Method for labeling a dataset for a supervised learner, supervised learner, state estimator

Abstract

To perform labeling according to a state to an intermediate state of two reference states at a low cost.SOLUTION: A method for labeling a data set of a supervised learning device, about a plurality of pieces of correspondence data including a case in which a plurality of state estimation objects are a reference state A (hereafter A) and a reference state B (hereafter, B), calculates deviance from the reference state A of each correspondence data (steps S1 to S4), can transfer the deviance from A in the case of A to a prescribed value that should be outputted by the supervised learning device in the case of A, can transfer deviance from A in the case of B to a prescribed value that should be outputted by the learning device in the case of B, and calculates normalization deviance by mapping the deviance from A of each correspondence data by determined mapping in each state estimation object so as to be monotonic mapping in a section between the deviance from A in the case of A and the deviance from A in the case of B (steps S5 to S6), and determines the normalization deviance as an output of training data corresponding to an input of the training data (step S7).SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a method for labeling a data set of a supervised learning device in a machine learning model used in machine tools and the like to diagnose the damage state of tools, etc., a supervised learning device created using the method, and a state estimation device equipped with the learning device. [Background technology]

[0002] In machine tools, if the bearings that support the rotating body are damaged and the machine continues to be used, the machining accuracy will deteriorate or the bearings will seize up, making the machine unable to operate. Also, if cutting tools are used with significant wear, the machining accuracy will deteriorate or the machine will continue to operate even after the cutting tools are broken, causing interference and damaging the machine tool. In the past, in order to avoid the occurrence of such troubles, standards such as time or number of times were set in advance for the cumulative rotation time of a bearing, the cumulative number of rotations of a bearing, the usage time of a tool, etc., and were sometimes managed. However, these did not correspond to the actual condition, and excessive costs were incurred by taking measures when there were no problems with continued use, or trouble occurred even though the specified time or number of times had not been reached. For this reason, methods have been considered that estimate the actual state rather than the time or number of times, detect changes in the state from normal to abnormal, and take action before the state reaches a completely abnormal state. Patent Document 1 discloses an apparatus that acquires machining information when there is sufficient tool life remaining, generates input data, and constructs a learning model in which clusters are created through unsupervised learning. It then performs inference on certain input data and determines that there is no tool life remaining if the input data does not belong to a cluster in which there is sufficient tool life remaining. Patent document 2 discloses a machine tool that is capable of responding according to the damage level by inputting data from actual machining into an unsupervised learning device created using data from normal machining, calculating how far the data from actual machining deviates from the data from normal machining, and comprising an abnormality score calculation means for calculating an abnormality score based on the difference between the output and the input, and a means for assigning the abnormality scores to a finite number of damage levels. Patent Document 3 discloses a tool condition estimation device equipped with a learning model memory unit that estimates the tool condition by classifying dynamic information around the tool when the tool condition is good from other information based on input data such as the chip scattering distance, the chip scattering speed, and the central angle of the chip scattering range, which are based on dynamic information around the tool. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2018-103284 A [Patent Document 2] JP 2018-25936 A [Patent Document 3] JP 2018-138327 A Summary of the Invention [Problem to be solved by the invention]

[0004] However, the device of Patent Document 1 uses a learning model that is learned using only machining information when the tool life is sufficiently remaining, so no one can predict what kind of inference result the learning model will produce for machining information when the tool life is no longer remaining, which is a state in which the learning model has not been learned. Therefore, it is difficult to determine a threshold value that allows the device to determine whether the tool life is no longer remaining. In addition, even if the inference result of the learning model for machining information when the tool life is no longer remaining is inappropriate, the learning model cannot be improved to produce a desired inference result. For this reason, there is a problem in that the states of various state estimation targets cannot be judged using a common criterion. In the machine tool of Patent Document 2, data from actual machining is input to an unsupervised learning device created using only data from normal machining, and the data from actual machining is calculated to see how far the data from actual machining is from the data from normal machining, that is, the abnormality score. Therefore, no one can predict what the magnitude of the abnormality score will be for data from abnormal machining under different machining conditions such as tools, workpieces, and cutting speeds. In addition, it is not possible to improve the abnormality score calculation means so as to calculate a desired abnormality score for data from abnormal machining. For this reason, in order to judge the states of various state estimation targets based on a certain standard, it is necessary to prepare in advance a means for assigning abnormality scores to a finite number of damage levels for each machining condition. Therefore, there is a problem that it is not possible to respond to new machining conditions. In the tool state estimation device of Patent Document 3, when unsupervised learning is used to create a cluster when the tool state is good, when machining conditions such as the tool, workpiece, and cutting speed are different, no one can predict how far the inference results of the machine learning model for input data under different machining conditions will deviate from the cluster when each tool state is good. For this reason, there is a problem that the state of various state estimation targets cannot be judged based on a certain criterion. When supervised learning is used, a method is shown in which each signal in the output layer is made to correspond to "good tool state" and "worsened tool state", and the teacher data is determined so that one of the signals is 1 in correspondence with the teacher data. With such supervised learning, it is possible to judge the state of various state estimation targets based on a certain criterion. However, in supervised learning, it is very difficult to appropriately label intermediate states between two reference states, such as a state between "good tool condition" and "worsened tool condition". Furthermore, if the intermediate states are not labeled, data on the intermediate states cannot be added to the training data. Therefore, even if a desirable inference result is not obtained in the intermediate states, the learning model cannot be improved. Patent Document 3 also shows another example in which teacher data is determined for each time period before the tool condition deteriorates in order to label each intermediate state, but this method is difficult to say as an appropriate labeling according to the degree of deterioration of the tool.

[0005] FIG. 1 is a conceptual diagram for explaining the problem of referring to time for labeling in an intermediate state. Here, the horizontal axis represents the tool use time, and the vertical axis represents the tool wear amount as the state to be estimated. The diagram shows the transition of the tool wear amount in three cases: a case where the tool wear progresses slowly, a case where the tool wear progresses rapidly, and a case where the tool wear condition is changed from a case where the tool wear progresses slowly to a case where the tool wear progresses rapidly. In FIG. 1, the change in the tool wear amount under the machining condition where the tool wear progresses slowly is shown by a solid line, and the change in the tool wear amount under the machining condition where the tool wear progresses rapidly is shown by a dashed line. Note that FIG. 1 is a conceptual diagram, and the tool wear amount does not necessarily increase in proportion to the tool use time in reality. As shown in FIG. 1, it is clear that the difference in the cumulative machining time required from a certain amount of tool wear to the time when the tool condition deteriorates is not the same in each of the above three cases. That is, in each of the above three cases, data acquired at the same tool wear amount, which should be labeled with the same label, is labeled with a different label indicating that it is a different time period before the tool condition deteriorates, because the cumulative machining time required from the same tool wear amount to the time when the tool condition deteriorates is different. Since a learning device cannot learn to produce different outputs for the same input, if different labels are attached to the same state, it will not be able to learn properly. In other words, when referring to time for labeling the intermediate state, there is a problem that the learning device will not be able to learn properly. On the other hand, in order to accurately label the intermediate state, it is possible to check the tool state every time between machining and grasp the changes in the state. However, in this case, it is necessary to spend a lot of time checking the tool state. As a result, there is an issue that the productivity decreases due to the increase in time unrelated to machining, and the cost of labeling is incurred.

[0006] Therefore, in consideration of the above problems, the present invention aims to provide a method for labeling a dataset of a supervised learner, a supervised learner, and a state estimation device, which are capable of labeling intermediate states between two reference states in a manner that is more in line with the actual state at low cost in labeling a dataset that determines the output of training data corresponding to the input of training data of a supervised learner for judging the states of various state estimation targets based on certain criteria. [Means for solving the problem]

[0007] In order to achieve the above object, the invention described in claim 1 is a method for labeling a dataset of a supervised learning device for determining an output of training data corresponding to an input of training data in training data for creating a supervised learning device used for estimating in which state a state estimation target is in between a case where the state estimation target is in a predetermined reference state A and a case where the state estimation target is in a predetermined reference state B different from the reference state A, the method comprising: A first physical quantity obtained when a state estimation target is in a predetermined state and, Data related to a state of a state estimation target obtained based on a first physical quantity and at least one of Let any data be the training data input, A second physical quantity that is acquired when the state of the state estimation target is the same as the state when the first physical quantity related to the input of the training data is acquired. and, Data related to a state of a state estimation target obtained based on a second physical quantity and at least one of Any data is called corresponding data. For two or more state estimation targets, a deviation degree from a reference state A is calculated for each of a plurality of corresponding data including at least corresponding data when the state estimation target is in a reference state A and corresponding data when the state estimation target is in a reference state B; As a mapping for mapping the deviation of each of the multiple corresponding data from the reference state A to the normalized deviation, a map capable of mapping the deviation from the reference state A when the reference state is A to a prescribed value to be output by the supervised learning device when the reference state is A, and mapping the deviation from the reference state A when the reference state is B to a prescribed value to be output by the supervised learning device when the reference state is B, and a map which is a monotonic map is set for each state estimation target in an interval between the deviation from the reference state A when the reference state is A and the deviation from the reference state A when the reference state is B, Using the set mapping, a normalized deviation is calculated by mapping the image of the deviation from the reference state A for each of the multiple corresponding data, The normalized deviation is determined as a training data output corresponding to a training data input. The invention described in claim 2 is, in the above configuration, when the state of the state estimation target transitions in one direction from either one of reference state A and reference state B to the other in chronological order, For each of the multiple state estimation targets, the deviation of each of the multiple corresponding data from a reference state A is calculated, and then the calculated deviations from the reference state A are arranged in chronological order and filtered to convert them into a smooth deviation from the reference state A; This method is characterized in that when setting a mapping for mapping the deviation of each of the multiple corresponding data from a reference state A to a normalized deviation and when calculating the normalized deviation, a deviation from a smooth reference state A is used instead of the deviation from the reference state A. In order to achieve the above object, the invention described in claim 3 is a method for labeling a dataset of a supervised learning device for determining an output of training data corresponding to an input of training data in training data for creating a supervised learning device used for estimating in which state a state estimation target is in between a case where the state estimation target is in a predetermined reference state A and a case where the state estimation target is in a predetermined reference state B different from the reference state A, the method comprising: The supervised learning device has a function of performing post-processing to calculate a representative value in an output group, which is a set of outputs calculated from each input, when a plurality of predetermined data belonging to a data group, which is a set of a plurality of data, are input; A first physical quantity obtained when a state estimation target is in a predetermined state and, Data related to a state of a state estimation target obtained based on a first physical quantity and at least one of Let any data be the training data input, A second physical quantity that is acquired when the state of the state estimation target is the same as the state when the first physical quantity related to the input of the training data is acquired. and, Data related to a state of a state estimation target obtained based on a second physical quantity and at least one of Any data is called corresponding data. For two or more state estimation targets, a deviation degree of the corresponding data belonging to each of the corresponding data groups from the reference state A is calculated for each of the corresponding data groups, the corresponding data groups including at least a first corresponding data group consisting of a plurality of corresponding data when the state estimation target is in a reference state A and a second corresponding data group consisting of a plurality of corresponding data when the state estimation target is in a reference state B; A set of deviations from the reference state A corresponding to a plurality of corresponding data groups to which corresponding data corresponding to each calculated deviation from the reference state A belongs is defined as a deviation group from the reference state A; For each of the plurality of deviation groups from the reference state A, a process identical to the post-processing performed by the supervised learning device is applied to calculate a representative deviation from the reference state A for each of the plurality of deviation groups from the reference state A; As a mapping for mapping the representative deviation from the reference state A corresponding to each of the multiple corresponding data groups to the normalized representative deviation, a representative deviation from reference state A when reference state A is assumed to be a specified value that the representative value calculated by the supervised learning device should take when reference state A is assumed, and a representative deviation from reference state A when reference state B is assumed to be a specified value that the representative value calculated by the supervised learning device should take when reference state B is assumed; and a map that is a monotonic map is set for each state estimation target in an interval between the representative deviation from reference state A when reference state A is assumed to be and the representative deviation from reference state A when reference state B is assumed; Using the set mapping, a normalized representative deviation is calculated by mapping an image of the representative deviation from the reference state A for each of the multiple corresponding data groups; A process that is the reverse of the post-processing is performed on each of the plurality of normalized representative deviations and each of the deviation groups from each reference state A corresponding to each of the plurality of normalized representative deviations, and a normalized individual deviation that will result in a normalized representative deviation when post-processing is performed is calculated for each of the plurality of corresponding data groups corresponding to the plurality of normalized representative deviations; The normalized individual discrepancy is determined as a training data output corresponding to a training data input. The invention described in claim 4 is, in the above configuration, when the state of the state estimation target transitions in one direction from either one of reference state A and reference state B to the other in chronological order, For each of a plurality of state estimation targets, a representative deviation from a reference state A for each of a plurality of corresponding data groups is calculated, and then the calculated representative deviations from the reference state A are arranged in chronological order and filtered to convert them into a smooth representative deviation from the reference state A; This method is characterized in that when setting a mapping for mapping the representative deviation from reference state A of each of a plurality of corresponding data groups to a normalized representative deviation and when calculating the normalized representative deviation, a representative deviation from a smooth reference state A is used instead of the representative deviation from reference state A. The invention described in claim 5 is a supervised learning machine, which is created using the method for labeling a dataset for a supervised learning machine described in claim 1 or 2. The invention recited in claim 6 is a state estimation device, characterized in that it is equipped with the supervised learning device recited in claim 5. The invention described in claim 7 is characterized in that, in the above configuration, the output of the supervised learning device, a specified value that the supervised learning device should output when the state estimation target is in reference state A, and a specified value that the supervised learning device should output when the state estimation target is in reference state B are simultaneously displayed as numerical values ​​or graphs. The invention described in claim 8 is a supervised learning machine, which is created using the method for labeling a dataset for a supervised learning machine described in claim 3 or 4. The invention recited in claim 9 is a state estimation device, characterized in that it is equipped with the supervised learning device recited in claim 8. The invention described in claim 10 is characterized in that, in the above configuration, a representative value in the output group output by the supervised learning device, a specified value that the representative value calculated by the supervised learning device should take when the state estimation target is in reference state A, and a specified value that the representative value calculated by the supervised learning device should take when the state estimation target is in reference state B are simultaneously displayed as numerical values ​​or graphs. Effect of the Invention

[0008] According to the present invention, the output of the training data determined by the present invention can be given a normalization effect to a supervised learning device that has learned using training data including the output determined by the present method, or a supervised learning device with post-processing function, so as to output a common value in reference state A and reference state B regardless of the state estimation target. Furthermore, since the output of the training data determined by the present invention reflects the timing of change in the deviation from reference state A even when the state of the state estimation target is intermediate between reference state A and reference state B, a supervised learning device that has learned using training data including the output determined by the present method, or a supervised learning device with a post-processing function, can be provided with a state estimation ability that is more in line with the actual state. In other words, it becomes possible to label a data set that is more in line with the actual state at low cost, even for an intermediate state between two reference states. [Brief description of the drawings]

[0009] [Figure 1] FIG. 1 illustrates the challenges of referencing time in labeling intermediate states. [Diagram 2] 1 is a flowchart of a labeling method according to a first embodiment. [Diagram 3]FIG. 13 is a graph showing the deviation from the reference state A. [Figure 4] FIG. 13 is a diagram showing the deviation from the reference condition A when there is measurement variability. [Diagram 5] FIG. 13 is a diagram showing normalized deviation. [Figure 6] 4 is a screen displayed by the state estimation device. [Figure 7] 4 is a screen displayed by the state estimation device. [Figure 8] 11 is a flowchart of a labeling method according to a second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. First, the first embodiment will be described. Figure 2 shows a flowchart of how to label a dataset for a supervised learner. In the following description of Example 1, the state estimation target is a bearing that supports a spindle in a machine tool. When two or more state estimation targets are involved, each state estimation target is a separate bearing. The state of the bearing immediately after it is put into use is referred to as reference state A. Furthermore, the state in which the bearing has deteriorated and should be replaced with a new one is referred to as reference state B. In data collection step S1, data required for labeling the data set of the supervised learning machine is collected. In the first embodiment, each time the machine tool is used for a predetermined period of time from the state immediately after the start of use, i.e., reference state A, to the state in which the bearing should be replaced due to deterioration, i.e., reference state B, signals are collected regarding physical quantities such as the spindle motor torque at that time, the vibration of an acceleration sensor attached near the bearing, and the temperature of a temperature sensor attached near the bearing, and the collected physical quantities are recorded as data.

[0011] Following the data collection step S1, a data generation step S2 is performed in which training data is input to the supervised learner and corresponding data is generated. Training data is data used to train a learning device to produce a desired output for a given input, and in this application, unless otherwise specified, is used to include both input and output. The input of training data for the supervised learning device refers to any data selected from a physical quantity (first physical quantity) obtained when the state estimation object is in a specified state and data related to the state of the state estimation object obtained based on the first physical quantity. The corresponding data refers to any data selected from a physical quantity (second physical quantity) acquired when the state of the state estimation target is the same as when the first physical quantity is acquired and data related to the state of the state estimation target obtained based on the second physical quantity. The same state here means that the state of the state estimation target at the timing of acquiring the second physical quantity can be regarded by a person skilled in the art as being the same as the state of the state estimation target when the first physical quantity is acquired, and it is acceptable even if the state does not completely match as the state of the state estimation target. In the description of the present invention, the corresponding data may be used to mean not only the input of training data but also the output. The data related to the state of the state estimation target obtained based on the first physical quantity and the second physical quantity refers to data generated by performing a predetermined process such as signal processing on the first physical quantity and the second physical quantity, respectively. However, the types of the first physical quantity and the second physical quantity, the signal processing method, the sampling frequency, etc. do not need to be the same between the input of the training data of the supervised learner and the corresponding data. In other words, the first physical quantity and the second physical quantity may be different physical quantities as long as the state of the state estimation target is acquired in the same state, and these physical quantities may be used as the input and corresponding data of the training data. Also, different types of data related to the state of the state estimation target corresponding to each physical quantity may be generated from the first physical quantity and the second physical quantity, which are different types of physical quantities, through different processes, and the generated different types of data related to the state of the state estimation target may be used as the input and corresponding data of the training data. In addition, multiple different types of data may be used as both the input and corresponding data of the training data. In addition, information that can be considered to have small changes depending on the state change of the state estimation target is collected in any state, These may then be stored and added to the inputs of the training data or correspondence data for a supervised learner. Examples of information that can be considered to have small changes depending on the state of the object of state estimation include the spindle mass and the susceptibility of an acceleration sensor attached near a bearing to vibration, i.e., the magnitude of the transfer function from the force generated by bearing damage at each frequency to the vibration acceleration at the acceleration sensor position. In the data generation step S2 of the first embodiment, the vibration amplitude for each frequency obtained by frequency analysis of the vibration of the acceleration sensor, the spindle mass, and the ease of vibration of the acceleration sensor attached near the bearing for each frequency are used as inputs of training data for the supervised learning device. An example of the ease of vibration of the acceleration sensor attached near the bearing is inertance. In addition, the time series waveform of the spindle motor torque, the time series waveform of the vibration, and the time series waveform of the temperature are used as corresponding data.

[0012] Next, a deviation calculation learning device creation step S3 is executed to create a deviation calculation learning device that calculates the deviation from the reference state A. Examples of learning machines that can be used to calculate the deviation from reference state A include unsupervised learning machines such as autoencoders, learning machines that perform predictive learning, and VAEs (variational autoencoders), which are created using corresponding data in which the state of the object to be estimated is considered to be close to reference state A as training data. Autoencoders are The number of input units and the number of output units are The same deep neural network uses the same training data for both input and output, and learns to minimize the error between the input and output. Therefore, for data obtained when the state of the state estimation target included in the training data can be considered to be close to reference state A, the error between the input corresponding data and the output when the corresponding data is input to the autoencoder is calculated to be small. On the other hand, as the state of the state estimation target at the time the data is obtained approaches reference state B from reference state A, the error between the input corresponding data and the output when the corresponding data is input to the autoencoder is calculated to gradually increase. This error can be regarded as the deviation from the reference condition A. The learning device that performs predictive learning uses a deep neural network that learns to input data that comes earlier in the training data and output data that comes later in the training data. For data included in the training data that can be considered to be close to the reference state A, Supported Data The output when the data ahead in time is input is Supported Data On the other hand, as the state of the state estimation target at the time when the data is obtained approaches reference state B from reference state A, Supported Data The output when the data ahead in time is input to a learning device that performs predictive learning is Between corresponding data and later data As the error gets gradually larger, This error can be regarded as the deviation from the reference condition A. VAE is a form of autoencoder, which calculates the mean and variance in particular in the intermediate layer. It is said that VAE has a more pronounced property of not reproducing features that are not included in the training data, and for data included in the training data where the state of the state estimation target can be considered to be close to reference state A, the error between the input corresponding data and the output when the corresponding data is input into the VAE is calculated to be small. On the other hand, as the state of the state estimation target at the time the data is obtained approaches reference state B from reference state A, the error between the input corresponding data and the output when the corresponding data is input into the VAE is calculated to gradually increase, This error can be regarded as the deviation from the reference condition A. All of the learning machines shown in the examples have the property that they can calculate the deviation from reference state A by learning using data that is considered to be close to reference state A, and they can be replaced with learning machines having similar properties. A learning device that calculates the deviation from a reference state A may be created for each state estimation object, may be created for each group of state estimation objects, or may be created just one for all state estimation objects.

[0013] In the deviation calculation learning machine creation step S3 in the first embodiment, the autoencoder is trained as a deviation calculation learning machine using corresponding data collected at a time that can be considered close to the state immediately after starting use (reference state A) for a bearing BR1 that supports the spindle of a machine tool model MA, and a bearing BR2 that supports the spindle of a model MB that has a different vibration tendency from model MA when an acceleration sensor attached near the bearing is installed near the bearing. In the next step, deviation calculation step S4, a plurality of corresponding data acquired from the state immediately after the start of use of the bearings BR1 and BR2 (reference state A) to the state when they should be replaced due to deterioration (reference state B) are input to the trained autoencoder created in deviation calculation learning machine creation step S3, and a corresponding output is calculated for each of the input corresponding data. Calculated Output Based on this, the error between the output calculated when each corresponding data is input to the autoencoder and the input corresponding data is calculated as the deviation from the reference state A. In the first embodiment, the average of the absolute values ​​of the error between the output when the corresponding data is input to the autoencoder and the input corresponding data is used as the deviation from the reference state A. In another embodiment, the deviation from the reference state A is calculated by inputting a certain corresponding data to an autoencoder and calculating the output. For example, use the average of the squared difference between the input corresponding data and the corresponding data. It may also be replaced with another value that can be considered as the error between the input and output, determined based on the output when input to the autoencoder and the corresponding input data.

[0014] In deviation calculation step S4, a plurality of corresponding data acquired from the state immediately after the start of use of bearings BR1 and BR2 (reference state A) to the state when they should be replaced due to deterioration (reference state B) are input to the autoencoder trained in deviation calculation learning machine creation step S3, and the deviation from reference state A is calculated. The calculated deviation from reference state A of bearings BR1 and BR2 is shown in Figure 3. The state of the bearings BR1 and BR2 either deteriorates or does not change, and does not return from a deteriorated state to a non-deteriorated state. In other words, the state of the bearings BR1 and BR2 should change in one direction, from the reference state A to the reference state B. Therefore, if the deviation from the reference state A obtained in the deviation calculation step S4 is not calculated as a monotonically increasing value as shown in FIG. 4, it is presumed that the deviation from the reference state A includes the influence of a phenomenon unrelated to the deterioration of the state of the bearings BR1 and BR2, such as measurement variation. If it is presumed that the deviation from the reference state A includes a component unrelated to the deterioration of the state of the bearings BR1 and BR2, it is possible to remove the component unrelated to the deterioration of the state of the bearings BR1 and BR2 by applying a moving average filter or a low-pass filter to the deviation from the reference state A in a state where the deviation from the reference state A is arranged in chronological order and converted into a smooth deviation from the reference state A. If the deviation from the reference state A has been subjected to filtering, the deviation from the smoothed reference state A is used in the subsequent steps instead of the deviation from the reference state A. In this manner, in deviation calculation step S4, the deviation from the reference state A or the deviation from the smooth reference state A is calculated.

[0015] In the normalization mapping determination step S5 executed next, a mapping (normalization mapping) is determined for mapping the deviations from the reference state A of each of the multiple corresponding data obtained from the states of the bearings BR1 and BR2 calculated in the deviation calculation step S4 or the deviations from the smooth reference state A as normalized deviations in the normalization deviation calculation step S6. In the following description of the first embodiment, for the purpose of facilitating understanding, the deviations from the reference state A and the deviations from the smooth reference state A are both expressed as the deviations from the reference state A. However, as described above, in the case where the deviations from the reference state A without filtering are used in the steps described later, only the deviations from the reference state A without filtering are used, and in the case where the deviations from the smooth reference state A are used, only the deviations from the smooth reference state A are used, and the deviations from the reference state A and the deviations from the smooth reference state A are not used in combination. Here, the deviation from the reference state A in the case of the reference state A is the deviation V A The deviation from the reference condition A when the reference condition B is the deviation V B The standard value that the supervised learning model should output when the reference state is A is the standard value W A The default value that the supervised learning model should output when the reference state is B is the default value W B If a map satisfies all of the following three properties, where n is a positive exponent, it can be used to map the deviation V of each of multiple corresponding data from the reference state A to the normalized deviation W. The first property is that the deviation V A The specified value W A The second property is that it is possible to map V B The specified value W B The third property is that it is possible to map V A And the deviation V B The mapping must be monotonic in the interval between . A mapping that satisfies these three properties is as follows: In addition, when determining the mapping for each desired state estimation target, not limited to bearings, the specified value W A and the specified value W B By making W and W the same for all state estimation targets, the supervised learning device trained using training data created by the labeling method of the present invention can incorporate a normalization effect. The normalization effect here means that when the state estimation target is in a range from reference state A to reference state B, the output of the supervised learning device is normalized to a specified value W regardless of the state estimation target. A From the specified value W B This refers to work that falls within a common scope.

[0016] V A <V B , W A <W B When n=1, the number 1 becomes a strictly monotonically increasing map. Deviation V A and deviation V B Since the values ​​of the bearings BR1 and BR2 are used, the mapping is determined for each of the bearings BR1 and BR2. However, the specified value WA and W B is a common value for bearings BR1 and BR2. When the mapping corresponding to bearings BR1 and BR2 is used, the deviation V (Fig. 3) of bearings BR1 and BR2 is mapped to the normalized deviation W (Fig. 5) of bearings BR1 and BR2. The normalized deviation shown in Fig. 5 is common to bearings BR1 and BR2, and is calculated by dividing the normalized deviation in the reference condition A by the specified value W A The normalized deviation in the case of the reference state is the specified value W B It is as follows.

[0017]

number

[0018] In addition, any mapping that satisfies the above-mentioned three properties, such as the following formula 2 obtained by adding formula 1 at any ratio, can also be adopted as a mapping that maps the deviation V from the reference state A to the normalized deviation W.

[0019]

number

[0020] In the case of a mapping created without using Equation 1 and Equation 2 (hereinafter referred to as a mapping candidate), if it is confirmed that all three of the above characteristics are satisfied, the normalized deviation calculation step S6 It can be adopted as a mapping that maps the deviation V from the reference state A used in to the normalized deviation W. Below, a method for checking whether or not a mapping candidate satisfies the above three properties will be described. To confirm that a mapping candidate satisfies the first property, the deviation V A The value obtained by mapping is the standard value W that the supervised learning model should output when the reference state is A. A Or, you can check whether the deviation V A The mapping value is the specified value W A It is only necessary to confirm that the error is within a predetermined range. To confirm that a candidate mapping satisfies the second property, the candidate mapping must have a deviation V B The mapping value is the specified value W B Or, you can check whether the deviation V B The mapping value is the specified value W B It is only necessary to confirm that the error is within a predetermined range. To confirm that a candidate mapping satisfies the third property, if the candidate mapping is a differentiable function, the first derivative of the function must be smaller than the deviation V A and deviation V B It is sufficient to confirm that in the interval between and is either always greater than or equal to 0. Depending on the purpose of state estimation, when it is acceptable for the value obtained by mapping the deviation from the reference state A using the mapping candidate to be constant in a portion of the deviation from the reference state A, it is sufficient to confirm whether the mapping candidate is monotonically increasing or decreasing in a broad sense, which includes a case where the first derivative of the mapping candidate is 0. When it is not acceptable for the value obtained by mapping the deviation from the reference state A using the mapping candidate to be constant in a portion of the deviation from the reference state A, it is sufficient to confirm whether the mapping candidate is monotonically increasing or decreasing in a narrow sense, which does not include a case where the first derivative of the mapping candidate is 0. On the other hand, if the mapping candidate is not a differentiable function, the deviation V A From, the deviation V B For example, the deviation V A Deviation from V B By gradually changing the deviation V from the reference state A by 1 / 10000 of the interval, for example, and calculating the mapping value for each deviation V from the reference state A that has been changed using the mapping candidates, it is possible to confirm that the mapping value either always increases or always decreases before and after changing the deviation V from the reference state A. If it is acceptable for the mapping value not to change, it is possible to confirm that the mapping value either never decreases or never increases. As described above, in the normalization mapping determination step S5, a mapping that maps the deviation V from the reference state A to the normalization deviation W is determined for each state estimation target.

[0021] Next, in normalized deviation calculation step S6, the deviation V from reference state A is mapped to normalized deviation W using the mapping determined in normalization mapping determination step S5 that maps the deviation V from reference state A to normalized deviation W. Then, in the label determination step S7, the normalized deviation W calculated using the corresponding data obtained when the input of training data and the corresponding state of the state estimation target are the same is set as the output of training data corresponding to the input of training data of the supervised learning device. Thereafter, a supervised learning device creation step S8 is executed in which a combination of the training data input and the normalized deviation degree W as the training data output is used as training data to train the supervised learning device. The supervised learning machine used here may be one that can perform regression so as to output the normalized deviation W for the input of training data. A supervised learning machine that performs regression may be, for example, a deep neural network. In addition, a label determination step may be performed. S7 It is also possible to use a supervised learning device that performs a process of discretizing the normalized deviation W as the output of training data corresponding to the input of training data of the supervised learning device, calculates the discretized normalized deviation, and learns into which discretized normalized deviation the input of training data should be classified. An example of a supervised learning device that performs such classification is a support vector machine.

[0022] The trained supervised learning machine created in the supervised learning machine creation step S8 can be used, for example, to estimate the state of bearing BR3 supporting the spindle of model MC, in which the susceptibility of an acceleration sensor attached near the bearing to vibration is different from that of both model MA and model MB. In other words, the inference result (output) obtained by inputting the same input as the training data of a supervised learning device generated by performing signal processing, etc. on physical quantities obtained for another state estimation target during actual operation, or physical quantities, to a trained supervised learning device, is equal to or greater than the specified value W A and the specified value WB The state of the bearing BR3 can be determined based on whether the state is closer to the state of the bearing BR3 or the state of the bearing BR3. Generate In step S2, the vibration amplitude for each frequency obtained by frequency analysis of the vibration of the acceleration sensor adopted as training data, the spindle mass, and the ease of vibration of the acceleration sensor attached near the bearing for each frequency are obtained. Therefore, as shown in Figure 6 or Figure 7, the default value W that the supervised learning device should output in the reference state A is A (0.00) and the default value W that the supervised learning model should output when in the reference state. B (1.00) as a numerical value or graph and simultaneously with the inference result value (0.21), the supervised learning device functions as a state estimation device. A , specified value W B and a predetermined display means capable of simultaneously displaying the values ​​of the inference results, thereby making it possible to form a state estimation device. Furthermore, the state estimation device determines the specified value W A and the specified value W B A value between (for example, 0.2 × W A +0.8×W B ) as a threshold value, a warning can be displayed on the screen when the threshold value is exceeded, and a notification means can be provided to notify the warning by email or the like to the machine manager or the machine tool manufacturer that will replace or repair the bearings. It should be noted that the values ​​exemplified in FIG. 6 and FIG. 7 and the values ​​exemplified as threshold values ​​are hypothetical values, and the values ​​exemplified are not actually calculated and set.

[0023] Next, a second embodiment will be described. In the second embodiment, a drill used for hole machining in a machining center is used as a state estimation target, and reference state A is a state immediately after starting to use the drill, and reference state B is a state in which the drill should be replaced due to deterioration. In this case, the two or more state estimation targets are different drills. Incidentally, when performing learning and inference with a supervised learning device, the number of data to be input to the supervised learning device must be uniform. In the case of collecting data of the same length by performing a predetermined diagnostic operation to diagnose the bearings of the main shaft as described above in Example 1, the number of data is uniform, and it is possible to input the data as it is. However, the processing time for drilling holes using a drill may differ depending on the processing content. In this case, in order to uniform the number of data to be input, information for a predetermined time is extracted while shifting the extraction start time little by little. After that, a post-processing is performed to determine a representative value from among the multiple outputs (output group) calculated by inference using each of the extracted data (data group) as input and the supervised learning device, thereby making it possible to show the state of each hole drilling. In the following, when the word "group" is added to the end of a noun, such as a data group or an output group, it indicates that it is a single set that serves as a parent for calculating a representative value that expresses a state. Therefore, for example, in order to diagnose the bearings of a spindle, it is possible to perform operations at different rotational speeds consecutively within a short period of time in which the state of the bearings is considered to be almost unchanged, and to collect and store a plurality of data including different types of data, which are collected and stored, as one data group. In addition, for the post-processing of calculating a representative value from the output group, an appropriate method may be selected depending on the type of state estimation target and the data collection method. For example, when the same measurement is repeated to reduce the effect of variation, the average value of the output group may be used as the representative value. In this application, the post-processing refers to a process of determining representative data from a set of predetermined data, and the process of determining the representative data may be set arbitrarily.

[0024] In the case of a series of operations for drilling holes, in which there is a non-machining section before reaching the workpiece and a machining section in which the workpiece is being cut, the spindle load and feed shaft load have similar values ​​in both the reference state A and the reference state B, since the drill and the workpiece are not in contact in the non-machining section. Therefore, the supervised learning machine should learn to calculate the same output as in the reference state A in the non-machining section, even if the state is close to the reference state B. Otherwise, it is necessary to learn to calculate different outputs for almost the same input, and appropriate learning will not be possible. However, if the average value of the output group is adopted as the representative value, for example, if the ratio of the non-machining section is large even when approaching the reference state B, a value close to the reference state A will be calculated. In such a case, by using the maximum value of the output group as the representative value, it is possible to capture the change in state without depending on the ratio of the non-machining section and the machining section. In addition, if the specified value that the supervised learning device should output in reference state B is greater than the specified value that the supervised learning device should output in reference state A, the maximum value of the output group is used as the representative value, but if the specified value that the supervised learning device should output in reference state B is smaller than the specified value that the supervised learning device should output in reference state A, the minimum value of the output group must be used as the representative value. In Example 2, post-processing is used in which the maximum output value in the output group of the supervised learning device is used as the representative value. Even in the case of drilling holes, a machine learning device that distinguishes between non-machined and machined sections can be provided separately to selectively infer only the data group of the machined section, so that it is possible to employ post-processing in which the average value of the output group is used as the representative value. Taking the moving average of multiple consecutive outputs It is also possible to take the maximum value.

[0025] FIG. 8 shows a flowchart of another method for performing labeling, and the second embodiment will be specifically described based on this flowchart. In data collection step S101, data required for labeling the data set of the supervised learning machine is collected. In the second embodiment, a series of operations for drilling holes is repeated from immediately after the start of use until the machine is deteriorated and needs to be replaced with a new one, and the spindle motor torque and feed axis motor torque at that time are collected and recorded. Next, in a data generation step S102, training data for the supervised learning device is input and corresponding data is generated. For the spindle motor torque and feed shaft motor torque in a series of operations for drilling holes collected in data collection step S101, the process of extracting data for a predetermined length of time T2 (e.g., T2=0.5 seconds) is repeated while shifting the extraction start time by a predetermined time T1 (e.g., T1=0.01 seconds), to generate multiple training data inputs with timings at which the state of the state estimation target can be considered to be the same. Here, when T1=0.01 seconds and T2=0.5 seconds, data from 0 seconds to 0.5 seconds from the start of processing, data from 0.01 seconds to 0.51 seconds from the start of processing, ... are obtained. In addition, in the second embodiment, the corresponding data is the same as the input of each training data. As described above, the corresponding data does not have to be the same as the input of the training data as long as it is data collected in the same state as the training data, and if the information is considered to have small changes depending on the state change of the state estimation target, it may be collected in any state and added to the input of the training data or the corresponding data.

[0026] Next, a deviation calculation learning device creation step S103 is executed to create a deviation calculation learning device that calculates the deviation from the reference state A. In Example 2, in deviation calculation learning machine creation step S103, the autoencoder is trained using the corresponding data collected for drill DR1 and drill DR2, which has a different tool diameter from drill DR1, in a state immediately after starting use (reference state A) as training data. In the second embodiment, the deviation from the reference state A is calculated by using the average of the absolute values ​​of the difference between the output when the corresponding data is input to the autoencoder and the input corresponding data. create In step S103, the autoencoder is trained to calculate the average of the absolute values ​​of the difference between the output of training data corresponding to the input of training data and the input of training data, based on the input of training data. In another embodiment, the deviation from the reference state A may be replaced with another value determined based on the error between the input and output of the autoencoder, such as the average of the squares of the difference between the output when certain corresponding data is input to the autoencoder and the input corresponding data. In deviation calculation step S104, a plurality of correspondence data acquired from the state immediately after the start of use of the drills DR1 and DR2 (reference state A) to the state when the drills DR1 and DR2 should be replaced due to deterioration (reference state B) are input to the autoencoder that has been trained in deviation calculation learning machine creation step S103, and deviations from the reference state A are calculated. At this time, the calculated deviations of the plurality of correspondence data from the reference state A are based on the correspondence data belonging to each of the plurality of correspondence data groups, each of which includes at least a first correspondence data group consisting of a plurality of correspondence data when the drills DR1 and DR2 are in the reference state A and a second correspondence data group consisting of a plurality of correspondence data when the drills DR1 and DR2 are in the reference state B. In addition, when a plurality of deviations from the reference state A calculated by inputting a plurality of correspondence data belonging to a certain correspondence data group are collectively handled, they are referred to as a deviation group from the reference state A.

[0027] In representative deviation calculation step S105, the deviations from reference state A calculated in deviation calculation step S104 are subjected to the same post-processing as that performed on the output group of the supervised learning device described above, to calculate a representative deviation from reference state A. For each group of deviations from the reference state A calculated from each of the multiple corresponding data groups, a process is performed in which the maximum deviation from the reference state A in the group of deviations from the reference state A is set as a representative deviation from the reference state A, and a representative deviation from the reference state A for each of the multiple corresponding data groups is calculated. Moreover, the states of the drills DR1 and DR2 either deteriorate or do not change, and do not return from a deteriorated state to a non-deteriorated state. In other words, the states of the drills DR1 and DR2 should change in one direction from the reference state A to the reference state B. Therefore, if the representative deviation from the reference state A obtained in the representative deviation calculation step S105 does not increase monotonically, it is presumed that the influence of a phenomenon unrelated to the deterioration of the states of the drills DR1 and DR2, such as measurement variation, is included. If it is presumed that a component unrelated to the deterioration of the states of the drills DR1 and DR2 is included, it is possible to remove the component unrelated to the deterioration of the states of the drills DR1 and DR2 by applying a moving average filter or a low-pass filter to the representative deviation from the reference state A in a state where the representative deviation from the reference state A is arranged in chronological order and converted into a smooth representative deviation from the reference state A. If the representative deviation from the reference state A has been subjected to filtering, the smooth representative deviation from the reference state A is used in the subsequent steps instead of the representative deviation from the reference state A. In this manner, in the representative deviation calculation step S105, the representative deviation from the reference state A or the representative deviation from the smooth reference state A is calculated.

[0028] In normalized mapping determination step S106 executed next, a mapping (normalized mapping) is determined for mapping the representative deviations from reference state A of each of the multiple corresponding data groups obtained from the states of drills DR1 and DR2 calculated in representative deviation calculation step S105 or the representative deviations from the smooth reference state A as normalized representative deviations in normalized representative deviation calculation step S107. Note that in the following explanation of Example 2, for the purpose of facilitating understanding, the representative deviations from reference state A and the representative deviations from the smooth reference state A are both represented as the representative deviations from reference state A. However, as described above, in the steps described below, when using a representative deviation from unfiltered reference state A, only the representative deviation from unfiltered reference state A is used, and when using a representative deviation from smooth reference state A, only the representative deviation from smooth reference state A is used; the representative deviation from reference state A and the representative deviation from smooth reference state A are not used in combination. Here, the representative deviation from the reference state A in the case of the reference state A is the representative deviation V A * The representative deviation from the reference state A when the reference state B is the representative deviation V B * The standard value that the representative value should take when the reference condition is A is the standard value W A * The standard value that the representative value should take when the reference condition is B is the standard value W B * Then, if the mapping satisfies all of the following three characteristics, the representative deviation V from the reference state A for each of the multiple corresponding data groups is * The normalized representative deviation W * The first characteristic is that the representative deviation V A * The specified value W A * The second property is that it is possible to map the representative deviation V B * The specified value W B * The third property is that it is possible to map the representative deviation V A* and the representative deviation V B * A map that satisfies the above three properties is called a representative deviation V from the reference state A. * The normalized representative deviation W * is determined as a mapping onto Next, the representative deviation V from the reference state A * The normalized representative deviation W * Using the mapping to the reference state A, the representative deviation V * The normalized representative deviation W * Copy it to.

[0029] Furthermore, in the normalized individual deviation calculation step S108, the normalized representative deviation V * and the normalized representative deviation V * The deviation group GP from the reference state A corresponding to * Calculate. Here, since the post-processing is a predetermined process for determining representative data from a set of predetermined data, the process having the reverse function to the post-processing refers to a predetermined process for deriving a set of predetermined data, which is representative of a certain data. Therefore, the process having the reverse function to the post-processing executed in the normalized individual deviation calculation step S108 in the second embodiment is the normalized individual deviation U * After post-processing, the normalized representative deviation W * In the second embodiment, the following formulas 3 and 4 are considered as suitable functions to be used in the process that has the opposite effect to the post-processing. The normalized representative deviation W corresponding to the machining of a hole using drill DR1 or DR2 * , deviation group GP from standard state A = {deviation from standard state A U1, deviation from standard state A U2, ..., deviation from standard state A U M}, Normalized Individual Deviation Group GP * ={Normalized individual deviation degree U1 * , normalized individual deviation degree U2 *, …, normalized individual deviation U M *} for (U and U * (The subscripts in indicate the number of the cutout section.) By applying functions such as Equation 3 and Equation 4, the normalized individual deviation U k * In other words, the U calculated in Equation 3 and Equation 4 can be calculated. k * is the normalized individual deviation group GP corresponding to the machining of a certain hole. * By post-processing, the normalized representative deviation W * It becomes. In another embodiment, when the post-processing involves taking an average, a function such as the following formula 5 or 6 may be used.

[0030]

number

number

number

number

[0031] Then, in the label determination step S109, the normalized individual deviation U k * will be adopted. The creation of a supervised learning machine is the same as in Example 1, where the normalized individual deviation U k * A deep neural network, support vector machine, etc. can be regressed to output the training data input and the training data output (normalized individual deviation U k *) as training data.

[0032] In this way, the output of the training data determined by the labeling method of Examples 1 and 2 above can be given a normalization effect to a supervised learning device that has learned using training data including the output determined by this method, or a supervised learning device with post-processing function, so as to output a common value in reference state A and reference state B regardless of the state estimation target. Furthermore, since the output of the training data determined by this method reflects the timing of changes in the deviation from reference state A even when the state of the state estimation target is intermediate between reference state A and reference state B, a supervised learning device that has learned using training data including the output determined by this method, or a supervised learning device with a post-processing function, can be endowed with a state estimation capability that is more in line with the actual state. In other words, it becomes possible to label a data set that is more in line with the actual state at low cost, even for an intermediate state between two reference states.

[0033] It is clear that a supervised learning machine trained with training data labeled according to the present invention is within the scope of the present invention, but a method of training another supervised learning machine using the output of this supervised learning machine itself or the output after appropriate processing as the output of the training data is also within the scope of the right, since the present invention is used in the process of creating another supervised learning machine. Even if such processing is repeated a finite number of times, the present invention is used in the process and is within the scope of the right. Each step may be performed across multiple computers or on a single computer. This labeling method may be executed multiple times to obtain multiple training data outputs corresponding to the same training data input, and the labels may be calculated by combining the multiple training data outputs using arithmetic operations, functions, etc. For example, it is also possible to calculate the training data output calculated when reference state A is the state immediately after the start of use and reference state B is the state in which the product should be replaced with a new one due to deterioration, and the training data output when reference state A is the state in which the product should be replaced with a new one due to deterioration and reference state B is the state immediately after the start of use, and to use the difference between these training data outputs as the training data output of the supervised learning device.

[0034] In the above-mentioned first and second embodiments, in the deviation calculation step, the deviation from the reference state A calculated by the learning device that calculates the deviation from the reference state A is used as is. On the other hand, it is also possible to create a separate learning device that calculates the deviation from the reference state B, and to regard the difference between the deviation from the reference state A calculated by the learning device that calculates the deviation from the reference state A and the deviation from the reference state B calculated by the learning device that calculates the deviation from the reference state B as the deviation from the reference state A. In this way, if the deviation from the reference state A can be obtained as a result, it is also possible to use multiple learning devices and combine the outputs of each learning device using arithmetic operations, functions, etc., and use them as the deviation from the reference state A. In the above Examples 1 and 2, embodiments in which data of two state estimation targets are learned are shown, but as long as two or more state estimation targets are learned, it is possible to learn data of any number of state estimation targets.

[0035] It is not necessary to use all of the training data outputs determined according to this labeling method. For example, if drill DR1 can drill 1000 holes before it needs to be replaced, and drill DR2 can drill 2000 holes before it needs to be replaced, learning all of the data would result in a high proportion of drill DR2, so it is possible to learn the training data for drill DR2 (training data input and training data output) with a 50 / 50 probability. In addition, in the above-mentioned first and second embodiments, the data called training data of the supervised learning device can be used not only as training data but also as evaluation data when evaluating the performance of the supervised learning device. In other words, it is possible to compare the output calculated by inputting the training data to a trained supervised learning device and the output of the training data, and if there is a good match, it is possible to determine that the performance of the supervised learning device is good.

[0036] The output of the training data determined by the present invention can provide a normalization function to a supervised learning device that has learned using training data including the output determined by the present method, or a supervised learning device with a post-processing function, so that the supervised learning device outputs a common value in the reference state A and the reference state B regardless of the state estimation target. In addition, the output of the training data determined by the present invention reflects the timing of the change in the deviation from the reference state A even when the state of the state estimation target is intermediate between the reference state A and the reference state B. Therefore, a state estimation ability that is more in line with the actual state can be provided to a supervised learning device that has learned using training data including the output determined by the present method, or a supervised learning device with a post-processing function. In other words, it is possible to label a data set that is more in line with the actual state at low cost, even for an intermediate state between two reference states. Furthermore, if a supervised learner, or a supervised learner with post-processing functionality, is trained using training data for a large number of state estimation targets under different conditions to output common values ​​in reference state A and reference state B regardless of the state estimation target, it is expected that it will acquire the property (generalization) of outputting common values ​​in reference state A and reference state B regardless of the state estimation target, even for state estimation targets under different conditions that have not yet been learned.

Claims

1. A method for labeling a data set of a supervised learning device for determining an output of training data corresponding to an input of training data in training data for creating a supervised learning device used for estimating in which state a state estimation target is in between a case where the state estimation target is in a predetermined reference state A and a case where the state estimation target is in a predetermined reference state B different from the reference state A, the method comprising: the training data includes at least any one of a first physical quantity acquired when the state estimation object is in a predetermined state and data related to a state of the state estimation object obtained based on the first physical quantity; and Any data including at least one of a second physical quantity acquired when the state of the state estimation target is the same as the state when the first physical quantity related to the input of the training data is acquired, and data related to the state of the state estimation target acquired based on the second physical quantity is referred to as corresponding data, For two or more state estimation objects, a deviation degree from a reference state A is calculated for each of a plurality of corresponding data including at least corresponding data when the state estimation object is in the reference state A and corresponding data when the state estimation object is in the reference state B; As a mapping for mapping the deviation degree from the reference state A of each of the plurality of correspondence data to a normalized deviation degree, a deviation from the reference state A when the state is the reference state A is mapped to a specified value to be output by the supervised learning device when the state is the reference state A, and a deviation from the reference state A when the state is the reference state B is mapped to a specified value to be output by the supervised learning device when the state is the reference state B, and the mapping is a monotonic mapping in an interval between the deviation from the reference state A when the state is the reference state A and the deviation from the reference state A when the state is the reference state B, Using the set mapping, an image of deviation from the reference state A is mapped for each of the plurality of corresponding data, thereby calculating the normalized deviation; determining the normalized deviance as an output of the training data corresponding to an input of the training data.

2. When the state of the state estimation target transitions in a unidirectional manner from one of the reference state A and the reference state B to the other in a chronological order, For each of the plurality of state estimation targets, a deviation degree from the reference state A of each of the plurality of corresponding data is calculated, and then the calculated deviation degrees from the reference state A are arranged in chronological order and filtered to convert them into a smooth deviation degree from the reference state A; 2. The method for labeling a dataset for a supervised learning machine according to claim 1, wherein when setting the mapping for mapping the deviation from the reference state A of each of the plurality of corresponding data to the normalized deviation and calculating the normalized deviation, the deviation from the smooth reference state A is used instead of the deviation from the reference state A.

3. A method for labeling a data set of a supervised learning device for determining an output of training data corresponding to an input of training data in training data for creating a supervised learning device used for estimating in which state a state estimation target is in between a case where the state estimation target is in a predetermined reference state A and a case where the state estimation target is in a predetermined reference state B different from the reference state A, the method comprising: the supervised learning device has a function of performing post-processing to calculate a representative value in an output group, which is a set of outputs calculated from each of the inputs, when a plurality of predetermined data belonging to a data group, which is a set of a plurality of data, are input; the training data includes at least any one of a first physical quantity acquired when the state estimation object is in a predetermined state and data related to a state of the state estimation object obtained based on the first physical quantity; and Any data including at least one of a second physical quantity acquired when the state of the state estimation target is the same as the state when the first physical quantity related to the input of the training data is acquired, and data related to the state of the state estimation target acquired based on the second physical quantity is referred to as corresponding data, For two or more state estimation objects, a deviation degree of the corresponding data belonging to each of the corresponding data groups from the reference state A is calculated for each of the corresponding data groups, the corresponding data groups including at least a first corresponding data group consisting of a plurality of corresponding data when the state estimation object is in the reference state A and a second corresponding data group consisting of a plurality of corresponding data when the state estimation object is in the reference state B; A set of deviations from the reference state A corresponding to the plurality of corresponding data groups to which the corresponding data corresponding to each of the calculated deviations from the reference state A belongs is defined as a deviation group from the reference state A; calculating a representative deviation from the reference state A in each of the plurality of deviation groups from the reference state A by applying the same processing as the post-processing executed by the supervised learning device to each of the plurality of deviation groups from the reference state A; As a mapping for mapping the representative deviation from the reference state A corresponding to each of the plurality of corresponding data groups to a normalized representative deviation, a representative deviation from the reference state A when the reference state is A is mapped to a prescribed value that a representative value calculated by the supervised learning device should take when the reference state is A, and a representative deviation from the reference state A when the reference state is B is mapped to a prescribed value that a representative value calculated by the supervised learning device should take when the reference state is B, and the mapping is a monotonic mapping in an interval between the representative deviation from the reference state A when the reference state is A and the representative deviation from the reference state A when the reference state is B, Using the set mapping, an image of a representative deviation from the reference state A is mapped for each of the plurality of corresponding data groups to calculate the normalized representative deviation; A process that is the reverse of the post-processing is performed on each of the plurality of normalized representative deviations and each of the deviation groups from the reference state A corresponding to each of the plurality of normalized representative deviations, and a normalized individual deviation is calculated for each of the plurality of corresponding data groups corresponding to the plurality of normalized representative deviations such that the normalized representative deviation is calculated when the post-processing is performed; determining the normalized individual discrepancy as an output of the training data corresponding to an input of the training data.

4. When the state of the state estimation target transitions in a unidirectional manner from one of the reference state A and the reference state B to the other in a chronological order, For each of the plurality of state estimation targets, a representative deviation from the reference state A of each of the plurality of corresponding data groups is calculated, and then the calculated representative deviation from the reference state A is arranged in chronological order and filtered to convert the calculated representative deviation from the reference state A into a smooth representative deviation from the reference state A; The method for labeling a dataset for a supervised learning machine according to claim 3, characterized in that when setting the mapping for mapping the representative deviation from the reference state A of each of the plurality of corresponding data groups to the normalized representative deviation and calculating the normalized representative deviation, the representative deviation from the smooth reference state A is used instead of the representative deviation from the reference state A.

5. A supervised learner, which is created using the method for labeling a data set for a supervised learner according to claim 1 or 2.

6. A state estimation device comprising the supervised learning device according to claim 5.

7. The state estimation device according to claim 6, characterized in that the output of the supervised learning device, a specified value that the supervised learning device should output when the state estimation object is in the reference state A, and a specified value that the supervised learning device should output when the state estimation object is in the reference state B are simultaneously displayed as numerical values ​​or graphs.

8. A supervised learner, which is created using the method for labeling a data set for a supervised learner according to claim 3 or 4.

9. A state estimation device comprising the supervised learning device according to claim 8.

10. The state estimation device according to claim 9, characterized in that a representative value in the output group output by the supervised learning device, a specified value that the representative value calculated by the supervised learning device should take when the state estimation object is in the reference state A, and a specified value that the representative value calculated by the supervised learning device should take when the state estimation object is in the reference state B are simultaneously displayed as numerical values ​​or graphs.

Images