Labeling method of dataset of teacher learner, teacher learner, state estimation device

Abstract

Provided are a label annotation method for a dataset of a teacher learner, the teacher learner, and a state estimation device. For a plurality of corresponding data including a plurality of state estimation objects in a reference state alpha and a reference state beta, a deviation from alpha of each of the corresponding data is calculated so as to be able to depict the deviation from alpha in the case of alpha as a prescribed value that the teacher learner should output in the case of alpha, depict the deviation from alpha in the case of beta as a prescribed value that the learner should output in the case of beta, and the interval of the deviation from alpha in the case of alpha and the deviation from alpha in the case of beta is monotonically mapped. A mapping is determined for each state estimation object in a manner in which the mapping, by which the deviation from alpha of each of the corresponding data is obtained, is monotonically mapped within the interval of the deviation from alpha in the case of alpha and the deviation from alpha in the case of beta. A normalized deviation is calculated, and the normalized deviation is determined as output data of training data corresponding to input of the training data.

Description

Technical Field

[0001] The present invention relates to a method for labeling a dataset with a teacher learner in a machine learning model used in machine tools and for diagnosing damage states of diagnostic tools, a teacher learner generated using the method, and a state estimation apparatus equipped with the learner. Background Technology

[0002] In machine tools, when bearings supporting rotating parts are used in a damaged state, machining accuracy deteriorates, or the bearings may stick together, rendering the machine inoperable. Furthermore, when cutting tools are used in a state of significant wear, machining accuracy deteriorates, or the tools may continue to operate even if broken, causing interference and damaging the machine tool.

[0003] In the past, to avoid these failures, management was sometimes based on preset time or number of cycles for bearings, such as cumulative rotation time, cumulative revolutions, and tool usage time. However, this approach failed to address actual conditions, resulting in unnecessary expenses when continued use was not required, or failures occurred even though the predetermined time or number of cycles were not reached.

[0004] Therefore, the following method was studied: instead of estimating time or number of times, we estimate the actual state, capture the state change from normal to abnormal state, and take action before it becomes completely abnormal.

[0005] Patent document 1 shows the following apparatus: obtaining processing information in a state where the tool life is sufficiently remaining, generating input data, constructing a learning model obtained by generating clusters through teacherless learning, and thereby inferring from a certain input data, and judging that it is in a state where the tool life is not remaining if it does not belong to a cluster of states where the tool life is sufficiently remaining.

[0006] Patent document 2 shows a machine tool having: an anomaly score calculation unit that takes data generated during normal processing as input to an unattended learner, and calculates the extent to which the data during actual processing deviates from the data during normal processing, and calculates an anomaly score based on the difference between the output and the input; and a unit that allocates the anomaly score according to a finite number of damage levels, thereby enabling responses corresponding to the damage levels.

[0007] Patent document 3 shows a tool state estimation device with a storage unit having the following learning model: the learning model classifies the dynamic information near the tool at a time when the tool is in good condition and other information based on input data such as dynamic information near the tool, such as chip scattering distance, chip scattering speed, and the center angle of the chip scattering range, thereby estimating the tool state.

[0008] Patent Document 1: Japanese Patent Application Publication No. 2018-103284

[0009] Patent Document 2: Japanese Patent Application Publication No. 2018-25936

[0010] Patent Document 3: Japanese Patent Application Publication No. 2018-138327

[0011] However, the device in Patent Document 1 uses a learning model obtained solely from processing information obtained under conditions where the tool life is sufficiently remaining. Therefore, for processing information under conditions where the learning model has not been learned (i.e., when the tool life is exhausted), it is impossible to predict what the inference result based on the learning model would be. Consequently, it is difficult to determine the threshold for the device to determine whether a state with exhausted tool life exists. Furthermore, even if the inference result of the learning model for processing information under conditions where the tool life is exhausted is inappropriate, it is impossible to improve the learning model to achieve the desired inference result. Therefore, there is a problem that the state of various state estimation objects cannot be determined using a universal benchmark.

[0012] In the machine tool described in Patent Document 2, actual machining data is input into an unattended learning device generated using only data from normal machining. The deviation of the actual machining data from the normal machining data is then calculated as an anomaly score. However, it is impossible to predict the magnitude of the anomaly scores for abnormal machining data with varying machining conditions such as tool type, material being cut, and cutting speed. Furthermore, the anomaly score calculation unit cannot be improved by calculating the desired anomaly score for abnormal machining data. Therefore, when judging the state of various state estimation objects using a fixed benchmark, it is necessary to pre-prepare anomaly scores for a finite number of damage levels according to each machining condition. This results in an inability to cope with new machining conditions.

[0013] In the tool state estimation apparatus of Patent Document 3, when using unsupervised learning to generate clusters of time points indicating good tool state, it is impossible to predict the extent to which the inference results of the machine learning model for different processing conditions (such as tool, material being cut, and cutting speed) deviate from the clusters of time points indicating good tool state relative to the input data under different processing conditions. Therefore, there is a problem that the state of various state estimation objects cannot be determined using a fixed benchmark. A method is shown below: when using supervised learning, each signal of the output layer is associated with "good tool state" and "deteriorating tool state," and the teacher data is determined by making any one signal equal to 1 corresponding to the teacher data. In this supervised learning, it is possible to determine the state of various state estimation objects using a fixed benchmark.

[0014] However, in teacher-led learning, it is very difficult to properly label intermediate states between two baseline states, such as "tool condition is good" and "tool condition is deteriorating." Furthermore, without labeling intermediate states, it is impossible to add intermediate state data to the training data. Therefore, without obtaining the desired inference results for the intermediate states, the learning model cannot be improved. Patent document 3 also shows another example of determining teacher data according to a time period before the tool's condition deteriorates, in order to label intermediate states separately; however, it is difficult to say that this method provides appropriate labeling corresponding to the degree of tool deterioration.

[0015] Figure 1 This is a conceptual diagram used to illustrate a problem where time is referenced in the labeling of intermediate states. Here, tool usage time is set on the horizontal axis, and tool wear amount, representing the state to be estimated, is set on the vertical axis. It shows the progression of tool wear under three conditions: machining conditions where tool wear deteriorates slowly, machining conditions where tool wear deteriorates rapidly, and machining conditions where the machining process is changed from slow to rapid tool wear. Figure 1 In the diagram, solid lines represent the changes in tool wear under machining conditions where tool wear deteriorates slowly, while dashed lines represent the changes in tool wear under machining conditions where tool wear deteriorates drastically. Additionally, Figure 1 This is a conceptual diagram; in reality, tool wear does not necessarily increase proportionally to tool usage time. For example... Figure 1As shown, in the three cases described above, the difference in cumulative processing time from the same tool wear level to the point where the tool condition deteriorates is different for each of them. That is, in the three cases, for the same tool wear level that should have been labeled with the same tag, the cumulative processing time from the same tool wear level to the point where the tool condition deteriorates is different. Therefore, different tags are used to indicate different time periods before the specified time when the tool condition deteriorates. Since the learner cannot learn by producing different outputs for the same input, it cannot learn properly when different tags are used for the same state. In other words, when time is referenced in the labeling of intermediate states, there is a problem that the learner cannot learn properly.

[0016] On the other hand, to accurately label intermediate states, it's also considered to check the tool status each time during processing intervals to track changes in status. However, in this case, checking the tool status requires a significant amount of time. As a result, time unrelated to processing increases, productivity decreases, and therefore, there is a problem of labeling incurring costs. Summary of the Invention

[0017] Therefore, the present invention was made in view of the above problems, and its object is to provide a method for labeling a dataset with a teacher learner, and a teacher learner and a state estimation device, wherein in labeling a dataset that determines the output of training data corresponding to the input of training data of a teacher learner used to judge the state of various state estimation objects with a fixed benchmark, labeling that is more in line with the actual state can be performed at low cost for the intermediate state between two benchmark states.

[0018] To achieve the above objective, the invention described in the first aspect is a labeling method for a dataset with a teacher learner, which is used to determine the output data of training data corresponding to the input data of the training data in relation to the generation of the teacher learner, wherein the teacher learner is used to estimate the state of a state estimation object between a predetermined reference state A and a predetermined reference state B different from reference state A. The method is characterized in that arbitrary data selected from a first physical quantity and data associated with the state of the state estimation object obtained based on the first physical quantity are designated as input data of the training data, wherein the first physical quantity is obtained when the state estimation object is in the predetermined state; arbitrary data selected from a second physical quantity and data associated with the state of the state estimation object obtained based on the second physical quantity are designated as corresponding data, wherein the second physical quantity is obtained when the state of the state estimation object is in the same state as when the input data of the training data is obtained; for two or more state estimation objects, for multiple corresponding data, a value relative to reference state A is calculated separately. The deviation, comprising multiple corresponding data, includes at least corresponding data for each state estimation object in the case of reference state A and corresponding data for each state estimation object in the case of reference state B. As a mapping used to depict the deviation of each of the multiple corresponding data relative to reference state A as a normalized deviation, it can depict the deviation of the state estimation object relative to reference state A as a predetermined value that the teacher learner should output in the case of reference state A, and can depict the deviation of the state estimation object relative to reference state A in the case of reference state B as a predetermined value that the teacher learner should output in the case of reference state B. This mapping is monotonic within the interval of the deviation of the state estimation object relative to reference state A in the case of reference state A and the deviation of the state estimation object relative to reference state A in the case of reference state B. A mapping is set for each state estimation object. Using the set mapping, images of the deviation relative to reference state A are depicted for each of the multiple corresponding data, thereby calculating the normalized deviation. The normalized deviation is determined as the output data of the training data corresponding to the input data of the training data.

[0019] The invention described in the second aspect is characterized in that, in the above structure, when the state of the state estimation object changes unidirectionally from either reference state A or reference state B in a temporal sequence, for multiple state estimation objects, after calculating the deviation of each of the multiple corresponding data relative to reference state A, filtering is performed on the state in which the calculated deviations relative to reference state A are arranged in a temporal sequence to convert them into smooth deviations relative to reference state A. When setting the mapping and calculating the normalized deviation in order to describe the deviations of the multiple corresponding data relative to reference state A as normalized deviations, the smooth deviations relative to reference state A are used instead of the deviations relative to reference state A.

[0020] To achieve the above objectives, the invention described in the third aspect is a labeling method for a dataset with a teacher learner, which is used to determine the output data of training data corresponding to the input data of the training data in relation to the generation of the teacher learner, wherein the teacher learner is used to estimate the state of a state estimation object between a predetermined reference state A and a predetermined reference state B different from reference state A, characterized in that the teacher learner has a function of performing post-processing, in which, when multiple predetermined data belonging to a data group as a set of multiple data are input, representative values ​​in the output data group, which is a set of output data calculated based on each input data, are calculated from the first physical The input data for training data is any data selected from the first physical quantity and the data associated with the state of the state estimation object obtained based on the first physical quantity, where the first physical quantity is obtained under the condition that the state estimation object is in a specified state. The input data for training data is any data selected from the second physical quantity and the data associated with the state of the state estimation object obtained based on the second physical quantity, where the second physical quantity is obtained under the condition that the state estimation object is in the same state as when the first physical quantity involved in obtaining the input data for training data is obtained. For two or more state estimation objects, for multiple corresponding data groups, the deviation of the corresponding data belonging to each of the multiple corresponding data groups relative to the baseline state A is calculated separately. The multiple corresponding data groups must have at least... The dataset includes a first corresponding data group and a second corresponding data group. The first corresponding data group consists of multiple corresponding data sets for each state estimation object when it is in reference state A, and the second corresponding data group consists of multiple corresponding data sets for each state estimation object when it is in reference state B. The set of deviations relative to reference state A corresponding to the multiple corresponding data groups to which the calculated deviations relative to reference state A belong is defined as a deviation group relative to reference state A. For each of the multiple deviation groups relative to reference state A, the same post-processing as performed by the teacher-assisted learner is applied to calculate the representative deviation relative to reference state A for each of the multiple deviation groups relative to reference state A, which is used to compare the deviations relative to reference state A with the target state B. Multiple corresponding data sets are respectively depicted as normalized representative deviations relative to baseline state A, and the representative deviations relative to baseline state A are depicted as the specified values ​​that the teacher learner should take when the representative value is in baseline state A. Similarly, the representative deviations relative to baseline state A in baseline state B are depicted as the specified values ​​that the teacher learner should take when the representative value is in baseline state B. Furthermore, within the intervals of the representative deviations relative to baseline state A and B, this mapping is monotonic. This mapping is set according to each state estimation object.Using a defined mapping, for multiple corresponding data sets, images of representative deviations relative to the baseline state A are drawn, thereby calculating normalized representative deviations. For each of the multiple normalized representative deviations and the corresponding deviation sets relative to the baseline state A, processing opposite to post-processing is performed. For each of the multiple corresponding data sets with multiple normalized representative deviations, normalized individual deviations are calculated. After post-processing the normalized individual deviations, normalized representative deviations are calculated. These normalized individual deviations are then used as the output data of the training data corresponding to the input data of the training data.

[0021] The invention described in aspect 4 is characterized in that, in the above structure, when the state of the state estimation object changes unidirectionally from either reference state A or reference state B in a temporal sequence, for multiple state estimation objects, after calculating the representative deviation of each of multiple corresponding data groups relative to reference state A, filtering is performed on the state in which the calculated representative deviations relative to reference state A are arranged in a temporal sequence to convert them into smooth representative deviations relative to reference state A. When setting the mapping and calculating the normalized representative deviations in order to depict the representative deviations of each of the multiple corresponding data groups relative to reference state A as normalized representative deviations, the smooth representative deviations relative to reference state A are used instead of the representative deviations relative to reference state A.

[0022] The invention described in the fifth aspect is a teacher-aided learner, characterized in that the teacher-aided learner is generated using a labeling method for a dataset of teacher-aided learners described in the first or second aspect.

[0023] The invention described in the sixth aspect is a state estimation device, characterized in that the state estimation device is equipped with the teacher learning device described in the fifth aspect.

[0024] The invention described in aspect 7 is characterized in that, in the above structure, the output of the teacher learner, the predetermined value that the teacher learner should output when the state estimation object is in baseline state A, and the predetermined value that the teacher learner should output when the state estimation object is in baseline state B are simultaneously displayed as numerical values ​​or graphs.

[0025] The invention described in aspect 8 is a teacher-aided learner, characterized in that the teacher-aided learner is generated using the labeling method of the dataset of the teacher-aided learner described in aspect 3 or aspect 4.

[0026] The invention described in aspect 9 is a state estimation device, characterized in that the state estimation device is equipped with the teacher learning device described in aspect 8.

[0027] The invention described in aspect 10 is characterized in that, in the above structure, the representative value in the output data set output by the teacher learner, the predetermined value that the representative value calculated by the teacher learner should take when the state estimation object is in baseline state A, and the predetermined value that the representative value calculated by the teacher learner should take when the state estimation object is in baseline state B are simultaneously displayed as numerical values ​​or graphs.

[0028] Invention Effects

[0029] According to the present invention, the output data of the training data determined by the present invention can be normalized to a teacher learner or a teacher learner with post-processing function that has learned using the training data containing the output data determined by the present method, so as to output common values ​​in reference state A and reference state B, independent of the state estimation object.

[0030] Furthermore, when the state of the object being estimated is an intermediate state between baseline state A and baseline state B, the output data of the training data determined by this invention also reflects the moment of change in deviation from baseline state A. Therefore, a more realistic state estimation capability can be given to a teachered learner or a teachered learner with post-processing capabilities that has learned from training data containing the output data determined by this method. That is, for the intermediate state between two baseline states, it is also possible to perform labeling of a dataset that more closely resembles the actual state at a low cost. Attached Figure Description

[0031] Figure 1 This is a diagram showing the topic of reference time in the labeling of intermediate states.

[0032] Figure 2 This is a flowchart of the labeling method in Example 1.

[0033] Figure 3 This is a graph showing the deviation from the baseline state A.

[0034] Figure 4 This is a graph showing the deviation from the reference state A under the condition of measurement bias.

[0035] Figure 5 This is a graph showing the normalized deviation.

[0036] Figure 6 This is the screen displayed by the state estimation device.

[0037] Figure 7 This is the screen displayed by the state estimation device.

[0038] Figure 8 This is a flowchart of the labeling method in Example 2. Detailed Implementation

[0039] The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0040] First, Example 1 will be described.

[0041] Figure 2 A flowchart illustrating a method for labeling datasets with teacher learners is provided.

[0042] In the following description relating to Example 1, the bearing supporting the spindle in the machine tool is defined as the state estimation object. Furthermore, when more than two state estimation objects are involved, each state estimation object becomes a different individual bearing. Additionally, the bearing's initial state after use is defined as reference state A. And, the state where the bearing should be replaced with a new one due to deterioration is defined as reference state B.

[0043] In data collection step S1, the data required for labeling the dataset of the teacher learner is collected. In Example 1, the following actions are performed in advance, from the initial state of the bearing (reference state A) to the state where the bearing should be replaced due to deterioration (reference state B): whenever the machine tool is used for a specified time, the spindle is rotated at a specified speed, and signals are collected for physical quantities such as the spindle motor torque, the vibration of the acceleration sensor installed near the bearing, and the temperature of the temperature sensor installed near the bearing. The collected physical quantities are recorded as data.

[0044] After the data collection step S1, the data generation step S2 is executed, in which the input data and corresponding data of the training data of the teacher learner are generated.

[0045] Training data is the data used to enable a learner to learn so that the learner produces the desired output in response to a given input. In this application, unless otherwise specified, training data is meant to include both input and output.

[0046] The input data for the training data of the teacher learner refers to any data selected from the physical quantity (the first physical quantity) obtained when the state estimation object is in a specified state and the data associated with the state of the state estimation object obtained from the first physical quantity.

[0047] Furthermore, corresponding data refers to any data selected from the physical quantity (second physical quantity) obtained when the state of the state estimation object is in the same state as when the first physical quantity was obtained, and the data associated with the state of the state estimation object obtained based on the second physical quantity. Here, "the same state" means that the state of the state estimation object at the moment the second physical quantity is obtained can be considered by those skilled in the art to be the same as the state of the state estimation object when the first physical quantity was obtained; cases where the states of the state estimation object are not completely identical are permissible. Additionally, in this description, corresponding data includes not only the input data of the training data but also the output data of the training data. Furthermore, the data associated with the state of the state estimation object obtained based on the first and second physical quantities refers to data generated separately by performing prescribed processing such as signal processing on the first and second physical quantities. However, the types of the first and second physical quantities, the signal processing methods, sampling frequencies, etc., do not need to be the same in the input data and corresponding data of the training data with a teacher learner. That is, for the first and second physical quantities, they can be obtained as long as the state of the object being estimated is the same; they can also be different physical quantities. These physical quantities can be used as input data and corresponding data for training data. Furthermore, based on the different types of physical quantities, i.e., the first and second physical quantities, different processing can be performed to generate different types of data associated with the state of the object being estimated, and these different types of data associated with the state of the object being estimated can be used as input data and corresponding data for training data. Moreover, the input data and corresponding data for training data can both use multiple different types of data.

[0048] Furthermore, information considered to have relatively small changes depending on the state changes of the state-estimated object is collected and stored in any state, and can also be added to the input and corresponding data of the training data of the teacher learner. Examples of information considered to have relatively small changes depending on the state changes of the state-estimated object include the spindle mass, the vibration susceptibility of the accelerometer mounted near the bearing, and the magnitude of the transfer function of the force generated by bearing damage at various frequencies to the vibration acceleration at the location of the accelerometer.

[0049] In the data generation step S2 of Example 1, the input data for the training data of the teacher learner includes the vibration amplitude at each frequency, the spindle mass, and the vibration ease of the accelerometer mounted near the bearing, obtained by frequency analysis of the vibration of the accelerometer. For example, inertia can be used as an example of the vibration ease of the accelerometer mounted near the bearing. Furthermore, the timing waveforms of the spindle motor torque, vibration, and temperature are used as corresponding data.

[0050] Next, step S3, which generates a deviation calculation learner, is executed. In step S3, a deviation calculation learner is generated to calculate the deviation relative to the reference state A.

[0051] Examples of learners that can be used to calculate the deviation from a reference state A include unsupervised learners that include autoencoders generated using corresponding data of states that are close to the reference state A as training data, learners that perform predictive learning, and VAEs (variable autoencoders).

[0052] An autoencoder is a deep neural network with the same number of input and output units. It learns by using the same training data for both input and output, minimizing the error between the input and output. Therefore, for data in the training data obtained when the state of the object being estimated is close to the reference state A, the error between the corresponding input data and the output data when that data is input to the autoencoder is calculated to be small. On the other hand, as the state of the object being estimated moves from the reference state A to the reference state B at the time the data is obtained, the error between the corresponding input data and the output data when that data is input to the autoencoder is calculated to gradually increase. Therefore, this error can be set as the deviation relative to the reference state A.

[0053] The learner performing predictive learning uses a deep neural network that learns by taking earlier time data from the training data as input and later time data from the training data as output. For data in the training data whose state, considered as the object of state estimation, is close to the baseline state A, the error between the output data when earlier time data is input and the corresponding later time data is calculated to be small. On the other hand, as the state of the object of state estimation at the time the data is obtained moves from the baseline state A to the baseline state B, the error between the output data when earlier time data is input into the learner performing predictive learning and the corresponding later time data is calculated to gradually increase. Therefore, this error can be defined as the deviation relative to the baseline state A.

[0054] A VAE is a form of autoencoder, specifically one that calculates the mean and variance in intermediate layers. In the case of a VAE, the property of not reproducing features not included in the training data becomes more pronounced. For data in the training data that is considered as the state of the state estimation object and is close to the baseline state A, the error between the input corresponding data and the output data when that corresponding data is input into the VAE is calculated to be small. On the other hand, as the state of the state estimation object at the time of data acquisition moves from the baseline state A to the baseline state B, the error between the input corresponding data and the output data when that corresponding data is input into the VAE is calculated to gradually increase. Therefore, this error can be set as the deviation relative to the baseline state A.

[0055] The illustrated learners all have the property of being able to calculate the degree of deviation relative to the reference state A by learning from data that is considered to be close to the reference state A, and can be replaced by learners with the same property.

[0056] The learner that calculates the deviation from the baseline state A can be generated separately for each state estimation object, or it can be created as a group of several state estimation objects, or only one can be generated for all state estimation objects.

[0057] In the deviation calculation learner generation step S3 of Example 1, the bearing BR1 that supports the spindle of machine type MA, which is a machine tool, and the bearing BR2 that supports the spindle of machine type MB, which has a different vibration susceptibility than machine type MA, will be used as training data for the automatic encoder that serves as the deviation calculation learner. The corresponding data collected at a time when it is considered to be close to the state after initial use (reference state A) will be used as training data.

[0058] In the next step, the deviation calculation step S4, for the autoencoder that has completed learning and was generated in the deviation calculation learner generation step S3, multiple corresponding data obtained during the period from the initial use state of bearings BR1 and BR2 (reference state A) to the state where they should be replaced with new bearings due to deterioration (reference state B) are input. For each input corresponding data, corresponding output data is calculated. Furthermore, based on the input corresponding data and the calculated output data, the error between the calculated output data and the input corresponding data when each corresponding data is input to the autoencoder is calculated as the deviation relative to reference state A. In Example 1, the deviation relative to reference state A is the average of the absolute values ​​of the differences between the output data when the corresponding data is input to the autoencoder and the input corresponding data.

[0059] Alternatively, in another approach, the deviation relative to reference state A can be replaced by another value that is considered as the error between the input and output data, determined by the output data and the input corresponding data when a certain corresponding data is input to the auto encoder. For example, the average of the squares of the difference between the output data and the input corresponding data when a certain corresponding data is input to the auto encoder can be used.

[0060] In the deviation calculation step S4, for the autoencoder that has completed learning in the deviation calculation learner generation step S3, multiple corresponding data obtained from the initial state of bearings BR1 and BR2 (reference state A) to the state where they should be replaced with new bearings due to deterioration (reference state B) are input respectively, and the deviation relative to reference state A is calculated. Figure 3 The calculated deviations of bearings BR1 and BR2 relative to reference state A are shown.

[0061] The state changes of bearings BR1 and BR2 are either deterioration or no change, and they will not return from a deteriorated state to a non-deteriorated state. That is, the state of bearings BR1 and BR2 should change unidirectionally from reference state A to reference state B. Therefore, the deviation relative to reference state A obtained in deviation calculation step S4 was not calculated. Figure 4 In the case of monotonically increasing deviations, it is speculated that this is due to the influence of phenomena unrelated to the deterioration of the condition of bearings BR1 and BR2, such as measurement deviations. If it is speculated that the deviation relative to reference state A contains components unrelated to the deterioration of the condition of bearings BR1 and BR2, then, with the deviation relative to reference state A arranged in time sequence, a moving average filter or low-pass filter is applied to the deviation relative to reference state A to convert it into a smooth deviation relative to reference state A. This removes components unrelated to the deterioration of the condition of bearings BR1 and BR2. In the case where the deviation relative to reference state A has been filtered, the smoothed deviation relative to reference state A is used instead of the original deviation in subsequent steps.

[0062] In this way, in the deviation calculation step S4, the deviation relative to the reference state A or the smooth deviation relative to the reference state A is calculated.

[0063] In the next normalization mapping determination step S5, a normalization mapping is determined. This mapping is used in the normalization deviation calculation step S6 to depict the deviation relative to the reference state A of each of the multiple corresponding data obtained in the deviation calculation step S4 based on the states of bearings BR1 and BR2, or the smooth deviation relative to the reference state A, as normalized deviations. Furthermore, in the following description related to Embodiment 1, for ease of understanding, the deviation relative to the reference state A and the smooth deviation relative to the reference state A are both represented as the deviation relative to the reference state A. However, as described above, in the later steps, when using the unfiltered deviation relative to the reference state A, only the unfiltered deviation relative to the reference state A is used; when using the smooth deviation relative to the reference state A, only the smooth deviation relative to the reference state A is used. The deviation relative to the reference state A and the smooth deviation relative to the reference state A are not used simultaneously.

[0064] Here, the deviation from the baseline state A is defined as deviation V. A Let the deviation from reference state A under reference state B be defined as deviation V. B Under the baseline state A, the specified value that the teacher learner should output will be set as the specified value W. A Under baseline state B, the specified value that the teacher learner should output will be set to the specified value W. B Let the positive exponent be n. Then, if the mapping satisfies all of the following three properties, it can be used to describe the deviation V of multiple corresponding data points relative to the baseline state A as a normalized deviation W. The first property is that it can describe the deviation V... A Described as a specified value W A The mapping. The second characteristic is that it can map the deviation V. B Described as a specified value W B The mapping. The third characteristic is that, in the deviation V A And deviation V B The interval is a monotonic mapping. As a mapping that satisfies these three properties, consider the following mathematical expression 1.

[0065] Furthermore, not limited to bearings, when determining the mapping according to each desired state estimation object, the specified value W will be used. A and the specified value W BBy setting the same value across all state estimation objects, the teacher-trained learner obtained using training data generated through the labeling method of this invention can include a normalization effect. This normalization effect refers to the following: when the state estimation object is in a state ranging from baseline state A to baseline state B, the output of the teacher-trained learner does not depend on the state estimation object, but becomes a value from the specified value W. A up to the specified value W B Such a common scope.

[0066] In V A <V B W A <W B When n=1, mathematical expression 1 becomes a narrowly monotonically increasing mapping. Deviation V A and deviation V B The values ​​for bearings BR1 and BR2 are used; therefore, mappings are determined separately for bearings BR1 and BR2. The specified value W... A and W B This is a common value for bearings BR1 and BR2. When using the mappings corresponding to bearings BR1 and BR2 respectively, the deviation V( ) of bearings BR1 and BR2 is... Figure 3 ) is described as the normalized deviation W( ) of bearings BR1 and BR2 respectively. Figure 5 ). Figure 5 The normalized deviation shown is the same for bearings BR1 and BR2. The normalized deviation under reference condition A becomes the specified value W. A The normalized deviation under the baseline state B becomes the specified value W. B .

[0067]

Mathematical Formula 1

[0068]

[0069] Furthermore, any mapping that satisfies the above three properties, such as the following mathematical formula 2 obtained by adding mathematical formula 1 at any ratio, can be used as a mapping that describes the deviation V relative to the reference state A as the normalized deviation W.

[0070]

Mathematical Formula 2

[0071]

[0072] In the case of a mapping generated without using mathematical formulas 1 and 2 (hereinafter referred to as a mapping candidate), if it is confirmed that all three characteristics described above are satisfied, it can be used as a mapping that depicts the deviation V relative to the reference state A as the normalized deviation W in the normalized deviation calculation step S5. The method for confirming whether the mapping candidate satisfies the above three characteristics will be explained below.

[0073] To confirm that the mapping candidate satisfies the first characteristic, it is only necessary to confirm that the deviation V is obtained through the mapping candidate. A Does the value obtained by solving the mapping have the specified value W that the teacher learner should output under the condition of baseline state A? A Consistency is sufficient. Alternatively, it is sufficient to confirm the deviation V of the candidate pairs mapped. A The value obtained by solving the mapping is relative to the specified value W. A As long as it falls within the specified error range, it is acceptable.

[0074] To confirm that the mapping candidate satisfies the second characteristic, it is only necessary to confirm that the deviation V is obtained through the mapping candidate. B Does the value obtained by solving the mapping correspond to the specified value W? B Consistency is sufficient. Alternatively, it is sufficient to confirm the deviation V of the candidate pairs mapped. B The value obtained by solving the mapping is relative to the specified value W. B As long as it falls within the specified error range, it is acceptable.

[0075] To confirm that the mapping candidate satisfies the third property, assuming that the mapping candidate is a differentiable function, it is sufficient to confirm that the first derivative of this function lies in the deviation V. A and deviation V B The value must always be either above or below 0 within the interval. Depending on the purpose of state estimation, if the value obtained by describing the deviation from the reference state A through the mapping candidate is allowed to be constant within a portion of the deviation from the reference state A, it is sufficient to confirm whether the mapping candidate is a generalized monotonically increasing or decreasing value where the first derivative of the mapping candidate is 0. If the value obtained by describing the deviation from the reference state A through the mapping candidate is not allowed to be constant within a portion of the deviation from the reference state A, it is sufficient to confirm whether the mapping candidate is a narrowly monotonically increasing or narrowly monotonically decreasing value where the first derivative of the mapping candidate is 0.

[0076] On the other hand, when the mapping candidate is not a differentiable function, from the deviation V A To deviation degree V BUntil, for example, the deviation V relative to the reference state A is adjusted according to the deviation V each time. A To deviation degree V B The values ​​gradually change in increments of 1 / 10,000 of the interval. The mapped values ​​relative to the baseline state A are calculated using the mapping candidate. It is sufficient to confirm that the mapped value either always increases or always decreases before and after the change in the deviation V relative to the baseline state A. Alternatively, if the mapped value is allowed to remain unchanged, it is sufficient to confirm that it either never decreases or never increases.

[0077] As described above, in the normalization mapping determination step S5, a mapping is determined for each state estimation object to depict the deviation V relative to the reference state A as the normalized deviation W.

[0078] Next, in the normalized deviation calculation step S6, the mapping determined in the normalized mapping determination step S5, which describes the deviation V relative to the reference state A as the normalized deviation W, is used to describe the deviation V relative to the reference state A as the normalized deviation W.

[0079] Then, in the label determination step S7, the normalized deviation W calculated using the input data of the training data and the corresponding data obtained when the state of the corresponding state estimation object is the same is set as the output data of the training data corresponding to the input data of the training data with the teacher learner.

[0080] Then, the teacher-led learner generation step S8 is executed. In this teacher-led learner generation step S8, the combination of the input data of the training data and the normalized deviation W of the output data used as training data is used as training data for the teacher-led learner to learn.

[0081] The teacher-trained learner used here can be any teacher-trained learner capable of performing regression by outputting a normalized deviation W based on the input data of the training data. For example, a deep neural network can be used as a teacher-trained learner for regression. Alternatively, a teacher-trained learner can be used that learns by discretizing the normalized deviation W, calculating the discretized normalized deviation, and using this discretized normalized deviation as the output data of the training data corresponding to the input data of the teacher-trained learner in step S8, thereby learning which discretized normalized deviation classifies the input data of the training data. For example, a support vector machine can be used as a teacher-trained learner for this classification.

[0082] The learned teacher-trained learner generated in step S8 can be used, for example, to estimate the state of bearing BR3, which supports the spindle of model MC, whose vibration susceptibility to accelerometers mounted near the bearing is different from that of models MA and MB.

[0083] That is, the inference result (output) obtained by inputting the same input data as the training data of the teacher-trained learner generated by signal processing of the physical quantity to be obtained for another state estimation object in actual application, is close to a predetermined value W. A and the specified value W B Which of the following is used to determine the state of bearing BR3? The data input to the teacher learner after the learning process is the training data used in data generation step S2, which includes the vibration amplitude at each frequency obtained by frequency analysis of the vibration of the accelerometer, the spindle mass, and the vibration ease of the accelerometer at each frequency mounted near the bearing.

[0084] Therefore, as Figure 6 or Figure 7 In that case, by having a specified value W that the teacher learner should output in the case of reference state A, A In the case of (0.00) and baseline state B, the teacher learner should output the specified value W. B (1.00) A display unit that simultaneously displays the value of the inference result (0.21) as a numerical value or graph, enabling the teacher learner to function as a state estimation device. Furthermore, a specified device can be configured to simultaneously display a specified value W. A Specified value W B A state estimation device is formed by a specified display unit that can display the values ​​of the inference results.

[0085] Furthermore, the state estimation device may also have a notification unit, which notifies the user by setting a predetermined value W. A With the specified value W B The values ​​between (e.g., 0.2 × W) A +0.8×W B This can be used as a threshold, so that a warning is displayed on the screen when the threshold is exceeded, or a warning is sent to the machine manager or the machine tool manufacturer who is replacing or repairing the bearing via email.

[0086] in addition, Figure 6 and Figure 7 The values ​​exemplified in the table and the values ​​exemplified as thresholds are hypothetical values; in practice, the exemplified values ​​are not calculated or set.

[0087] Next, Example 2 will be described.

[0088] In Example 2, the following situation will be described: the state estimation object is set as a drill bit used in hole machining in a machining center, the reference state A is set as the state after initial use, and the reference state B is set as the state that should be replaced with a new part due to deterioration. In this case, as two or more state estimation objects, they become different individual drill bits.

[0089] However, when learning and inference with a teacher-learner, the amount of data input to the teacher-learner needs to be consistent. For diagnosing the bearing of the spindle described above as Example 1, when collecting data of the same length during a prescribed diagnostic operation, the amount of data is consistent, and therefore, data can be directly input. However, in drill-based hole machining, the machining time may vary depending on the machining content. In this case, to ensure a consistent amount of input data, the cutting start time is offset slightly each time, while cutting out information for a predetermined amount of time. Then, by performing post-processing, the state of machining one hole at a time can be shown, in which a representative value is determined from multiple outputs (output groups) calculated using the teacher-learner with each of the multiple cut data (data groups) as input data. Furthermore, below, when "group" is appended to the end of terms such as "data group" and "output group," a set is shown as the parent used to calculate the representative value for the performance state. Therefore, for example, to diagnose the bearings of a spindle, it is possible to continuously operate at different speeds for a short period of time while assuming the bearing's state remains almost unchanged, and to set up a data set containing multiple data points, including different types of data collected and stored separately. Furthermore, regarding the post-processing for calculating a representative value from the output set, an appropriate method can be selected based on the type of state estimation object, the data collection method, etc. For example, when the same measurement is repeatedly performed to reduce the influence of deviations, the average value of the output set can be set as the representative value. In this application, post-processing refers to the process of determining the representative data from a given set of data, and the method of determining the representative data can be arbitrarily set.

[0090] In a series of actions similar to drill-based hole machining, there exists a non-machining interval before reaching the workpiece and a machining interval for cutting the workpiece. In the non-machining interval, the drill bit and workpiece do not contact each other. Therefore, the spindle load and feed axis load are the same in any state from reference state A to reference state B. Thus, a teacher learner should learn in the non-machining interval to calculate the same output as in reference state A, even when the state is close to reference state B. If this is not the case, learning would require calculating different outputs for almost identical inputs, making proper learning impossible. However, when using the average of the output set as a representative value, for example, even when the ratio of the non-machining interval is large, a value close to reference state A might be calculated. In this case, setting the maximum value of the output set as the representative value allows for capturing state changes independently of the ratio of the non-machining interval to the machining interval. Furthermore, although the maximum value of the output group is set as the representative value when the teacher learner's output value is greater in baseline state B than in baseline state A, the minimum value of the output group is set as the representative value when the teacher learner's output value is smaller in baseline state B than in baseline state A. In Example 2, as post-processing, the maximum output value in the output group of the teacher learner is set as the representative value.

[0091] Furthermore, even in the case of drill-based hole machining, by additionally setting up a machine learning machine to identify non-machining and machining intervals, it is possible to select only the data set of the machining interval for inference. Therefore, post-processing that sets the average of the output set as the representative value can also be employed. In addition, as a post-processing step for calculating the representative value based on the output set, it is also possible to take the maximum value after obtaining the moving average of multiple consecutive output sets.

[0092] Figure 8 A flowchart of another method for labeling is shown, and Example 2 is described in detail based on this flowchart.

[0093] In data collection step S101, the data required for labeling the dataset with the teacher learner is collected. In Example 2, from the time of initial use until the state where a new part should be replaced due to deterioration, a series of actions for drilling based on the drill bit are repeatedly performed, and the spindle motor torque and feed axis motor torque at this time are collected and recorded.

[0094] Next, as the data generation step S102, input data and corresponding data for the training data of the teacher learner are generated.

[0095] For the spindle motor torque and feed axis motor torque collected in the data collection step S101 during the series of actions of drilling, the data is repeatedly processed to offset the start time of cutting by a predetermined time T1 (e.g., T1 = 0.01 seconds) while cutting a predetermined length for a time T2 (e.g., T2 = 0.5 seconds). This generates input data for multiple training data at the same time when the state of the object being estimated is the same. Here, with T1 = 0.01 seconds and T2 = 0.5 seconds, data from 0 seconds to 0.5 seconds from the start of machining, data from 0.01 seconds to 0.51 seconds from the start of machining, and so on, are obtained.

[0096] Furthermore, in Example 2, the corresponding data is the same as the input data of each training data set. Additionally, as mentioned above, the corresponding data can be any data collected in the same state as the training data, or it can be different from the input data of the training data, as long as it represents information that is considered to depend on the state changes of the state estimation object and has relatively small variations. It can also be collected in any state and added to both the input data and the corresponding data of the training data.

[0097] Next, the deviation calculation learner generation step S103 is executed, in which a deviation calculation learner is generated to calculate the deviation relative to the reference state A.

[0098] In Example 2, in the deviation calculation learner generation step S103, for drill bit DR1 and drill bit DR2 with a different tool diameter than drill bit DR1, the corresponding data collected in the initial use state (baseline state A) are used as training data for learning by the autoencoder.

[0099] In Example 2, the deviation relative to reference state A is calculated using the average of the absolute values ​​of the differences between the output data and the input corresponding data when the corresponding data is input to the autoencoder. Therefore, in the deviation calculation learner calculation step S103, the autoencoder learns by calculating the average of the absolute values ​​of the differences between the output data of the training data corresponding to the input data of the training data and the input data of the training data, based on the input data of the training data. Alternatively, in another form, the deviation relative to reference state A can be replaced with other values ​​determined by the error between the input and output of the autoencoder, such as the average of the squares of the differences between the output data and the input corresponding data when a certain corresponding data is input to the autoencoder.

[0100] In the deviation calculation step S104, for the autoencoder that has completed learning in the deviation calculation learner generation step S103, multiple corresponding data obtained from the initial use state of drill bits DR1 and DR2 (reference state A) to the state where new bearings should be replaced due to deterioration (reference state B) are input, and the deviation relative to reference state A is calculated. At this time, for multiple corresponding data groups, including at least a first corresponding data group consisting of multiple corresponding data of drill bits DR1 and DR2 in reference state A and a second corresponding data group consisting of multiple corresponding data of drill bits DR1 and DR2 in reference state B, the calculated deviation of multiple corresponding data relative to reference state A is based on each corresponding data belonging to each of these multiple corresponding data groups. Furthermore, the multiple deviations relative to reference state A calculated by inputting multiple corresponding data belonging to a certain corresponding data group are collectively referred to as a deviation group relative to reference state A.

[0101] In the representative deviation calculation step S105, the deviation group calculated in the deviation calculation step S104 relative to the reference state A is processed in the same way as the post-processing for the output group with the aforementioned teacher learner, and the representative deviation relative to the reference state A is calculated.

[0102] For each set of deviations relative to the reference state A calculated from multiple corresponding data sets, the largest deviation relative to the reference state A in the set of deviations relative to the reference state A is set as the representative deviation relative to the reference state A, and the representative deviations relative to the reference state A for each of the multiple corresponding data sets are calculated.

[0103] Furthermore, the states of drill bits DR1 and DR2 are either deteriorated or unchanged, and they will not revert from a deteriorated state to an undeteriorated state. That is, the states of drill bits DR1 and DR2 should change unidirectionally from reference state A to reference state B. Therefore, if the representative deviation obtained in the representative deviation calculation step S105 is not monotonically increasing relative to reference state A, it is presumed that it includes the influence of phenomena unrelated to the deterioration of the states of drill bits DR1 and DR2, such as measurement deviation. If it is presumed that it includes components unrelated to the deterioration of the states of drill bits DR1 and DR2, then, by applying a moving average filter or a low-pass filter to the representative deviation relative to reference state A in a time-series arrangement, a smoothed representative deviation relative to reference state A can be obtained. This removes components unrelated to the deterioration of the states of drill bits DR1 and DR2. In subsequent steps, after filtering the representative deviation relative to reference state A, a smoothed representative deviation relative to reference state A is used instead of the representative deviation relative to reference state A.

[0104] Thus, in the representative deviation calculation step S105, the representative deviation relative to the reference state A or the smoothed representative deviation relative to the reference state A is calculated.

[0105] In the next normalization mapping determination step S106, a mapping (normalization mapping) is determined. This mapping is used in the normalization representative deviation calculation step S107 to depict the image using the representative deviation relative to the reference state A for each of the multiple corresponding data sets obtained in the representative deviation calculation step S105 based on the states of drill bits DR1 and DR2, as normalized representative deviation. Furthermore, in the following description related to Embodiment 2, for ease of understanding, the representative deviation relative to the reference state A and the smoothed representative deviation relative to the reference state A are both represented as the representative deviation relative to the reference state A. However, as described above, in the steps described later, when using the unfiltered representative deviation relative to reference state A, only the unfiltered representative deviation relative to reference state A is used; when using the smoothed representative deviation relative to reference state A, only the smoothed representative deviation relative to reference state A is used. The representative deviation relative to reference state A and the smoothed representative deviation relative to reference state A are not used together.

[0106] Here, the representative deviation relative to the reference state A is set as the representative deviation V. A * Let the representative deviation degree relative to the reference state A under the reference state B be set as the representative deviation degree V. B * The specified value that the representative value should take under the reference state A is set as the specified value W. A * The specified value that the representative value should take under the reference state B is set as the specified value W. B * At this point, if the mapping satisfies all of the following three characteristics, it can be used to represent the deviation V of each of the multiple corresponding data sets relative to the baseline state A. * Described as normalized representation of deviation W * The mapping. The first characteristic is the ability to map the deviation V. A * Described as a specified value W A * The mapping. The second characteristic is the ability to map the deviation V. B * Described as a specified value W B *The mapping. The third characteristic is in the representation of the deviation V. A * And represents the degree of deviation V B * The interval is a monotonic mapping. A mapping satisfying the above three properties is determined by defining the representative deviation V relative to the reference state A. * Described as normalized representation of deviation W * The mapping.

[0107] Next, the representative deviation V relative to the reference state A is used. * Described as normalized representation of deviation W * The mapping will represent the deviation V relative to the baseline state A. * Described as normalized representation of deviation W * .

[0108] Furthermore, in the normalized individual deviation calculation step S108, for the normalized representative deviation V * The sum and normalization represent the deviation V. * The corresponding deviation group GP relative to the baseline state A is processed in the opposite way to the post-processing, and the normalized individual deviation U is calculated. * .

[0109] Here, post-processing is a prescribed process used to determine the data that represents the set from a given set of data. Therefore, the opposite of post-processing is a prescribed process used to derive the given set of data represented by a given data. Therefore, the opposite of post-processing performed in the normalized individual deviation calculation step S108 of Example 2 is as follows: The normalized individual deviation U calculated through the process opposite to post-processing... * During post-processing, it will become the normalized representation of the deviation W. * In Example 2, as a suitable function used in a process that has the opposite effect to post-processing, consider the following mathematical formula 3 or mathematical formula 4.

[0110] The normalized representative deviation W corresponds to the machining of a hole using drill bit DR1 or DR2. * The deviation group GP relative to reference state A is defined as follows: {Deviation U1 relative to reference state A, Deviation U2 relative to reference state A, ..., Deviation U...} M}, Normalized Individual Deviation Group GP * ={Normalized Individual Deviation U1} * Normalized individual deviation U2 * Normalized individual deviation U M*}(U and U * The subscript indicates which cut-out interval it is. Using a function like mathematical formula 3 or 4, the normalized individual deviation U of the k-th cut-out interval of a given hole can be calculated. k * That is, by taking the normalized individual deviation group GP corresponding to the machining of a certain hole. * The maximum such post-processing, using formula 3 or formula 4 to calculate U k * Becoming the normalization representative of deviation W * .

[0111] Alternatively, in another form, where the post-processing involves averaging, a function such as mathematical formula 5 or mathematical formula 6 can be used.

[0112]

Mathematical Expression 3

[0113]

[0114]

Mathematical Expression 4

[0115]

[0116]

Mathematical Expression 5

[0117]

[0118]

Mathematical Expression 6

[0119]

[0120] Then, in the label determination step S109, the output data of the training data corresponding to the input data of the training data with the teacher learner, which is the k-th cutout interval of a certain hole, is calculated using the normalized individual deviation U of the k-th cutout interval of a certain hole. k * .

[0121] Similar to Example 1, it enables the output of a normalized individual deviation U based on the input of the training data. k * Deep neural networks and support vector machines, which perform regression in a manner that combines the input and output of training data (normalized individual deviation U), k * The combination of these data is used as training data for learning, thereby generating a teacher-trained learner.

[0122] Thus, the output data of the training data determined by the labeling methods of Embodiments 1 and 2 above can be normalized to a teacher-trained learner or a teacher-trained learner with post-processing capabilities that has learned using the training data containing the output determined by this method, and outputs common values ​​in reference state A and reference state B that are independent of the state estimation object.

[0123] Furthermore, when the state of the object being estimated is an intermediate state between baseline state A and baseline state B, the output data of the training data determined by this method will also reflect the moment of change in deviation from baseline state A. Therefore, it is possible to endow a teachered learner or a teachered learner with post-processing capabilities that has learned using training data containing the output data determined by this method with a state estimation capability that is more in line with the actual state. That is, for the intermediate state between two baseline states, it is also possible to perform labeling of the dataset that is more in line with the actual state at low cost.

[0124] Furthermore, it is obvious that a teacher-trained learner that learns using training data labeled according to this invention is within the scope of this invention. However, for a method of using the output data of the teacher-trained learner itself, or data obtained by appropriately processing the output data, as output data for training another teacher-trained learner, this invention is also used in the process of generating the other teacher-trained learner, and therefore this method is also included within the scope of the claims. Even if this process is performed repeatedly a limited number of times, the invention is used in the process and is within the scope of the claims.

[0125] Each step can be performed by multiple computers simultaneously, or by a single computer.

[0126] This labeling method can also be executed multiple times to obtain the output data of multiple training data corresponding to the same input data. The content calculated by combining the output data of multiple training data using arithmetic operations or functions is then set as the label. For example, the output data of the training data can be calculated separately under the following conditions: setting baseline state A to the initial state and baseline state B to the state where the component should be replaced due to degradation; and setting baseline state A to the state where the component should be replaced due to degradation and baseline state B to the initial state. The difference between the outputs of these training data is set as the output data of the training data with the teacher learner.

[0127] In Examples 1 and 2 above, the following example is shown: In the deviation calculation step, the deviation relative to reference state A calculated by the learner that calculates the deviation relative to reference state A is directly used. On the other hand, a learner that calculates the deviation relative to reference state B can also be generated separately, and the difference between the deviation relative to reference state A calculated by the learner that calculates the deviation relative to reference state A and the deviation relative to reference state B calculated by the learner that calculates the deviation relative to reference state B can be regarded as the deviation relative to reference state A. As a result, if the deviation relative to reference state A is obtained, the data obtained by combining the output data using multiple learners through arithmetic operations or functions can also be used as the deviation relative to reference state A.

[0128] In the above embodiments 1 and 2, an implementation method for learning data from two state estimation objects is shown. However, it is possible to learn data from any number of state estimation objects as long as there are more than two state estimation objects.

[0129] The data used does not necessarily have to be all output data of the training data determined by this labeling method. For example, if drill bit DR1 can perform 1000 holes before it needs to be replaced, and drill bit DR2 can perform 2000 holes before it needs to be replaced, if all the data is learned, the proportion of drill bit DR2 will increase. Therefore, the training data (input data and output data) of drill bit DR2 can also be learned with a 50% probability.

[0130] Furthermore, in embodiments 1 and 2 above, the training data referred to as the teacher-trained learner can be used not only as training data but also as evaluation data when evaluating the performance of the teacher-trained learner. That is, the output data calculated by inputting the training data into the teacher-trained learner that has completed learning is compared with the output of the training data. If they are very consistent, it can be judged that the performance of the teacher-trained learner is good.

[0131] The output data of the training data determined by this invention can normalize the output of a common value in both baseline state A and baseline state B, independent of the state estimation object, to a teacher-trained learner or a teacher-trained learner with post-processing capabilities that has learned from the training data containing the output data determined by this method. Furthermore, when the state of the state estimation object is an intermediate state between baseline state A and baseline state B, the output data of the training data determined by this invention also reflects the moment of change in deviation relative to baseline state A. Therefore, a more realistic state estimation capability can be given to a teacher-trained learner or a teacher-trained learner with post-processing capabilities that has learned from the training data containing the output data determined by this method. That is, even for intermediate states between two baseline states, labeling of a dataset that more closely resembles the actual state can be performed at low cost. Furthermore, it is expected that when a teacher-trained learner or a teacher-trained learner with post-processing capabilities learns to output common values ​​in reference state A and reference state B independently of the state estimation object using training data from a wide variety of state estimation objects with different conditions, it can also obtain the characteristic (generality) of outputting common values ​​in reference state A and reference state B independently of the state estimation object for state estimation objects with different conditions that have not been learned.

Claims

1. A method for labeling a dataset with a teacher learner, used to determine output data of training data corresponding to input data in training data related to the generation of the teacher learner, wherein the teacher learner is used to estimate what state an object for state estimation, such as a bearing or cutting tool used in a machine tool, is in between a predetermined reference state A and a predetermined reference state B different from said reference state A, wherein the predetermined reference state A is the state of the object for state estimation when it is first used, and the predetermined reference state B is the state in which the object for state estimation should be replaced with a new component due to deterioration, characterized in that, Arbitrary data selected from the first physical quantity and the data associated with the state of the state estimation object obtained based on the first physical quantity are set as the input data of the training data. The first physical quantity is obtained under the condition that the state estimation object is in a specified state. Any data selected from the second physical quantity and the data associated with the state of the state estimation object obtained based on the second physical quantity is called corresponding data. The second physical quantity is obtained when the state of the state estimation object is in the same state as when the input data of the training data is related to the first physical quantity. For two or more state estimation objects, the deviation from the reference state A is calculated for each of the following corresponding data. The multiple corresponding data include at least the corresponding data for each state estimation object when it is in the reference state A and the corresponding data for each state estimation object when it is in the reference state B. As a mapping used to depict the deviation of each of the corresponding data points relative to a reference state A as a normalized deviation, the deviation relative to the reference state A can be depicted as a predetermined value that the teacher learner should output in the case of the reference state A, and the deviation relative to the reference state A in the case of the reference state B can be depicted as a predetermined value that the teacher learner should output in the case of the reference state B. The mapping is set according to each state estimation object, wherein the mapping is a monotonic mapping within the interval of the deviation relative to the reference state A in the case of the reference state A and the deviation relative to the reference state A in the case of the reference state B. Using the defined mapping, for each of the corresponding data points, an image of the deviation relative to the reference state A is drawn, thereby calculating the normalized deviation. The normalized deviation is determined as the output data of the training data corresponding to the input data of the training data. Based on the output data, the state of the state estimation object between the reference state A and the reference state B is estimated.

2. The labeling method for a dataset with a teacher learner according to claim 1, characterized in that, When the state of the state estimation object changes unidirectionally from either reference state A or reference state B in a temporal sequence to the other. For multiple state estimation objects, after calculating the deviation of each of the corresponding data relative to the reference state A, filtering is performed on the calculated deviations relative to the reference state A in chronological order to convert them into smooth deviations relative to the reference state A. When setting the mapping and calculating the normalized deviation in order to depict the deviation of each of the multiple corresponding data relative to the reference state A as the normalized deviation, the smooth deviation relative to the reference state A is used instead of the deviation relative to the reference state A.

3. A method for labeling a dataset with a teacher learner, used to determine output data of training data corresponding to input data in training data related to the generation of the teacher learner, wherein the teacher learner is used to estimate what state an object for state estimation, such as a bearing or cutting tool used in a machine tool, is in between a predetermined reference state A and a predetermined reference state B different from said reference state A, wherein the predetermined reference state A is the state of the object for state estimation when it is first used, and the predetermined reference state B is the state in which the object for state estimation should be replaced with a new component due to deterioration, characterized in that, The teacher-learning device has a post-processing function, in which, given multiple specified data belonging to a set of multiple data groups as input, representative values ​​in the output data group, which is a set of output data calculated based on each of the input data, are calculated. Arbitrary data selected from the first physical quantity and the data associated with the state of the state estimation object obtained based on the first physical quantity are set as the input data of the training data. The first physical quantity is obtained under the condition that the state estimation object is in a specified state. Any data selected from the second physical quantity and the data associated with the state of the state estimation object obtained based on the second physical quantity is called corresponding data. The second physical quantity is obtained when the state of the state estimation object is in the same state as when the input data of the training data is related to the first physical quantity. For two or more state estimation objects, for multiple corresponding data groups, the deviation of the corresponding data belonging to each of the multiple corresponding data groups relative to the reference state A is calculated respectively. The multiple corresponding data groups include at least a first corresponding data group and a second corresponding data group. The first corresponding data group consists of multiple corresponding data of each state estimation object when it is in the reference state A, and the second corresponding data group consists of multiple corresponding data of each state estimation object when it is in the reference state B. The set of deviations relative to reference state A corresponding to the corresponding data of the multiple corresponding data groups to which the deviations relative to reference state A are calculated and belong is defined as the deviation group relative to reference state A. For each of the multiple sets of deviations relative to the reference state A, the same processing as the post-processing performed by the teacher-assisted learner is applied to calculate the representative deviation of each of the multiple sets of deviations relative to the reference state A. As a mapping used to depict the representative deviations relative to reference state A corresponding to the plurality of corresponding data groups as normalized representative deviations, the mapping can depict the representative deviations relative to reference state A in the case of reference state A as a predetermined value that the representative value calculated by the teacher learner should take in the case of reference state A, and can depict the representative deviations relative to reference state A in the case of reference state B as a predetermined value that the representative value calculated by the teacher learner should take in the case of reference state B. This mapping is set according to each state estimation object, wherein the mapping is a monotonic mapping within the interval of the representative deviations relative to reference state A in the case of reference state A and the representative deviations relative to reference state A in the case of reference state B. Using the defined mapping, for each of the multiple corresponding data sets, an image of the representative deviation relative to the reference state A is drawn, thereby calculating the normalized representative deviation. For each of the multiple normalized representative deviations and each deviation group relative to the reference state A corresponding to the multiple normalized representative deviations, a process opposite to the post-processing is performed. For each of the multiple corresponding data groups corresponding to the multiple normalized representative deviations, a normalized individual deviation is calculated. Specifically, when the post-processing is performed on the normalized individual deviations, the normalized representative deviation is calculated. The normalized individual deviation is determined as the output data of the training data corresponding to the input data of the training data. Based on the output data, the state of the state estimation object between the reference state A and the reference state B is estimated.

4. The labeling method for a dataset with a teacher learner according to claim 3, characterized in that, When the state of the state estimation object changes unidirectionally from either reference state A or reference state B in a temporal sequence to the other. For multiple state estimation objects, after calculating the representative deviation of each of the corresponding data groups relative to the reference state A, filtering is performed on the calculated representative deviations relative to the reference state A in chronological order to convert them into smooth representative deviations relative to the reference state A. When setting the mapping and calculating the normalized representative deviation in order to depict the representative deviation of each of the multiple corresponding data groups relative to the reference state A as the normalized representative deviation, the smoothed representative deviation relative to the reference state A is used instead of the representative deviation relative to the reference state A.

5. A teacher-learning device, characterized in that, The teacher-trained learner is generated using the labeling method for the dataset of the teacher-trained learner as described in claim 1 or 2.

6. A state estimation device, characterized in that, The state estimation device is equipped with the teacher-assisted learning device as described in claim 5.

7. The state estimation device according to claim 6, characterized in that, The output of the teacher-assisted learner, the specified value that the teacher-assisted learner should output when the state estimation object is in the baseline state A, and the specified value that the teacher-assisted learner should output when the state estimation object is in the baseline state B are simultaneously displayed as numerical values ​​or graphs.

8. A teacher learning device, characterized in that, The teacher-trained learner is generated using the labeling method for the dataset of the teacher-trained learner as described in claim 3 or 4.

9. A state estimation device, characterized in that, The state estimation device is equipped with the teacher-assisted learning device as described in claim 8.

10. The state estimation device according to claim 9, characterized in that, The representative values ​​from the output data set of the teacher-learner, the specified values ​​that the representative values ​​calculated by the teacher-learner should take when the state estimation object is in the baseline state A, and the specified values ​​that the representative values ​​calculated by the teacher-learner should take when the state estimation object is in the baseline state B are simultaneously displayed as numerical values ​​or graphs.

Images