JP6859381B2 - State change detection device and state change detection program - Google Patents

Description

The present invention relates to a state change detection device and a state change detection program.

For example, when a state change such as an abnormality is detected by measuring the current, voltage, and vibration of a machine, the abnormality is generally determined based on the past experience of a person or the like. That is, the threshold value of the state change is set by a person, and it is almost impossible for a machine having no abnormal data to perform the state change.

Patent Document 1 discloses a highly versatile abnormality detection device that does not depend on the type and number of objects and events that detect abnormal situations, and the space (place, time zone, etc.) that monitors those objects and events. ing.

Specifically, it is a device that detects that an abnormality has occurred in a target event. In this device, the time series of the input pattern, which is the data that changes depending on the event, is acquired, the characteristics of the transition of the acquired input pattern in the time series are analyzed, and the analyzed characteristics of the transition and the characteristics of the transition in advance are analyzed. It is compared with the determined reference values, and if they are not close to each other within a certain range, it is determined that an abnormality has occurred in the event. In this way, when it is determined that an abnormality has occurred as a result of the above comparison, an output means for performing an output operation to that effect is provided.

In other words, a time series is acquired with data that changes according to changes in events such as the movement of a person or object to be monitored as an input pattern, and the transition of the pattern in that time series is focused on, and the feature amount in the transition is extracted. However, the occurrence of an abnormal situation is detected by comparing it with that in a normal case. Here, the "event" is a phenomenon that can be expressed as computer-processable data by a signal from a device sensor or the like, a report (data input) from a person, or the like, and "human behavior" in the house. It is also applicable to.

Patent Document 2 discloses a data prediction method or data prediction system that uses a prediction model that realizes highly accurate prediction and that allows the prediction process to be understood.

The invention of Patent Document 2 is a data prediction method for predicting the future based on past achievements, wherein the prediction model includes a first prediction model that outputs a prediction value for at least one feature amount on the prediction target date. It is composed of a second prediction model that includes the prediction value output from the first prediction model as an input factor and outputs the prediction value for each predetermined time on the prediction target day.

Then, a prediction model construction means for constructing a prediction model using the collected nearest actual data and past actual data, and prediction execution for obtaining a predicted value by inputting input data for prediction into the constructed prediction model. Means, predictive error calculation means for calculating prediction error or model error from collected closest actual data and predicted value, calculation of correction coefficient or correction amount to correct the predicted value based on prediction error or model error, and correction It is realized by the predicted value correction means for obtaining the predicted value.

References 3 include rainwater inflow forecasts, inflow water quality forecasts, river water level forecasts, dam storage volume forecasts, demand forecasts for water, gas, electricity, etc., natural phenomena and artificial phenomena / events. A prediction model system used for making online predictions is disclosed. In Patent Document 3, the method using a prediction model can be roughly classified into a method called a white box approach using a prediction model based on a mathematical formula based on physical and chemical laws, and actually available data. , There is a description that it can be divided into a method called a black box approach that constructs a prediction model from actual data by performing statistical processing, learning (for example, neural network, etc.), system identification, and so on.

The invention of Patent Document 3 considers the physical law (conservation law) by the white box approach in the prediction by the system identification method using the model of the discrete time system, which is a typical method of the black box approach. Provide a predictive model system that can build a model. The prediction model system of Patent Document 3 includes a prediction model including a plurality of parameters and predicts a prediction target, a parameter value storage unit that stores parameter values, and input / output data that stores input and output time series data. Provide storage means. Constraints on parameters are determined in the static parameter identification means based on stationary input data and output data. A parameter that satisfies this constraint condition is determined by the dynamic parameter identification means based on the non-stationary input data and the output data stored in the input / output data storage means, and is sent to the parameter value storage unit.

Patent Document 4 forecasts electricity demand, water demand, heat demand, gas demand, steam demand and other demands, various loads, various sales volumes, various economic indicators, etc., and past results or related information. A data prediction method and a data prediction system for predicting the future are disclosed using.

Patent Document 4 solves the following problems in the prior art. In the first method using the similar pattern in the past, it is necessary to make a correction according to the situation of the prediction target day, and it is difficult to make an accurate correction. If the second method of creating a forecast model for each forecast target time is adopted, it is necessary to create a model for each forecast target time, which requires a great deal of modeling work. The prediction accuracy is poor because it cannot be modeled. In the third method of collectively predicting the prediction target by the neural network prediction model, the neural network is usually used as the prediction model, but it is difficult to analyze the inside of the neural network because the inside is a black box. , There is a problem that it is difficult to explain the reason for prediction about the prediction result.

In view of the above, the invention of Patent Document 4 aims to reduce a large amount of work and work time for constructing a prediction model, and is highly accurate in consideration of the time-series transition of daily data. The purpose is to provide a prediction method.

The invention of Patent Document 4 is a data prediction method for predicting the future based on past achievements.
The prediction model to be used includes a first prediction model that outputs a prediction value for at least one feature quantity on the prediction target date and a prediction value output from this first prediction model as input factors, and a predetermined prediction target date is used. It is composed of a second prediction model that outputs a prediction value for each time. The invention of Patent Document 4 is a prediction model construction means for constructing a prediction model using the collected closest actual data and past actual data, inputting input data for prediction into the constructed prediction model, executing prediction, and making a prediction. Prediction execution means for obtaining a value, prediction error calculation means for calculating a prediction error or model error from the collected nearest actual data and the predicted value, a correction coefficient or a correction amount for correcting the predicted value based on the prediction error or the model error. Is provided, and the predicted value correction means for obtaining the corrected predicted value is provided.

Patent Document 5 discloses a technique for predicting a synthesis target in speech synthesis, and in particular, a method for creating a prediction model for an acoustic feature amount of a synthesis target during speech synthesis. An object of the present invention of Patent Document 5 is to provide a method for creating a prediction model capable of constructing a prediction model of acoustic features of a synthesis target of speech synthesis more efficiently.

According to the invention of Patent Document 5, the method of creating a prediction model is a method of creating a prediction model for predicting a synthesis target value of an acoustic feature amount for a predetermined speech unit in speech synthesis, and is a predetermined prosody. -A step of preparing computer-readable training data in which prosodic information and linguistic information are attached for each predetermined voice unit and a label for a predetermined acoustic feature amount is attached, and a gradient boosting algorithm using the training data. This includes a step of training a predetermined prediction model.

Further, the invention described in Patent Document 6 relates to a prediction error evaluation device, a prediction error evaluation method, and a prediction error evaluation program. The invention described in Patent Document 6 describes, for example, how the error between the predicted value and the measured value changes (for example, in time series) in relation to the series of attribute values of the phenomenon to be observed. Take this into account so that the error can be evaluated.

In the invention described in Patent Document 6, for example, a predicted value predicted to be obtained by observing the phenomenon for each attribute value of a predetermined phenomenon and the phenomenon observed each time the attribute value of the phenomenon changes. The present invention includes a storage device that stores in advance error data indicating an error from the actually measured value obtained in association with the attribute value of the phenomenon. The invention described in Patent Document 6 has a statistical processing unit, and in a series of attribute values stored in the storage device, one sub-series and at least one of the beginning and the end of the sub-series are used. A plurality of sub-series obtained by shifting by a predetermined number is set as a processing target series, and for each processing target series, error data corresponding to the attribute value included in the processing target series is read from the storage device, and the read error data is obtained. The predetermined statistical processing is executed by the processing device. Further, the output unit includes a result output unit that outputs the result of the statistical processing executed by the statistical processing unit by the output device.

Japanese Unexamined Patent Publication No. 2003-256957 Japanese Unexamined Patent Publication No. 2015-127914 Japanese Unexamined Patent Publication No. 2004-62440 Japanese Unexamined Patent Publication No. 2004-94437 Japanese Unexamined Patent Publication No. 2006-89467 Japanese Unexamined Patent Publication No. 2011-95946

However, as described above, the conventional prediction model aims at improving the accuracy and reliability of prediction and further improving efficiency, and only from the data when there is no state change or it is unknown whether the state change occurs. A state change detection device capable of accurately capturing state change has not been known.

The present invention has been made in view of the current state of the state change detection device as described above, and an object of the present invention is to obtain state change from only data when there is no state change, or only data in which state change occurs. It is an object of the present invention to provide a state change detection device and a state change detection program capable of detecting no state change.

The state change detection device according to the present embodiment uses data collected from a plurality of collection sources for a target for which state change should be detected as test data, and data collected from one first collection source among the plurality of collection sources. By machine learning using the teacher data for the first prediction model, in which the data collected from the second collection source other than the one first collection source among the plurality of collection sources is used as the explanatory variable. A first prediction model that predicts and outputs the data collected by the first collection source as an objective variable from the created explanatory variables that are the data collected from the second collection source.
The objective variable, which is the data of the prediction output result obtained by giving the data collected from the second collection source to the first prediction model as explanatory variables, and the data of the explanatory variables given to the first prediction model. The explanatory variables, the explanatory variables that are the data of the explanatory variables given to the first prediction model, and the label data indicating the presence or absence of the state change corresponding to the objective variable that is the prediction output result of the first prediction model. A second prediction model that predicts and outputs the ratio of state fluctuations from the explanatory variables of the data collected from the second collection source as the objective variable, which was created by machine learning using the teacher data for the second prediction model. Is a state fluctuation detection device that detects state fluctuations by inputting data collected from the second collection source into the first prediction model and the second prediction model as explanatory variables using the above-mentioned first prediction. Based on the comparison main value, which is a value based on the error between the first prediction information which is the prediction output result obtained by the model and the actual measurement data obtained from the first collection source by the actual measurement, and the second prediction model. It is characterized in that state fluctuation detection is performed based on both the obtained predicted output result and the second predicted value.

The block diagram which shows the structure of the computer system which realizes the state change detection apparatus which concerns on this invention. The functional block diagram of the embodiment of the state change detection apparatus which concerns on this invention. The figure which shows the means realized by the state change detection program provided in the computer system realized by using the 1st prediction model of the state change detection apparatus which concerns on this invention. The flowchart for demonstrating the operation of the structure realized by using the 1st prediction model of the state change detection apparatus which concerns on this invention. The figure which shows the teacher data used for the composition using the 1st prediction model of the state change detection apparatus which concerns on this invention, the prediction data obtained by applying it to the 1st prediction model, and the value of a standard error. The figure which shows the procedure until the first prediction model is made using the teacher data in the state change detection apparatus which concerns on this invention. The figure which shows the procedure from inputting the teacher data to the 1st prediction model in the state change detection apparatus which concerns on this invention, and obtaining the standard error. Measurement data of three times used in the configuration realized by using the first prediction model of the state fluctuation detection device according to the present invention, and prediction data and errors obtained by applying them to the first prediction model. The figure which shows the value. The figure which shows the procedure until the error is obtained by inputting the measurement data of three times used for the configuration realized by using the 1st prediction model of the state change detection apparatus which concerns on this invention into the 1st prediction model. .. The figure which conceptually showed the magnification of the error with respect to the standard error obtained from the measurement data of three time in the configuration realized by using the 1st prediction model of the state change detection apparatus which concerns on this invention. In the configuration realized by using the first prediction model of the state change detection device according to the present invention, it is shown by a graph that the difference between the peaks, which is the degree of dissociation, is obtained by using the standard error distribution and the error distribution. Figure. In the configuration realized by using the first prediction model of the state change detection device according to the present invention, the state change is detected based on the dissociation degree and the deviation value which are the values obtained by standardizing the error for each measurement. The figure explaining. The flowchart which shows the presence / absence determination operation of the state change using both the predicted value (ratio) and the error amount from the creation of the teacher data for the 2nd prediction model in the state change detection apparatus which concerns on this invention. The figure which shows the procedure of making the 2nd prediction model in the state change detection apparatus which concerns on this invention. The figure which shows the procedure which obtains the predicted value from the normal data by the 2nd prediction model in the state change detection apparatus which concerns on this invention. The figure which shows the procedure of obtaining the predicted value from the abnormality data by the 2nd prediction model in the state change detection apparatus which concerns on this invention. The figure which shows the determination result separated by the threshold value used in the state change presence / absence determination process using both the predicted value (ratio) and the error amount in the state change detection apparatus which concerns on this invention.

Hereinafter, embodiments of the state change detection device and the state change detection program according to the present invention will be described with reference to the accompanying drawings. In each figure, the same components are designated by the same reference numerals, and duplicate description will be omitted. The state change detection device according to the embodiment of the present invention can be configured by, for example, a personal computer, a workstation, or other computer system as shown in FIG. This computer system operates as a state change detection device by controlling each part based on a program or data stored in the main memory 11 or read into the main memory 11 by the CPU 10 and executing necessary processing. is there.

An external storage interface 13, an input interface 14, a display interface 15, and a data input interface 16 are connected to the CPU 10 via a bus 12. A program such as a state change detection program and an external storage device 23 in which necessary data and the like are stored are connected to the external storage interface 13. An input device 24 such as a keyboard as an input device for inputting commands and data and a mouse 22 as a pointing device are connected to the input interface 14.

A display device 25 having a display screen such as an LED or an LCD is connected to the display interface 15. Sensors 26-1, 26-2, ..., 26-m for obtaining measurement data are connected to the data input interface 16. This computer system may be provided with other configurations, and the configuration of FIG. 1 is only an example.

An embodiment of the state change detection device 50 includes, for example, as shown in FIG. 2, a first prediction model 30, a second prediction model 40, and a state change determination unit 53. For the first prediction model 30 and the second prediction model 40, the data collected from the test data collection unit 54 for the target for which the state change should be detected is given as test data. The test data collecting unit 54 corresponds to the sensors 26-1, 26-2, ..., 26-m in the configuration example of FIG. The sensors 26-1, 26-2, ..., 26-m can be collected from different objects such as temperature, humidity, vibration, etc., and may be a part or a plurality of parts in the machine to be installed. It may be provided in one or more of the machines. State detection is performed using a device to be collected and a group of test data having a plurality of locations / positions.

The first prediction model 30 generates the first prediction information based on the input test data, and the second prediction model 40 generates the second prediction value based on the input test data. It is a thing. The state change determination unit 53 includes a comparison subject value created from the first prediction information obtained by the first prediction model 30 and a second prediction value obtained by the second prediction model 40. State change detection is performed using both.

The determination result information obtained by the state change determination unit 53 is output to the output device 55. The output device 55 can be the display device 25 in FIG. 1, and a printer or the like may be provided in addition to the display device 25 to function as the output device 55 as well.

The first prediction model 30 is created by the first prediction model creation means 31, and the second prediction model 40 is created by the second prediction model creation means 41. The second prediction model 40 is created by machine learning using teacher data based on the first prediction information obtained by the first prediction model 30.

Next, teacher data for creating the second prediction model 40 performed by the first prediction model 30, the first prediction model creating means 31, and the state change determination unit 53 (teacher data for the second prediction model). Will be described.

The CPU 10 shown in FIG. 1 realizes each means and the like shown in FIG. 2 by a state change detection program in the external storage device 23. That is, the prediction model creating means 31, the prediction model 30, the teacher prediction data acquisition means 32, the standard error acquisition means 33, the measurement prediction data acquisition means 34, the error acquisition means 35, and the state change detection means 36 are realized. Further, teacher data is stored in the external storage device 23. The prediction model creating means 31 creates the prediction model 30 using the teacher data. The prediction model 30 predicts the objective variable from the explanatory variables by machine learning. Here, as a machine learning algorithm, a random forest of pattern mining can be mentioned, but in addition to this, an algorithm such as regression analysis or regression tree may be adopted.

Regarding the explanatory variable and objective variable prediction model, when y is the objective variable, xi is the explanatory variable, and f is the prediction model, y = f (x1, x2, ..., Xi, ..., Xn). Can be represented by. The teacher prediction data acquisition means 32 predicts the teacher explanatory variable by using the teacher data composed of the teacher objective variable which is the objective variable of the fixed value and the teacher explanatory variable which is the explanatory variable of the fixed value. It is input to the model 30 to obtain teacher prediction data. The standard error acquisition means 33 acquires a standard error which is a difference between the teacher prediction data and the teacher objective variable.

The measurement prediction data acquisition means 34 inputs the measurement explanatory variable into the prediction model 30 by using the measurement data composed of the measurement objective variable which is the measured objective variable and the measurement explanatory variable which is the measured explanatory variable. To obtain measurement prediction data. The measurement data can be the data obtained by the sensors 26-1, 26-2, ..., 26-m. The error acquisition means 35 acquires an error which is a difference between the measurement prediction data and the measurement objective variable.

The state change detecting means 36 obtains the degree of dissociation between the standard error and the error, and detects the state change based on the degree of dissociation.

Since the state change detection device configured by the above means and the like executes the processing operation according to the flowchart shown in FIG. 4, the operation will be described with reference to this flowchart. First, the prediction model 30 is created using the teacher data (STEP 1). For example, as shown in FIG. 5, the teacher data includes the objective variable data to be obtained by the sensor A, the first explanatory variable data obtained by the sensor B, and the second explanatory variable obtained by the sensor C. It is composed of variable data. The number of times data is acquired (the number of data in the vertical direction in FIG. 5) is arbitrary. As shown in FIG. 6, the prediction model creating means 31 creates the prediction model 30 using the teacher data T composed of the numerical values as illustrated in FIG.

Next, as shown in the flowchart of FIG. 4 and the operation sequence diagram of FIG. 7, explanatory variables (data of sensor B and data of sensor C of FIG. 5) are input to the prediction model 30 using the teacher data T, and prediction is performed. The data (teacher prediction data TF (teacher data prediction result in FIG. 7)) is obtained, and the error (standard error) which is the difference between the prediction data (teacher prediction data TF) and the objective variable of the teacher data T is obtained (standard error). STEP2). Here, the explanatory variables are the data of the sensor B and the sensor C in the teacher data of FIG. As shown in FIG. 7, if the objective variable of the teacher data T is TT and the objective variable TT is represented by, for example, a perfect circle, the teacher prediction data TF is distorted with respect to this objective variable TT. Therefore, the part that protrudes or retracts from the perfect circle becomes an error. Actually, as shown in the column of "predicted value / standard error of teacher data" in FIG. 5, the predicted data (teacher prediction data) and the error (standard error STER) are obtained as numerical values. Since these are obtained by calculating the same number as the number of times the data is acquired, the same quantity as the number of times the data is acquired can be obtained. Therefore, the average of these is calculated and obtained, and this is held as the average standard error (AVSTER).

Next, as shown in the flowchart of FIG. 4, the explanatory variables M1 and M2 are input to the prediction model 30 using the measurement data M by the sensors B and C, and the prediction data (measurement prediction data) is obtained, and the prediction data is obtained. The error, which is the difference between (measurement prediction data) and the measured objective variable of the measurement data M, is obtained (STEP 3).

Here, the measurements t1, t2, and t3 are performed at different times to obtain the explanatory variables t1M1 and t1M2 of the measurement t1, the explanatory variables t2M1 and t2M2 of the measurement t2 are obtained, and the explanatory variables t3M1 and t3M1 of the measurement t3 are obtained. Obtain t3M2 (Fig. 8). As shown in FIG. 9, these are input to the same prediction model 30 to obtain prediction data (measurement prediction data) t1F, t2F, and t3F (in the right column in FIG. 8). In FIG. 9, it is drawn as if three prediction models 30 are provided, but one prediction model 30 may be used for sequential prediction. Of course, three prediction models 30 may be provided. As shown in FIGS. 8 and 9, the difference between the predicted data (measurement predicted data) t1F, t2F, t3F and the measured objective variables t1MT, t2MT, t3MT of the measured data M is an error (individual error) t1ER, t2ER, t3ER. Ask as. Since the errors t1ER, t2ER, and t3ER are obtained by calculating the same number as the number of times the data is measured, the same quantity as the number of times of acquisition can be obtained. It is held as an average error (t1AVER, t2AVER, t3AVER).

Next, as shown in STEP 4 of FIG. 4, it is calculated how many times the amount of error obtained from each measurement data is compared with the standard error obtained from the teacher data. We conclude that the larger this magnification is, the more the state changes. In FIG. 10, as the comparison main body, “measurement t1 error (average) (t1AVER)”, “measurement t2 error (average) (t2AVER)”, “measurement t3 error (average) (t3AVER)” are shown. It is assumed that there is a value corresponding to the area of the circle, and there is a value corresponding to the area of the circle as shown in "Standard error (mean) (AVSTER)" as a comparison target. Assuming that the magnification is 1.2, 2.3, and 3.8, (measurement t1 error) <(measurement t2 error) <(measurement t3 error) holds, so measurement 3 has the most state fluctuation. It can be concluded that the state change is occurring, then the state change is occurring in the measurement 2, and the measurement 1 is the least probable that the state change is occurring.

Alternatively, instead of calculating the average of the standard error and the error, a distribution graph of the standard error value is created, a distribution graph of the error value is created, and the state variation is detected using these distribution graphs. The distribution graph of the standard error value is as shown in FIG. 11 (a), the distribution graph of the error value of the measurement t1 is as shown in FIG. 11 (b), and the distribution graph of the error value of the measurement t2 is as shown in FIG. It is assumed that the distribution graph of the error value of the measurement t3 is as shown in FIG. 11 (d).

In the above case, the difference between the peaks of the distribution graph is adopted as the degree of dissociation. The difference between the peak of the distribution graph of the standard error value and the peak of the distribution graph of the error value of the measurement t1 is as shown in d1 shown in FIG. 11 (b). The difference between the peak of the distribution graph of the standard error value and the peak of the distribution graph of the error value of the measurement t2 is as shown in d2 shown in FIG. 11 (c). The difference between the peak of the distribution graph of the standard error value and the peak of the distribution graph of the error value of the measurement t3 is as shown in d3 of FIG. 11 (d). Since the difference between the peaks is d ₁ <d ₂ <d ₃ , the determination of the state change is as described with reference to FIG.

Further, as shown in STEP 4 of FIG. 4, a threshold value for abnormality determination is set for the magnification obtained above or for the value σ obtained by standardizing the error for each measurement number, so that the abnormality is detected. Is also good. For example, in the example of FIG. 10, if the threshold value is "3.5", it can be determined that the measurement t3 is abnormal.

An example of using the deviation value σ is shown in FIG. It is assumed that the above-mentioned error distribution is as shown in the graph of FIG. The average value of this error is obtained (S11). Using this average value, the deviation value σ is obtained (S12). Next, the state change is detected using the deviation value σ obtained above (S13). For example, since the standard error is also obtained for each measurement, the deviation value σ _{ST of the} standard error can be obtained from these distributions. On the other hand, the deviation values σ _t1 , σ _t2 , and σ _t3 are also obtained for the measurements t1, t2, and t3. Of the deviation values σ _t1 , σ _t2 , and σ _t3 , the larger the difference or ratio or magnification of the standard error deviation value σ _ST , the higher the probability of a state change. _{The threshold values σ SH} for the deviation values σ _t1 , σ _t2 , and σ _t3 can be set, and if they are exceeded, the state change (abnormality) can be considered.

In the above, one objective variable is obtained from two explanatory variables, but it can be applied to a prediction model in which one objective variable is obtained from one or more explanatory variables. Further, the state change is not limited to the change to an abnormality, and can be applied to various state change detections such as an excess state, a shortage state, a low state, and a high state.

In the above embodiment, when the data is measured by the sensors A, B, and C, the sensor A prediction model in which the objective variable is the data obtained by the sensor A and the explanatory variable is the data obtained by the sensors B and C. Was made to configure only one prediction model. However, in such a system that measures data of three or more systems, a plurality of prediction systems can be constructed.

For example, when data is measured by sensors A, B, and C, sensor A predicts that the objective variable is the data measured by sensor A and the explanatory variable is the data measured by sensors B and C. The case where the model is called a model, the objective variable is the data measured by the sensor B, and the explanatory variables are the data measured by the sensors A and C is called the sensor B prediction model, and the objective variable is the data measured by the sensor C. When the explanatory variables are the data measured by the sensors A and B and are referred to as the sensor C prediction model, three prediction models such as the sensor A prediction model, the sensor B prediction model, and the sensor C prediction model are used. Can be created.

It is assumed that the time-series measured values obtained by measuring the data of the sensors A, B, and C are t1, t2, and t3. Each of t1, t2, and t3 is composed of three types of data, sensors A, B, and C. In the sensor A prediction model, the measured values t1, t2, and t3 of the sensors B and C can be input to obtain three types of average errors of the sensor A prediction model, A_t1AVER, A_t2AVER, and A_t3AVER. In the sensor B prediction model, the measured values t1, t2, and t3 of the sensors A and C can be input to obtain three types of average errors of the sensor B prediction model, B_t1AVER, B_t2AVER, and B_t3AVER. In the sensor C prediction model, the measured values t1, t2, and t3 of the sensors A and B can be input to obtain three types of average errors of the sensor C prediction model, C_t1AVER, C_t2AVER, and C_t3AVER.

When the standard error of each model obtained from the teacher data is A_AVSTER, B_AVSTER, C_AVSTER, the magnification and error distribution (deviation) of A_t1AVER, A_t2AVER, and A_t3AVER based on the standard error A_AVSTER can be obtained. can, standard error versus relative to the B_AVSTER B_t1AVER, B_t2AVER, can be obtained each magnification and error distribution of B_t3AVER the (deviation) vs. relative to the standard error C_AVSTER C_t1AVER, C_t2AVER, each magnification and error of C_t3AVER The distribution (deviation) can also be obtained.

When the deviation is obtained in the above, for example, the deviation value σA_t1 of the measured value t1 in the sensor A prediction model, the deviation value σB_t1 of the measured value t1 in the sensor B prediction model, and the deviation value t1 of the measured value t1 in the sensor C prediction model. By obtaining the average of σC_t1 and the like, a new feature amount of the measured value t1 can be generated, and by setting a threshold value for this, it is possible to detect an abnormality. The above is an example of the measured value t1, but the measured values t2 and t3 can be handled in the same manner.

As described above, the first prediction information is obtained by using the first prediction model 30 based on the measurement data (test data), and the error amount and the deviation value are obtained from the first prediction information and compared with the threshold value. Then, it is determined whether or not a state change (abnormality in this case) has occurred. After that, the process described in the flowchart shown in FIG. 13 is performed. That is, the measurement data (test data) is labeled from the judgment result based on the first prediction information, and the teacher data (the teacher data for the second prediction model is created (S21). As a result of the above judgment, the judgment is made. The measurement data (test data) should be labeled as "state fluctuation (abnormal here) has occurred" and "state fluctuation (abnormal here) has not occurred (normal)". The measurement data (test data) labeled in this way is used as the teacher data (teacher data for the second prediction model) for creating the second prediction model.

As described above, the teacher data based on the first prediction information includes the test data used in the past and the label data indicating the presence or absence of the state change which is the state change detection result for the test data. be able to. The teacher data for the second prediction model may include a magnification, an error amount, and a deviation value obtained from the first prediction information in addition to the measurement data (test data). This second prediction model teacher data targets only the measurement data (test data) selected by the operator, etc. in the operation of the system, assuming that there is a clear state change or no state change. Can be. In this way, the teacher data for the second prediction model created in step S21 is given to the second prediction model creation means 41, and the second prediction model 40 is created by machine learning such as deep learning or random forest (S22). ).

That is, as shown in FIG. 14, from the teacher data T-2 for the second prediction model including the data labeled as normal and abnormal, the second prediction model creation means 41 is such as deep learning or random forest. A second prediction model 40 is created by performing machine learning.

Next, as shown in step S23 of FIG. 13, the test data is input to the second prediction model 40 to obtain the second prediction value (S23). As shown in FIG. 15A, for test data test1, which is composed of data in which normal data is generally indicated by a large circle and data excluding anomalous data indicated by a small black circle, the predicted anomaly value is less than 0.5. On the other hand, as shown in FIG. 15B, for the test data test2 composed of only the abnormal data shown by the small black circles, the predicted value of the abnormality is obtained with a value of 0.5 or more. In this embodiment, the predicted value is from 0 to 1.

Before the processing of the flowchart of FIG. 13 is performed, it is assumed that the predicted value of the error amount is obtained by the processing shown in FIG. 9 for the same test data. Therefore, it is compared with the threshold value TH-A for the predicted value (ratio) obtained in step S23, compared with the threshold value TH-B for the amount of error (S24), and there is a state change when both exceed the threshold value (S24). It is determined (S25).

Regarding the determination result, as shown in FIG. 16, the case where only the predicted value (ratio) exceeds the threshold value TH-A and the case where only the error amount exceeds the threshold value TH-B are regarded as "quasi-abnormal", and the predicted value ( When the ratio) exceeds the threshold TH-A and the error amount also exceeds the threshold TH-B, it is regarded as "abnormal", and the predicted value (ratio) is lower than the threshold TH-A and the error amount also exceeds the threshold TH-. If it is less than B, it can be regarded as "normal".

Here, the comparison main value is used as the error amount, but the error average value or the deviation value σ described above may be used. According to the state change detection device according to this embodiment, it is possible to accurately capture the state change only from the data when there is no state change or it is unknown whether the state change has occurred.

10 CPU
11 Main memory 12 Bus 13 External storage interface 14 Input interface 15 Display interface 16 Data input interface 22 Mouse 23 External storage device 24 Input device 25 Display device 26 Sensor 30 Prediction model 31 Prediction model creation means 32 Teacher Prediction data acquisition means 33 Standard error Acquisition means 34 Measurement prediction data acquisition means 35 Error acquisition means 36 State change detection means 40 Prediction model 41 Prediction model creation means 50 State change detection device 53 State change determination unit 54 Test data collection unit 55 Output device

Claims (12)

Data collected from a plurality of collection sources for a target for which a state change should be detected is used as test data, and data collected from one first collection source among the plurality of collection sources is used as an objective variable. the created by machine learning using one first first teacher data predicted model data and the explanatory variables collected from the collection source other than the second collection source, collected from the second collection source A first prediction model that predicts and outputs the data collected by the first collection source from the explanatory variables that are the obtained data as the objective variable, and
The objective variable, which is the data of the prediction output result obtained by giving the data collected from the second collection source to the first prediction model as an explanatory variable, and the data of the explanatory variable given to the first prediction model. The explanatory variable, which is the data of the explanatory variable given to the first prediction model, and the label data indicating the presence or absence of the state change corresponding to the objective variable, which is the prediction output result of the first prediction model. A second prediction model that predicts and outputs the ratio of state fluctuations from the explanatory variables of the data collected from the second collection source as the objective variable, which was created by machine learning using the teacher data for the second prediction model.
Is a state change detection device that detects state changes by inputting the data collected from the second collection source into the first prediction model and the second prediction model as explanatory variables.
The comparison main value, which is a value based on the error between the first prediction information which is the prediction output result obtained by the first prediction model and the actual measurement data obtained from the first collection source by the actual measurement, and the first. A state change detection device characterized in that state change detection is performed based on both the second predicted value, which is the predicted output result obtained by the second prediction model. The state change detection device according to claim 1, wherein the state change detection is performed using the device to be collected and the group of the test data having a plurality of places / positions. The state change detection device according to claim 1 or 2, wherein the state change is whether or not it is an abnormal state. Computer,
Data collected from a plurality of collection sources for a target for which a state change should be detected is used as test data, and data collected from one first collection source among the plurality of collection sources is used as an objective variable. Collected from the second collection source created by machine learning using the teacher data for the first prediction model using the data collected from the second collection source other than the one first collection source of the above as an explanatory variable. A first prediction model that predicts and outputs the data collected by the first collection source from the explanatory variables that are the obtained data as the objective variable.
The objective variable, which is the data of the prediction output result obtained by giving the data collected from the second collection source to the first prediction model as an explanatory variable, and the data of the explanatory variable given to the first prediction model. The explanatory variable, which is the data of the explanatory variable given to the first prediction model, and the label data indicating the presence or absence of the state change corresponding to the objective variable, which is the prediction output result of the first prediction model. The second prediction model, which is created by machine learning using the teacher data for the second prediction model, predicts and outputs the ratio of the state change as the objective variable from the explanatory variables of the data collected from the second collection source.
The state change detection that detects the state change by inputting the data collected from the collection sources other than the one collection source into the computer that functions as the first prediction model and the second prediction model as an explanatory variable. Program for
The computer
A means for creating the first prediction model,
A means for creating the second prediction model,
The comparison main value, which is a value based on the error between the first prediction information which is the prediction output result obtained by the first prediction model and the actual measurement data obtained from the first collection source by the actual measurement, and the first. A state change detection program characterized in that it functions as a means for performing state change detection based on both the second predicted value, which is the predicted output result obtained by the second prediction model. The program for detecting state fluctuations according to claim 4 , wherein the state change detection is performed using the device to be collected and the group of the test data having a plurality of places / positions. The program for detecting a state change according to claim 4 or 5 , wherein the state change is whether or not it is an abnormal state. The error between the prediction output result obtained by giving the explanatory variable of the teacher data for the first prediction model to the first prediction model and the objective variable corresponding to the explanatory variable of the teacher data for the first prediction model. The standard error, and the predicted output result obtained by giving the data collected from the second collection source as explanatory variables to the first prediction model, and the measured data obtained from the first collection source by actual measurement. A means for obtaining an error amount which is an error between the above and the above error amount and a comparison value between the error amount and the standard error is provided.
Any of claims 1 to 3, wherein the state change detection is performed by combining the determination based on the comparison value and the first threshold value, and the predicted value of the state change rate and the determination based on the second threshold value. The state change detection device according to item 1. The prediction output result obtained by giving a plurality of explanatory variables of the teacher data for the first prediction model to the first prediction model in chronological order and the plurality of explanations of the teacher data for the first prediction model. Obtained by giving a plurality of standard errors, which are errors from the objective variables corresponding to the variables in time series, and the data collected in time series from the second collection source to the first prediction model as explanatory variables. A means for obtaining a plurality of error amounts, which is an error between the predicted output result and the actually measured data obtained from the first collection source by actual measurement in a time series synchronized with the time series, is provided.
A claim characterized in that state fluctuation detection is performed using the difference between the peak of the standard error distribution obtained in time series and the peak of the distribution of the error amount obtained in time series as a comparison subject value. The state change detection device according to any one of 1 to 3. The prediction output result obtained by giving a plurality of explanatory variables of the teacher data for the first prediction model to the first prediction model in chronological order and the plurality of explanations of the teacher data for the first prediction model. Obtained by giving a plurality of standard errors, which are errors from the objective variables corresponding to the variables in time series, and the data collected in time series from the second collection source to the first prediction model as explanatory variables. A means for obtaining a plurality of error amounts, which is an error between the predicted output result and the actually measured data obtained from the first collection source by actual measurement in a time series synchronized with the time series, is provided.
It is characterized in that state fluctuation detection is performed using the difference between the standard deviation value of the standard error distribution obtained in time series and the standard deviation value of the error amount distribution obtained in time series as a comparison main value. The state change detection device according to any one of claims 1 to 3. The computer
The error between the prediction output result obtained by giving the explanatory variable of the teacher data for the first prediction model to the first prediction model and the objective variable corresponding to the explanatory variable of the teacher data for the first prediction model. The standard error, and the predicted output result obtained by giving the data collected from the second collection source as explanatory variables to the first prediction model, and the measured data obtained from the first collection source by actual measurement. A means for obtaining an error amount, which is an error between the two, and a comparison value between this error amount and the standard error.
Means for detecting state fluctuations by combining the comparison value and the determination based on the first threshold value, and the predicted value of the rate of state change and the determination based on the second threshold value.
The program for detecting state fluctuations according to any one of claims 4 to 6, wherein the program functions as. The computer
The prediction output result obtained by giving a plurality of explanatory variables of the teacher data for the first prediction model to the first prediction model in chronological order and the plurality of explanations of the teacher data for the first prediction model. Obtained by giving a plurality of standard errors, which are errors from the objective variables corresponding to the variables in time series, and the data collected in time series from the second collection source to the first prediction model as explanatory variables. A means for obtaining a plurality of error amounts, which are errors between the predicted output result and the measured data obtained from the first collection source by actual measurement in a time series synchronized with the time series.
It is characterized in that it functions as a means for detecting state fluctuations by using the difference between the peak of the distribution of the standard error obtained in the time series and the peak of the distribution of the amount of error obtained in the time series as the comparison main value. The program for detecting state fluctuations according to any one of claims 4 to 6. The computer
The prediction output result obtained by giving a plurality of explanatory variables of the teacher data for the first prediction model to the first prediction model in chronological order and the plurality of explanations of the teacher data for the first prediction model. Obtained by giving a plurality of standard errors, which are errors from the objective variables corresponding to the variables in time series, and the data collected in time series from the second collection source to the first prediction model as explanatory variables. A means for obtaining a plurality of error amounts, which are errors between the predicted output result and the measured data obtained from the first collection source by actual measurement in a time series synchronized with the time series.
It functions as a means for detecting state fluctuations using the difference between the standard deviation value of the standard error distribution obtained in time series and the standard deviation value of the error amount distribution obtained in time series as a comparison main value. The state change detection program according to any one of claims 4 to 6, wherein the program is used.

Images