JP6257986B2 - Interactive facility failure prediction method and interactive facility failure prediction system - Google Patents

Description

  The present invention relates to an IT system and method for supporting maintenance work such as power poles (telegraph poles), manholes, bridges, tunnels, and pavements in roads and railway operations in power distribution, communication, and water and sewage operations.

  Due to the development of IT technology, a large amount of information related to equipment ledgers and inspection work records related to a large amount of equipment to be maintained has been accumulated in the maintenance work of infrastructure companies. Moreover, it is considered that environmental information related to the maintenance target facility can be easily obtained by open data or the like.

  Under such circumstances, for the purpose of improving inspection efficiency and reducing inspection costs by preferentially inspecting objects that are predicted to be defective, such as failure of maintenance objects, There is an increasing need for defect prediction that predicts defects by analyzing causal relationships with equipment defects.

  As an example of failure prediction, Patent Document 1 discloses a method for obtaining an optimal inspection plan using historical data such as past inspections and failures, or equipment conditions such as equipment specifications and installation environment, Among them, the failure prediction result is used as information necessary for creating the inspection plan. In this patent, a failure is predicted from the installation time of the equipment or the number of days since the previous replacement date, the statistics regarding the past failure occurrence interval of the maintenance target, and the environmental attributes given manually.

PPCT / JP11 / 0666206

  In Patent Document 1, failure prediction is performed by comparing the number of installation days of the equipment and the past failure occurrence interval. However, since other factors such as the installation environment are not considered, the past is not missed. The defect occurrence interval is estimated to be twice as short as the standard deviation from the average value of the defect occurrence intervals, and there is a problem that the prediction accuracy is lowered by predicting that the maintenance target is free of defects.

  In order to solve the above problem, it is conceivable to perform failure prediction using all related data items. However, there are actually some data items that are not related to equipment failures, so such data items Therefore, there is a problem that the prediction accuracy is lowered.

  In order to solve the above problem, it is conceivable to select only data items that improve prediction accuracy from related data items and use them for prediction. Generally, such data item combinations are obtained by combination optimization. There was a problem that even if a computer was used, a solution could not be obtained with an effective calculation time.

  A typical example of the invention disclosed in the present application is as follows. That is, a factor analysis for determining the relationship between the presence of past equipment failures and other maintenance-related data is performed, the relationships between the data items are illustrated, and the equipment failures selected by the maintenance staff based on the illustrated relationship diagram Using the data item combinations of the data items having the causal relationship of each of them, predicting the defect respectively, calculating the prediction accuracy and fixing the data item combination of the portion, and the selected data item and the selected data item combination Predict failure by using various data item combinations with low-relevance data items, calculate the prediction accuracy, fix the data item combination to be added, and then add the remaining data items to each data item combination. Data item combination that calculates the prediction accuracy by performing defect prediction, and uses the data item combination with the highest prediction accuracy to predict the failure of uninspected equipment To.

  According to the typical embodiment of the present invention, since failure prediction is performed using data items that are closely related based on the knowledge of the maintenance staff, failure prediction with high accuracy is possible.

  Further, by creating an inspection plan based on the failure prediction according to the representative embodiment of the present invention, it is possible to expect improvement in inspection efficiency and reduction in inspection cost.

  Further, according to the representative embodiment of the present invention, the maintenance person selects a data item that has a causal relationship with the equipment failure and fixes the data item combination of that part, and then selects the maintenance person from the selection of the maintenance person. Precisely add the data items that are likely to improve the accuracy of defect prediction from the remaining data items, and check the prediction accuracy by each data item combination, so the maintenance personnel acquire knowledge about the causal relationship In addition, data item combinations used for predicting defects of uninspected equipment can be determined efficiently at a realistic calculation speed.

It is a figure which shows an example of an interactive facility defect prediction process. It is a figure showing the example of an interactive facility defect prediction system. It is a figure which shows an example of an important causal relationship input process. It is a figure which shows an example of maintenance related data. It is a figure which shows an example of the relevance data between data items. It is a figure showing an example of the related figure between data items. It is a figure which shows an example of causal relationship data. It is a figure which shows an example of a learning data defect prediction process.

  Embodiments of the present invention will be described below with reference to the drawings.

  First, the method of the present invention will be described with reference to FIGS. FIG. 1 shows an example of the processing flow of the present invention and input / output data. FIG. 2 shows an example of a system in which the processing of the present invention is executed. First, the maintenance-related data 100 including the equipment failure record recorded in the input data recording unit 201 of the storage unit 200 is input, and the factor analysis program recorded in the program recording unit 203 is processed by the processing unit 211 to obtain data items. The interrelationship 110 is output to the generated data recording unit 202 (step 101). Next, the important causal relationship input program recorded in the program recording unit 203 is processed by the processing unit 211 using the data item relationship 110 as an input, the data item relationship is displayed on the display unit 212, and the user input unit The causal relationship 111 is input from 213 and output to the generated data recording unit 202 (step 102). Next, the learning data failure prediction program recorded in the program recording unit 203 is processed by the processing unit 211 using the maintenance related data 100 and the causal relationship 111 as inputs, and failure prediction is performed using the data items recorded in the causal relationship. The learning result 112 is output to the generated data recording unit 202 (step 103). Next, the learning data defect prediction program recorded in the program recording unit 203 by the processing unit 211 is processed using the maintenance related data 100 and the data item relationship 110, the causal relationship 111, and the learning result 112 output in step 103 as inputs. The failure prediction is performed by adding the data item recorded in the causal relationship to the data item combination of the learning result and the data item having low data item relevance, and the learning result 112 is output to the generated data recording unit 202 (Step 104). Next, the learning data failure prediction program recorded in the program recording unit 203 is processed by the processing unit 211 using the maintenance-related data 100 and the learning result 112 output in step 104 as input, and the data item combination of the learning result is processed. Then, the failure prediction is performed by adding the remaining data items, and the learning result 112 is output to the generated data recording unit 202 (step 105). Next, the result with the highest prediction accuracy is selected from the learning results (112) output in the processing of steps 103 to 105, and is output to the learning results 112 (step 106). Finally, the failure prediction program recorded in the program recording unit 203 is processed by the processing unit 211 from the maintenance-related data 100 and the learning result 112 selected in step 106, the equipment failure is predicted, and the prediction result 113 is generated data. The data is output to the recording unit 202 (step 107).

  Next, an example of maintenance related data is shown in FIG. As maintenance data, data 411 to 424 are recorded for each facility, and the data in each row is the facility failure data 400 and other maintenance data items 401 that are good / bad judgment results determined by the maintenance staff inspecting the equipment. To 404. Here, the other maintenance data items include equipment specifications, environmental conditions, and inspection records other than equipment failures described in the equipment management ledger. In addition, the maintenance-related data includes learning data 411 to 414 that has been inspected and the equipment defect data 400 is known, and prediction data 421 to 424 that have not been inspected and the quality of the equipment is unknown. Hereinafter, learning data is used in factor analysis (step 101) and learning data failure prediction (steps 103 to 105), and prediction data is used in failure prediction (step 107).

  Next, factor analysis (step 101) will be described with reference to FIGS. In the factor analysis, the maintenance related data 100 is read, correlation coefficients are obtained for all combinations of all data items, and recorded in the data item related data 110. FIG. 5 shows an example of output data obtained by performing factor analysis (step 101) from these data. Four combinations of 510 to 519 are conceivable from the five data items in FIG. 4, and a correlation coefficient (column 500) is recorded for each combination shown in columns 501 and 502. In the present embodiment, the absolute value of Pearson's correlation coefficient is used as the correlation coefficient, and regarding the equipment defect data item, “good” is 1 and “bad” is 0.

  FIG. 3 shows a detailed flow of the important causal relationship input (step 102) in the processing system of this embodiment. The important causal relationship input 102 will be described with reference to FIGS. 3 and 5 to 7. In the relationship display step of step 301, the processing unit 211 creates a relationship diagram between data items from the relationship (110) between the data items indicated by the result of the factor analysis as shown in FIG. indicate. An example of the related diagram is shown in FIG. First, in the relevance of FIG. 5, the value of the correlation coefficient 500 is included in a data item combination (here, combinations 510, 511, 512, and 514) that is larger than a predetermined threshold value (eg, 0.7 here). The data items to be displayed are displayed as islands, and the combinations of the data items in the columns 501 and 502 are displayed as straight lines (600 to 603) connecting the islands indicating the items. At this time, when equipment failure is included in the data item of the combination, the straight line is set to arrows 601, 602, 603. In this embodiment, it is considered that the correlation coefficient is larger or more relevant when the correlation coefficient is larger than the threshold value 0.7, and the correlation coefficient larger than the threshold value is underlined in the column 500 of FIG. Next, a maintenance person who has business knowledge thinks that the causal relationship between the equipment failure and the equipment to be checked is large from the arrows 601, 602, 603... On the displayed related diagram (FIG. 6). Select maintenance data items that are considered appropriate for automatic extraction. The causal relation input step 302 receives the input of the maintenance data item selected as appropriate for the automatic extraction of defective equipment from the maintenance staff and records it in the causal relation data 111. An example of the selected maintenance data item is shown in FIG. In FIG. 7, maintenance data item 1, maintenance data item 5, and maintenance data item 8 are data items that are considered to be the cause of equipment failure selected by the person in charge of the equipment. For example, click arrow 601 on the display screen of FIG. 6. Then, “maintenance data item 1” is recorded in the causal relationship (111).

Next, an example of learning data failure prediction (steps 103 to 105) will be described with reference to FIGS. FIG. 8 shows an example of a detailed flow of learning data failure prediction corresponding to any of steps 103, 104, and 105. Hereinafter, the data item ID is represented by a number. That is, the equipment failure data item is set to 0, and the maintenance data item N is set to N. The learning data failure prediction takes as input the learning data (800) in the maintenance-related data, the initial data item combination (801), and the additional data item combination (802). Here, the data item combination is a set of data item IDs. The initial data item combination A (801) includes equipment failure data items, and {0} and {0, 1} are examples. The additional data item combination B (802) is a non-empty set that does not include equipment failure data items, and {1} and {1, 2} are examples. First, all 1 to N _B data item combinations are obtained from _B (step 811). Here, N _B is the number of elements in the set B. For example, if B is {1,2,3}, {1} {2} {3} {1,2} {1,3} {2,3} and {1,2,3} are generated The Next, the following processing is executed for each generated combination B _i (step 810).

First, and the union of A and B _i data combination d (804), extracts the data D corresponding from the learning data (803) (step 812). For example, if A is {0} and B _i is {1,3}, d is {0,1,3} and D is the equipment failure (400) and data of the learning data (411 to 414) in FIG. It becomes the data for three columns of items 1, 3 (401, 403). Next, discriminant analysis is performed using the obtained first column of D as an objective variable and the remaining columns as explanatory variables (step 813). As the discriminant analysis, a known analysis method such as linear discriminant analysis, canonical discriminant analysis, SVM or the like can be considered. Regardless of which method is used, the discriminator (805) for performing prediction is output. Next, data E (814) excluding the first column of D and the discriminator (805) are used as inputs to perform defect prediction (step 815). Next, the prediction result and the equipment failure data which is the first column of D are compared, and the prediction accuracy (806) of the discriminator is calculated (step 816). Here, the prediction accuracy represents the proportion of data that has been predicted correctly, and is a value obtained by dividing the number of data in which the prediction result is equal to the equipment failure data by the total number of data. The above processing is performed for all combinations generated in step 811, and the prediction accuracy (806), data combination d (804), and discriminator (805) are recorded as learning results (820) for the combination with the maximum prediction accuracy. (Step 817).

Here, the description returns to the process flow of FIG. When the important causal relationship input from the maintenance staff in step 102 described in detail with reference to FIG. 3 and the like is completed, learning data failure prediction using the selected data item (step 103) is performed, and then the selection is performed. Learning data failure prediction (step 104) is performed by adding the data item and the unrelated data item. The learning data failure prediction using the selected data item (step 103) is recorded in the causal relationship data 111 among the relationships between the combination items recorded in the data item relationship data 110 as B and {0} as A. enter the integrity data items selected from maintenance personnel performs learning data fault predicting union of a and B _i as the data combination d (804). For example, when the maintenance data item selected from the maintenance staff is shown in FIG. 7, B is {1, 5, 8,. Moreover, the learning data defect prediction (step 104) which added the data item (unrelated data item) with low relevance with the selection data item is set to A, and the data combination d output to the learning result is set to A as a result of performing 103, enter the data item combinations excluding the data items that have been maintenance data items and high correlation selected from the conservation personnel as B, and the learning data fault predicting union of a and B _i as the data combination d (804) Do. For example, when data as shown in FIGS. 5 and 7 is recorded, 1, 5, 8 and maintenance data item 2 (see combination 514 in FIG. 5) that is highly correlated with data item 1, that is, data item ID = The data item combination excluding 2 is B.

  Further, learning data failure prediction (step 105) with the remaining data items added is performed as follows. As a result of performing Step 104, the data combination d output in the learning result is set as A, and the data item combination excluded in Step 104 is learned as B because there is a high correlation with the maintenance data item selected from the maintenance staff. Predict data failure. For example, when data as shown in FIGS. 5 and 7 is recorded, {2} is B.

  Next, prediction accuracy maximum data item combination selection (step 106) and defect prediction (step 107) will be described. In 103 to 105, three sets of prediction accuracy (806), data combination d (804), and discriminator (805) are recorded in the learning result (112). The operation of step 106 is to find a learning result having the maximum prediction accuracy and select a data combination d corresponding to the learning result. In step 107, the selected data combination d and the discriminator are input, and the failure prediction is performed on the data (prediction data) of the equipment that has not been inspected in the maintenance-related data 100, that is, the equipment whose quality is unknown. Run. The result of the failure prediction is data that estimates and rejects the good / failure of each uninspected facility, and is output to the prediction result 113 in FIG. This prediction result is used for grasping the status of the entire equipment to be maintained, and in particular, for developing a facility inspection inspection plan and inspection inspection route by the maintenance staff, and is useful for improving the efficiency of facility maintenance.

  In the equipment failure prediction processing of the embodiment described above, maintenance data that gives the maximum prediction accuracy after completing the calculation of learning data failure prediction for all combinations of an arbitrary number of maintenance data items that are enormous numbers. Rather than finding the item combination, the maintenance data items are divided into three groups, and the learning data failure prediction calculation is performed on the combination from the maintenance data items selected from the maintenance person who is the first group in the first stage. The data item combination of the part is fixed, and then in the second stage, the combination of the maintenance data items in the second group is added to perform the learning data failure prediction calculation, and the data item combination that gives the maximum prediction accuracy up to that point Maintenance data to obtain the maximum prediction accuracy by fixing the third group and adding the union back from the remaining maintenance data items in the third stage to perform the learning data failure prediction calculation And finally to determine the eye combination. Therefore, it is possible to determine the combination that gives the maximum prediction accuracy in the learning data failure prediction calculation of the data item combinations whose number is significantly reduced from the number of the arbitrary number of item combinations from all the maintenance data items. Moreover, since the first group is maintenance data items selected by the maintenance staff, knowledge about the causal relationship of the maintenance staff can be acquired, and the maintenance data item combination with high prediction accuracy can be reached at an early stage. Further, since the second group is a data item excluding data items highly correlated with the maintenance data item selected from the maintenance personnel, and the third group is the remaining maintenance data items, any number of maintenance data items can be selected. There is a high possibility of reaching a data item combination that gives the same maximum prediction accuracy as when all the combinations of learning data failure prediction are performed.

  As described above, according to the procedure of the present embodiment, a data item combination that gives the maximum prediction accuracy to be used for predicting a defect of uninspected equipment can be determined at a realistic calculation speed. However, in practice, it is not always necessary to perform the learning data failure prediction calculation in which candidates for the data item combination are sequentially added to the first group, the second group, and the third group. Since the third group of maintenance data items is a maintenance data item highly correlated with one of the first group of maintenance data items selected by the maintenance staff, the prediction accuracy can be further increased by adding to these data item combinations. The property is low compared to the combination addition of the maintenance data items of the second group. Therefore, in some cases, the learning data failure prediction procedure including the remaining data items in step 105 in FIG. 1 may be omitted. An arrow that bypasses step 105 in FIG. 1 indicates an operation flow when step 105 is omitted.

  That is, the important point in the above embodiments is that the maintenance data items selected by the maintenance staff are set as the first group, and their combination is first fixed, and then the maintenance data items highly correlated with the first group. Is the second group, and finally the remaining maintenance data items are the third group.

  The omission of the learning data failure prediction procedure is not limited to the omission of the entire step 105 as described above, and various methods can be employed. For example, a method for monitoring the passage of time and deciding the learning data failure prediction in the middle of step 104 or 105 to determine a data item combination that gives the maximum prediction accuracy up to that point, or a data item with a predetermined satisfactory prediction accuracy For example, a method of aborting the learning data failure prediction when the combination is reached is possible. Further, the method of further dividing the maintenance data items (third group maintenance data items) to be added to the combination in step 105 of the above embodiment into two groups is also effective. That is, among the remaining data items for which the prediction accuracy was confirmed by the first group, the second group and the learning data defect prediction calculation, the maintenance data items selected by the maintenance staff but not remained in the learning result of step 103 Maintenance data items having high relevance are set as the third group of maintenance data items, and maintenance data items still remaining, that is, maintenance data items having high relevance with the maintenance data items in the data item combination first fixed in step 103 are stored. What is necessary is just to perform the learning data defect prediction calculation which determines the data item added to a combination with the priority of a 3rd group and a 4th group after that by making an item into the last 4th group maintenance data item. Further, the third group in this modification is treated as the same priority as the second group, that is, among the data items not selected by the maintenance staff, the maintenance data item in the data item combination first fixed in step 103 Another modification is possible in which all except highly related data items are candidates for additional combinations in step 104. Also in these modified examples, the learning data failure prediction calculation of the second group to the fourth group is aborted by calculation at an arbitrary time point, and the data item combination that gives the maximum prediction accuracy up to that point is a facility (unchecked). It is possible to determine the data item used for the prediction of equipment failure.

100: maintenance-related data 101: factor analysis 102: factor-causal relationship input 103: learning data failure prediction using selected data items 104: learning data failure prediction adding selected data items and unrelated data items 105: remaining data items Data prediction 105: Prediction accuracy maximum data item combination selection 107: Defect prediction 110: Relevance between data items 111: Causal relationship 112: Learning result 113: Prediction result

Claims (7)

An interactive equipment failure prediction method for predicting equipment failure of a maintenance object using data related to the maintenance object,
Is recorded for each maintenance target equipment, among the data items related to the conservation target, the correlation for all combinations of data items associated with the object of maintenance including equipment failure data is inspection result of the equipment Calculate the number,
Displaying data items whose correlation coefficient is equal to or higher than a predetermined threshold among all combinations of data items related to the maintenance target, and displaying the data items connected with lines ,
In said data items displayed, recorded as a data item in the first group accepts a selection of the key related data items by maintenance personnel,
A first fixing unit of a combination that is a data item combination that performs defect prediction using an arbitrary number of combinations of the data items of the first group and gives the best prediction accuracy among the data items of the first group Seeking
A combination of the remaining data items not selected as the important related data items is added to the first immobilization unit to perform defect prediction, and data item combinations used for defect prediction of equipment whose quality is unknown An interactive facility failure prediction method characterized by deciding. The interactive facility failure prediction method according to claim 1,
An interaction between the data items including all the combinations of the data items related to the maintenance target, the correlation coefficient of which is equal to or greater than a predetermined threshold value and including the equipment failure data, is displayed with an arrow. Mold facility failure prediction method. The interactive facility failure prediction method according to claim 1 or 2,
The step of determining a data item combination to be used for predicting a failure of an equipment whose unknown quality is unknown includes adding a combination from a second group of data items in the displayed data item to the first immobilization unit. A first sub-step is performed in which a data item combination that performs prediction and gives the best prediction accuracy at that stage is a second fixing unit of the data item combination, and combinations from the remaining third group of data items are A second sub-step for performing defect prediction in addition to the two immobilization unit,
An interaction characterized in that a combination of data items that gives the best prediction accuracy through the first sub-step and the second sub-step is determined as a data item combination to be used for failure prediction of an equipment whose quality is unknown. Mold facility failure prediction method. The interactive facility failure prediction method according to claim 3 ,
2. The interactive facility failure prediction method according to claim 2, wherein the second group of data items is a data item analyzed as having a low relevance to the important related data item by analyzing the relevance between the data items. The interactive facility failure prediction method according to claim 3 ,
The second group of data items is a data item analyzed as having low relevance with respect to the data item adopted as the first immobilization unit among the important related data items. Interactive facility failure prediction method. In the interactive facility failure prediction method according to any one of claims 1 to 5 ,
Defect prediction using a combination of data items is a discriminant analysis in which data items related to equipment defects are objective variables and a plurality of combinations of the data items are explanatory variables among the data items related to the maintenance target. interactive equipment failure prediction how to characterized in that by performing.   An interactive equipment failure prediction system that predicts equipment failure of a maintenance object using data related to the maintenance object,
  Correlation coefficient for all combinations of data items related to the maintenance object, including equipment defect data, which is the inspection result of the equipment, among the data items related to the maintenance object recorded for each maintenance target equipment A processing unit for calculating
  A display unit for displaying a data item in which the correlation coefficient is equal to or greater than a predetermined threshold in the combination, and a line connecting the data items;
  A recording unit that accepts selection of important related data items by a maintenance person among the displayed data items and records as a first group of data items;
  The processing unit performs defect prediction using an arbitrary number of combinations of the data items of the first group, respectively, and is a combination of data items that gives the best prediction accuracy among the data items of the first group. Seeking the first immobilization part,
  A combination of the remaining data items not selected as the important related data items is added to the first immobilization unit to perform defect prediction, and data item combinations used for defect prediction of equipment whose quality is unknown An interactive facility failure prediction system to be determined.

Images