Machine learning device and machine learning method

Abstract

A machine learning device (1) learns the cause of an alarm for a motor drive device (2). The machine learning device (1) includes: a state observation unit (10) that acquires, as state variables, feature quantities including at least one of a current detection value detected from a motor obtained from the motor drive device (2), a speed command value that specifies the rotational speed of the motor, an output voltage value output to the motor, a speed estimation value of the motor, and a speed detection value of the motor, and acquires, as label data, an alarm cause corresponding to the feature quantity; and a learning unit (11) that generates a learning model for estimating a new alarm cause corresponding to a new feature quantity based on a data set created from a combination of the state variables and the label data.

Description

Technical Field

[0001] The present invention relates to a machine learning device and a machine learning method for learning the cause of an alarm of a motor drive device that drives a motor. Background Art

[0002] When a motor drive device that drives a motor enters an alarm state, after identifying the cause of the alarm, parameters of the motor drive device are changed, a system including the motor drive device is adjusted, and the like.

[0003] The fault diagnosis device described in Patent Document 1 observes data of the fault time of the motor drive device as state variables, obtains repair component data indicating components that have been repaired in the motor drive device as label data, and learns the correlation between the state variables and the label data. The fault diagnosis device uses the learning results to estimate which part of the motor drive device has a fault.

[0004] Patent Document 1: Japanese Patent Application Publication No. 2019-144174 Summary of the invention

[0005] However, in the technology of the above-mentioned patent document 1, since the fault location is learned when a component fails, the cause of the failure can be estimated when the component fails, but the cause of the alarm caused by the usage problem rather than the physical failure of the component cannot be estimated.

[0006] The present invention has been proposed in view of the above-mentioned problems, and an object of the present invention is to obtain a machine learning device capable of generating a learning model for estimating the cause of an alarm generated due to a problem in the usage pattern.

[0007] In order to solve the above-mentioned problems and achieve the purpose, the machine learning device of the present invention for learning the alarm cause of the motor drive device that drives the motor has a state observation unit and a learning unit. The state observation unit obtains the feature quantity obtained from the motor drive device as a state variable, and obtains the alarm cause corresponding to the feature quantity as label data. The feature quantity includes at least one of the current detection value detected from the motor, the speed command value specifying the rotation speed of the motor, the output voltage value output to the motor, the speed estimation value indicating the estimated rotation speed of the motor, and the speed detection value indicating the detected rotation speed of the motor. The learning unit generates a learning model, which is used to estimate a new alarm cause corresponding to a new feature quantity based on a data set created based on a combination of state variables and label data.

[0008] Effects of the Invention

[0009] The machine learning device according to the present invention achieves the following effect, that is, it can generate a learning model for inferring the cause of an alarm generated due to a problem in the usage method. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 It is a diagram for explaining the processing in the learning stage performed by the machine learning device according to the embodiment.

[0011] Figure 2 It is a diagram for explaining the processing in the inference stage performed by the machine learning device according to the embodiment.

[0012] Figure 3 It is a diagram showing a schematic structural example of a decision tree used by the machine learning device according to the embodiment.

[0013] Figure 4 It is a diagram for explaining gradient boosting used by the machine learning device according to the embodiment.

[0014] Figure 5 It is a diagram for explaining random forest used by the machine learning device according to the embodiment.

[0015] Figure 6 It is a diagram showing the structure of the machine learning device according to the embodiment.

[0016] Figure 7 It is a diagram for explaining an example of the cause of an alarm used by the machine learning device according to the embodiment.

[0017] Figure 8 It is a flowchart for explaining the processing flow of inferring the cause of an alarm performed by the machine learning device according to the embodiment.

[0018] Figure 9 It is a diagram showing a schematic hardware structure example for implementing the machine learning device according to the embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] Hereinafter, the machine learning device and the machine learning method according to the embodiment of the present invention will be described in detail with reference to the drawings. In addition, the present invention is not limited to this embodiment.

[0020] Embodiment

[0021] Figure 1 It is a diagram for explaining the processing in the learning stage performed by the machine learning device according to the embodiment. Figure 2 It is a diagram for explaining the processing in the inference stage performed by the machine learning device according to the embodiment.

[0022] The machine learning device 1 learns the state of the motor drive device that drives the motor. The machine learning device 1 is a computer that estimates the cause of the alarm generated by the motor drive device (hereinafter referred to as the alarm cause) through artificial intelligence (AI) technology, and notifies the operator of the alarm cause, the parameters for clearing the alarm cause, etc.

[0023] like Figure 1 As shown, the machine learning device 1 acquires information indicating the state of the motor drive device at which an alarm has been generated in the learning phase. The information indicating the state of the motor drive device at which an alarm has been generated is the alarm cause 32A and the drive data feature 31A.

[0024] The drive data feature 31A is a feature extracted by the motor drive device from a physical quantity acquired by the motor drive device when an alarm is generated in the learning phase. The alarm cause 32A is an alarm cause in the learning phase and corresponds to the drive data feature 31A. The alarm cause 32A is set by an operator or the like.

[0025] The machine learning device 1 learns a learning model representing the relationship between the alarm cause 32A and the drive data feature 31A as the state of the motor drive device. In other words, the machine learning device 1 learns the alarm cause 32A corresponding to the drive data feature 31A.

[0026] like Figure 2 As shown, the machine learning device 1 uses the learning model in the estimation phase to estimate the alarm cause 32B indicating the state of the motor drive device that has generated the alarm based on the drive data feature 31B indicating the state of the motor drive device that has generated the alarm. In addition, the machine learning device 1 obtains the design data 33 in the estimation phase.

[0027] The drive data feature 31B is a feature extracted by the motor drive device from a physical quantity obtained by the motor drive device when an alarm is generated in the estimation phase. The alarm cause 32B is the alarm cause in the estimation phase and corresponds to the drive data feature 31B. The alarm cause 32B is derived by applying the drive data feature 31B to the learning model.

[0028] In the following description, when there is no need to distinguish between the drive data feature quantities 31A and 31B, at least one of the drive data feature quantities 31A and 31B may be referred to as a feature quantity or a drive data feature quantity. In addition, when there is no need to distinguish between the alarm causes 32A and 32B, at least one of the alarm causes 32A and 32B may be referred to as an alarm cause.

[0029] For example, when an alarm is generated, drive data characteristic quantities 31A and 31B are obtained by the motor drive device. Among the physical quantities obtained by the motor drive device, at least one of a current detection value, a speed command value, an output voltage value, and a speed estimated value is included. In addition, the motor drive device may obtain a speed detection value instead of the speed estimated value, or obtain a speed detection value together with the speed estimated value. The drive data characteristic quantities 31A and 31B include at least one of a characteristic quantity of the current detection value, a characteristic quantity of the speed command value, a characteristic quantity of the output voltage value, a characteristic quantity of the speed estimated value, and a characteristic quantity of the speed detection value.

[0030] The characteristic quantities of the respective physical quantities are the maximum value, minimum value, peak value, average value, etc. of the respective physical quantities. The physical quantities obtained by the motor drive device are represented by waveform data, for example.

[0031] The current detection value is the current value detected from the motor when the motor is driven, the speed command value is the command value for designating the rotational speed of the motor. The output voltage value is the voltage value output to the motor when the motor is driven, and the speed estimated value is the rotational speed of the motor estimated when the motor is driven. The speed estimated value is estimated based on various command values, detection values, etc. The speed detection value is the rotational speed of the motor detected when the motor is driven.

[0032] The in-design data 33 includes at least one of a parameter setting value set in the motor drive device, the drive state of the motor drive device, and the internal data of an amplifier that amplifies the power output from the motor drive device. An example of the parameter setting value is a lift setting value of the motor torque. An example of the drive state is the control cycle of the motor. An example of the internal data of the amplifier is the number of bridge arms of the switching bridge arm.

[0033] The alarms include overcurrent, overvoltage, overspeed, motor overheat, inverter overheat, etc. Overcurrent means that the current flowing through the motor exceeds the allowable current, and overvoltage means that the voltage applied to the amplifier exceeds the allowable voltage. Overspeed means that the rotational speed of the motor exceeds the allowable speed. Motor overheat means that the temperature of the motor exceeds the allowable temperature, and inverter overheat means that the inverter circuit that supplies voltage to the motor exceeds the allowable temperature.

[0034] Here, the machine learning performed by the machine learning device 1 will be described. The machine learning device 1 extracts useful rules, knowledge representations, judgment criteria, etc. existing in the set of data input to the machine learning device 1 through analysis, outputs the extraction results, and performs learning of knowledge, thereby realizing machine learning. The methods of machine learning are various and are roughly classified into "supervised learning", "unsupervised learning", and "reinforcement learning".

[0035] The machine learning device 1 of this embodiment adopts the "supervised learning" algorithm. When the machine learning device 1 executes "supervised learning", it obtains a large number of data sets of a certain input and result (label data to be described later), and learns the characteristics of these data sets. Moreover, the machine learning device 1 obtains a model for estimating the result based on the input (hereinafter referred to as the learning model), that is, the correlation between the input and the result. The machine learning device 1 uses algorithms such as extreme gradient boosting to be described later to implement a learning model for estimating the result based on the input.

[0036] The machine learning device 1 can perform learning based on explicit data. By executing supervised learning that can determine the alarm cause of the motor drive device based on the learning result, the correlation between the feature quantity and the alarm cause is learned.

[0037] The processing performed by the machine learning device 1 that performs supervised learning can be divided into two stages: a learning stage and an estimation stage. In the learning stage, the machine learning device 1 learns the correlation between the feature quantity and the alarm cause based on teacher data including the value of the state variable used as the input data and the value of the target variable used as the output data. If the machine learning device 1 is given teacher data, it performs the following learning, that is, when the value of the state variable is input, the value of the target variable is output based on the learned correlation. The machine learning device 1 constructs a learning model, which is an estimation model for outputting the value of the target variable corresponding to the value of the state variable, by giving a large amount of teacher data.

[0038] The state variable is the drive data feature quantity 31A or the drive data feature quantity 31B, and the target variable is the alarm cause 32A or the alarm cause 32B. When the state variable is the drive data feature quantity 31A, the target variable is the alarm cause 32A, and when the state variable is the drive data feature quantity 31B, the target variable is the alarm cause 32B.

[0039] Examples of the state variable are the feature quantity of the current detection value, the feature quantity of the speed command value, etc. The value of the target variable is, for example, a value indicating that "the acceleration time of the motor is set short for the load".

[0040] In this way, when the machine learning device 1 in the learning stage is given teacher data (a group of feature quantity and alarm cause), that is, the state variable, it learns a learning model that represents the correlation between the feature quantity and the alarm cause.

[0041] When the machine learning device 1 in the estimation stage is newly input with the value of the state variable, it uses the learning model to output the value of the target variable corresponding to the new state variable. Specifically, as Figure 2As shown, the machine learning device 1 receives the value of a new state variable, i.e., a new drive data feature amount 31B (st1). Then, based on the value of the new state variable, i.e., the new drive data feature amount 31B, and the trained learning model, the machine learning device 1 estimates the value of a new target variable, i.e., a new alarm cause 32B (st2), and outputs it to a display device (not shown) or the like.

[0042] In addition, the feature amount may also include the feature amount extracted by the motor drive device during normal operation. In this case, there is no alarm cause corresponding to the feature amount during normal operation.

[0043] In addition, the machine learning device 1 has a correction unit for the learning model, i.e., a unique inference unit 14. The unique inference unit 14 receives the estimated result obtained by supervised learning, i.e., the alarm cause 32B, from the learning model (st3). In addition, the unique inference unit 14 receives the acquired data at the time when the motor drive device issues an alarm, i.e., the in-design data 33, from the motor drive device or the like (st4).

[0044] The unique inference unit 14 corrects the estimated result, i.e., the alarm cause 32B, based on the alarm cause 32B received from the learning model and the in-design data 33 received from the motor drive device, thereby improving the accuracy of the estimated result obtained by machine learning. Specifically, the unique inference unit 14 corrects the alarm cause 32B based on the alarm cause 32B and the in-design data 33, and outputs the corrected alarm cause, i.e., the alarm cause 32C (st5).

[0045] The unique inference unit 14 corrects the estimated result, i.e., the alarm cause 32B, so that the machine learning device 1 can estimate with high accuracy even for alarm causes that are difficult to characterize in the state variable. An example of the alarm cause 32C that is difficult to characterize in the state variable is a mis-setting of a parameter setting value or the like.

[0046] In addition, the in-design data 33 may also include change information indicating the parameter setting value that was changed before abnormal operation among the parameter setting values. In this case, the unique inference unit 14 corrects the alarm cause estimated by the learning model based on the change information.

[0047] Thereby, in the case where the alarm cause is the setting of a wrong parameter setting value caused by an operator, the machine learning device 1 can easily determine the alarm cause. Even if the alarm cause is a mis-setting of a parameter setting value, if the in-design data 33 includes change information, the unique inference unit 14 can also grasp which parameter setting value was changed before abnormal operation, so that the determination accuracy for the alarm cause of mis-setting can be improved.

[0048] In addition, the machine learning device 1 may generate or update the learning model by increasing the weight of the alarm cause corresponding to the parameter setting value of the change information. In addition, the machine learning device 1 may generate or update the learning model when the abnormality is resolved by changing the parameter setting value after the abnormality occurs.

[0049] In addition, the machine learning device 1 may store the value of the estimation result corrected by the independent inference unit 14 as an error variable, and correct the learning model based on the error variable. In this case, the machine learning device 1 corrects the learning model so that the learning model can calculate the same estimation result as the estimation result corrected by the independent inference unit 14. In other words, the machine learning device 1 corrects the learning model so that even in the absence of correction by the independent inference unit 14, the learning model can calculate the same estimation result as the case where the independent inference unit 14 performs correction.

[0050] The machine learning device 1 obtains the target variable from the state variable through classification using the tree structure of the decision tree. Figure 3 This is a diagram showing a schematic structural example of a decision tree used by a machine learning device according to an embodiment of the present invention. The machine learning device 1 adjusts the learning result by adjusting the depth of the tree structure of the decision tree 40A. The decision tree 40A is a regression tree that shows how to conditionally branch the input feature quantity to reach label data that is a correct answer.

[0051] In the decision tree 40A, for the state variable, i.e., input data 41, a determination is made in the conditional branch of the first layer whether x1>0. At the first layer, if the input data 41 is x1>0, a determination is made in the conditional branch of the second layer whether x2>0, and if the input data 41 is x1≤0, a determination is made in the conditional branch of the second layer whether x3>0.

[0052] In the second layer, if the input data 41 is x2>0, the input data 41 regresses to the data y1, and if the input data 41 is x2≤0, the input data 41 regresses to the data y2.

[0053] In the second layer, if the input data 41 is x3>0, the input data 41 regresses to the data y3, and if the input data 41 is x3≤0, the input data 41 regresses to the data y4.

[0054] In this case, for example, data y2 is output data 45. That is, data y2 is a target variable (label data). The machine learning device 1 generates a learning model by setting a decision tree 40A that can regress the state variable to the data y2.

[0055] In addition, as an application of the method of using decision trees for supervised learning, the machine learning device 1 can also use a method such as Extreme Gradient Boosting (XGBoost). Extreme Gradient Boosting is an ensemble learning that combines Gradient Boosting and Random Forests.

[0056] Figure 4 It is a diagram for explaining the gradient boosting used by the machine learning device related to the embodiment. Gradient boosting is a method of learning in such a way that the error between the prediction result of a weak learner (a shallow decision tree, etc.) and the correct answer information is minimized. Just by a weak learner such as the decision tree 40A, an error occurs between the prediction result, i.e., the output data 45, obtained by the decision tree 40A and the actual correct answer information (label data).

[0057] When the machine learning device 1 uses gradient boosting, it learns with this error as the target variable and corrects the weak learner, thereby improving the accuracy. For example, the machine learning device 1 reduces the error between the output data 45 and the actual correct answer information by correcting the decision tree 40A to the decision tree 40B.

[0058] For example, the machine learning device 1 generates the decision tree 40B from the decision tree 40A by setting the decision tree 40B in which the conditional branch x1>0 as the first layer in the decision tree 40A is corrected to x4>0, and the conditional branch x2>0 as the second layer is corrected to x5>0. After that, the machine learning device 1 generates a new decision tree in which the conditional branch x4>0 as the first layer in the decision tree 40B is corrected to other branch conditions, and the conditional branch x5>0 as the second layer is corrected to other branch conditions.

[0059] The machine learning device 1 repeats the process of generating a new decision tree with corrected branch conditions to reduce the error between the output data 45 and the actual correct answer information. Thus, the machine learning device 1 can obtain a decision tree with a small difference between the output data 45 and the actual correct answer information.

[0060] The method of repeatedly learning the error to improve the accuracy of the weak learner in such a way that the error between the output data 45 output from the generated decision tree and the actual correct answer information becomes infinitely minimized is gradient boosting, and the gradient boosting related to decision trees is called Gradient Tree Boosting.

[0061] Figure 5This is a diagram showing the random forest used by the machine learning device involved in the implementation. When the machine learning device 1 uses a random forest, it uses multiple decision trees mentioned above and adopts the majority decision principle to improve the accuracy of the estimation result. In Figure 5 the case where the machine learning device 1 uses four decision trees 51 to 54 is described. The decision trees 51 to 54 are the same as the decision tree 40A.

[0062] When the machine learning device 1 uses a random forest, it repeatedly and randomly obtains the input data 41 and creates decision trees. In this way, the decision trees 51 to 54 are created. Thus, by using the decision trees 51 to 54, the machine learning device 1 can select classification methods for various patterns, and therefore the estimation result becomes more accurate than when using a single decision tree for classification. The machine learning device 1 applies the decision trees 51 to 54 to the input data 41 and outputs the result based on the majority decision principle in the decision trees 51 to 54 as the output data 45.

[0063] In addition, as an algorithm for supervised learning, the machine learning device 1 can also use any supervised learning algorithms such as the least squares method, the stepwise method, SVM (Support Vector Machine), and neural networks. Since these other supervised learning algorithms and the above-mentioned learning using decision trees are well-known, the detailed descriptions of the above algorithms are omitted.

[0064] Here, the specific structure of the machine learning device 1 and the process of estimating the alarm cause are described. Figure 6 This is a diagram showing the structure of the machine learning device involved in the implementation. The machine learning device 1 is connected to the motor drive device 2 and the alarm management device 6. The machine learning device 1 obtains the drive data feature amount from the motor drive device 2 and obtains information such as the alarm cause from the alarm management device 6.

[0065] The machine learning device 1 is installed as an information processing device such as a computer connected to the motor drive device 2, for example. The machine learning device 1 has a state observation unit 10, a learning unit 11, a learning result storage unit 12, an alarm estimation unit 13, a unique inference unit 14, a model update unit 15, and an estimation result output unit 16.

[0066] The motor drive device 2 is a device for driving a motor (not shown). The motor drive device 2 has a data acquisition unit 21, a feature amount extraction unit 22, and a non-volatile memory 23. The alarm management device 6 is a device for receiving the alarm cause input by the operator. The alarm management device 6 has an input unit 61, a storage unit 62, and an output unit 63.

[0067] The data acquisition unit 21 of the motor drive device 2 acquires physical quantities (data representing physical quantities) such as current detection value, speed command value, output voltage value, speed estimation value, etc. from the amplifier, etc., and sends them to the feature quantity extraction unit 22. The data acquisition unit 21 acquires the alarm generation date and time, the alarm name, and the alarm content together with the physical quantity when the alarm is generated. In addition, the data acquisition unit 21 acquires the design internal data from the motor drive device 2 and stores it in the non-volatile memory 23.

[0068] The feature extraction unit 22 extracts the drive data feature from the physical quantity acquired by the data acquisition unit 21. The feature extraction unit 22 may convert the drive data feature into numerical values ​​using a common statistical method, or may extract the drive data feature by reducing the dimension through convolution.

[0069] The nonvolatile memory 23 stores the driving data feature extracted by the feature extraction unit 22. Specifically, the nonvolatile memory 23 stores feature information in which the driving data feature, the alarm generation date and time, the alarm name, and the alarm content are associated. In addition, the nonvolatile memory 23 stores design internal information in which the design internal data, the alarm generation date and time, the alarm name, and the alarm content are associated.

[0070] The nonvolatile memory 23 has a function of outputting feature information including drive data feature quantities to the machine learning device 1 and the alarm management device 6. In addition, the nonvolatile memory 23 has a function of outputting design information including design data to the machine learning device 1. In addition, the nonvolatile memory 23 may also be a cloud server configured outside the motor drive device 2.

[0071] The input unit 61 of the alarm management device 6 reads the characteristic information including the driving data characteristic amount from the nonvolatile memory 23. Alternatively, the driving data characteristic amount may be sent from the nonvolatile memory 23 to the input unit 61.

[0072] In addition, the input unit 61 of the alarm management device 6 receives the alarm cause and the alarm cause cancellation parameter input by the operator. The alarm cause cancellation parameter is a parameter of the cause for canceling the alarm. In other words, the alarm cause cancellation parameter is a measure that the operator should perform in order to cancel the alarm.

[0073] The operator inputs the alarm cause and the alarm cause cancellation parameter into the input unit 61 while referring to the characteristic information stored in the alarm management device 6. The input unit 61 stores the alarm information in which the alarm cause, the alarm cause cancellation parameter, and the information for identifying the alarm are associated with each other in the storage unit 62. An example of the information for identifying the alarm, i.e., the alarm identification information, is the alarm generation date and time included in the characteristic information.

[0074] The storage unit 62 is a memory or the like that stores alarm information. The output unit 63 sends the alarm information stored by the storage unit 62 to the state observation unit 10. In addition, the machine learning device 1 can also read the alarm information from the storage unit 62.

[0075] The state observation unit 10 observes the drive data feature amount as a state variable by reading the feature information including the drive data feature amount from the non-volatile memory 23. In addition, it is also possible that the non-volatile memory 23 sends the feature information to the state observation unit 10.

[0076] In addition, the state observation unit 10 reads the alarm information corresponding to the drive data feature amount from the alarm management device 6. At this time, the state observation unit 10 reads the alarm information corresponding to the alarm identification information included in the feature information from the alarm management device 6. Thus, the state observation unit 10 observes the alarm cause included in the alarm information as a target variable. The state observation unit 10 associates the drive data feature amount and the alarm cause based on the alarm identification information. The state observation unit 10 sends the data set of the drive data feature amount and the alarm cause to the learning unit 11 as teacher data.

[0077] The state observation unit 10 can also change the acquired drive data feature amount for each control mode of the motor. That is, the state observation unit 10 can also acquire the drive data feature amount corresponding to the control mode of the motor. The control modes of the motor include vector control, sensorless vector control, V / f (Voltage / frequency) control, AD (Advanced Flux Vector) flux control, etc. V / f control is a control that outputs a voltage corresponding to the frequency.

[0078] When the control mode of the motor is vector control or sensorless vector control, the state observation unit 10 acquires the drive data feature amount including at least one of the speed detection value, the speed estimation value, and the speed command value. In addition, when the control mode of the motor is neither vector control nor sensorless vector control, the state observation unit 10 acquires the drive data feature amount including at least one of the output voltage value and the current detection value.

[0079] The learning unit 11 performs supervised learning. When the machine learning device 1 is set to the learning stage, the learning unit 11 performs supervised learning and stores the learning result, that is, the learning model, in the learning result storage unit 12. The learning unit 11 in the present embodiment performs supervised learning with the state variable, that is, the drive data feature amount, and the target variable, that is, the alarm cause, as inputs.

[0080] The learning unit 11 learns the cause of the alarm using the drive data feature extracted from the physical quantity obtained when the alarm is generated by the motor drive device 2. Since the drive data feature is stored in the non-volatile memory 23, the learning unit 11 may also learn the cause of the alarm using the past drive data feature. In this case, the learning unit 11 learns the cause of the alarm using the past cause of the alarm corresponding to the past drive data feature.

[0081] Figure 7 FIG. 1 is a diagram for explaining an example of an alarm cause used by a machine learning device according to an embodiment of the present invention. Figure 7 , the corresponding relationship information in which the alarm generation date and time, the alarm generation, the alarm content, the alarm cause, and the alarm cause release parameter are associated with each other is shown.

[0082] The alarm generation date and time is the date and time when the alarm is generated, the generated alarm is the name of the alarm, and the alarm content is the content of the alarm.

[0083] An example of alarm content is overcurrent when the motor is accelerated (overcurrent during acceleration). An example of alarm cause is that the acceleration time of the motor is set short for the load. An example of an alarm cause release parameter is the acceleration time of the motor. For example, in the case where an overcurrent is generated when the motor is accelerated because the acceleration time of the motor is set short for the load, the alarm can be released by setting the acceleration time of the motor to be long.

[0084] When the motor drive device 2 generates an alarm, the operator who responds to the alarm inputs the alarm cause and the alarm cause release parameter into the input unit 61 of the alarm management device 6. In addition, the alarm management device 6 can also be configured in the machine learning device 1 or in the motor drive device 2. The input of the alarm cause, the alarm cause release parameter, etc. to the alarm management device 6 by the operator can be performed when responding to the alarm, or can be input afterwards when analyzing the alarm. The corresponding relationship information can be created by the state observation unit 10, or it can be created by the learning unit 11.

[0085] The learning unit 11 sets the alarm cause and the driving data feature quantity generated based on the physical quantity obtained when the alarm corresponding to the alarm cause is generated as one group and inputs them into the learning model. More specifically, the learning unit 11 sets the driving data feature quantity extracted from the physical quantity of a specific period obtained when the alarm corresponding to the alarm cause is generated as a trigger and the alarm cause as one group and inputs them into the learning model.

[0086] In this way, the learning unit 11 can adopt a teacher-assisted learning method when the driving data feature quantity and the alarm cause are input. In this case, the learning unit 11 uses extreme gradient enhancement to estimate the alarm cause with respect to the driving data feature quantity. The learning unit 11 stores the learning result, that is, the learning model, in the learning result storage unit 12.

[0087] The learning result storage unit 12 is a memory or the like that stores the result (learning model) obtained by the learning unit 11 through learning based on the teacher data. The learning result stored in the learning result storage unit 12 can be installed in an external device other than the machine learning device 1, or can be sent to an external device. By inputting the driving data feature quantity into the learning result obtained by the learning unit 11 by the external device, the external device can infer the cause of the alarm. The learning result storage unit 12 can also store the learning result of the learning unit 11. Figure 7 The corresponding relationship information shown is stored.

[0088] In the estimation phase, the alarm estimation unit 13 estimates the cause of the alarm from the drive data feature quantity based on the learning result stored in the learning result storage unit 12. In this case, the alarm estimation unit 13 obtains the drive data feature quantity from the motor drive device 2. The drive data feature quantity can be read from the non-volatile memory 23 by the alarm estimation unit 13, or can be sent from the non-volatile memory 23 to the alarm estimation unit 13. The alarm estimation unit 13 outputs the estimated cause of the alarm to the independent estimation unit 14.

[0089] The alarm cause estimated by the alarm estimation unit 13 and the design internal information stored in the nonvolatile memory 23 are input to the independent estimation unit 14. The independent estimation unit 14 determines whether it is necessary to correct the estimated alarm cause based on the alarm cause estimated by the alarm estimation unit 13 and the data at the time of the alarm, that is, the design internal data. If it is necessary to correct the alarm cause, the independent estimation unit 14 corrects the estimated alarm cause based on the estimated alarm cause and the data at the time of the alarm, that is, the design internal data. The independent estimation unit 14 sends the corrected alarm cause to the model updating unit 15.

[0090] The model updating unit 15 stores, as the estimated error, the corrected position of the alarm cause corrected by the unique estimation unit 14. The model updating unit 15 feeds back the estimated error to the learning unit 11 to cause the learning unit 11 to update the learning model.

[0091] When the estimation result output unit 16 obtains the alarm cause (estimation result) corrected by the independent estimation unit 14 from the independent estimation unit 14 , it obtains the alarm cause cancellation parameter corresponding to the obtained alarm cause based on the correspondence information in the learning result storage unit 12 .

[0092] The estimation result output unit 16 outputs the corrected alarm cause and alarm cause cancellation parameter to a display device, etc. Thus, the alarm cause and the alarm cause cancellation parameter are displayed on the display device, etc. The operator can cancel the alarm by setting the displayed alarm cause cancellation parameter to the motor drive device 2, etc.

[0093] The alarm cause obtained by the state observation unit 10 is the first alarm cause. Figure 1 The alarm cause estimated by the learning model generated by the learning unit 11, that is, the alarm cause estimated by the alarm estimation unit 13, is the second alarm cause, which corresponds to the alarm cause 32A described in Figure 2 The alarm cause estimated by the alarm estimation unit 13 is a new alarm cause corresponding to a new feature amount of a newly generated alarm.

[0094] The alarm cause corrected by the independent inference unit 14 based on the change information is the second alarm cause, and the corrected alarm cause is the third alarm cause. Figure 2 This corresponds to alarm cause 32C described in .

[0095] In addition, when the learning unit 11 generates or updates the learning model by increasing the weight of the alarm cause corresponding to the parameter setting value of the change information, the alarm cause is the third alarm cause, and the learning unit 11 generates or updates the learning model with the third alarm cause as the first alarm cause. In addition, when the abnormality that becomes the third alarm cause occurs and the abnormality is resolved by changing the parameter setting value, the learning unit 11 generates or updates the learning model.

[0096] Next, the flow of the process of estimating the cause of the alarm executed by the machine learning device 1 will be described. Figure 8 This is a flowchart for explaining the processing flow of estimating the cause of an alarm by the machine learning device involved in the embodiment.

[0097] When an alarm is generated, the state observation unit 10 of the machine learning device 1 acquires a feature value corresponding to the alarm as a state variable (step S10 ). In addition, the state observation unit 10 acquires an alarm cause corresponding to the alarm as a target variable (step S15 ).

[0098] The learning unit 11 performs teacher learning by taking the state variable, i.e., the feature quantity, and the target variable, i.e., the alarm cause, as input. That is, the learning unit 11 learns the alarm cause corresponding to the feature quantity (step S20). Specifically, the learning unit 11 learns the alarm cause by setting the dataset of the feature quantity and the alarm cause as teacher data in the learning model.

[0099] After the learning model has been learned, if a new alert is generated, the alert estimation unit 13 receives new feature quantities corresponding to the new alert (step S30). Then, the alert estimation unit 13 estimates the cause of the alert corresponding to the new feature quantity by inputting the new feature quantity into the trained learning model (step S40). The machine learning device 1 performs the processes of steps S10, S15, and S20 in the learning phase and the processes of steps S30 and S40 in the estimation phase.

[0100] In this way, the machine learning device 1 obtains the feature quantities and alert causes extracted from the physical quantities, and thus can quickly generate a learning model for estimating the alert cause before a failure occurs. That is, the machine learning device 1 can quickly generate a learning model for estimating the alert cause due to problems in the usage method rather than due to physical failures of components or the like. Thereby, the machine learning device 1 can estimate the alert cause before a failure occurs. That is, the machine learning device 1 can estimate the alert cause due to problems in the usage method. Therefore, the operator can perform the processing of the alert based on the alert cause. As a result, the machine learning device 1 can shorten the downtime of the device when an alert is generated.

[0101] However, as a method for estimating the alert cause, there are methods such as the operator confirming the state or parameters of the motor drive device and the operator monitoring the input / output information of the motor drive device. In this method, since it depends on the experience of the operator for estimating the alert cause, the speed or accuracy of determining the alert cause varies according to the experience of the operator and the like. In addition, in order to determine whether the alert is caused by a misconfiguration of the motor drive device or an alert cause such as high load or equipment abnormality, sufficient input data is required, so a considerable amount of time is needed.

[0102] In addition, there is also a method for estimating abnormalities based on sound, vibration, and current values. In this method, it is difficult to obtain sufficient input data for judging the condition of a rotating body such as a motor drive device, so high-precision estimation cannot be performed. In addition, there is a method of performing machine learning by using state variables obtained when an abnormality occurs and state variables obtained when no abnormality occurs, thereby obtaining the internal parameters of the machine learning device. In this method, the machine learning device uses the internal parameters for abnormality diagnosis. In this method, it is also difficult to obtain sufficient input data for judging the condition of a rotating body such as a motor drive device.

[0103] Here, the hardware structure of the machine learning device 1 will be described. Figure 9 It is a diagram showing an example of the hardware structure of the machine learning device according to the embodiment.

[0104] The machine learning device 1 can be implemented by an input device 151, a processor 152, a memory 153, and an output device 154. Examples of the processor 152 are a CPU (Central Processing Unit, also known as a central processing unit, a processing unit, an arithmetic unit, a microprocessor, a microcomputer, a DSP (Digital Signal Processor)) or a system LSI (Large Scale Integration). Examples of the memory 153 are a RAM (Random Access Memory) and a ROM (Read Only Memory).

[0105] The machine learning device 1 is implemented by the processor 152 reading and executing a computer-executable learning program for executing the operation of the machine learning device 1 stored in the memory 153. The program for executing the operation of the machine learning device 1, i.e., the learning program, can also be said to be a process or method for causing a computer to execute the machine learning device 1.

[0106] The learning program executed by the machine learning device 1 is a module structure including a state observation unit 10, a learning unit 11, an alarm estimation unit 13, an independent inference unit 14, and a model updating unit 15, which are loaded into the main storage device and are generated on the main storage device.

[0107] The input device 151 receives characteristic information including the driving data characteristic quantity, alarm information including the alarm cause, design information, etc. and sends them to the processor 152. The memory 153 is also used as a temporary memory when the processor 152 performs various processes. The memory 153 stores the learning model, the corresponding relationship information, etc. The output device 154 outputs the alarm cause and the alarm cause release parameter corresponding to the alarm to an external device such as a display device.

[0108] The learning program may also be provided as a computer program product in the form of a file in an installable format or an executable format, stored in a computer-readable storage medium. In addition, the learning program may also be provided to the machine learning device 1 via a network such as the Internet. In addition, the functions of the machine learning device 1 may be partially implemented by dedicated hardware such as a dedicated circuit, and partially implemented by software or firmware.

[0109] In addition, in the present embodiment, the following processing has been described. That is, for convenience, the machine learning device 1 learns and makes inferences using the physical quantities, characteristic quantities, and teacher data related to one motor drive device 2. However, the present embodiment is not limited to such processing. For example, the machine learning device 1 collects the physical quantities, characteristic quantities, and teacher data related to a plurality of motor drive devices 2 within a factory or multiple sites, and performs learning and inference, so that learning and inference can be performed efficiently. In other words, by taking a plurality of motor drive devices 2 as the learning objects, the machine learning device 1 can efficiently learn the alarm causes.

[0110] In addition, in the present embodiment, the case where the machine learning device 1 has a learning unit 11 has been described. However, when the learning result storage unit 12 of the machine learning device 1 stores the learning results obtained by other machine learning devices, the learning unit 11 may not be installed. In other words, the machine learning device having the learning unit 11 and the machine learning device having the alarm inference unit 13 may also be separate machine learning devices.

[0111] As described above, in the embodiment, the state observation unit 10 obtains the characteristic quantities of physical quantities such as the current detection value obtained from the motor drive device 2 as state variables, and obtains the alarm cause corresponding to the characteristic quantity as the target variable. Moreover, the learning unit 11 generates a learning model for inferring the alarm cause based on a data set created from the combination of the state variable and the target variable (label data). Therefore, the machine learning device 1 can generate a learning model for inferring the alarm cause due to problems in the usage mode rather than physical failures of components or the like.

[0112] In addition, the machine learning device 1 has an independent inference unit 14 that corrects the alarm cause inferred by the alarm inference unit 13 based on the in-design data, so that the alarm cause corresponding to the drive data characteristic quantity can be diagnosed with high accuracy.

[0113] In addition, if all the characteristic quantities at the time of generating an alarm are stored in the non-volatile memory 23, it will become huge. Therefore, by limiting the characteristic quantities obtained for each control mode of the motor, the number of teacher data that can be stored in the non-volatile memory 23 can be increased. As a result, the machine learning device 1 can improve the inference accuracy of the alarm cause.

[0114] The structure shown in the above embodiment is an example, and it can be combined with other known technologies, or the embodiments can be combined with each other. Within the scope not departing from the gist, a part of the structure can also be omitted or changed.

[0115] Description of reference numerals

[0116] 1 Machine learning device, 2 Motor drive device, 6 Alarm management device, 10 State observation unit, 11 Learning unit, 12 Learning result storage unit, 13 Alarm estimation unit, 14 Independent inference unit, 15 Model update unit, 16 Estimation result output unit, 21 Data acquisition unit, 22 Feature quantity extraction unit, 23 Non-volatile memory, 31A, 31B Drive data feature quantity, 32A, 32B, 32C Alarm cause, 33 In-design data, 40A, 40B, 51 to 54 Decision trees, 41 Input data, 45 Output data, 61 Input unit, 62 Storage unit, 63 Output unit, 151 Input device, 152 Processor, 153 Memory, 154 Output device.

Claims

1. A machine learning device that learns the cause of an alarm for a motor drive device that drives a motor. Among them, This machine learning device includes: A state observation unit that obtains a feature quantity including at least one of a current detection value detected from the motor obtained from the motor drive device, a speed command value that designates the rotational speed of the motor, an output voltage value output to the motor, a speed estimation value indicating the estimated rotational speed of the motor, and a speed detection value indicating the detected rotational speed of the motor as a state variable, and obtains the cause of the alarm corresponding to the feature quantity as label data. A learning unit that generates a learning model for estimating the cause of a new alarm corresponding to a new feature quantity based on a data set created from a combination of the state variable and the label data. An alarm estimation unit that, if a new feature quantity is obtained, applies the learning model to the new feature quantity and estimates the cause of a new alarm corresponding to the new feature quantity. And A correction unit that corrects the cause of the alarm estimated by the alarm estimation unit based on design-in data including at least one of a parameter setting value set in the motor drive device, the driving condition of the motor drive device, and internal data of an amplifier that amplifies the power output from the motor drive device. In the design-in data, change information indicating a parameter setting value that was changed before abnormal operation among the parameter setting values is included. The correction unit corrects the cause of the alarm estimated by the alarm estimation unit based on the change information. The state observation unit obtains a feature quantity corresponding to the control method of the motor. When the control method of the motor is vector control or sensorless vector control, the state observation unit obtains a feature quantity including at least one of the speed detection value, the speed estimation value, and the speed command value. When the control method of the motor is neither vector control nor sensorless vector control, the state observation unit obtains a feature quantity including at least one of the output voltage value and the current detection value.

2. The machine learning device according to claim 1, characterized in that It further has a model update unit that updates the learning model based on the correction result obtained by the correction unit.

3. The machine learning device according to claim 1, characterized in that The learning unit generates or updates the learning model by increasing the weight of the cause of the alarm corresponding to the parameter setting value of the change information.

4. The machine learning device according to claim 3, characterized in that After an abnormality that causes the cause of the alarm occurs, when the abnormality is eliminated by changing the parameter setting value, the learning unit generates or updates the learning model.

5. A machine learning method that performs processing divided into a learning stage and an estimation stage, and the learning stage learns the cause of an alarm for a motor drive device that drives a motor. This machine learning method is characterized by including: State observation step: In the learning stage, obtain a feature quantity including at least one of a current detection value detected from the motor obtained from the motor drive device, a speed command value for specifying the rotational speed of the motor, an output voltage value output to the motor, a speed estimation value representing the estimated rotational speed of the motor, and a speed detection value representing the detected rotational speed of the motor as a state variable, and obtain an alarm cause corresponding to the feature quantity as label data; Learning step: In the learning stage, generate a learning model for performing the following learning, that is, learn the correlation between the feature quantity and the alarm cause based on a data set created based on the combination of the state variable and the label data, and output the alarm cause based on the correlation; Alarm estimation step: In the estimation stage, if a new alarm occurs, obtain a new feature quantity corresponding to the new alarm, and input the new feature quantity into the trained model to estimate a new alarm cause corresponding to the new feature quantity; And Correction step: In the estimation stage, correct the alarm cause estimated by the trained model in the alarm estimation step based on design-in data including at least one of a parameter setting value set in the motor drive device, a driving condition of the motor drive device, and internal data of an amplifier that amplifies the power output from the motor drive device; The design-in data includes change information indicating a parameter setting value changed before abnormal operation among the parameter setting values; In the correction step, correct the alarm cause estimated by the alarm estimation step through the trained model in the estimation stage based on the change information; In the state observation step, Obtain a feature quantity corresponding to the control mode of the motor; When the control mode of the motor is vector control or sensorless vector control, obtain a feature quantity including at least one of the speed detection value, the speed estimation value, and the speed command value; When the control mode of the motor is neither vector control nor sensorless vector control, obtain a feature quantity including at least one of the output voltage value and the current detection value.

Images