JP7134062B2 - Machine tool abnormality diagnosis system, abnormality diagnosis method, abnormality diagnosis program - Google Patents

Description

TECHNICAL FIELD The present invention relates to an abnormality diagnosis system, an abnormality diagnosis method, and an abnormality diagnosis program for a machine tool that moves in space by a drive shaft and performs processing, and is intended to detect an abnormality in order to suppress the occurrence or spread of a defect. .

At present, most of the machine tools that perform removal and lamination processing are automated by NC units, perform predetermined operations according to processing programs, and can be processed unmanned.
However, depending on the state of the cutting tool and the state of the pre-machining, there are cases where the workpiece does not meet the required specifications even if the machining is performed according to the program. In addition, even in a situation where the workpiece cannot be removed due to damage to the tool, the feed axis continues to operate and the tool and workpiece collide, resulting in damage not only to the tool and workpiece but also to the machine. Therefore, it is common practice to monitor the load of the spindle motor, which is considered to represent the state of machining, and stop the feed axis.
However, if the difference in the load on the spindle motor between normal machining and abnormal machining is smaller than the change in the load on the spindle motor during normal machining, it is difficult to determine an abnormality by exceeding a certain threshold value. be. Therefore, in Patent Document 1, the waveform of the pressure change during normal press-fitting is recorded as a base waveform, and a threshold value offset from the base waveform is set. Anomalies are detected depending on whether

Japanese Patent No. 2673656

However, the method of Patent Document 1 is based on the premise that the same processing is repeatedly performed, and cannot be used for processing different from the processing for which the base waveform was obtained, such as objects with different press-fit distances and shapes. .

SUMMARY OF THE INVENTION Accordingly, the present invention has been made in view of the above problems, and aims to provide an abnormality diagnosis system, an abnormality diagnosis method, and an abnormality diagnosis program capable of universally detecting machining abnormalities.

In order to achieve the above object, the invention according to claim 1 has a rotary shaft for rotating a tool or a work, and a feed shaft for relatively moving the tool and the work, and a programmed command A system for diagnosing an abnormality in a machine tool that processes the workpiece with the tool based on
a machine information acquisition unit that acquires at least one of operation information of the machine tool and output information of various sensors attached to the machine tool as machine information;
a storage unit for recording the machine information;
a model creation unit that creates an estimation model for estimating the degree of abnormality from the recorded mechanical information;
The degree of abnormality in the machine information is estimated at any time by the created estimation model, and the presence or absence of abnormality is determined by comparing the time change of the estimated value with a threshold set so as to change with time according to the machining mode. an abnormality diagnosis unit for judging;
In the above configuration, a machining mode determination unit that determines the sameness between machining modes is included, and the abnormality diagnosis unit is configured to, when repeatedly performing machining determined to be the same by the machining mode determination unit, determine the diagnosis target. The threshold value is set based on the time-series waveform of the estimated value obtained by estimating the machine information when the machining is normally completed by the estimation model .
According to a second aspect of the present invention, in the configuration of claim 1 , the model creating section extracts data, as teacher data, in which the change in the machining mode determined by the machining mode determining section falls within a predetermined range. and creating the estimation model from the training data.
According to a third aspect of the present invention, in the configuration of the first aspect, the abnormality diagnosis unit uses a time-series waveform obtained by estimating the machine information during normal operation with the estimation model as a master waveform, and the storage unit and records an interval in which the master waveform is below a predetermined identifiable interval setting threshold in the storage unit as an identifiable interval in which data processing can be performed using an algorithm, and changes the estimated value with time. The comparison with the threshold value is performed only in the identifiable section.
The invention according to claim 4 is the configuration according to claim 3 , wherein the model creation unit extracts the machine information of the identifiable section as teacher data, and updates the estimation model using the teacher data. characterized by performing
In order to achieve the above object, the invention according to claim 5 has a rotating shaft for rotating a tool or a work and a feed shaft for relatively moving the tool and the work, and a programmed command A method for diagnosing an abnormality in a machine tool that processes the workpiece with the tool based on
a machine information acquisition step of acquiring at least one of operation information of the machine tool and output information of various sensors attached to the machine tool as machine information;
a recording step of recording the machine information;
a model creation step of creating an estimation model for estimating the degree of abnormality from the recorded mechanical information;
The degree of abnormality in the machine information is estimated at any time by the created estimation model, and the presence or absence of abnormality is determined by comparing the time change of the estimated value with a threshold set so as to change with time according to the machining mode. while executing an abnormality diagnosis step for determining ,
In the configuration, in the model creation step,
a labeled data creation step of creating labeled data in which an operation result indicating whether or not the machine tool is operating normally and the machine information are linked;
a master data creation step of creating master data by extracting from the labeled data only portions containing features corresponding to the operation result;
a master model setting step of creating a master model for estimating the operation result from the master data or adopting an existing learned master model;
an additional training data creation step of estimating the labeled data using the master model and extracting from the estimation result a location containing a feature corresponding to the operation result to create additional training data;
a relearning model creation step of creating a relearning model for estimating the motion result from the master data and the additional teacher data;
a model performance evaluation step of performing an estimation performance evaluation of the relearning model using the master model;
If the estimated performance is inferior to the master model, the master model is adopted as the estimated model, and if the estimated performance exceeds the master model, the master model is updated with a re-learned model, and the additional teacher is used. and a relearning execution determination step of repeating the data creating step and the relearning model creating step .
According to a sixth aspect of the present invention, in the configuration of the fifth aspect, in the additional training data creation step executed after the re-learning execution determination step, the extraction The threshold value is reset, and the extraction threshold value is changed according to the evaluation result in the model performance evaluation step.
In order to achieve the above object, the invention according to claim 7 is an abnormality diagnosis program for a machine tool, which causes a computer to execute the abnormality diagnosis method for a machine tool according to claim 5 or 6 . and

According to the present invention, the degree of abnormality in the machine information is estimated at any time by the estimation model, and the change in the estimated value over time is compared with the threshold set so as to change over time in accordance with the machining mode. Since the presence/absence is determined, it is possible to effectively detect a machining abnormality even in machining in which the machining mode is not constant and changes greatly. Moreover, even if the processing mode of the model creation range, which is used as teacher data when creating the estimation model, does not sufficiently cover the range of abnormality monitoring, it is possible to effectively diagnose anomalies.
In particular, according to the first aspect of the present invention, in addition to the above effects, the threshold can be easily determined because the effective range and setting of the threshold need not be performed for each section. Moreover, even if an estimated model whose effective working range is unclear is used, the threshold value is determined from the actual identification result, so that it can be easily and effectively performed.
In particular, according to the second aspect of the present invention, in addition to the above effect, it is not necessary to create a model from data of various variations in a large amount of data, so model creation is facilitated.
In particular, according to the third aspect of the invention, in addition to the effects described above, when using an estimation model, it is possible to perform identification without grasping the scope of application.
In particular, according to the fourth aspect of the present invention, in addition to the above effects, it is possible to extract data from locations where effective learning can be performed even with a small amount of data as teacher data. data can be effectively used. In addition, since the calculation cost can be reduced, the estimation model can be tuned without using a large-scale computer. Furthermore, since fine adjustment can be performed according to the abnormality detection target, abnormality detection can be performed with higher accuracy.
In particular, according to the fifth aspect of the invention, in addition to the effects described above, an estimation model with an expanded range of application while maintaining the estimation performance can be created without performing detailed labeling. In addition, it is possible to reduce the cost required for selecting effective learning data and reduce the increase in calculation cost due to using a large amount of data that does not contain the feature value corresponding to the operation result.
In particular, according to the sixth aspect of the invention, in addition to the effects described above, the resetting of the threshold value for extraction narrows the range of the additional training data, and an improvement in estimation accuracy can be expected.

1 is a block diagram of a machine tool provided with an abnormality diagnosis device of form 1; FIG. 4 is a flow chart showing a procedure of abnormality diagnosis of form 1; FIG. 4 is an explanatory diagram of target machining for which model creation is performed; 4 is a graph showing changes in the spindle load of the target machining for which the model was created. FIG. 10 is an explanatory diagram of target machining for which abnormality diagnosis is performed; 7 is a graph showing changes in the spindle load of the target machining for which abnormality diagnosis was performed; It is a graph which shows the identification result of the object process which performed abnormality diagnosis. It is a graph of the threshold setting for abnormality diagnosis when processing is performed a plurality of times. FIG. 11 is a block diagram of a machine tool provided with an abnormality diagnosis device of form 2; 10 is a flow chart showing a procedure of abnormality diagnosis of form 2. FIG. 5 is a graph showing identifiable intervals; 7 is a graph showing abnormality diagnosis based on identifiable intervals; FIG. 11 is a graph showing abnormality diagnosis based on identifiable intervals (thresholds are given margins); FIG. 10 is a flow chart showing another procedure of abnormality diagnosis of form 2. FIG. 4 is a flow chart showing a procedure for creating an estimation model; It is a graph which shows the range used as additional learning data.

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[Mode 1]
FIG. 1 is an explanatory diagram showing an example of a machine tool 1 related to the present invention.
A machine tool 1 has a spindle 2 as a rotating shaft, and a workpiece 3 is mounted on a table. A tool attached to the spindle 2 rotates and moves relative to the workpiece 3 by the feed shaft, thereby performing cutting.
An abnormality diagnosis device 4 that constitutes an abnormality diagnosis system is provided in the control device of the machine tool 1 . Inside this abnormality diagnosis device 4, there are operational information such as the drive load, speed, vibration value, temperature, etc. of the spindle 2, feed shaft, tool change shaft, and peripheral equipment of the machine tool 1, and output information of various attached sensors. A machine information acquisition unit 5 is provided for acquiring the .

In addition, the abnormality diagnosis device 4 includes a model creation unit 6 that creates an estimated model for detecting an abnormality, a machining mode determination unit 7 that determines the sameness between machining modes, and a machine information acquisition unit 5. A storage unit 8 for recording machine information, an estimated model created by the model creating unit 6, estimated model creating conditions, etc., a model creating condition input unit 9 for inputting the creating conditions for the estimated model, etc., and an abnormality diagnosis A monitoring setting input unit 10 for inputting monitoring settings, an interpretation unit 11 for interpreting the input contents of the model creation condition input unit 9 and the monitoring setting input unit 10, a machine operation command unit 12 for issuing operation commands to the machine, and an abnormality diagnosis. An abnormality diagnosis unit 13 is provided.

Although FIG. 1 shows all the components of the abnormality diagnosis device 4 as if they were in the control device, the present invention is not limited to this. 6. The abnormality diagnosis unit 13 may be implemented and executed in another terminal to configure an abnormality diagnosis system. For example, an external terminal such as a workstation or server on a local network connected to the machine tool, or a cloud service connected via the Internet may be used. In addition, network connection is not essential, as estimation models and data can be exchanged regardless of interfaces.

The procedure of abnormality diagnosis by this abnormality diagnosis device 4 will be described based on the flowchart shown in FIG. Here, S1 to S3 are steps for creating an estimation model, and S4 to S5 are steps for diagnosing abnormality using the estimation model.
First, in S1, the machine information acquired by the machine information acquisition section 5 is recorded as record information in the storage section 8 (mechanical information acquisition step and recording step). As the machine information, it is preferable to record information that can determine the processing mode together. For example, tool information used for machining, a rotation speed command for the spindle 2 during machining, a speed command for relative motion, and command information for machining such as the amount of cutting into the workpiece 3 are preferable.

Next, in S2, the model generating unit 6 extracts teacher data to be used for generating an estimation model from the recorded information recorded in the storage unit 8. FIG. When creating an estimation model, all data that clearly indicates whether the processing state is normal or abnormal may be used as training data. It is desirable to limit the variation of the data because it may not exist or become too complicated, requiring time and computational costs. Therefore, here, the machining mode determination unit 7 determines whether or not the machining modes such as machining conditions, jigs and tools, program commands, and coordinate information are the same.

As an example, a case where a convex portion having a height of 50 mm is side-machined by an end mill T as shown in FIG. FIG. 4 shows changes over time in the spindle load when the workpiece W shown in FIG. 3 is machined. The period from 0 to 5 seconds in FIG. 4 is the air cut portion, and the spindle load does not increase. Also, during 5 to 8 seconds, the tool shown in FIG. 3 approaches from the left side to the right side, so the cutting amount changes until the center of the tool passes the left end of the workpiece. Since the cutting conditions change, the spindle load is not constant, and in some cases chattering occurs temporarily and the spindle load changes greatly. If the above range is used as training data, an estimation model cannot be created efficiently. Therefore, the range of 10 to 20 seconds in which the processing is stable is extracted as training data, as indicated by the dotted line. The extraction range can be set by observing the target waveform of the teacher data, as in the method described above, or by artificially determining it from the processing details, performing numerical processing on the target waveform, or using a program command to perform machine operation. , coordinate information, machining conditions, tools, workpiece material type, workpiece shape, operator, cutting fluid type, supply pressure, and other information may be combined and limited.

For example, when performing numerical processing on the target waveform, the amount of change within a unit time may be within a certain value range, or frequency analysis may be performed and the characteristic frequency component may exceed a certain value. Also, when using a program command or coordinate information for machine operation, it is possible to set a range for the rotation speed command or feed speed, or limit it to a cycle operation, etc., or perform machining with a constant depth of cut. You can extract. When using the information on jigs and tools, it may be cut according to the size of the turning chuck, or a range may be set for the shape such as the long diameter of the tool, the number of blades, the coating, and the like. It is preferable to change the extraction range according to the abnormal target to be monitored in combination with the above-mentioned information, the desired generalization performance, and the amount of teacher data. For example, it is possible to extract data where the machining mode is the same regardless of the drill diameter, or to extract data where the depth of cut changes within a certain range even in end mill side machining as described above. By doing so, the teacher data may be used to create an estimation model with higher generalization performance.

In S3, the model creating unit 6 constructs an estimation model for identifying anomalies from the teacher data extracted in the process of S2 (S2, S3: model creating steps). In this example, an autoencoder is constructed by a neural network, and a normal state is learned only from data that has been normally processed from the teacher data, and a trained model is created. An estimation model for abnormality diagnosis is defined by expressing a change from a normal state by a difference between input and output of the learned model. The method of defining the estimation model is not limited to the above-mentioned method, and it is preferable to use an algorithm such as machine learning or deep learning that has high performance for the target data. For example, it is possible to identify abnormalities by defining the distribution of normal states by clustering with Gaussian Mixture Model, Approximate Gaussian Mixtures, etc., and outputting the degree of divergence from normal features. In addition, a support vector machine, nearest neighbor method, random forest method, naive Bayes method, etc. may be used.

In S4, a threshold value for performing an abnormality diagnosis in the abnormality diagnosis unit 13 is defined. FIG. 5 shows a workpiece W1 to be processed for abnormality identification. Although the shape is different from that of the work W shown in FIG. 3, the work includes the same machining mode, such as the height of the protrusions being 50 mm.
FIG. 6 shows the change over time of the spindle load during machining of the workpiece shown in FIG. Depending on the input from the monitoring setting input unit 10, a lower threshold value is applied to areas where stable machining can be performed, and a higher threshold value is applied to areas where machining is likely to be unstable. artificially define the threshold for Further, instead of artificially giving a value that changes with time as described above, if the estimation model created in S3 is used to identify the time change of the spindle load, the output value as shown in FIG. can get. By using a value obtained by offsetting this output value as a threshold value, effective monitoring can be performed. The machining shown in FIG. 5 includes machining different from the model creation range shown in FIG. can be done. The offset value may be given from the monitoring setting input unit 10, or may be obtained from the average value or variance value of the estimation results.

Next, in S5, the machine information obtained from the machine information obtaining section 5 is subjected to abnormality estimation by the abnormality diagnosis section 13 using the estimation model, and the output result and the threshold defined in S4 are compared. The presence or absence of abnormality is diagnosed by comparison (S4, S5: abnormality diagnosis step). When an abnormality is diagnosed, the machine operation command unit 12 issues an alarm, notifies of the abnormality, and performs a retraction operation. Diagnosis results may be displayed on a monitor.

As described above, according to the abnormality diagnosis device 4, the abnormality diagnosis method, and the abnormality diagnosis program for the machine tool 1 of Embodiment 1, the machine information acquisition unit 5 acquires the operation information and the output information as machine information, and the acquired A model creation unit 6 that creates an estimation model for estimating the degree of abnormality from the machine information, and the created estimation model is used to estimate the degree of abnormality of the machine information at any time, and the estimated value changes over time according to the time change and the machining mode. and an abnormality diagnosis unit 13 for judging the presence or absence of an abnormality by comparing with a threshold value set as follows. Therefore, it is possible to detect various processing abnormalities. Moreover, even if the processing mode of the model creation range, which is used as teacher data when creating the estimation model, does not sufficiently cover the range of abnormality monitoring, it is possible to effectively diagnose anomalies.

In particular, here, a machining mode judgment unit 7 for judging the sameness between machining forms to be diagnosed is further provided. Since the threshold is set based on the waveform of the time change of the estimated value that is estimated at any time by the estimation model for the machine information when the target machining is normally completed, the effective range and setting of the threshold must be set for each section. Since there is no , the threshold can be easily determined. Moreover, even if an estimated model whose effective working range is unclear is used, the threshold value is determined from the actual identification result, so that it can be easily and effectively performed.
In addition, the model creating unit 6 extracts the data in which the change in the processing mode determined by the processing mode determination unit 7 is within a predetermined range as teacher data, and creates an estimation model from the teacher data. Since there is no need to create a model from data of various variations, model creation is facilitated.

It should be noted that although both the operation information and the output information are acquired as the machine information in the above-described form 1, only one of them may be acquired.

[Mode 2]
Next, another form of the present invention will be described. However, the same reference numerals are given to the same components as those in the first embodiment, and overlapping descriptions are omitted.
The abnormality diagnosis device 4A shown in FIG. 9 does not include a machining mode determination unit, but includes an existing data input unit 14 for inputting existing data acquired by the machine tool 1 or another machine tool. Different from form 1.
The procedure of abnormality diagnosis processing in this abnormality diagnosis device 4A will be described based on the flowchart shown in FIG.
First, an estimation model is created in S11. The machine information acquired by the machine information acquisition unit 5 is recorded as record information in the storage unit 8 (machine information acquisition step and recording step). Existing data may be input to the storage unit 8 from the existing data input unit 14 in addition to the machine information acquisition unit 5 . The existing data may be data acquired during machining under the same machining conditions, or may be data acquired under different machining conditions, jigs and tools, or other machines in order to improve the performance of the estimation model.
Then, the model creating unit 6 creates an estimation model using the information recorded in the storage unit 8 as teacher data (model creating step). Here, the points in which the range in which the processing is stable are extracted as training data and the specific method of creating an estimation model are the same as in the first mode.

Next, in S12, test machining is performed before repeated machining, machine information when machining is normally completed is recorded in the storage unit 8 as record information, and estimation is performed as needed using the estimation model created in S11. The estimated waveform obtained in 1) is recorded in the storage unit 8 as a master waveform. When the workpiece W1 shown in FIG. 5 is machined, the master waveform obtained by estimating the change in the spindle load in the normal state shown in FIG. 6 is shown in FIG.
Furthermore, in S13, an identifiable section is set. Here, an identifiable interval setting threshold value is set for the master waveform, and an interval in which the master waveform falls below the identifiable interval setting threshold value is defined as an identifiable interval. For example, as indicated by the dashed line in FIG. 11, a set threshold value (0.2) for the identifiable interval is provided in the master waveform, and the lower ranges of 10 to 26 sec, 42 to 53 sec, and 65 to 75 sec are defined as identifiable intervals. do.

Next, in S14, abnormality diagnosis is performed when repeated machining is performed (S12 to S14: abnormality diagnosis steps). Abnormality is detected depending on whether or not the estimated waveform obtained by estimating the mechanical information by the estimation model as shown in FIG. 12 exceeds the threshold for abnormality diagnosis in the identifiable interval determined in S13. Here, the threshold value for abnormality diagnosis to be determined may be the same value as the identifiable section setting threshold value (0.2) as shown in FIG. This has the effect of preventing erroneous detection of abnormalities. When an abnormality is detected, the machine operation command unit 12 issues an alarm, notifies of the abnormality, and performs a retraction operation.

Also, a different method of abnormality diagnosis processing will be described with reference to the flowchart of FIG. 14 . Steps S21 to S23 in FIG. 14 are the same as steps S11 to S13 in FIG. 10, and identifiable sections are set. Here, the abnormality diagnosis steps after S24 are different.
At S24, teacher data is extracted in order to update the estimation model. Extraction of teacher data is performed by identifiable sections of data whose processing state is confirmed to be normal. In the case of the estimation results shown in FIG. 12, the ranges of 10 to 26 sec, 42 to 53 sec, and 65 to 75 sec are used as teacher data for re-learning.
Next, in S25, the estimation model is updated based on the teacher data extracted in S24. For example, when a neural network is used, by using the weights and biases of the estimation model described above as initial values, it is possible to perform fine tuning according to the current diagnostic target while retaining the characteristics of the original estimation model.

Next, in S26, an abnormality diagnosis is performed in the same manner as in S14 of FIG. The machine information is estimated at any time by the updated estimation model updated in S25, and abnormality is detected depending on whether or not the threshold value is exceeded in the identifiable section defined in S23.
Here, since the estimation model is updated in accordance with the identifiable interval in S25, the degree of anomaly in the entire identifiable interval is lowered, and there is no need to provide a margin for the threshold value, so that anomalies can be discriminated with higher sensitivity. can be done.

As described above, even in the abnormality diagnosis device 4A of the second embodiment, it is possible to effectively detect a machining abnormality even in machining in which the machining mode is not constant and changes greatly. Moreover, even if the processing mode of the model creation range, which is used as teacher data when creating the estimation model, does not sufficiently cover the range of abnormality monitoring, it is possible to effectively diagnose anomalies. Also, when using the estimation model, the identification can be done without knowing the scope of application.
In particular, here, even with a small amount of data, it is possible to extract data from locations where effective learning can be performed as teacher data, so data can be effectively used to improve the accuracy of the estimation model. In addition, since the calculation cost can be reduced, the estimation model can be tuned without using a large-scale computer. Furthermore, since fine adjustment can be performed according to the abnormality detection target, abnormality detection can be performed with higher accuracy.

In the second embodiment, the method of creating an estimation model is incorporated into a series of flows, but if an estimation model already exists, that estimation model may be used. Furthermore, the estimation model creator and the estimation model user who performs anomaly detection need not be the same person.

Next, a method of creating an estimation model by the model creating unit 6 in the forms 1 and 2 will be described based on the flow chart of FIG.
First, in S31, labeled data is created by adding an operation result indicating whether the operation of the machine tool 1 was normal to the record information recorded in the storage unit 8 (labeled data creating step).
Next, in S32, master data is created (master data creation step). From the master data, only the locations that contain the features linked to the operation results are extracted. By creating data that excludes portions that do not contain features that are effective in discriminating motion results, it is possible to efficiently construct an estimation model. Taking the example of creating an estimation model based on the data obtained by machining the side surface of the work W shown in FIG. Features are often absent, complicating model construction. Therefore, in order to construct the estimation model efficiently, it is effective to use as master data a portion where the machining state is steady, such as between 10 and 20 seconds in FIG.

Next, in S33, a master model is created (master model setting step). Here, a method of constructing a model that outputs the degree of abnormality by an autoencoder will be described as an example. A master model is created using only those with normal machining results from the master data. The change from the normal state is expressed by the difference between the input and output of the trained estimation model. At this time, instead of creating a master model using all master data, master learning data for learning and master evaluation data for evaluation are created. The method of defining the estimation model is not limited to the above-mentioned method, and it is preferable to use an algorithm such as machine learning or deep learning that has high performance for the target data. For example, it is possible to perform clustering using Gaussian Mixture Model, Approximate Gaussian Mixtures, etc., define the distribution of normal states, and output the degree of deviation from normal features to estimate abnormalities. In addition, a support vector machine, nearest neighbor method, random forest method, naive Bayes method, etc. may be used.

Next, in S34, machining conditions different from the master data and machine information when machining the work material are estimated by the master model. An extraction threshold is set for the estimation result, and machine information corresponding to a range below the extraction threshold is used as additional training data (additional training data creation step). An example of the procedure for creating a model for estimating the degree of machining abnormality based on the spindle load (FIG. 6) obtained when side cutting a workpiece W1 having a shape as shown in FIG. 5 with an end mill T will be described. .
When the aforementioned spindle load is estimated using the master model created in S33, an estimation result (abnormality degree) as shown in FIG. 7 is obtained. Therefore, an extraction threshold is provided as shown in FIG. 16, and 10 to 26 sec, 42 to 53 sec, and 65 to 75 sec in which the estimation results are below the extraction threshold are extracted as additional learning data.

Next, in S35, the additional teacher data is added to the master data and relearning is performed to create a relearning model (relearning model creating step). The re-learning procedure is the same as in S33.
Next, in S36, the created master model is evaluated (model performance evaluation step). The evaluation of the master model is performed by estimating the master data and determining whether it matches the actual operation result.
Next, in S37, based on the evaluation result in S36, it is determined whether or not to re-learn the master model (re-learning execution determination step). Here, if the estimation performance of the relearning model is inferior to that of the master model, it is determined that the additional teacher data was not effective, and in S38, the master model is adopted instead of the relearning model (employed model determination step). On the other hand, if relearning can be performed without degrading the estimation performance of the master model, it is determined that the expansion of applicable machining has succeeded, the master model is updated with the relearning model, and steps S34 to S37 are repeated. The scope of application can be expanded by creating additional training data again. In addition, since there is a trade-off between the expansion of the scope of application and the accuracy of estimation, if the expansion of the scope of application is more important than strict monitoring, set a certain amount of allowance for performance degradation from the master model. Also good.

Here, when re-learning is performed in the determination of S37, the re-learning condition is changed (S39). That is, when the relearning model does not reach the estimated performance of the master model in S37, the extraction threshold value is lowered in the additional training data creation step of S34. By resetting the threshold value for extraction, the range of additional training data is changed to be smaller, and improvement in estimation accuracy can be expected.
In this example, the process of creating a master model is included in a series of flows, but an existing estimation model that matches the estimation target may be used.
As described above, according to the method of creating an estimation model of the above embodiment, it is possible to create an estimation model with an expanded range of application while maintaining the estimation performance without performing detailed labeling. In addition, it is possible to reduce the cost required for selecting effective learning data and reduce the increase in calculation cost due to using a large amount of data that does not contain the feature value corresponding to the operation result.
In particular, since the extraction threshold value for determining the estimation result is set at the same time, it is possible to set an effective extraction threshold value while performing learning.

In each form, the change in the spindle load in side milling with an end mill was explained as an example, but the machining form is not limited to this, and can also be applied to drilling, tap milling, and turning. The sensor output value is not limited to this, and any sensor such as the load of the feed shaft or the output of the vibration sensor may be used.

DESCRIPTION OF SYMBOLS 1... Machine tool, 2... Spindle, 3... Workpiece, 4, 4A... Abnormal diagnosis device, 5... Machine information acquisition part, 6... Model preparation part, 7... Machining form judgment part, 8 Storage unit 9 Model creation condition input unit 10 Monitoring setting input unit 11 Interpretation unit 12 Machine operation command unit 13 Abnormal diagnosis unit 14 Existing data input unit , W, W1... work, T... end mill.

Claims (7)

In a machine tool having a rotary shaft for rotating a tool or a work and a feed shaft for relatively moving the tool and the work, the work is machined by the tool based on a programmed command. A system for diagnosing
a machine information acquisition unit that acquires at least one of operation information of the machine tool and output information of various sensors attached to the machine tool as machine information;
a storage unit for recording the machine information;
a model creation unit that creates an estimation model for estimating the degree of abnormality from the recorded mechanical information;
The degree of abnormality in the machine information is estimated at any time by the created estimation model, and the presence or absence of abnormality is determined by comparing the time change of the estimated value with a threshold set so as to change with time according to the machining mode. an abnormality diagnosis unit for judging;
and a processing form judgment unit that judges the identity between the processing forms ,
When the machining determined to be the same by the machining mode determining part is repeatedly performed, the abnormality diagnosing part uses the estimated value obtained by estimating the machine information when the machining to be diagnosed is normally completed by the estimation model as needed. An abnormality diagnosis system for a machine tool , wherein the threshold value is set based on a series waveform . The model creation unit is characterized in that it extracts, as teacher data, data in which the change in the machining mode determined by the machining mode determination part falls within a predetermined range, and creates the estimated model from the teacher data. An abnormality diagnosis system for a machine tool according to claim 1 . The abnormality diagnosis unit records a time-series waveform obtained by estimating the machine information during normal operation with the estimation model in the storage unit as a master waveform, and sets a predetermined identifiable section for the master waveform. recording an interval below the threshold in the storage unit as an identifiable interval that can be processed using an algorithm;
2. An abnormality diagnosis system for a machine tool according to claim 1, wherein the time variation of said estimated value and said threshold value are compared only in said identifiable interval. 4. The machine tool abnormality according to claim 3 , wherein the model creation unit extracts the machine information of the identifiable section as teaching data, and updates the estimation model using the teaching data. diagnostic system. In a machine tool having a rotary shaft for rotating a tool or a work and a feed shaft for relatively moving the tool and the work, the work is machined by the tool based on a programmed command. A method of diagnosing
a machine information acquisition step of acquiring at least one of operation information of the machine tool and output information of various sensors attached to the machine tool as machine information;
a recording step of recording the machine information;
a model creation step of creating an estimation model for estimating the degree of abnormality from the recorded mechanical information;
The degree of abnormality in the machine information is estimated at any time by the created estimation model, and the presence or absence of abnormality is determined by comparing the time change of the estimated value with a threshold set so as to change with time according to the machining mode. while executing an abnormality diagnosis step for determining,
In the model creation step,
a labeled data creation step of creating labeled data in which an operation result indicating whether or not the machine tool is operating normally and the machine information are linked;
a master data creation step of creating master data by extracting from the labeled data only portions containing features corresponding to the operation result;
a master model setting step of creating a master model for estimating the operation result from the master data or adopting an existing learned master model;
an additional training data creation step of estimating the labeled data using the master model and extracting from the estimation result a location containing a feature corresponding to the operation result to create additional training data;
a relearning model creation step of creating a relearning model for estimating the motion result from the master data and the additional teacher data;
a model performance evaluation step of performing an estimation performance evaluation of the relearning model using the master model;
If the estimated performance is inferior to the master model, the master model is adopted as the estimated model, and if the estimated performance exceeds the master model, the master model is updated with a re-learned model, and the additional teacher is used. a relearning execution determination step that repeats the data creation step and the relearning model creation step;
A method for diagnosing an abnormality in a machine tool, comprising: In the additional training data creation step executed after the re-learning execution determination step, an extraction threshold value for extracting a portion containing a feature in the estimation result is reset, and the evaluation result in the model performance evaluation step is set. 6. The machine tool abnormality diagnosis method according to claim 5 , wherein said extraction threshold value is changed in accordance with . An abnormality diagnosis program for causing a computer to execute the machine tool abnormality diagnosis method according to claim 5 or 6 .

Images