Root cause analysis system, method, and program

Abstract

To improve the accuracy of root cause analysis.SOLUTION: A root cause analysis system 1 that recommends a root cause of an abnormality occurring at an analysis target based on normal data and abnormal data obtained by measuring the analysis target with a plurality of sensors includes: a causal relationship information processing part 351, a significance level calculation part 352, and a recommendation result calculation part 3531. The causal relationship information processing part, based on the normal data, calculates a causal structure that links a plurality of sensors through causal relationships and an influence level of each of the plurality of sensors. The significance level calculation part, based on the normal data and the abnormal data, calculates the significance level of each of the plurality of sensors. The recommendation result calculation part, based on the causal structure and the significance level, extracts a plurality of recommended sensors as a target of recommendation, and determines a recommendation ranking of each of the plurality of recommended sensors based on the causal structure and the influence level to calculate recommendation results in which the plurality of recommended sensors are arrayed in order of the recommendation ranking.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a root cause analysis system, method and program. [Background technology]

[0002] In recent years, the use of big data and AI (Artificial Intelligence) has been increasing across industries. In particular, in the manufacturing industry, there is a growing need to quickly identify root causes by mathematically analyzing sensing data indicating operation logs obtained by sensors and product quality data.

[0003] Patent Document 1 describes a method for determining a judgment criterion based on a feature value generated based on time-series data, calculating an abnormality level based on the determined judgment criterion, and identifying a feature value that is a cause of a processing defect from among a plurality of feature values ​​based on the calculated abnormality level.

[0004] Patent Document 2 describes a method for generating a dependency graph based on time series data and precursor patterns, and generating and training a probabilistic network for a set of time spans based on the dependency graph and signal representation. This method uses causal structure and importance to find root causes. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2021-149727 [Patent Document 2] Japanese Patent Application Publication No. 2023-17888 Summary of the Invention [Problem to be solved by the invention]

[0006] The importance values ​​in Patent Documents 1 and 2 are determined without considering whether they are the root cause. As a result, the methods in Patent Documents 1 and 2 may recommend sensors that are not the root cause at the top, or may result in a decrease in analytical accuracy if there is insufficient abnormality data.

[0007] The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a technique for improving the accuracy of root cause analysis. [Means for solving the problem]

[0008] In order to achieve the above object, the present invention provides a root cause analysis system that recommends the root cause of an abnormality that has occurred in an analysis target based on normal data and abnormal data obtained by measuring the analysis target by multiple sensors, and includes a causal relationship information processing unit that calculates a causal structure that causally links the multiple sensors and the degree of influence of each of the multiple sensors based on the normal data; an importance calculation unit that calculates the importance of each of the multiple sensors based on the normal data and the abnormal data; and a recommendation result calculation unit that extracts multiple recommendation target sensors that are the subject of recommendations based on the causal structure and the importance, determines a recommendation ranking for each of the multiple recommendation target sensors based on the causal structure and the degree of influence, and calculates a recommendation result in which the multiple recommendation target sensors are arranged in order of the recommendation ranking. [Effects of the Invention]

[0009] According to the present invention, the accuracy of root cause analysis is improved. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a functional block diagram showing an example of the configuration of a root cause analysis system. [Figure 2] 10 is a flowchart illustrating an example of a root cause analysis process. [Figure 3] 10 is a flowchart showing an example of an initial consideration recommendation process. [Figure 4]10 is a flowchart showing an example of a causal relationship information calculation process. [Figure 5] 10 is a flowchart showing an example of a recommendation process after recalculation. [Figure 6] 10 is a flowchart showing an example of a process for creating a recalculated recommendation table. [Figure 7] 10 is a flowchart showing an example of a recommendation display process after recalculation. [Figure 8] 10 is a flowchart showing an example of an influence degree calculation process. [Figure 9] 10 is a flowchart showing an example of importance calculation processing. [Figure 10] 10 is a flowchart showing an example of a recommendation result calculation process. [Figure 11] FIG. 10 is a diagram showing an example of a time-series data table. [Figure 12] FIG. 10 is a diagram showing an example of a time-series data table after preprocessing. [Figure 13] FIG. 10 is a diagram showing an example of a feature amount table. [Figure 14] FIG. 10 is a diagram showing an example of an error sensor label. [Figure 15] FIG. 10 is a diagram showing an example of a causal structure table. [Figure 16] FIG. 10 is a diagram showing an example of a direct influence degree table. [Figure 17] FIG. 10 is a diagram showing an example of an importance table. [Figure 18] FIG. 10 is a diagram showing an example of a recommendation table. [Figure 19] FIG. 10 is a diagram showing an example of a recalculated direct influence degree table. [Figure 20] FIG. 10 is a diagram showing an example of a screen at the time of initial consideration. [Figure 21] FIG. 10 is a diagram showing an example of a screen after recalculation. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, specific examples of the root cause analysis system according to the embodiment of the present invention will be described with reference to the drawings. Note that the present invention is not limited to the examples, but is defined by the claims.

[0012] FIG. 1 is a functional block diagram illustrating an example of the configuration of a root cause analysis system.

[0013] The root cause analysis system 1 is a computer including a communication unit 2, a control unit 3, a storage unit 4, and a display unit 5. The root cause analysis device 1 executes a root cause analysis process that recommends the root cause of an abnormality that has occurred in an analysis target. The analysis target may be, for example, an industrial product such as a factory plant facility, an automobile part, or a home appliance.

[0014] The root cause analysis system 1 can be connected to a storage medium 6. The storage medium 6 is configured as, for example, a memory device, a hard disk device, an optical disk device, a magneto-optical disk device, or a magnetic tape device, and non-temporarily stores computer programs and data. The storage medium 6 can transfer and store computer programs and data to the storage unit 4 of the root cause analysis system 1. The storage medium 6 can also transfer and store computer programs and data from the storage unit 4 to the storage medium 6. A computer program that realizes the functions of the root cause analysis system 1 can be stored in the storage medium 6, and by connecting the storage medium 6 to another computer and installing the computer program stored in the storage medium 6 on the other computer, the other computer can function as the root cause analysis system 1.

[0015] The communication unit 2 can communicate with an analysis target database (hereinafter referred to as DB) 7. The analysis target DB 7 stores normal data and abnormal data obtained by measuring the analysis target using a plurality of sensors in chronological order. The sensors may be, for example, weight sensors, rotation sensors, temperature sensors, opening sensors, vibration sensors, etc. Abnormal data may be classified as non-normal data by the user 8 or the control unit 3. The communication unit 2 acquires the normal data and abnormal data from the analysis target DB 7 and stores the acquired normal data and abnormal data in chronological order in a time-series DB 41 (described later) in the storage unit 4. Furthermore, the communication unit 2 acquires the time-series data of the normal data and abnormal data from the time-series DB 41 (communication processing).

[0016] The control unit 3 includes a data acquisition unit 31, a data preprocessing unit 32, a visualization result calculation unit 33, an error code occurrence condition extraction unit 34, and a cause identification unit 35.

[0017] The data acquisition unit 31 acquires normal data and abnormal data from the analysis target DB 7 via the communication unit 2 (data acquisition process). The data acquisition unit 31 stores the acquired normal data and abnormal data in the time-series data DB 41.

[0018] The data pre-processing unit 32 denoises the normal data and the abnormal data using a moving average or the like, and extracts a feature quantity X that indicates the characteristics of the sensor (data pre-processing). The data pre-processing unit 32 stores the extracted feature quantity X in the feature quantity DB 42.

[0019] The visualization result calculation unit 33 calculates visualization results for the normal data and abnormal data (visualization result calculation process). The visualization results may include visualized images of the time-series data of the normal data and abnormal data, and images of the feature distribution comparison results between the normal data and abnormal data. The visualization result calculation unit 33 stores the calculated visualization results in the visualization DB 43.

[0020] The error code generation condition extraction unit 34 extracts an error sensor label from the abnormal data (error code generation condition extraction process), and stores the extracted error sensor label in the error code DB 44.

[0021] The cause identification unit 35 includes a causal relationship information processing unit 351 , an importance calculation unit 352 , and a root cause identification unit 353 .

[0022] The causal relationship information processing unit 351 includes a causal structure generating unit 3511 and an influence degree calculating unit 3512, and executes the causal relationship information calculation process shown in FIG. 4, which will be described later.

[0023] The causal structure generation unit 3511 acquires the error sensor label from the error code DB 44, creates a causal structure table using a Bayesian network or the like, and extracts a causal structure that connects the error sensors by a causal relationship based on the error sensor label (causal structure generation process). The causal structure generation unit 3511 stores the extracted causal structure in the causal structure DB 45.

[0024] The influence calculation unit 3512 calculates the influence for each of the multiple sensors (influence calculation process in FIG. 8). The influence indicates the degree of association between a parent sensor and a target sensor among the multiple sensors. The influence may be the objective influence if the parent sensor is a target variable, or the direct influence if the parent sensor is a child node. The influence may be the probabilistic rate of change in the value of the parent sensor when the value (category) of the target sensor is shifted. In the case of LinGAM (Linear Non-Gaussian Model), which assumes a non-Gaussian distribution in a linear model, the influence may also be a regression coefficient. The influence calculation unit 3512 stores the calculated influence in the influence DB 46.

[0025] The importance calculation unit 352 generates an anomaly detection model based on the normal data, and calculates the importance R based on the normal data and the abnormal data (importance calculation process in FIG. 9). The importance calculation unit 352 stores the calculated importance in the importance DB 47. If the importance is greater than 0, the sensor is classified as an abnormal sensor. An abnormal sensor means a sensor that exhibits behavior that is statistically different from normal. The importance may be the amount of change in classification certainty due to randomly changing the sensor value from the normal data value.

[0026] The root cause identification unit 353 includes a recommendation result calculation unit 3531 , a reconsideration candidate calculation unit 3532 , and a cause reconsideration unit 3533 .

[0027] The recommendation result calculation unit 3531 extracts multiple recommendation target sensors based on the causal structure and importance, and determines the recommendation ranking for each of the multiple recommendation target sensors based on the causal structure and influence. As a result, the recommendation result calculation unit 3531 calculates a recommendation table as an example of a "recommendation result" in which the multiple recommendation target sensors are arranged in order of recommendation ranking (recommendation result calculation process in FIG. 10). The recommendation result calculation unit 3531 may select multiple sensors based on the magnitude of importance, and extract sensors with a chain of causal relationships from the selected multiple sensors as recommendation target sensors.

[0028] The reconsideration candidate calculation unit 3532 excludes the recommendation table and the excluded sensor table from the abnormal sensor table (reconsideration candidate calculation process).

[0029] The cause reexamination unit 3533 performs a causal search under the condition that the sensors in the reexamination candidate sensor table are causally connected to other sensors, and creates a causal structure and a conditional probability table. Next, the cause reexamination unit 3533 calculates the direct influence level based on the created causal structure and conditional probability table, and creates a recalculated recommendation table in the same way as the root cause analysis process (cause reexamination process). The cause reexamination unit 3533 stores the created recommendation table in the recommendation DB 48.

[0030] The storage unit 4 includes a time series DB 41, a feature DB 42, a visualization DB 43, an error code DB 44, a causal structure DB 45, an impact DB 46, an importance DB 47, and a recommendation DB 48. The time series DB 41 stores a time series data table (FIG. 11). The feature DB 42 stores a feature table (FIG. 13). The error code DB 44 stores an error sensor label (FIG. 14). The causal structure DB 45 stores a causal structure table (FIG. 15). The impact DB 46 stores a direct impact table (FIG. 16). The importance DB 47 stores an importance table (FIG. 17). The recommendation DB 48 stores a recommendation table (FIG. 18).

[0031] The display unit 5 includes a visualization result display unit 51, a causal structure display unit 52, a recommendation result display unit 53, a root cause management unit 54, and a reexamination result management unit 55. The display unit 5 may be a display or the like that displays a screen that can be viewed by the user 8.

[0032] The visualization result display unit 51 displays to the user 8 the visualized image of the time-series data of the normal data and the abnormal data, and the image of the comparison result of the feature amount distributions of the normal data and the abnormal data (visualization result display process).

[0033] The causal structure display unit 52 displays to the user 8 a causal structure that links the sensors by causal relationships (causal structure display processing).

[0034] The recommendation result display unit 53 displays the recommendation table or the recalculated recommendation table to the user 8 (recommendation result display process).

[0035] The root cause management unit 54 checks the recommendation table, the causal structure, and the visualization results, and if the root cause cannot be determined, the process proceeds to the process of the cause reexamination unit 3533 (root cause management process).

[0036] The reexamination result management unit 55 checks the recalculated recommendation table, causal structure, and visualization results. If the reexamination result management unit 55 cannot identify the root cause, it either reviews the anomaly detection model, causal structure, and learning data, or determines that the cause cannot be identified using only the current sensors (reexamination result management process).

[0037] FIG. 2 is a flowchart illustrating an example of a root cause analysis process.

[0038] The control unit 3 executes the initial study recommendation process of FIG. 3 (S21), which will be described later. The control unit 3 checks the initial study recommendation table (S22). The control unit 3 determines whether or not the root cause can be guessed (S23). If the determination result of S23 is true (S23: YES), the control unit 3 ends the root cause analysis process. If the determination result of S23 is false (S23: NO), the control unit 3 executes the recalculated recommendation process of FIG. 5 (S24), which will be described later. The control unit 3 checks the recalculated recommendation table (S25). The control unit 3 determines whether or not the root cause can be guessed (S26). If the determination result of S26 is true (S26: YES), the control unit 3 ends the root cause analysis process. If the determination result of S26 is false (S26: NO), the control unit 3 reviews the anomaly detection model, the causal structure, and the learning data, or determines that the cause cannot be identified using only the current sensors (S27).

[0039] FIG. 3 is a flowchart showing an example of the initial consideration recommendation process.

[0040] In the initial consideration recommendation process, a data preparation process (S31), an importance calculation process (S32), a causal relationship information calculation process (S33) shown in FIG. 4 (described later), and an initial consideration recommendation display process (S34) are executed in this order.

[0041] Specifically, in the data preparation process (S31), the data acquisition unit 31 acquires normal data and abnormal data set by the user 8 from the analysis target DB 7 via the communication unit 2, and the data pre-processing unit 32 denoises the normal data and abnormal data and extracts the feature X.

[0042] In the importance calculation process (S32), the importance calculation unit 352 generates an anomaly detection model and calculates the importance R.

[0043] In the initial consideration recommendation display process (S34), the recommendation result calculation unit 3531 creates visualization results and displays the recommendation table, the causal structure, and the visualization results to the user 8.

[0044] FIG. 4 is a flowchart showing an example of the causal relationship information calculation process.

[0045] The error code occurrence condition extraction unit 34 extracts error code occurrence conditions (S41). The causal structure generation unit 3511 creates a causal structure table and a conditional probability table through causal search (S42). The influence calculation unit 3512 calculates the influence F (S43). The recommendation result calculation unit 3531 creates a recommendation table, an abnormal sensor table, and an excluded sensor table (S44).

[0046] FIG. 5 is a flowchart showing an example of the recommendation process after recalculation.

[0047] In the recalculated recommendation process, the reconsideration candidate calculation unit 3532 executes the recalculated recommendation table creation process shown in FIG. 6 (S51), and the reconsideration result management unit 55 executes the recalculated recommendation display process shown in FIG. 7 (S52).

[0048] FIG. 6 is a flowchart showing an example of a process for creating a recalculated recommendation table.

[0049] The reconsideration candidate calculation unit 3532 creates a reconsideration candidate sensor table based on the recommendation table, the abnormal sensor table, and the excluded sensor table (S61). The reconsideration candidate calculation unit 3532 acquires a feature quantity X from the feature quantity DB 42 (S62). The reconsideration candidate calculation unit 3532 performs a causal search assuming that the sensors in the reconsideration candidate sensor table are connected to other sensors by a causal relationship, and creates a recalculated causal structure table and a conditional probability table (S63). The reconsideration candidate calculation unit 3532 calculates the recalculated impact degree F (S64). The reconsideration candidate calculation unit 3532 creates a recalculated recommendation table (S65).

[0050] FIG. 7 is a flowchart showing an example of a recommendation display process after recalculation.

[0051] The recommendation result display unit 53 acquires the feature X from the feature DB 42 and acquires the error sensor label from the error code DB 44 (S71). The recommendation result display unit 53 acquires the recalculated recommendation table from the recommendation DB 48 (S72). The recommendation result display unit 53 acquires the recalculated causal structure table from the causal structure DB 45 (S73). The recommendation result display unit 53 acquires the visualization result from the visualization DB 43 (S74). The recommendation result display unit 53 displays the recommendation table, the causal structure, and the visualization result to the user 8 (S75).

[0052] FIG. 8 is a flowchart illustrating an example of the influence degree calculation process.

[0053] The influence calculation unit 3512 creates a condition probability table based on the causal structure table output from the causal structure generation unit 3511 (S81). The influence calculation unit 3512 starts a loop with a sensor in the causal structure (S82). The influence calculation unit 3512 creates a direct influence list sensor{i}_directlist[ ] for each sensor i (S83). The influence calculation unit 3512 starts a loop the number of times equal to the parameter α (S84). The influence calculation unit 3512 stores the change rate of the child node when the value of sensor i is shifted in the direct influence list (S85). The influence calculation unit 3512 completes a loop the number of times equal to the parameter α (S86). The influence calculation unit 3512 aggregates the values ​​in the list for each sensor i by averaging or the like (S87). The influence calculation unit 3512 completes a loop with a sensor in the causal structure (S88).

[0054] FIG. 9 is a flowchart illustrating an example of the importance calculation process.

[0055] The importance calculation unit 352 generates an anomaly detection model based on normal data (S91). The importance calculation unit 352 calculates the degree of anomaly using abnormal data as input (S92). The importance calculation unit 352 starts a loop a specified number of times (S93). The importance calculation unit 352 swaps the value of the abnormal data with the value of the normal data (S94). The importance calculation unit 352 calculates a recalculated degree of anomaly using the swapped abnormal data as input (S95). The importance calculation unit 352 calculates the importance based on the difference between the degree of anomaly and the recalculated degree of anomaly (S96). The importance calculation unit 352 completes the loop a specified number of times (S97).

[0056] FIG. 10 is a flowchart illustrating an example of the recommendation result calculation process.

[0057] The recommendation result calculation 3531 extracts sensors to be recommended based on the importance table and the causal structure (S101). The recommendation result calculation 3531 defines and lists abnormal subtrees for the number of sensors that are causally connected to the root node, and creates an abnormal subtree list k[i] (S102). The recommendation result calculation 3531 starts a loop for each abnormal subtree k[i] (S103). The recommendation result calculation 3531 determines whether there is a node in the abnormal subtree (S104). If the determination result in S104 is false (S104: NO), the recommendation result calculation 3531 advances the loop by one and proceeds to S108. If the determination result in S104 is true (S104: YES), the recommendation result calculation 3531 preferentially searches the abnormal subtree k[i], giving priority to sensors with a large direct influence (S105). The recommendation result calculation 3531 stores the root node in a list rec[ ] (S106). The recommendation result calculation 3531 deletes the root node (S107). The recommendation result calculation 3531 completes a loop for each abnormal subtree k[i] (S108).

[0058] FIG. 11 is a diagram illustrating an example of a time-series data table.

[0059] The time series data table is a table showing time series data for each sensor. The time series data table is the output of the communication process by the communication unit 2, is stored in the time series data DB, and is used as the input for the data acquisition process by the data acquisition unit 31.

[0060] The time-series data table stores, as item values ​​(column values), a sensor ID, time, weight sensor, fan rotation amount, temperature sensor, and the like.

[0061] The sensor ID is an identifier that uniquely identifies the sensor. The time is the time when the output value from each sensor was acquired. The weight sensor is the measurement value at each time by the weight sensor. The fan rotation amount is the measurement value at each time by the rotation sensor. The temperature sensor is the measurement value at each time by the temperature sensor.

[0062] As an example, let's look at the first row of the time series data table. The first row is related to the sensor ID "1." This first row indicates that the time is "2023 / 9 / 6 18:36:33," the weight sensor is "60," the fan rotation rate is "6000," and the temperature sensor is "20."

[0063] FIG. 12 is a diagram illustrating an example of a time-series data table after preprocessing.

[0064] The preprocessed time series data table is a table showing the preprocessed time series data for each sensor. The preprocessed time series data table is the output of denoising processing such as moving average by the data preprocessing unit 32, stored in the time series data DB 41, and used as input for the causal structure generation unit 3511's causal search processing and the visualization result calculation unit 33's processing for calculating visualization results of time series data of normal data and abnormal data.

[0065] The preprocessed time-series data table stores, as item values ​​(column values), the sensor ID, time, weight sensor, fan rotation amount, temperature sensor, and the like, in the same way as the time-series data table.

[0066] As an example, let's look at the first row of the preprocessed time series data table. The first row is for sensor ID "1." This row indicates that the time is "2023 / 9 / 6 18:36:33," the weight sensor is "62," the fan rotation rate is "6000," and the temperature sensor is "20." In this way, the weight sensor value has been changed from "60" in the preprocessed time series data table to "62" through denoising processes such as moving average.

[0067] FIG. 13 is a diagram illustrating an example of the feature amount table.

[0068] The feature table is a table that represents the feature for each sensor. The feature table is the output of the feature creation process by the data preprocessing unit 32, is stored in the time-series data DB 41, and serves as input to the feature distribution comparison result image creation process by the visualization result calculation unit 33 and the importance calculation process by the importance calculation unit 352.

[0069] The feature table stores, as item values ​​(column values), a sensor data ID, an average value _ave of the weight sensor, a standard deviation _std of the fan rotation amount, a maximum value _max of the temperature sensor, and the like.

[0070] The sensor data ID is an identifier that uniquely identifies one operation of the analysis target. The weight sensor average value _ave is the average value of the measurement values ​​at each time by the weight sensor when the analysis target is operating once. The fan rotation amount standard deviation _std is the standard deviation of the measurement values ​​at each time by the rotation sensor when the analysis target is operating once. The temperature sensor maximum value _max is the maximum value of the measurement values ​​at each time by the temperature sensor when the analysis target is operating once.

[0071] As an example, we will explain the first row of the feature table. The first row is related to the sensor data ID "1". This first row means that the time is "2023 / 9 / 6 18:36:33", the weight sensor is "60", the fan rotation rate is "6000", and the temperature sensor is "20".

[0072] FIG. 14 is a diagram showing an example of an error sensor label.

[0073] The error sensor label indicates the error sensor for each error code. The error sensor label is the output of the error code occurrence condition extraction process by the error code occurrence condition extraction unit 34, is stored in the error code DB 44, and is the input of the error sensor label acquisition process by the causal structure generation unit 3511.

[0074] The error sensor label stores the sensor ID, error code, and error sensor as item values ​​(column values).

[0075] The sensor ID is an identifier for uniquely identifying the sensor. The error code is a code for each type of error. The error sensor is the sensor in which an error occurred.

[0076] As an example, let's look at the first line of the error sensor label. The first line is for the sensor data ID "1". This first line means that the error code is "E01" and the error sensor is the "temperature sensor".

[0077] FIG. 15 is a diagram illustrating an example of the causal structure table.

[0078] The causal structure table is a table that represents causes and effects that are connected by causal relationships in a brute-force format. The causal structure table is the output of the causal structure generation process by the causal structure generation unit 3511, is stored in the causal structure DB 45, and serves as input to the causal structure display process by the causal structure display unit 52, and the recommendation target sensor extraction process and recommendation ranking determination process by the recommendation result calculation unit 3531.

[0079] The causes are the fan rotation amount, current value, opening sensor, vibration sensor, etc. The results are the temperature sensor (error code), fan rotation amount, opening sensor, vibration sensor, etc.

[0080] As an example, let's look at the first line of the feature table. The first line indicates that the cause is related to the fan rotation amount. This line indicates that there is a causal relationship between the fan rotation amount, which is the cause, and the temperature sensor (error code), which is the result.

[0081] FIG. 16 is a diagram illustrating an example of a direct influence degree table.

[0082] The direct influence table is a table that shows the influence of causes and effects that are connected by a causal relationship in a round-robin format.

[0083] The direct influence table is the output of the influence calculation process by the influence calculation unit 3512, is stored in the influence DB 46, and is input to the recommendation ranking determination process by the recommendation result calculation unit 3531.

[0084] The causes are the fan rotation amount, current value, opening sensor, vibration sensor, etc., as in the causal structure table, and the results are the temperature sensor (error code), fan rotation amount, opening sensor, vibration sensor, etc.

[0085] As an example, let's look at the first row of the direct influence table. The first row is related to the fan rotation amount. This row means that the direct influence between the fan rotation amount (cause) and the temperature sensor (error code) (result) is "0.70."

[0086] FIG. 17 is a diagram illustrating an example of the importance table.

[0087] The importance table is a table that shows the importance of each sensor in descending order of importance. The importance table is an output of the importance calculation process by the importance calculation unit 3512, is stored in the importance DB 47, and is an input to the recommendation result calculation unit 3531's process of extracting recommendation target sensors.

[0088] The importance table stores the sensor ID, the sensor, and the importance as item values ​​(column values). The sensors include a weight sensor, a fan rotation amount sensor, and a vibration sensor.

[0089] As an example, the first row of the importance table will be explained. The first row is related to the weight sensor. This first row means that the importance of the weight sensor is "0.9".

[0090] FIG. 18 is a diagram illustrating an example of a recommendation table.

[0091] The recommendation table is a table that shows sensors in order of recommendation rank. The recommendation table is the output of the recommendation result calculation process of the recommendation result calculation unit 3531, is stored in the recommendation DB 48, and serves as input to the recommendation result display process by the recommendation result display unit 53 and the reconsideration candidate creation process by the reconsideration candidate calculation unit 3532.

[0092] The recommendation table stores rankings and sensor names as item values ​​(column values).

[0093] As an example, the first row of the recommendation table will be explained. The first row is related to a weight sensor. This first row indicates that the weight sensor is ranked "1."

[0094] FIG. 19 is a diagram illustrating an example of the recalculated direct influence degree table.

[0095] The recalculated direct influence table is a table that shows the recalculated influence between causes and effects that are connected by a causal relationship in a round-robin format.

[0096] The recalculated direct impact table is the output of the direct impact calculation process by the cause reexamination unit 3533, is stored in the impact DB 46, and is used as input to the recommendation ranking determination process by the cause reexamination unit 3533.

[0097] The causes are the fan rotation amount, current value, opening sensor, vibration sensor, etc., as in the causal structure table and direct influence table, and the results are the temperature sensor (error code), fan rotation amount, opening sensor, vibration sensor, etc.

[0098] As an example, we will explain the first row of the recalculated direct impact table. The first row indicates that the cause is related to the fan rotation amount. This first row indicates that the recalculated direct impact between the fan rotation amount as the cause and the temperature sensor (error code) as the result is "0.34."

[0099] FIG. 20 is a diagram showing an example of a screen at the time of initial consideration.

[0100] During the initial investigation, the screen of the display unit 5 displays the causal structure, the recommendation table, and the visualization results. In this example, the visualization results are expressed as line graphs or the like, showing the time-series data of each sensor. The direct influence of the causal structure is greater for the fan rotation rate than for the weight sensor, and greater for the current value than for the motor rotation rate. Therefore, the causal relationships between the current value, the fan rotation rate, and the temperature sensor are linked in order. As a result, the ranking of the recommendation table is current value, motor rotation rate, fan rotation rate, and water level sensor in that order.

[0101] FIG. 21 is a diagram showing an example of the screen after recalculation.

[0102] After the recalculation, the causal structure, the recommendation table, and the visualization results are displayed on the screen of the display unit 5. In this example, the visualization results, like those in the initial study, are represented by line graphs or the like showing the time-series data of each sensor. The direct influence of the causal structure is such that the fan rotation amount is greater than the weight sensor, the current value is greater than the motor rotation speed, and sensor T is greater than sensor C. Therefore, the causal relationships between sensor T, the current value, the fan rotation amount, and the temperature sensor are each linked in order. As a result, the ranking of the recommendation table is sensor T, sensor C, ..., and sensor N. In this case, because sensors T and C are each connected to the current value as a causal relationship, the ranking of the parent sensors is the current value, ..., and sensor G.

[0103] According to this configuration, the root cause analysis system 1, which recommends the root cause of an abnormality occurring in an analysis target based on normal data and abnormal data obtained by measuring the analysis target by multiple sensors, includes a causal relationship information processing unit 351, an importance calculation unit 352, and a recommendation result calculation unit 3531. The causal relationship information processing unit 351 calculates a causal structure causally linking multiple sensors and an influence level for each of the multiple sensors based on the normal data. The importance calculation unit 352 calculates the importance level for each of the multiple sensors based on the normal data and the abnormal data. The recommendation result calculation unit 3531 extracts multiple recommendation target sensors that are the target of recommendations based on the causal structure and importance levels, and determines a recommendation ranking for each of the multiple recommendation target sensors based on the causal structure and direct influence level. The recommendation result calculation unit 3531 calculates a recommendation result in which the multiple recommendation target sensors are arranged in order of recommendation ranking.

[0104] This improves the accuracy of root cause analysis, allowing for early identification of abnormality factors.

[0105] The importance calculation unit 3512 generates an anomaly detection model based on normal data, and inputs normal data and abnormal data into the generated anomaly detection model to calculate the importance. This makes it possible to narrow down the sensors to be recommended. At this time, because the anomaly detection model generated from only normal data is used, the accuracy of calculating the importance can be improved even for analysis target data with little abnormal data.

[0106] The recommendation result calculation unit 3531 selects multiple sensors based on the level of importance, and extracts sensors that are linked in a causal relationship from the selected multiple sensors as sensors to be recommended, thereby further improving the accuracy of root cause analysis.

[0107] The recommendation result calculation unit 3531 determines the recommendation ranking by depth-first search that prioritizes the sensor with the greatest influence, thereby making it possible to determine the recommendation ranking that prioritizes the magnitude of the influence.

[0108] The system includes a reconsideration candidate calculation unit that performs a causal search on sensors other than the recommendation target sensors among the multiple sensors and adds them to the recommendation target sensors. This makes it possible to expand the causal search range for sensors.

[0109] The system includes a visualization result calculation unit 33 that calculates visualization results for normal data and abnormal data, and a display unit 5 that displays the causal structure, recommendation results, and visualization results to the user, allowing the user 8 to check the basis for the analysis results.

[0110] The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations. [Explanation of symbols]

[0111] 1...Root cause analysis system, 33...Visualization result calculation unit, 351...Causal relationship information processing unit, 352...Importance calculation unit, 3531...Recommendation result calculation unit, 3532...Reconsideration candidate calculation unit, 5...Display unit, 8...User

Claims

1. A root cause analysis system that recommends a root cause of an abnormality that has occurred in an analysis target based on normal data and abnormal data measured by a plurality of sensors on the analysis target, a causal relationship information processing unit that calculates a causal structure that causally links the plurality of sensors and an influence degree for each of the plurality of sensors based on the normal data; an importance calculation unit that calculates an importance of each of the plurality of sensors based on the normal data and the abnormal data; and a recommendation result calculation unit that extracts a plurality of recommendation target sensors that are the subject of recommendations based on the causal structure and the importance, determines a recommendation ranking for each of the plurality of recommendation target sensors based on the causal structure and the influence level, and calculates a recommendation result in which the plurality of recommendation target sensors are arranged in order of the recommendation ranking.

2. the importance calculation unit generates an anomaly detection model based on the normal data, and inputs the normal data and the abnormal data into the generated anomaly detection model to calculate the importance. The root cause analysis system of claim 1 .

3. the recommendation result calculation unit selects the plurality of sensors based on the magnitude of the importance, and extracts sensors in which the causal relationship is linked as the recommendation target sensors from the selected plurality of sensors; The root cause analysis system of claim 1 .

4. the recommendation result calculation unit determines the recommendation ranking by a depth-first search that prioritizes the sensor with the greatest influence; The root cause analysis system of claim 1 .

5. a reconsideration candidate calculation unit that performs a causal search on sensors other than the recommendation target sensors among the plurality of sensors and adds the sensors to the recommendation target sensors; The root cause analysis system of claim 1 .

6. a visualization result calculation unit that calculates visualization results of the normal data and the abnormal data; a display unit that displays the causal structure, the recommendation result, and the visualization result to a user; The root cause analysis system of claim 1 .

7. A root cause analysis method in which a root cause analysis system recommends a root cause of an abnormality occurring in an analysis target based on normal data and abnormal data measured on the analysis target by a plurality of sensors, a causal relationship information processing step of calculating a causal structure linking the plurality of sensors by causal relationships and an influence degree for each of the plurality of sensors based on the normal data; an importance calculation step of calculating an importance for each of the plurality of sensors based on the normal data and the abnormal data; and a recommendation result calculation step of extracting a plurality of recommendation target sensors that are the subject of recommendations based on the causal structure and the importance, determining a recommendation ranking for each of the plurality of recommendation target sensors based on the causal structure and the influence level, and calculating a recommendation result in which the plurality of recommendation target sensors are arranged in order of the recommendation ranking.

8. the importance calculation step generates an anomaly detection model based on the normal data, and inputs the normal data and the abnormal data into the generated anomaly detection model to calculate the importance; The root cause analysis method of claim 7.

9. the recommendation result calculation step includes selecting the plurality of sensors based on the magnitude of the importance, and extracting sensors in which the causal relationship is linked from the selected plurality of sensors as the recommendation target sensors; The root cause analysis method of claim 7.

10. the recommendation result calculation step determines the recommendation ranking by a depth-first search that prioritizes the sensor with the greatest influence; The root cause analysis method of claim 7.

11. a reconsideration candidate calculation step of performing a causal search on sensors other than the recommendation target sensors among the plurality of sensors after the recommendation result calculation step, and adding the sensors to the recommendation target sensors; The root cause analysis method of claim 7.

12. A root cause analysis program that is executed by a computer to recommend a root cause of an abnormality that has occurred in an analysis target based on normal data and abnormal data measured by a plurality of sensors on the analysis target, Calculating a causal structure linking the plurality of sensors by a causal relationship and an influence degree for each of the plurality of sensors based on the normal data; Calculating the importance of each of the plurality of sensors based on the normal data and the abnormal data; A root cause analysis program that extracts multiple recommendation target sensors that are the subject of recommendations based on the causal structure and the importance, determines a recommendation ranking for each of the multiple recommendation target sensors based on the causal structure and the influence level, and calculates a recommendation result in which the multiple recommendation target sensors are arranged in order of the recommendation ranking.

Images