JP7584376B2 - Abnormality determination method, abnormality determination system, and program - Google Patents

Description

This disclosure relates to an abnormality determination method, an abnormality determination system, and a program.

In recent years, methods have been proposed that use machine learning to determine abnormalities in machinery and work. When there is sufficient data on normal and abnormal conditions, a model that distinguishes between normal and abnormal conditions can be constructed using supervised learning. However, when it comes to machinery and work, there is generally little data on abnormal conditions, so models are often constructed by learning from normal data (for example, Patent Document 1).

JP 2020-160618 A

By using a trained model built using normal data, it is possible to determine whether the subject being evaluated is normal or abnormal. However, if it is determined to be abnormal, it is unclear how to return the subject to a normal state.

The present disclosure provides an abnormality determination method, an abnormality determination system, and a program that can solve the above problems.

The anomaly determination method disclosed herein includes the steps of: comparing a determination model created by learning normal data including a plurality of variables with data to be evaluated including the plurality of variables; calculating a distance between the data to be evaluated and the determination model for each of the plurality of variables; and calculating the sum of the distances as the degree of anomaly of the data to be evaluated; determining whether the data to be evaluated is abnormal or not based on the degree of anomaly; calculating a direction of adjustment of a value for reducing the degree of anomaly for at least one of the plurality of variables; and outputting the adjustment direction, wherein in the step of calculating the adjustment direction, a predetermined number of variables are selected in descending order of the distance, and a direction in which the distance is reduced for the selected variables is calculated as the adjustment direction, and in the step of outputting the adjustment direction, the selected variables and the adjustment direction for each variable are output.

The anomaly determination system disclosed herein includes means for comparing a determination model created by learning normal data including a plurality of variables with data to be evaluated including the plurality of variables, calculating a distance between the data to be evaluated and the determination model for each of the plurality of variables, and calculating the sum of the distances as a degree of anomaly of the data to be evaluated, means for determining whether the data to be evaluated is anomalous or not based on the degree of anomaly, means for calculating a direction in which to adjust a value to reduce the degree of anomaly for at least one of the plurality of variables, and means for outputting the adjustment direction , wherein the means for calculating the adjustment direction selects a predetermined number of variables in descending order of the distance, and calculates a direction in which to reduce the distance for the selected variables as the adjustment direction, and the means for outputting the adjustment direction outputs the selected variables and the adjustment direction for each of the variables.

The program disclosed herein has a step of causing a computer to execute a process of comparing a judgment model created by learning normal data including a plurality of variables with data to be evaluated including the plurality of variables, calculating a distance between the data to be evaluated and the judgment model for each of the plurality of variables, and calculating the sum of the distances as the degree of abnormality of the data to be evaluated; a step of determining whether the data to be evaluated is abnormal based on the degree of abnormality; a step of calculating a direction of adjustment of a value to reduce the degree of abnormality for at least one of the plurality of variables; and a step of outputting the adjustment direction, wherein in the step of calculating the adjustment direction, a predetermined number of variables are selected in descending order of the distance, and a direction in which the distance is reduced for the selected variables is calculated as the adjustment direction, and in the step of outputting the adjustment direction, a process of outputting the selected variables and the adjustment direction for each variable .

The above-mentioned anomaly determination method, anomaly determination system, and program can provide not only the results of the anomaly determination, but also information on the cause of the anomaly and how to return to a normal state.

1 is a block diagram showing an example of an abnormality determination system according to an embodiment; 1 is a flowchart illustrating an example of a method for constructing a determination model according to an embodiment. FIG. 1 is a first diagram showing an example of a determination method according to an embodiment. 11A and 11B are diagrams illustrating a method for calculating an abnormality degree and a direction according to an embodiment. 1A to 1C are diagrams illustrating a determination method based on a probability density distribution according to an embodiment. 1A to 1C are diagrams illustrating a determination method based on reproducibility according to an embodiment. 5 is a flowchart illustrating an example of an abnormality determination process according to the embodiment. 1 is a flowchart illustrating an example of a parameter evaluation method according to an embodiment. FIG. 2 is a diagram illustrating an example of a hardware configuration of the abnormality determination system according to the embodiment.

Hereinafter, the abnormality determination method of the present disclosure will be described with reference to FIGS.
<Embodiment>
(composition)
FIG. 1 is a block diagram illustrating an example of an abnormality determination system according to an embodiment.
The abnormality determination system 1 is a system that determines abnormalities in monitored machinery and equipment, work, etc. For example, the abnormality determination system 1 acquires measurement values measured by sensors installed in the machinery and equipment, control signals issued by a control device, and the like, and determines whether or not an abnormality has occurred in the machinery and equipment during operation. Alternatively, the abnormality determination system 1 acquires measurement values measured by sensors of tools and equipment used by workers, or measurement results of work results, and determines whether or not an abnormality has occurred in the work. In addition to such abnormality determination, when the abnormality determination system 1 determines that there is an abnormality, it provides guidance information that indicates the cause of the abnormality and a method for improving it.

The abnormality determination system 1 includes an abnormality determination device 10 and a simulator 20. The abnormality determination device 10 includes a data acquisition unit 11, a setting reception unit 12, a control unit 13, an output unit 14, and a memory unit 15.

The data acquisition unit 11 acquires learning data used to construct a judgment model for abnormality judgment. The data acquisition unit 11 also acquires evaluation target data that is the subject of abnormality judgment. The learning data and evaluation target data include multiple parameters such as sensor measurement values such as temperature and pressure, control signals such as command values for valve opening and rotation speed, and information on the time when each parameter was measured, commanded, etc.

The setting reception unit 12 receives settings for normal state intervals for the learning data acquired by the data acquisition unit 11. The normal state interval is setting information that specifies the range of learning data in normal times, such as the time period when the operating state of the machinery and equipment is normal. The setting reception unit 12 also receives settings such as a threshold value for determining an abnormality. Furthermore, the setting reception unit 12 receives settings for whether or not to present parameters determined to be the cause of an abnormality as guidance information, or whether or not to present them preferentially. For example, the guidance information presents the parameters that are the cause of the abnormality and how to improve them, but parameters that cannot be operated by the user can be excluded from the presented information. Conversely, parameters that are effective in returning the state of the machinery and equipment to normal and that can be controlled and operated by the user can be presented preferentially as guidance information (weighting coefficients described later).

The control unit 13 controls each of the processes of constructing a judgment model, judging an abnormality, and outputting guidance information. The control unit 13 includes a model construction unit 131, an abnormality judgment unit 132, and a factor evaluation unit 133.
The model construction unit 131 learns learning data (normal data) in the normal state section and constructs a judgment model.
The abnormality determination unit 132 determines whether the mechanical equipment or the work is normal or abnormal based on the determination model and the evaluation target data.
When the abnormality determination unit 132 determines that an abnormality has occurred, the factor evaluation unit 133 selects a predetermined number of parameters that have a large influence from among the parameters that caused the abnormality determination, and evaluates how the values of those parameters should be (for example, by increasing or decreasing them) to bring them closer to the normal state.

The output unit 14 outputs the normal/abnormal determination result obtained by the abnormality determination unit 132. The output unit 14 creates guidance information based on the evaluation result obtained by the factor evaluation unit 133, and outputs the guidance information.
The storage unit 15 stores various information such as learning data, a judgment model, and settings of normal state intervals.

The simulator 20 is a computational model that simulates the operation of the monitored machinery and equipment and work such as welding. For example, when parameters contained in the learning data are given, the simulator 20 simulates the operation and work of the machinery and equipment, and outputs values indicating the state of the machinery and equipment after a certain period of operation, evaluation values for the work results, etc.

(Building a decision model)
FIG. 2 is a flowchart showing an example of a method for constructing a judgment model according to the embodiment. First, the data acquisition unit 11 acquires learning data (step S1) and records the learning data in the storage unit 15. For example, the learning data is time-series data on a plurality of parameters such as temperature, pressure, current, and voltage, and most of the learning data is data on normal conditions. Next, the user sets a normal state section (step S2). The setting reception unit 12 receives the setting of the normal state section and records it in the storage unit 15. For example, if a certain mechanical equipment is operating normally from time t1 to t2, the user sets the time t1 to t2 as a normal state section. Next, the user instructs the abnormality determination device 10 to construct a judgment model. Then, the model construction unit 131 starts constructing the judgment model. First, the model construction unit 131 extracts learning data belonging to the normal state section (step S3). For example, the model construction unit 131 extracts data of the normal state section from the learning data from time t1 to t2. Next, the model construction unit 131 learns the extracted learning data in normal times and constructs a judgment model (step S4). Most of the parameters that can be collected during operation of the machinery and equipment or during actual work are normal data. Therefore, a normal state section is set, normal data is extracted from all the learning data, and the extracted data is learned to construct a judgment model in which a normal value range for each parameter is determined. The construction method of the judgment model is not particularly limited. For example, a range that envelopes the distribution of normal data may be used as the judgment model, or a probability density distribution of data in normal times may be calculated to use the probability density distribution as a judgment model. For example, a probability density distribution (probability density function) for each parameter included in the normal data is calculated and used as the judgment model. Alternatively, a model that performs an abnormality judgment based on the reproducibility of the value of each parameter in data in normal times may be constructed. Specifically, a model that reproduces each parameter of the normal data to be learned (for example, an autoencoder that outputs a similar value when normal data is input) may be constructed, and this model may be used as the judgment model. Alternatively, a judgment model may be constructed by learning the correlation between multiple parameters in normal times.

Figure 3A shows an overview of the process of determining anomalies using a determination model constructed from normal learning data. The determination model M shown in Figure 3A is a schematic diagram of a determination model in which normal data is plotted in space and the distribution range is taken as the normal region.

(Example of a decision model)
Point x and point k shown in FIG. 3A are points that conceptually indicate values calculated from a plurality of parameters included in the learning data and the evaluation target data, where point x indicates the evaluation target data, and point k indicates the point closest to point x among the points included in the judgment model M (not limited to the nearest point, and may be, for example, any point within a predetermined range including the nth nearest point or the nearest point). When performing an abnormality judgment of the evaluation target data (point x), the abnormality judgment unit 132 identifies a nearby point k included in the judgment model M and calculates the distance (degree of abnormality) between point x and point k. If the distance is equal to or greater than a threshold value, the abnormality judgment unit 132 judges that the evaluation target data indicated by point x is abnormal. When point x is judged to be abnormal, the factor evaluation unit 133 identifies a parameter that has a high contribution to the distance among the parameters constituting point x and point k, compares the identified parameters with each other, and calculates how to make the value of the identified parameter at point x approach normal. These processes will be specifically described below with reference to FIG. 3B.

(Calculation method of abnormality level, abnormality factor, and direction)
3B shows specific examples of a method for calculating the distance (degree of anomaly), a method for identifying the cause of anomaly, and a method for calculating the direction of value adjustment. As an example, the learning data and the evaluation target data include three parameters, which are parameters 1, 2, and 3, respectively. Furthermore, the evaluation target data x is represented as x=(x1, x2, x3), and the neighboring point k of the evaluation target data x included in the judgment model M is represented as k=(k1, k2, k3). Note that the number of parameters (three) is just an example, and there is no limit to the number.

・Distance (degree of abnormality)
The abnormality determination unit 132 calculates the distance D(x) between the evaluation target data x and the neighboring point k by the formula (1a) in Fig. 3B. The distance D(x) is also called the degree of abnormality. As shown in the formula (1a), the degree of abnormality is a single value obtained by adding up the differences between the respective parameters. The abnormality determination unit 132 determines that the distance D(x) is abnormal if it is equal to or greater than the threshold Th, and determines that the distance D(x) is normal if it is less than the threshold Th.
Furthermore, the abnormality determination unit 132 may calculate the distance D(x) by weighting each of the parameters 1 to 3. For example, the abnormality determination unit 132 multiplies each of the parameters 1 to 3 by the weighting coefficients A1 to A3, and calculates the distance D(x) by the formula (1b) in FIG. 3B. If it is known that parameter 1 has a large effect on the state of the equipment or work, and the user places importance on parameter 1, the user sets a large value for A1. This can improve the accuracy of abnormality detection. Note that the distance may be calculated using Mahalanobis distance or Manhattan distance in addition to Euclidean distance.

Factor The factor evaluation unit 133 selects a predetermined number of parameters from (k ₁ -x ₁ ) ² , (k ₂ -x ₂ ) ² , and (k ₃ -x ₃ ) ² in descending order. That is, the factor evaluation unit 133 selects a parameter that contributes greatly to distance as a factor of an abnormality. The factor evaluation unit 133 may assign weights to each of the parameters 1 to 3 to identify a parameter that is a factor of an abnormality. For example, the factor evaluation unit 133 may select a predetermined number of parameters in descending order of the value obtained by multiplying each of the parameters 1 to 3 by the weighting coefficients A1 to A3. When the parameter 1 is a parameter that is effective in terms of control (easily controllable, easily affects the driving state, etc.), the user can set a large value to A1 to make the parameter 1 more likely to be selected as an abnormality factor.

Direction The factor evaluation unit 133 calculates the direction of adjustment of the value for the parameter identified as the abnormality factor to bring it closer to normal using a formula for calculating the difference before squaring. For example, when parameter 1 is identified as the abnormality factor, the factor evaluation unit 133 calculates (k ₁ - x ₁ ) and evaluates the direction of bringing (k ₁ - x ₁ ) closer to 0. Specifically, if (k ₁ - x ₁ ) is a positive value, the factor evaluation unit 133 evaluates that the direction of increasing x ₁ is the direction of bringing it closer to normal, and if (k ₁ - x ₁ ) is a negative value, the factor evaluation unit 133 evaluates that the direction of decreasing x ₁ is the direction of bringing it closer to normal.

Other examples of the judgment model are shown in Figures 4A and 4B. Figure 4A shows an example in which a probability density distribution is used as the judgment model, and Figure 4B shows an example of a judgment model in which anomaly judgment is performed based on the degree of reproducibility.
The upper left diagram of FIG. 4A shows the probability density distribution of parameter 1 among the multiple parameters included in the normal data. The probability density distribution calculated from the normal data illustrated in the upper left diagram is the judgment model. The lower left diagram of FIG. 4A shows the probability density distribution of parameter 1 among the multiple parameters included in the evaluation target data. The vertical axis of each graph shows the probability density, and the horizontal axis shows the value of parameter 1. The abnormality judgment unit 132 calculates the probability density distribution using a data set of the evaluation target data measured during a predetermined period (for example, t1 to t2), and calculates the density ratio with respect to the probability density distribution of the normal data (= judgment model) calculated in advance. More specifically, the abnormality judgment unit 132 calculates the density ratio (the ratio of the probability density function representing the curve of the probability density distribution) of the probability density distribution of the normal data and the evaluation target data in the range v1 to v2 in which the value of the evaluation target data exists as the degree of abnormality, and judges that the parameter 1 is abnormal if the density ratio deviates from the allowable range centered on "1". An example of the calculated density ratio is shown in the right diagram of FIG. 4A. In the illustrated example, since the density ratio deviates from "1", parameter 1 of the evaluation target data is determined to be abnormal. The abnormality determination unit 132 calculates a density ratio (degree of abnormality) for each of a plurality of parameters included in the evaluation target data, and determines that the evaluation target data measured in a predetermined period is abnormal if there is even one parameter whose density ratio deviates from an allowable range centered on "1". Alternatively, the abnormality determination unit 132 may be configured to determine that there is an abnormality when there are a predetermined number or more parameters whose density ratio deviates from the allowable range.

The factor evaluation unit 133 may also perform logarithmic conversion of the density ratios (degrees of abnormality) calculated for each parameter that deviate from the allowable range and identify a predetermined number of parameters as abnormal factors in ascending order of the density ratio, or identify a predetermined number of parameters as abnormal factors in descending order of the variance of the probability density distribution, or identify a predetermined number of parameters as abnormal factors in descending order of the area of the range that deviates from the allowable range centered on the density ratio "1", or calculate the distance L between the average value of the probability density distribution of normal data and the average value of the probability density distribution of the evaluation target data, and identify a predetermined number of parameters as abnormal factors in descending order of the distance L. The factor evaluation unit 133 may also identify abnormal factors by combining these. For example, the factor evaluation unit 133 may identify parameters identified as abnormal factors by multiple methods as final abnormal factors, or may identify all parameters identified as abnormal factors by any of multiple methods as final abnormal factors. In addition, for each parameter identified as an abnormality factor, the factor evaluation unit 133 evaluates that the direction of the value with the highest probability density of normal data (average value of the probability density distribution) (the direction of the bold arrow in the lower left diagram of FIG. 4A) is the direction approaching normal.

Next, referring to FIG. 4B, the process of performing an anomaly judgment using a judgment model based on reproducibility will be described. (Step 1) The model construction unit 131 constructs a judgment model (e.g., an autoencoder) that uses normal data including parameters 1 to 3 as learning data, and learns to output the input data itself regardless of the input of any normal data, and reproduces the normal data. (Step 2) The anomaly judgment unit 132 inputs parameters 1 to 3 (Q1 to Q3) of the evaluation target data to the constructed judgment model. If the values of parameters 1 to 3 output by the judgment model are Q1' to Q3', the anomaly judgment unit 132 calculates |Q1-Q1'|, |Q2-Q2'|, and |Q3-Q3'|, and sets these values as the reproducibility (degree of anomaly). The smaller the reproducibility (degree of anomaly), the better, and the anomaly judgment unit 132 judges the evaluation target data to be abnormal if the reproducibility for at least one parameter exceeds a threshold value. Alternatively, the anomaly determination unit 132 may be configured to determine an anomaly when there are a predetermined number or more parameters whose reproducibility exceeds a threshold. Furthermore, the factor evaluation unit 133 identifies a predetermined number of parameters as anomaly factors in descending order of reproducibility (anomaly level). For example, when |Q1-Q1'|=0, |Q2-Q2'|=2, and |Q3-Q3'|=5, and only one anomaly factor is identified, the factor evaluation unit 133 identifies parameter 3 as an anomaly factor. Furthermore, the factor evaluation unit 133 evaluates the direction in which reproducibility (anomaly level)=0 is a direction approaching normal. For example, for Q3' of the evaluation target data, the direction in which |Q3-Q3'|=5 becomes |Q3-Q3'|=0 is evaluated as a direction approaching normal.

In addition, when a judgment model is created by learning the correlation between multiple parameters and the correlation is expressed by a correlation coefficient calculated by a predetermined method, for example, the abnormality judgment unit 132 calculates the difference between the correlation coefficient between parameters 1 and 2 of the evaluation target data and the correlation coefficient between parameters 1 and 2 in the judgment model. The abnormality judgment unit 132 also calculates the difference in correlation coefficients between other parameters, and if the sum of these is equal to or greater than a threshold, judges the evaluation target data to be abnormal. When the evaluation target data is judged to be abnormal, the factor evaluation unit 133 selects a predetermined number of parameters in descending order of the difference between the parameters, and calculates how to make the parameter values approach normal. For example, if the correlation coefficient between parameters 1 and 2 and the correlation coefficient between parameters 3 and 4 are significantly higher or lower than the correlation coefficients of those parameters in the normal state, the factor evaluation unit 133 selects the set of parameters 1 and 2 and the set of parameters 3 and 4. The factor evaluation unit 133 then compares the correlation coefficient between corresponding parameters in the normal data with the correlation coefficient in the evaluation target data, and calculates the direction of adjustment for each parameter required to bring the correlation closer to normal, for example, by increasing parameter 1 and decreasing parameter 2, or decreasing parameter 3 and increasing parameter 4.

By presenting to the user the judgment results of the anomaly judgment unit 132, as well as the evaluation results of the anomaly factors identified by the factor evaluation unit 133 and the direction to move toward normalcy, the user can understand how to return the state of the machinery and equipment and operations to a normal state.

(Abnormality Judgment Processing)
Next, the flow of anomaly determination processing by the anomaly determination system 1 will be described using an example in which the anomaly determination method, the anomaly cause identification method, and the method of evaluating the direction toward normality described with reference to Figures 3A and 3B are applied.
FIG. 5 is a flowchart illustrating an example of an abnormality determination process according to the embodiment.
Prior to the abnormality determination, thresholds and the like are set (step S10). For example, the user sets the abnormality degree threshold Th used to determine whether the condition is normal or abnormal, a weighting factor such as A1 when formula (1b) is used, the maximum number of factors that will cause an abnormality to be presented in the guidance information (e.g., three), and parameters to be excluded from the parameters to be presented in the guidance information. Regarding the setting of the maximum number of abnormality factors to be presented, a number that can be handled by a person is set when the user visually recognizes the guidance information and uses it as a reference for subsequent control. Alternatively, if it is used in cooperation with a control device or the like and applied to computer control, it may be set to present several tens of parameters. Regarding the setting of parameters to be excluded from the parameters to be presented in the guidance information, parameters that cannot be controlled or parameters that may cause problems when controlled are set. The reason for excluding parameters that cannot be controlled is that the guidance information is presented with a direction to adjust the value to return to a normal state, and therefore only parameters that can be controlled by the user are presented. Parameters that may cause problems will be described next with reference to FIG. 6. Furthermore, instead of setting parameters to be excluded, parameters that can be presented as guidance information may be set. The setting reception unit 12 receives these settings and records the set information in the storage unit 15 .

Next, the data acquisition unit 11 acquires the evaluation target data (step S11). For example, the abnormality determination device 10 may be communicably connected to the machine equipment to be monitored, and the data acquisition unit 11 may acquire the evaluation target data online. Alternatively, the data acquisition unit 11 may acquire evaluation target data previously collected in an offline state. As an example, the evaluation target data is assumed to be x. Furthermore, x is composed of three parameters 1 to 3, and x = (x1, x2, x3). The data acquisition unit 11 outputs the acquired evaluation target data to the abnormality determination unit 132.

Next, the anomaly determination unit 132 calculates the degree of anomaly (step S12). For example, the anomaly determination unit 132 calculates the distance between the evaluation target data and the determination model for each parameter, and calculates the sum of the distances as the degree of anomaly. The degree of anomaly is the distance D(x) described with reference to FIG. 3B. The anomaly determination unit 132 calculates the neighboring point k in the determination model, and calculates the degree of anomaly using formula (1a) or formula (1b) in FIG. 3B.

Next, the abnormality determination unit 132 compares the calculated degree of abnormality with the threshold value Th set in step S10 to determine whether it is normal or abnormal (step S13). If the degree of abnormality is equal to or greater than the threshold value Th, the abnormality determination unit 132 determines it to be abnormal, and if the degree of abnormality is less than the threshold value Th, it determines it to be normal. If it determines it to be normal (step S13; No), the abnormality determination unit 132 outputs the determination result "normal" to the output unit 14. The output unit 14 outputs the determination result "normal" to a display device or the like (step S17).

When it is determined that there is an abnormality (step S13; Yes), the abnormality determination unit 132 outputs the determination result "abnormal" to the output unit 14, and instructs the factor evaluation unit 133 to calculate which parameter is the cause of the high abnormality level and the direction to reduce the abnormality level. The factor evaluation unit 133 identifies the cause of the abnormality (step S14). The factor evaluation unit 133 focuses on the distance (such as (k ₁ -x ₁ ) in formula (1a)) for each parameter calculated when calculating the abnormality level, and selects a predetermined number of parameters in descending order of distance (excluding those set as parameters to be excluded in step S10, or selecting from among parameters that can be presented if they are set). For example, if it is set to select the top two in step S10, the factor evaluation unit 133 selects two parameters from parameters 1 to 3 in descending order of distance (parameters 1 and 2). Next, the factor evaluation unit 133 calculates the direction for moving the parameter selected in step S14 closer to the normal state (step S15). For example, the factor evaluation unit 133 calculates (k ₁ -x ₁ ) for the direction for moving parameter 1 closer to normal, and calculates that if the calculation result is negative, the value of parameter 1 should be reduced, and if the calculation result is positive, the value of parameter 1 should be increased. The factor evaluation unit 133 also calculates (k ₂ -x ₂ ) for parameter 2, and calculates that if the calculation result is negative, the value of parameter 2 should be reduced, and if the calculation result is positive, the value of parameter 2 should be increased. The factor evaluation unit 133 outputs to the output unit 14 the parameter that is the cause of the identified abnormality and information on the direction for moving the parameter closer to normal.

Next, the output unit 14 outputs the judgment result and guidance information (step S16). For example, the output unit 14 displays on the display device the judgment result of "abnormality has occurred" and guidance information indicating that the cause of the abnormality is parameter 1 and parameter 2, and instructing the operator to decrease the value of parameter 1 and increase the value of parameter 2. This allows the operator of the machinery and equipment to not only know that the operating condition of the machinery and equipment has become abnormal, but also to see the displayed guidance information and know that the abnormality is caused by parameters 1 and 2. Furthermore, parameters 1 and 2 can be adjusted so that the operating condition of the machinery and equipment becomes normal.

In general, when machine learning is used to judge anomalies in machinery equipment such as processing machines and in work such as welding, data during abnormality is scarce, so a judgment model is often constructed using only data during normal times. Usually, such a judgment model only outputs the level of anomaly as an anomaly degree, and cannot guide the cause of the anomaly or a method for reducing the anomaly degree. However, in machinery equipment and work, in order to ensure quality, cost, and lead time, it is necessary to suppress rework as much as possible, and responsiveness is required in dealing with anomalies. For this reason, it is necessary to know the cause of the anomaly and the method of dealing with it as soon as possible when an anomaly occurs. In contrast, according to this embodiment, the distance between the evaluation target data and the judgment model for each parameter is calculated, and the sum of the distances is calculated as the anomaly degree. Then, when the anomaly degree exceeds a threshold value, it is possible to present which parameter is the cause of the increase in the anomaly degree based on the distance calculated for each parameter. In addition, by evaluating the method (direction) for reducing the anomaly degree, it is possible to present how to move the parameter to reduce the anomaly degree.

As described above, the information on the parameters that are abnormal factors and the direction to bring the parameters closer to normal, which is presented as guidance information, has the effect of quickly normalizing the state of work, etc., particularly when applied to online work or the operation of machinery and equipment. However, for example, there is a possibility that the accuracy of the judgment model is insufficient due to a lack of learning data, etc., and that inappropriate parameters are presented, or that changing a certain parameter affects other parameters and does not have a beneficial effect overall. In order to prevent such problems, the abnormality judgment system 1 of this embodiment can use the simulator 20 to confirm in advance whether it is appropriate to adjust the parameters identified as abnormal factors. Next, an example of a process for evaluating whether it is appropriate to adjust (bring closer to normal) the parameters identified in step S14 will be described with reference to FIG. 6.

(Method of Evaluating Abnormality Factor Parameters)
6 is a flowchart showing an example of a parameter evaluation method according to the embodiment. The same steps as those in FIG. 5 are denoted by the same reference numerals and will be briefly described.
The abnormality determination device 10 and the simulator 20 are connected to be able to communicate with each other. The simulator 20 is a calculation model that simulates the operation of a mechanical equipment, and is configured to simulate the operation of the mechanical equipment when a user sets calculation conditions such as initial conditions and instructs the simulation of the operation. The simulator 20 outputs calculation results (temperature, pressure, valve opening, etc.) calculated during the simulation, which become evaluation target data. The data acquisition unit 11 acquires the evaluation target data output by the simulator 20 during the simulation, and the abnormality determination device 10 performs abnormality determination, etc. on this data. The control unit 13 can output control information to the simulator 20 through the output unit 14. Alternatively, the user may input the control information to the simulator 20. When the simulator 20 receives the control information, it reflects the contents of the control information in the simulation calculation. For example, when the control unit 13 outputs control information to the simulator 20 to reduce the value of parameter 1 by a predetermined amount, the simulator 20 simulates, based on the control information, an operation that would occur when an operation is performed on the mechanical equipment to reduce parameter 1, and calculates a state in which the value of parameter 1 has been reduced by the predetermined amount. The following description will be given on the assumption that the abnormality determination device 10 and the simulator 20 are configured in this manner.

First, the user sets the thresholds and the like (step S10). Next, the user sets the calculation conditions and starts the simulator 20. The simulator 20 starts simulating the mechanical equipment (step S101). The data acquisition unit 11 acquires the evaluation target data output by the simulator 20 (step S11). The abnormality determination unit 132 calculates the degree of abnormality (step S12) and performs abnormality determination (step S13). In the case of an abnormality (step S13; Yes), the factor evaluation unit 133 identifies a predetermined number of parameters that are the cause of the abnormality (step S14) and calculates the direction of approaching normal for each parameter (step S15). The control unit 13 creates control information for adjusting one of the parameters identified by the factor evaluation unit 133 as the cause of the abnormality in a direction approaching normal by a predetermined amount, outputs this control information to the simulator 20, and reflects it in the simulation content (step S18). The control unit 13 executes the process of step S18 for the predetermined number of parameters identified in step S14. The simulator 20 simulates the operation of the mechanical equipment after the parameter adjustment. The data acquisition unit 11 records the calculation results calculated by the simulator 20 after the parameter adjustment in the storage unit 15 (step S19). Note that, in regard to steps S14 to S18, if multiple parameter adjustments are reflected at once, it becomes unclear which parameter adjustment has had an effect on the mechanical equipment. Therefore, it is possible to select only one parameter (for example, the parameter that contributes most to the degree of abnormality), adjust only that parameter value in a direction closer to normal, and reflect this in the simulator 20.

The user evaluates whether it is appropriate to adjust the parameter identified in step S15 based on the simulation results by the simulator 20 recorded in the storage unit 15 (step S20). For example, if the calculation result after the parameter adjustment indicates that the operating state of the mechanical equipment has returned to normal, it is considered that it was appropriate to adjust the parameter in the direction calculated in step S15. On the other hand, if the calculation result after the adjustment indicates that the operating state of the mechanical equipment remains abnormal even after the adjustment, it is considered that the adjustment of the parameter identified in step S14 may not be appropriate. If the adjustment of the parameter is not appropriate, the user can set the parameter to be excluded from the parameters presented as guidance information. For example, in step S10 of FIG. 5, the parameter is set as a parameter to be excluded from the evaluation target. Then, in step S14, the factor evaluation unit 133 excludes the excluded parameter from the identification target. As a result, the excluded parameter is no longer presented as guidance information, and it is possible to prevent the parameter from being erroneously adjusted when an abnormality is determined.

FIG. 7 is a diagram showing an example of the hardware configuration of the anomaly determination system according to the embodiment. The computer 900 includes a CPU 901, a main memory device 902, an auxiliary memory device 903, an input/output interface 904, and a communication interface 905. The above-mentioned anomaly determination system 1 is implemented in the computer 900. The above-mentioned functions are stored in the auxiliary memory device 903 in the form of a program. The CPU 901 reads the program from the auxiliary memory device 903, expands it in the main memory device 902, and executes the above-mentioned processing according to the program. The CPU 901 also secures a memory area in the main memory device 902 according to the program. The CPU 901 also secures a memory area in the auxiliary memory device 903 for storing data being processed according to the program.

A program for implementing all or part of the functions of the anomaly determination system 1 may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed to perform processing by each functional unit. The term "computer system" as used herein includes hardware such as an OS and peripheral devices. In addition, if a WWW system is used, the term "computer system" also includes a home page providing environment (or display environment). In addition, the term "computer-readable recording medium" refers to portable media such as CDs, DVDs, and USBs, and storage devices such as hard disks built into a computer system. In addition, if the program is distributed to a computer 900 via a communication line, the computer 900 that receives the program may expand the program into the main storage device 902 and execute the above processing. In addition, the program may be for implementing part of the functions described above, and may further be capable of implementing the functions described above in combination with a program already recorded in the computer system.

As described above, several embodiments of the present disclosure have been described, but all of these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope of the invention and its equivalents as described in the claims, as well as in the scope and gist of the invention.

<Additional Notes>
The abnormality determination method, the abnormality determination system, and the program described in each embodiment can be understood, for example, as follows.

(1) The anomaly judgment method according to the first aspect includes a step of comparing a judgment model created by learning normal data (data when mechanical equipment or work conditions are normal) including a plurality of variables (parameters) with data to be evaluated including the plurality of variables, calculating a distance between the data to be evaluated and the judgment model for each of the variables, and calculating the sum of the distances as the degree of abnormality of the data to be evaluated (S12), a step of judging whether the data to be evaluated is abnormal or not based on the degree of abnormality (S13), a step of calculating a direction of adjustment of a value for at least one of the plurality of variables to reduce the degree of abnormality (S14, S15), and a step of outputting the adjustment direction (S16).
This allows you to not only determine if there is an abnormality, but also how to return the system to a normal state.

(2) A second aspect of the abnormality determination method is the abnormality determination method of (1), in which, in the step of calculating the adjustment direction, a predetermined number of the variables are selected in order from the largest distance (S14), and a direction in which the distance is reduced for the selected variables is calculated as the adjustment direction (S15). In the step of outputting the adjustment direction, the selected variables and the adjustment direction for each variable are output.
This makes it possible not only to determine an abnormality, but also to understand which parameter is causing the abnormality to become greater and how to return that parameter to a normal state.

(3) A third aspect of the abnormality determination method is the abnormality determination method of (2), in which when selecting the predetermined number of variables, controllable variables are selected.
This makes it possible to control the parameters presented as guidance information and restore the state of the machinery and equipment or work.

(4) A fourth aspect of the anomaly determination method is an anomaly determination method according to any one of (1) to (3), wherein the adjustment direction is, for one of the variables, from the value of the variable contained in the evaluation target data to the value of the same variable contained in the determination model that is closest to the variable.
This makes it possible to calculate the direction in which a parameter identified as the cause of an abnormality should be brought closer to normal. By adjusting the parameter in the calculated direction, it may be possible to return the monitored machinery and equipment or work conditions to their normal state.

(5) An abnormality determination method according to a fifth aspect is an abnormality determination method according to any one of (1) to (4), further comprising the steps of: setting a normal interval for data of the plurality of variables; and learning the data included in the normal interval as the normal data, and creating the determination model that specifies a normal range for each of the variables.
Even if only normal data can be obtained from the monitored object and only a normal model learned from the normal data can be constructed, the normal model can be used not only to determine abnormalities but also to understand how to return the object to a normal state.

(6) A sixth aspect of the abnormality determination method is the abnormality determination method according to any one of (1) to (5), in which a weight is set for each of the variables to calculate the distance.
This makes it possible to set a desired weighting for the relationship between the parameters, thereby allowing guidance information regarding parameters that meet the user's intention (for example, parameters that are effective in terms of control) to be presented with priority.

(7) An abnormality determination method according to a seventh aspect is an abnormality determination method according to any one of (1) to (6), further comprising the steps of: inputting the value of the variable adjusted in a direction to reduce the degree of abnormality into a simulator that simulates the operation of a specified device or a specified task, and causing the simulator to simulate the operation or the task (S18); and evaluating, based on the results of the simulation, whether to adjust the variable for the purpose of reducing the degree of abnormality (S20).
This makes it possible to check in a simulation environment whether the parameters presented in the guidance information and the methods for improving them are appropriate.

(8) An anomaly determination method according to an eighth aspect is an anomaly determination method according to any one of (1) to (7), wherein the determination model is a probability density distribution of the variable contained in the normal data, and the calculating step of the degree of anomaly calculates a density ratio between the probability density distribution of the variable contained in the data to be evaluated and a probability density distribution of the variable based on the normal data, and the calculating step of the adjustment direction calculates the direction of adjustment of the value based on the positional relationship between the probability density distribution of the data to be evaluated and the probability density distribution of the normal data.
This makes it possible to calculate the degree of abnormality of the evaluation target data based on the likelihood of occurrence of the values of multiple variables that constitute normal data (probability density distribution).

(9) A ninth aspect of the anomaly determination method is an anomaly determination method according to any one of (1) to (8), wherein the determination model is a model that reproduces normal data constructed to output normal data when the normal data is input, and the step of calculating the degree of anomaly calculates a reproducibility indicating the degree to which data output by the determination model when the data to be evaluated is input into the determination model reproduces the input data to be evaluated, and the step of calculating the adjustment direction calculates a direction to adjust the value based on the reproducibility.
This makes it possible to calculate the degree of abnormality of the evaluation target data based on the degree of reproduction of the evaluation target data by a model that reproduces normal data.

(10) The anomaly determination system 1 according to the tenth aspect has a means (anomaly determination unit 132) for comparing a determination model created by learning normal data including a plurality of variables with data to be evaluated including the plurality of variables, calculating a distance between the data to be evaluated and the determination model for each of the variables, and calculating the sum of the distances as the degree of anomaly of the data to be evaluated, a means (anomaly determination unit 132) for determining whether the data to be evaluated is abnormal based on the degree of anomaly, a means (factor evaluation unit 133) for calculating a direction of adjustment of a value to reduce the degree of anomaly for at least one of the plurality of variables, and a means (output unit 14) for outputting the adjustment direction.

(11) The program according to the eleventh aspect causes a computer 900 to execute the steps of: comparing a judgment model created by learning normal data including a plurality of variables with data to be evaluated including the plurality of variables; calculating a distance between the data to be evaluated and the judgment model for each of the variables; and calculating the sum of the distances as the degree of abnormality of the data to be evaluated; determining whether the data to be evaluated is abnormal based on the degree of abnormality; calculating a direction of adjustment of a value for at least one of the plurality of variables to reduce the degree of abnormality; and outputting the direction of adjustment.

Reference Signs List 1: Abnormality determination system 10: Abnormality determination device 11: Data acquisition unit 12: Setting reception unit 13: Control unit 131: Model construction unit 132: Abnormality determination unit 133: Factor evaluation unit 14: Output unit 15: Memory unit 20: Simulator 900: Computer 901: CPU
902: Main memory device 903: Auxiliary memory device 904: Input/output interface 905: Communication interface

Claims (10)

a step of comparing a determination model created by learning normal data including a plurality of variables with evaluation target data including the plurality of variables, calculating a distance between the evaluation target data and the determination model for each of the variables, and calculating a sum of the distances as an abnormality degree of the evaluation target data;
determining whether the evaluation target data is abnormal or not based on the degree of abnormality;
calculating a direction of adjustment of a value of at least one of the plurality of variables to reduce the degree of abnormality;
outputting the adjustment direction;
having
In the step of calculating the adjustment direction,
Select a predetermined number of the variables in descending order of the distance;
Calculating a direction in which the distance is reduced for the selected variable as the adjustment direction;
In the step of outputting the adjustment direction,
outputting the selected variables and the adjustment direction for each variable;
Method for determining abnormality. When selecting the predetermined number of variables, the controllable variables are selected.
The method for determining an abnormality according to claim 1 . The adjustment direction is a direction from a value of the variable included in the evaluation target data to a value of the same variable included in the judgment model that is closest to the variable.
The method for determining an abnormality according to claim 1 or 2 . setting a normal interval for data of the plurality of variables;
A step of learning the data included in the normal interval as the normal data and creating the determination model that specifies a normal range for each of the variables;
The method for determining an abnormality according to any one of claims 1 to 3 , further comprising: A weight is set for each of the variables to calculate the distance.
The method for determining an abnormality according to any one of claims 1 to 3 . inputting the value of the variable adjusted in a direction to reduce the degree of abnormality into a simulator that simulates an operation of a predetermined device or a predetermined task, and causing the simulator to simulate the operation or the task;
evaluating whether to adjust the variables to reduce the degree of anomaly based on the results of the simulation;
The abnormality determination method according to any one of claims 1 to 5 , further comprising: the determination model is a probability density distribution of the variables contained in the normal data,
In the step of calculating the degree of anomaly, a density ratio between a probability density distribution of the variable included in the evaluation target data and a probability density distribution of the variable based on the normal data is calculated;
In the step of calculating the adjustment direction, a direction of adjustment of the value is calculated based on a positional relationship between a probability density distribution of the evaluation target data and a probability density distribution of the normal data.
The method for determining an abnormality according to any one of claims 1 to 6 . the judgment model is a model that reproduces the normal data and is constructed so as to output the normal data when the normal data is input,
In the step of calculating the degree of anomaly, a reproducibility indicating to what extent data output from the judgment model when the evaluation target data is input to the judgment model reproduces the input evaluation target data is calculated;
In the step of calculating the adjustment direction, an adjustment direction of the value is calculated based on the reproducibility.
The method for determining an abnormality according to any one of claims 1 to 7 . a means for comparing a determination model created by learning normal data including a plurality of variables with evaluation target data including the plurality of variables, calculating a distance between the evaluation target data and the determination model for each of the variables, and calculating a sum of the distances as an abnormality degree of the evaluation target data;
a means for determining whether the evaluation target data is abnormal or not based on the degree of abnormality;
means for calculating a direction of adjustment of a value for reducing the degree of abnormality for at least one of the plurality of variables;
A means for outputting the adjustment direction;
having
The means for calculating the adjustment direction includes:
Select a predetermined number of the variables in descending order of the distance;
Calculating a direction in which the distance is reduced for the selected variable as the adjustment direction;
The means for outputting the adjustment direction includes:
outputting the selected variables and the adjustment direction for each variable;
Anomaly detection system. On the computer,
a step of comparing a determination model created by learning normal data including a plurality of variables with evaluation target data including the plurality of variables, calculating a distance between the evaluation target data and the determination model for each of the variables, and calculating a sum of the distances as an abnormality degree of the evaluation target data;
determining whether the evaluation target data is abnormal or not based on the degree of abnormality;
calculating a direction of adjustment of a value of at least one of the plurality of variables to reduce the degree of abnormality;
outputting the adjustment direction;
having
In the step of calculating the adjustment direction,
Select a predetermined number of the variables in descending order of the distance;
Calculating a direction in which the distance is reduced for the selected variable as the adjustment direction;
In the step of outputting the adjustment direction,
outputting the selected variables and the adjustment direction for each variable;
A program that executes the following.

Images