermined storage unit,
Extracting from the normal data and the pre-failure data an event pattern that appears frequently only before the failure, and creating a plurality of predictive rules indicating a sign of the failure of the device;
A computer program for running.   A computer-readable storage medium storing the computer program according to any one of claims 18 to 21.

Images