Causal search combination abnormality prediction method, causal search combination abnormality prediction device, and causal search combination abnormality prediction program

Abstract

To provide a causal search combination abnormality prediction method capable of achieving accurate search and prediction, a causal search combination abnormality prediction device, and a causal search combination abnormality prediction program.SOLUTION: A causal search combination abnormality prediction method for estimating a cause and predicting an abnormality from time-series data includes the steps of: executing a causal search for the time-series data and estimating a causal relationship between different kinds of data in time-series data; automatically setting an explanatory variable based on the estimated causal relationship after selection of an objective variable; generating one or a plurality of time-series prediction models based on the combination of feature amounts; selecting the generated time-series prediction model using a prediction accuracy evaluation exponent; calculating a deterioration tendency index from the feature amount of the time-series data, estimating a feature amount with the highest deterioration tendency index, and predicting the future of an abnormality tendency index; searching for a feature amount indicative of a precursory phenomenon of abnormality occurrence; and calculating the abnormality tendency index.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a method for abnormality prediction combined with causal search, an apparatus for abnormality prediction combined with causal search, and a program for abnormality prediction combined with causal search. [Background technology]

[0002] When manufacturing products or maintaining a system, if a defective product or an abnormality occurs, it may be that the abnormality is the result of an abnormality occurring in a specific part inside the device or a specific location inside the system. In such cases, if it is possible to identify the location where the abnormality occurs and estimate the time until the abnormality occurs, it is possible to perform efficient maintenance before the abnormality occurs.

[0003] Technologies for performing such causal exploration and prediction have been disclosed (for example, see Patent Documents 1 to 3). Patent Document 1 discloses a method for acquiring observation data corresponding to a plurality of observation variables and determining a causal relationship between the plurality of observation variables. Patent Document 2 discloses a method for estimating the direction and strength of causality between two logs based on the time of occurrence of a spike between the two logs. Patent Document 3 discloses a method for utilizing a causal inference model in a manufacturing site. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6969637 specification [Patent Document 2] Patent No. 6196196 Specification [Patent Document 3] Patent No. 6990626 specification Summary of the Invention [Problem to be solved by the invention]

[0005] However, there is a demand for even higher accuracy in prediction, for example, by improving prediction accuracy.

[0006] In consideration of the above problems, the present invention aims to provide a method for predicting anomalies in combination with causal search, a device for predicting anomalies in combination with causal search, and a program for predicting anomalies in combination with causal search, which are capable of performing accurate searches and predicting anomalies using the results of the searches. [Means for solving the problem]

[0007] A first aspect of the present invention is a method for anomaly prediction combined with causal search for estimating a cause and predicting an anomaly from time series data, comprising the steps of: performing a causal search on the time series data to estimate a causal relationship between different types of data included in the time series data; selecting a target variable and then automatically setting an explanatory variable based on the estimated causal relationship; generating one or more time series prediction models based on a combination of feature quantities; selecting one or more of the generated time series prediction models using a prediction accuracy index indicating which model, among the generated one or more time series prediction models, is most correct for making a future prediction; calculating, from the feature quantities in the time series data, a deterioration trend index indicating the degree of change in the feature quantities that changes toward the time point at which an anomaly occurs, estimating the feature quantity with the highest deterioration trend index score, and predicting the future of the deterioration trend index and an anomaly trend index based on the time series data; searching for a feature quantity that indicates a precursor phenomenon of an anomaly occurrence from the feature quantities; and calculating the anomaly trend index.

[0008] In a first aspect of the present invention, when the time series data are discontinuous values, the estimating step may further include a step of selecting one of the values ​​that the time series data can take and classifying the time series data into the selected value and values ​​other than the selected value prior to performing a causal search on the time series data, a step of converting the time series data into probability values ​​by machine learning as a binary classification problem, and a step of sigmoid transforming the probability values.

[0009] The first aspect of the present invention may further comprise the step of obtaining dominant periodic information of the time series data by converting the time series data into a power spectrum by Fourier transform, and adding the obtained dominant periodic information to the explanatory variables.

[0010] The first aspect of the present invention may further comprise, prior to the causal relationship inferring step, a step of inferring a generation process of the time-series data by clustering or a classification problem.

[0011] In the first aspect of the present invention, the step of selecting one or more generated time series prediction models may include a step of determining a time series prediction model having a highest prediction accuracy index.

[0012] In a first aspect of the present invention, the step of searching for a feature indicative of a precursor phenomenon of an abnormality from among the feature values ​​may include a step of calculating an abnormality trend index by combining a deterioration trend index and a slope of a linear regression of the time series data, and a step of searching for a feature value that monotonically increases or monotonically decreases toward the occurrence of an abnormality using the abnormality trend index.

[0013] The first aspect of the present invention may further comprise the step of calculating an abnormality tendency index and determining the maximum abnormality tendency index as the evaluation index.

[0014] The first aspect of the present invention may further include a step of predicting the future of the abnormal trend index using the ensemble prediction model of the time series prediction model having the highest prediction accuracy index.

[0015] The first aspect of the present invention may further comprise a step of estimating a causal relationship from other factors to the abnormality trend index at the predicted position and calculating the contribution from each factor.

[0016] The first aspect of the present invention may further include a step of providing a user with a predicted result of an anomaly based on the calculated anomaly trend index.

[0017] The first aspect of the present invention may further include a step of providing a user with a predicted result of an anomaly based on the calculated contribution of each factor.

[0018] In the first aspect of the present invention, the time series data may be written in CSV format.

[0019] A second aspect of the present invention is an anomaly prediction device using causal search to estimate causes and predict anomalies from time series data, the device comprising: a storage unit for storing the time series data; an input unit for receiving input from a user; a calculation unit, the causal search unit performing a causal search on the time series data and estimating a causal relationship between a plurality of data included in the time series data; a time series prediction model generation unit for selecting a response variable, setting explanatory variables based on the estimated causal relationship, and generating one or more time series prediction models based on a combination of feature quantities; The present invention relates to a calculation unit that includes a prediction accuracy index evaluation unit that uses a prediction accuracy index indicating whether a model will make a correct prediction for the most future, selects one or more generated time series prediction models, calculates a deterioration trend index indicating the degree of change in the feature amount toward the time point of the abnormality occurrence from the feature amount in the time series data, estimates the feature amount having the highest deterioration trend index score, and predicts the future of the abnormality trend index based on the deterioration trend index and the slope of a linear regression of the time series data, and an abnormality trend index calculation unit that searches for a feature amount indicating a precursor phenomenon of the occurrence of an abnormality from the feature amounts, and calculates the abnormality trend index.

[0020] A third aspect of the present invention provides an anomaly prediction device for use in causal search and anomaly prediction, which comprises a storage unit for storing time series data, an input unit for receiving input from a user, and a calculation unit, and which estimates causes and predicts anomalies from time series data. The causal search and prediction device includes a causal search function for performing a causal search on the time series data and estimating a causal relationship between a plurality of data included in the time series data, a time series prediction model generation function for selecting a response variable, setting explanatory variables based on the estimated causal relationship, and generating one or more time series prediction models based on a combination of feature quantities, and a time series prediction model generation function for generating an actual prediction model based on the estimated causal relationship. The gist of the present invention is to provide a method for selecting one or more generated time series prediction models using a prediction accuracy index, which is the sum of a weighted accuracy index based on measurement and a weighted accuracy index based on evaluation, calculating a deterioration trend index that indicates the degree of change in the feature amount toward the time point of an abnormality from the feature amount in the time series data, estimating the feature amount with the highest deterioration trend index score, and realizing a prediction accuracy index evaluation function based on the deterioration trend index and the time series data, and an abnormality trend index calculation function that searches for the feature amount that indicates a precursor phenomenon of the occurrence of an abnormality from among the feature amounts and calculates an abnormality trend index. Effect of the Invention

[0021] According to the present invention, it is possible to provide a method for abnormality prediction combined with causal search, an abnormality prediction device for combined with causal search, and an abnormality prediction program for combined with causal search that are capable of performing an accurate search and an abnormality prediction using the search results. [Brief description of the drawings]

[0022] [Figure 1] 1 is a block diagram showing an example of a configuration of an anomaly prediction device with causal exploration according to an embodiment of the present invention. [Diagram 2] 1 is a block diagram illustrating an overall flow of an example of an anomaly prediction method combined with causal exploration according to an embodiment of the present invention. [Diagram 3] 1 is a block diagram showing an example of causal search in an anomaly prediction method combined with causal search according to an embodiment. FIG. [Figure 4] FIG. 2 is a block diagram showing an example of time series prediction in the anomaly prediction method combined with causal exploration according to the embodiment. [Diagram 5] 1 is a block diagram showing an example of anomaly prediction in an anomaly prediction method combined with causal exploration according to an embodiment; [Figure 6] 1 is a graph showing predicted values ​​obtained by the anomaly prediction method combined with causal exploration according to the embodiment; [Figure 7] 1 is a diagram illustrating an example of a result of causal exploration obtained by the anomaly prediction method combined with causal exploration according to an embodiment. FIG. [Figure 8] 1 is a diagram illustrating an example of a prediction result obtained by the anomaly prediction method combined with causal exploration according to an embodiment. FIG. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] Next, an embodiment of the present invention will be described with reference to the drawings. In the description of the drawings relating to the embodiment, the same or similar parts are denoted by the same or similar reference numerals. However, the drawings are schematic.

[0024] Furthermore, the embodiments are merely examples of devices and methods for embodying the technical idea of ​​the present invention, and the technical idea of ​​the present invention does not specify the configuration, arrangement, layout, etc. of each component to those described below. The technical idea of ​​the present invention can be modified in various ways within the technical scope defined by the claims.

[0025] (First embodiment) Causal search refers to searching for a connection between a certain event and the cause of the occurrence of the event, and estimating a causal relationship. The anomaly prediction method combined with causal search according to the present embodiment is an anomaly prediction method that uses causal search in combination to predict anomalies that may occur, and predicts anomalies that may occur in the future based on time series data. Time series data is data that indicates the state of the target of causal search, shown in a time series. In making the prediction, the anomaly prediction method combined with causal search according to the present embodiment searches for causal relationships between different types of data among multiple types of data included in the time series data. Furthermore, the search result is combined with the time series data to predict the occurrence of an anomaly that may occur in the future.

[0026] The anomaly prediction method combined with causal discovery according to the present embodiment described above can be realized by executing a program describing the processing contents for realizing the anomaly prediction method combined with causal discovery according to the present embodiment in an anomaly prediction device combined with causal discovery described below.

[0027] The anomaly prediction device with causal exploration according to this embodiment will be described with reference to Fig. 1. As shown in Fig. 1, the anomaly prediction device with causal exploration 10 according to this embodiment is composed of a CPU 11 for executing various calculations, a storage unit 12 for storing processing programs, data including time-series data, and the like, an I / O (input-output interface) 13, a display unit 14, and an input unit 15.

[0028] The causal search and anomaly prediction device 10 may be various electronic computers (computing resources) such as a personal computer (PC), a mainframe, a workstation, a cloud computing system, etc. The I / O 13 is an interface, a buffer, etc. for communication (transmission and reception).

[0029] The display unit 14 is a display device such as a display, and displays the results of calculations by the CPU 11. The input unit 15 is an input device such as a keyboard, a mouse, etc., and is realized by a device that receives input from a user and transmits it to the CPU 11.

[0030] 1 shows a block diagram of the CPU 11 and functional units in the storage unit 12. When each functional unit of the CPU 11 is realized by software, the CPU 11 realizes it by executing commands of a program, which is software that realizes each function. In detail, the CPU 11 includes a data classification unit 111, a causal search unit 112, a time series prediction model generation unit 113, a prediction accuracy index evaluation unit 114, an abnormal trend index calculation unit 115, an ensemble prediction model calculation unit 116, an abnormal trend index tracking unit 117, etc.

[0031] The data classification unit 111 judges whether the time series data acquired by the I / O 13 or the input unit 15 is a continuous value or a discontinuous value, and if it is a discontinuous value, performs a process of replacing the time series data with a continuous value and converts the data format of the time series data into a format that can perform causal search. Here, if the time series data is, for example, a variable, the variable is a continuous variable and the time series data is a continuous value. If the time series data is a flag such as YES / NO or 0, 1, etc., the time series data is a discontinuous value.

[0032] The causal search unit 112 executes a causal search on the time series data and estimates a causal relationship between a plurality of data included in the time series data.

[0033] The time series prediction model generation unit 113 sets explanatory variables based on the objective variables received by the input unit 15 and the causal relationships estimated by the causal search unit 112, and generates one or more time series prediction models based on a combination of features.

[0034] The prediction accuracy index evaluation unit 114 uses the prediction accuracy index to select one or more time series prediction models that have been generated, calculates a deterioration trend index from the features in the time series data, estimates the feature with the highest deterioration trend index score, and predicts the future abnormality trend index.

[0035] The abnormality tendency index calculation unit 115 searches for a feature amount that indicates a precursor phenomenon of an abnormality from among the feature amounts, and calculates an abnormality tendency index.

[0036] The ensemble prediction model calculation unit 116 predicts the future of the abnormal trend indicators using an ensemble prediction model of time series prediction models.

[0037] The abnormal trend index tracking unit 117 estimates the causal relationship from other factors to the abnormal trend index at the future time point to be predicted, and calculates the contribution of each factor.

[0038] The method for abnormality prediction combined with causal exploration according to the first embodiment of the present invention will be described below. Fig. 2 shows a diagram for explaining the overall flow of an example of the method for abnormality prediction combined with causal exploration according to this embodiment.

[0039] In step S201, time series data is acquired.

[0040] In step S202, the acquired time series data is clustered and classified. When performing causal search, the time series data needs to be continuous values. When the time series data is a discontinuous value, for example, a flag such as YES / NO or 0 or 1, the causal search does not work. In the anomaly prediction method combined with causal search according to the present embodiment, the discontinuous value is replaced with the probability value of the value to make it a continuous variable, so that the causal search works even if the time series data is a discontinuous value.

[0041] In step S203, a causal search is performed based on the time series data. If the time series data is a continuous value, a causal search is performed based on the time series data that is a continuous value. If the time series data is a discontinuous value, a causal search is performed based on the probability value replaced in step S202.

[0042] Details of steps S202 and S203 will be described later with reference to FIG.

[0043] In step S204, the time series data is converted into frequency-dependent data by performing FFT (Fast Fourier Transformation) on the time series data, and a dominant frequency is obtained from the frequency-dependent data, and the dominant frequency is set as a feature of the time series data.

[0044] In step S205, the causal factors of the cause of the occurrence of the event, i.e., explanatory variables, are automatically set based on the results of the causal search executed in step S203. If the causal relationships between the data among the multiple data of the time series data are clarified by the causal search, the explanatory variables for the target, i.e., the objective variable, are estimated. As a result, the explanatory variables are automatically set in step S205.

[0045] In step S206, future values ​​of the response variable are estimated based on the dominant frequency determined in step S204 and the explanatory variables set in step S205 to generate prediction models for a plurality of time series.

[0046] In step S207, after generating multiple time series prediction models in step S206, a selection is made to determine which of the multiple generated time series prediction models is most accurate in predicting the future. In this selection, if an evaluation index that is generally used in data science is used among various evaluation indexes, the evaluation index used may not be appropriate, for example, the evaluation index may have a good score but the prediction result may be significantly different from the actual situation. In contrast, in the anomaly prediction method combined with causal exploration according to this embodiment, a prediction accuracy evaluation index is used.

[0047] Details of steps S204, S205, S206, and S207 will be described later with reference to FIG.

[0048] In step S208, a deterioration trend index is calculated from the feature amounts in the time series data, the feature amount having the highest deterioration trend index score is estimated, and the future abnormality trend index is predicted.

[0049] In step S209, the occurrence of an abnormality is predicted. In this embodiment, the prediction of the occurrence of an abnormality includes prediction of the time when the abnormality will occur and prediction of the position where the abnormality will occur.

[0050] Details of steps S208 and S209 will be described later with reference to FIG.

[0051] With reference to FIG. 3, steps S202 and S203 of the anomaly prediction method combined with causal exploration according to this embodiment will be described in detail.

[0052] In step S301, explanatory variables are selected. Here, explanatory variables are variables that explain the cause of the occurrence of a certain event.

[0053] In step S302, a response variable is selected. Here, the response variable is an explanatory variable, that is, a variable that represents an event that occurs in response to a cause.

[0054] In step S303, it is determined whether the time series data is a continuous value or a discontinuous value. Generally, time series data can be used for causal exploration only when it is a continuous value. However, in this embodiment, the time series data may be either a continuous value or a discontinuous value. For example, the time series data may be written in CSV format. If the time series data is a continuous value, the process proceeds to step S308. If the time series data is a discontinuous value, the process proceeds to step S304.

[0055] In step S304, it is determined whether the time series data is binary or ternary or more. Here, binary refers to the time series data taking one of two values, such as YES / NO, ON / OFF, or 0 / 1. If the time series data takes one of three or more values, it is considered to be ternary or more. If the time series data is binary, the process moves to step S305. If the time series data is ternary or more, the process moves to step S306.

[0056] In step S305, if the time series data is a discontinuous value, such as a categorical variable, the discontinuous value is replaced with the probability of each discrete value, thereby making the time series data continuous. For example, when there is a probability of 0 and a probability of 1, the discrete values ​​0 and 1 are converted into the probabilities that each of them can take, and by applying a sigmoid transformation to these probability values, the stepped gap shape of 0 and 1 can be replaced with a smooth shape. In this way, the time series data that was a discontinuous value is converted into a continuous time series data that can be applied to existing causal discovery algorithms such as ICA-LiNGAM and CAM-UV.

[0057] In step S306, when the time-series data is, for example, a categorical variable, if the category indicates two values ​​such as YES or NO, OK or NG, machine learning is performed for a binary classification problem, and a probability is obtained for each binary value. Then, the process proceeds to step S307.

[0058] In step S305, if the categorical variables have three or more values, a category number of each categorical variable is selected, and the categorical variables are divided into categorical variables having the selected category number and categorical variables having an unselected category number, thereby converting the categorical variables into two values. Machine learning is performed as a binary classification problem, and a probability is obtained for each of the two values. The process proceeds to step S307.

[0059] In step S307, a sigmoid transformation is performed on the probability values.

[0060] In step S308, a selection is made as to whether or not to set an upper limit for the explanatory variables. If no selection is made, the process moves to step S309. If a selection is made, the process moves to step S310.

[0061] In step S309, a causal search between time series data is performed using ICA-LiNGAM, and then the process proceeds to step S204.

[0062] In step S310, a causal search between time series data is performed using CAM-UV, and then the process proceeds to step S204.

[0063] With reference to FIG. 4, steps S204, S205, S206, and S207 of the anomaly prediction method combined with causal exploration according to this embodiment will be described in detail.

[0064] In step S401, a response variable is selected.

[0065] In step S402, feature amounts are added. Examples of feature amounts to be added here include basic statistics, lags, feature amounts obtained from date and time (day of the week, month, date, time, etc.), periodic components, seasons, trend components, etc.

[0066] In step S403, the time series data is subjected to FFT (Fourier transform) to obtain frequency-dependent data, i.e., a power spectrum, from the time series data to obtain a dominant frequency, and the dominant frequency is set as a feature of the time series data. This step corresponds to step S204 in FIG. 2.

[0067] In step S404, causal factors, i.e., explanatory variables, are automatically set based on the results of the causal search executed in step S203. This step corresponds to step S205 in FIG.

[0068] In step S405, event information is added. In this embodiment, event information is external information such as break times, holidays, year-end and New Year holidays, festivals, anniversaries, and sale days.

[0069] In step S406, a time series prediction model is generated. At this time, the addition of features takes into consideration combinations. A prediction model is created once for each combination of features, and combinations that do not improve prediction performance are discarded. Furthermore, features whose contribution is smaller than a predetermined value are excluded from the explanatory variables. This step corresponds to step S206 in FIG. 2.

[0070] In step S407, a prediction accuracy index is evaluated. Only patterns of time series prediction models corresponding to combinations of various features are generated, and finally, the time series prediction model with the highest prediction accuracy index is determined to be highly accurate. The prediction accuracy index according to this embodiment is calculated using the following formula.

[0071]

number

[0072] Here, valid is the accuracy index based on the actual prediction, test is the accuracy index based on the evaluation, w1 and w2 are weights, w1 × valid is the contribution of the actual prediction to the above formula, and w2 × test is the contribution of the evaluation to the above formula. The above formula is the accuracy index for the validation interval and the accuracy index when the future prediction from the data terminal is actually predicted between time 1 and time n. In the validation interval, model training is performed using 50% to 90% of the data from the beginning of the time series data, and the model predicts the remaining part of the time series data, that is, 50% to 10% (validation interval) from the end of the input data, and the validity of the model is evaluated by comparing it with the already observed data from the beginning of the time series data in the 50% to 10% (validation interval). The weights w1 and w2 are expressed by the following formulas.

[0073]

number

[0074]

number

[0075] Here, taking into consideration the weighting of the accuracy index in the test interval, the number of data in the validation interval is set to nv and the number of data in the test interval is set to nt. A prediction accuracy index is calculated for all time series prediction models, and the time series prediction model with the highest prediction accuracy index is determined to have high prediction accuracy. The time series prediction model determined to have high prediction accuracy is used in the calculation of the ensemble prediction model pred(t,x) of time series prediction models when predicting future abnormal trend indicators, as described below.

[0076] In step S408, the models are sorted in order of high accuracy according to accuracy indices such as RMSE, MAE, MER, and Rsquared. RMSE, MAE, MER, and Rsquared are objective accuracy indices of a model. The smaller the RMSE value, the more accurate the model, the closer the Rsquared value to 1, the more accurate the model, and the closer the MAE and MER values ​​to 0, the more accurate the model. Based on these accuracy indices, an optimal model is selected. Alternatively, for example, an accuracy index for each model may be displayed on the display unit 14, allowing the user to select the optimal model based on the accuracy index.

[0077] With reference to FIG. 5, steps S208 and S209 of the anomaly prediction method combined with causal exploration according to this embodiment will be described in detail.

[0078] In step S501, a feature amount is added. The feature amount added here is a feature amount that can be an index of the degree of abnormality, such as a basic statistic, a Mahalanobis distance, or a UMAP value.

[0079] In step S502, a physical quantity that reflects the change in the process of deterioration or anomaly occurring is searched for among the multiple feature quantities. In predicting anomalies, it is necessary to capture precursor phenomena, and a physical quantity that indicates a change toward the state at the time of anomaly occurrence is searched for among the multiple feature quantities. In this embodiment, the following formula is used to search for a feature quantity that monotonically increases or decreases toward the occurrence of anomaly, and to predict anomalies. x may be actual observation data, but a value converted to the result of moving average smoothing or LOWESS (locally weighed scatterplot smoothing) is used. In the subsequent processes, this converted x is used.

[0080]

number

[0081]

number

[0082] Here, f(x) is an abnormality trend index, linear_regression(x).gradient() is a monotonically increasing feature such as the slope of the linear regression of previously sampled data, and Monotonicity(x) is a deterioration trend index, which is an index showing the degree of monotonous increase or decrease in a feature that, among multiple feature values, monotonically increases or decreases toward the point at which an abnormality occurs. # is a function that counts the number of data that meets the conditions. The deterioration trend index is calculated from the feature values ​​in the time series data, and the feature value with the highest deterioration trend index score is searched for.

[0083] In step S503, the abnormality tendency index is calculated, and the abnormality tendency index having the maximum value is set as the evaluation index.

[0084] In step S504, the ensemble prediction model pred(t,x) of the time series prediction model is used to predict the future of the abnormal trend indicator. pred(t,x) is expressed as follows:

[0085]

number

[0086]

number

[0087] where time_series_forecasting_model(x) is the time series forecasting model determined to have high accuracy in step S407. h(t) is the degradation function expressed as a function of time, φ is the intercept term that is taken as a constant, θ and β are random parameters that determine the gradient of the model, where θ is a log-normal distribution and β is a Gaussian distribution. u and v are adjustable parameters, respectively. At each time step t, the distributions of θ and β are updated ex post based on the latest observed value of h(t). ε is Gaussian white noise and N(0, σ 2 ) according to the exponential function -σ 2 The / 2 term ensures that the expectation of h(t) satisfies

[0088]

number

[0089] The time when an abnormality will occur can be predicted using pred(t,x). The causal relationship from other factors to the abnormality trend index at the predicted location is estimated, and the contribution from each factor is calculated. Places that require priority maintenance can be visualized, and information about this can be provided to the user via communication means such as email.

[0090] 6 is a graph showing the degradation function obtained by the anomaly prediction method with causal exploration according to the embodiment with respect to time. The threshold in FIG. 6 is the value at which an anomaly occurs. The predicted data increases monotonically toward the threshold with time, and the graph shows that the obtained predicted data and observed data generally match.

[0091] In step S505, the abnormality tendency index is tracked. For example, information on the location of an abnormality inside a machine, such as a machine where an abnormality occurs, is calculated to clarify which sensor value, among the sensor values ​​obtained from multiple sensors connected to the machine, contributed to the trend of the predicted value of the deterioration function, such as that shown in FIG. 6. When the deterioration function, such as that shown in FIG. 6, exceeds a threshold, the contribution of each of the multiple sensors to the deterioration function fluctuation is calculated, and it is considered that the abnormality is caused by a part inside the machine related to a sensor with a large contribution. Therefore, maintenance is performed, such as inspection and replacement of parts inside the machine near the position where a sensor with a large contribution is installed and parts that affect the value measured by the sensor.

[0092] According to the anomaly prediction method combined with causal exploration according to the embodiment, it is possible to predict an approximate period until an anomaly occurs and an occurrence location based on a fluctuation in the degradation function as shown in FIG. 6, and therefore it is possible to inform a user of information regarding the period until an anomaly occurs and the occurrence location.

[0093] An example of CSV format data applicable to this embodiment is shown in the table below.

[0094] [Table 1]

[0095] According to the method for abnormality prediction combined with causal exploration according to this embodiment, causal exploration and prediction are possible even for data in CSV format such as that shown in Table 1.

[0096] Fig. 7 shows an example of the result of causal search obtained by the method for anomaly prediction combined with causal search according to this embodiment. Fig. 7 shows, as an example, the result of causal search performed on the temperature of cooling water in a certain system. Fig. 7 shows the temperature of the cooling water, variables A, B, C, D, E, and F that may be the cause of an anomaly in the temperature of the cooling water, and arrows showing the causal relationships between them. The direction of the arrows shows the cause and result of the causal relationship, and the direction of the arrows is from the cause to the result.

[0097] 7, of variables A, B, C, D, E, and F that can be the cause of the cooling water temperature abnormality, the causes discovered by the method for abnormality prediction combined with causal exploration according to this embodiment are B and F. The degree of influence from the causal variables and the causal relationship of the causal variables can be expressed as probability values, and as an example, the degrees of influence from variables B and F on "abnormal cooling water temperature" are 0.4997 and -0.2639, respectively. The probability values ​​of the causal relationship of variables B and F on "abnormal cooling water temperature" are 95% or more and 5% or less, respectively.

[0098] FIG. 8 shows an example of a prediction result obtained by the method for anomaly prediction combined with causal exploration according to this embodiment. FIG. 8 shows, as an example, the predicted number of customers for each date at a certain store, and based on this result, for example, the number of staff members to be assigned to the store is determined, or in the case of a restaurant, the amount of ingredients to be purchased is adjusted. By the method for anomaly prediction combined with causal exploration according to this embodiment, it is possible to obtain a predicted value and probability for a desired date. Furthermore, it is possible to obtain the degree to which each feature contributed to the predicted value. All of these results are provided to the user, and the user can make decisions based on these results.

[0099] As described above, the present invention naturally includes various embodiments not described here. Therefore, the technical scope of the present invention is defined only by the invention-specifying matters related to the scope of the claims appropriate from the above description. [Explanation of symbols]

[0100] 10. Causal search and anomaly prediction device 11 CPU 12 Storage section 13 I / O 14 Display section 15 Input section 111 Data Classification Section 112 Causal Exploration Department 113 Time series prediction model generation unit 114 Prediction Accuracy Index Evaluation Section 115 Abnormal trend indicator calculation section 116 Ensemble Prediction Model Calculation Unit 117 Abnormal trend indicator tracking section

Claims

1. A method for predicting an anomaly in combination with causal exploration, which estimates a cause and predicts an anomaly from time-series data, comprising: performing a causal search on the time series data to estimate a causal relationship between different types of data included in the time series data; After selecting a response variable, an explanatory variable is automatically set based on the estimated causal relationship; generating one or more time series prediction models based on a combination of features; A step of selecting the one or more time series prediction models generated using a prediction accuracy index indicating which model is most correct in predicting the future, from among the one or more time series prediction models generated; calculating a deterioration trend index indicating a degree of change in the feature amount toward a time point when an abnormality occurs from the feature amount in the time series data, estimating the feature amount having the highest deterioration trend index score, and predicting a future of the deterioration trend index and an abnormality trend index based on the time series data; A step of searching for a feature quantity indicating a precursor phenomenon of an abnormality from among the feature quantities; Calculating an anomaly trend indicator; A method for predicting anomalies in combination with causal exploration, comprising:

2. When the time series data has discontinuous values, the estimating step includes, prior to performing causal search on the time series data, selecting one of the values ​​that the time series data can take and classifying the time series data into the selected value and values ​​other than the selected value; A step of converting the time series data into a probability value by machine learning as a binary classification problem; a sigmoid transformation of the probability values; The method for abnormality prediction combined with causal exploration according to claim 1, further comprising:

3. 2. The method for abnormality prediction with causal exploration according to claim 1, further comprising an additional step of obtaining dominant periodic information of the time series data by converting the time series data into a power spectrum by a Fourier transform, and adding the obtained dominant periodic information to an explanatory variable.

4. 2. The method for predicting anomalies combined with causal exploration according to claim 1, further comprising a step of estimating a generation process of the time series data by clustering or a classification problem prior to the step of estimating a causal relationship.

5. The step of selecting the one or more generated time series prediction models includes: determining a time series prediction model having the highest prediction accuracy index; The method for abnormality prediction combined with causal exploration according to claim 1, further comprising:

6. The step of searching for the feature quantity indicating a precursor phenomenon of an abnormality from among the feature quantities includes: calculating an anomaly trend indicator by combining the degradation trend indicator with a slope of a linear regression of the time series data; 6. The method for predicting anomalies combined with causal exploration according to claim 5, further comprising a step of searching for a feature quantity that monotonically increases or monotonically decreases toward the occurrence of an anomaly using the anomaly tendency index.

7. 7. The method for abnormality prediction combined with causal exploration according to claim 6, further comprising the step of calculating the abnormality tendency index, and determining the abnormality tendency index having a maximum value as an evaluation index.

8. The method for abnormality prediction combined with causal exploration according to claim 7, further comprising a step of predicting a future of the abnormality trend indicator using an ensemble prediction model of time series prediction models having the highest prediction accuracy index.

9. 2. The method for predicting anomalies with causal exploration according to claim 1, further comprising a step of estimating a causal relationship from other factors to the anomaly tendency index at the predicted position and calculating the degree of contribution from each factor.

10. The method for abnormality prediction combined with causal exploration according to claim 1 , further comprising a step of providing a user with a prediction result of the abnormality based on the calculated abnormality tendency index.

11. The method for abnormality prediction combined with causal exploration according to claim 9 , further comprising a step of providing a user with a prediction result of the abnormality based on the calculated contribution degree from each of the factors.

12. The method for abnormality prediction combined with causal exploration according to claim 1 , wherein the time series data is recorded in CSV format.

13. A causal exploration and prediction device that estimates causes and predicts anomalies from time-series data, A storage unit that stores the time series data; an input unit for accepting input from a user; A calculation unit, a causal search unit that performs a causal search on the time series data and estimates a causal relationship between a plurality of data included in the time series data; a time series prediction model generation unit that, after selecting a response variable, sets explanatory variables based on the estimated causal relationship and generates one or more time series prediction models based on a combination of feature quantities; a prediction accuracy index evaluation unit that selects the one or more generated time series prediction models using a prediction accuracy index that indicates which model among the one or more generated time series prediction models makes the most correct prediction for future prediction, calculates a deterioration trend index that indicates a degree of change in the feature value that changes toward a time point when an abnormality occurs from the feature value in the time series data, estimates the feature value with the highest deterioration trend index score, and predicts the future of the deterioration trend index and the abnormality trend index based on the time series data; an abnormality tendency index calculation unit that searches for the feature amount indicating a precursor phenomenon of an abnormality occurrence from the feature amount and calculates an abnormality tendency index; A calculation unit including: A causal search and abnormality prediction device comprising:

14. A causal exploration and prediction device that estimates causes and predicts abnormalities from the time-series data, the causal exploration and prediction device comprising: a storage unit that stores time-series data; an input unit that accepts input from a user; and a calculation unit. The causal exploration and prediction device further comprises: A causal exploration function that performs a causal exploration on the time series data and estimates a causal relationship between a plurality of data included in the time series data; a time series prediction model generation function for selecting a response variable, setting explanatory variables based on the estimated causal relationship, and generating one or more time series prediction models based on a combination of feature quantities; a prediction accuracy index evaluation function that uses a prediction accuracy index indicating which model among the one or more generated time series prediction models is most correct in making future predictions, calculates a deterioration trend index indicating a degree of change in the feature value toward the time point of an abnormality occurrence from the feature value in the time series data, estimates the feature value with the highest deterioration trend index score, and predicts the future of the deterioration trend index and the abnormality trend index based on the time series data; and A program for combined cause-and-effect search and anomaly prediction, which searches for, from among the feature quantities, the feature quantities that indicate precursor phenomena of an anomaly occurrence, and realizes an anomaly tendency index calculation function.

Images