Model learning guide method for machine abnormality detection

Abstract

A model learning guiding method for machine abnormality detection. The model learning guiding method comprises: receiving running data of a machine to be detected; determining, by a new behavior detection model, whether a new behavior event occurs in the machine to be detected based on the running data; in a case where it is determined that the new behavior event occurs in the machine to be detected, determining, based on the running data, whether the new behavior event belongs to a first type of event or a second type of event; in response to determining that the new behavior event belongs to the first type of event, transmitting the running data to an abnormality detection model to retrain the abnormality detection model; and in response to determining that the new behavior event belongs to the second type of event, transmitting the running data to the new behavior detection model to retrain the new behavior detection model.

Description

Technical Field

[0001] This invention relates to a method for model training, and more particularly to a model learning guidance method for machine tool anomaly detection. Background Technology

[0002] Machine malfunctions can cause losses for businesses. Therefore, to prevent unexpected equipment failures, with the development of technologies such as sensors, IoT, and big data analytics, fault prediction and health management have shifted from relying on regular inspections and experienced technicians' judgments to establishing detection models for real-time monitoring, objective data analysis, and lifespan prediction. However, in the initial stages of establishing a detection model, the lack of training data with anomaly labels can prolong the actual operation time of the model. Even after actual operation, the detection model requires frequent retraining to address data drift and model failure caused by factors such as changes in the tested equipment. Furthermore, different users' subjective perceptions of data normality and abnormality can lead to inconsistencies in the logic and quality of labeling, thus affecting model effectiveness. Summary of the Invention

[0003] This invention provides a model learning guidance method for machine tool anomaly detection, which can automatically update and maintain the model.

[0004] The present invention provides a model learning guidance method for machine anomaly detection, executed by a processor. The method includes: receiving operational data of the machine under test; determining, based on the operational data, whether a new behavioral event has occurred on the machine under test using a new behavior detection model; if a new behavioral event is determined to have occurred, determining, based on the operational data, whether the new behavioral event belongs to a first type of event or a second type of event; in response to determining that the new behavioral event belongs to a first type of event, transmitting the operational data to at least one anomaly detection model to retrain the at least one anomaly detection model; and in response to determining that the new behavioral event belongs to a second type of event, transmitting the operational data to a new behavior detection model to retrain the new behavior detection model.

[0005] In one embodiment of the present invention, when a new behavioral event is determined to have occurred on the machine under test, the method includes: displaying a user interface through a new behavior detection model, wherein the user interface includes multiple options corresponding to a first type of event and a second type of event; and after one of the options is selected, displaying a tagging page through the user interface to receive tagging selections through the tagging page, wherein the tagging selections are used to tag the running data with a normal tag or an abnormal tag.

[0006] In one embodiment of the present invention, the anomaly detection model includes a health detection model and a test object replacement detection model. The health detection model is used to determine whether the operating data is abnormal, and the test object replacement detection model is used to determine whether the test object has been replaced based on the operating data. The first type of event includes abnormal replacement events, planned replacement events, and data anomaly events. The second type of event includes maintenance events and no-abnormality events. The options include a first option, a second option, a third option, a fourth option, and a fifth option, corresponding to abnormal replacement events, planned replacement events, data anomaly events, maintenance events, and no-abnormality events, respectively. The first option is used to mark the operating data obtained before replacing the test object with an abnormal label, mark the operating data obtained after replacing the test object with a normal label, and transmit the operating data marked with normal and abnormal labels to the health detection model; and transmit the operating data marked with normal labels to the test object replacement detection model. The second option is used to mark the operating data obtained after replacing the test object with a normal label, and transmit the operating data marked with normal labels to the test object replacement detection model. The third option is used to label the running data as abnormal and then send the labeled running data to the health detection model. The fourth and fifth options are used to label the running data as normal and then send the labeled running data to the new behavior detection model to retrain the new behavior detection model.

[0007] In one embodiment of the present invention, the anomaly detection model is a health detection model. When a new behavioral event is determined to occur on the machine under test, the step of determining whether the new behavioral event belongs to a first type of event or a second type of event based on the operating data includes: determining whether the operating data is abnormal through the health detection model; in response to determining that the operating data is abnormal, determining that the new behavioral event belongs to a first type of event, and sending the abnormal operating data marked with an anomaly label to the health detection model for retraining; and in response to determining that the operating data is normal, determining that the new behavioral event belongs to a second type of event, and sending the normal operating data marked with a normal label to the new behavior detection model for retraining.

[0008] In one embodiment of the present invention, the anomaly detection model is a test object replacement detection model. When a new behavioral event is determined to occur on the test machine, the step of determining whether the new behavioral event belongs to a first type of event or a second type of event based on the operating data includes: determining whether the test object has been replaced based on the operating data using the test object replacement detection model; in response to determining that the test object has been replaced, determining that the new behavioral event belongs to a first type of event, and marking the operating data obtained after replacing the test object with a normal label before transmitting it to the test object replacement detection model for retraining; and in response to determining that the test object has not been replaced, determining that the new behavioral event belongs to a second type of event, and marking the operating data of the test object that has not been replaced with a normal label before transmitting it to the new behavior detection model for retraining.

[0009] In one embodiment of the present invention, responding to the determination that the tested object has been replaced includes: determining whether the tested object has been abnormally replaced based on user selection. If the tested object has been abnormally replaced: marking the operating data obtained before the replacement with an abnormal label, and marking the operating data obtained after the replacement with a normal label; transmitting the operating data marked with normal and abnormal labels to a health detection model; and transmitting the operating data marked with normal labels to a tested object replacement detection model. If the tested object has not been abnormally replaced, marking the operating data obtained after the replacement with a normal label, and transmitting the operating data marked with normal labels to a tested object replacement detection model.

[0010] In one embodiment of the present invention, the anomaly detection model includes a health detection model and a test object replacement detection model. The health detection model is used to determine whether the operating data is abnormal, and the test object replacement detection model is used to determine whether the test object has been replaced based on the operating data. The first type of event includes test object replacement events and data anomaly events, and the second type of event includes data normal events. When a new behavioral event is determined to occur on the test machine, the step of determining whether the new behavioral event belongs to the first type or the second type of event based on the operating data includes: using the health detection model and the test object replacement detection model, based on the operating data, determining whether the new behavioral event belongs to the first type or the second type of event. Specifically, operating data determined to be test object replacement events are labeled as normal and then transmitted to the test object replacement detection model for retraining; operating data determined to be data anomaly events are labeled as abnormal and then transmitted to the health detection model for retraining; and operating data determined to be data normal events are labeled as normal and then transmitted to the new behavioral detection model for retraining.

[0011] In one embodiment of the present invention, when a health detection model is first triggered to determine whether the operating data is abnormal, and then a test object replacement detection model is triggered to determine whether the test object has been replaced, the method includes: in response to determining that the operating data is abnormal and that the test object has been replaced, determining that the new behavioral event belongs to the test object replacement event; in response to determining that the operating data is abnormal and that the test object has not been replaced, determining that the new behavioral event belongs to the data abnormality event; in response to determining that the operating data is normal and that the test object has been replaced, determining that the new behavioral event belongs to the test object replacement event; and in response to determining that the operating data is normal and that the test object has not been replaced, determining that the new behavioral event belongs to the data normality event.

[0012] In one embodiment of the present invention, when a test object replacement detection model is first triggered to determine whether the test object has been replaced, and then a health detection model is triggered to determine whether the operating data is abnormal, the method includes: in response to determining that the test object has been replaced and the operating data is normal, determining that the new behavioral event belongs to the test object replacement event; in response to determining that the test object has been replaced and the operating data is abnormal, determining that the new behavioral event belongs to the data abnormal event; in response to determining that the test object has not been replaced and the operating data is abnormal, determining that the new behavioral event belongs to the data abnormal event; and in response to determining that the test object has not been replaced and the operating data is normal, determining that the new behavioral event belongs to the data normal event.

[0013] In one embodiment of the present invention, when it is determined that no new behavioral event has occurred on the machine under test, the method further includes: if the anomaly detection model includes a health detection model, determining whether the operating data is abnormal through the health detection model; if the anomaly detection model includes a test object replacement detection model, determining whether the test object has been replaced based on the operating data through the test object replacement detection model.

[0014] Based on the above, this disclosure uses a new behavior detection model to identify behaviors that have never occurred, so that the anomaly detection model and / or the new behavior detection model can be retrained and the model can be automatically updated and maintained. Attached Figure Description

[0015] Figure 1 This is a block diagram of an electronic device according to an embodiment of the present invention.

[0016] Figure 2 This is a flowchart of the model learning guidance method according to the first embodiment of the present invention.

[0017] Figure 3 This is a flowchart of a model learning guidance method according to the second embodiment of the present invention.

[0018] Figure 4A This is a flowchart of a model learning guidance method according to the third embodiment of the present invention.

[0019] Figure 4B This is a flowchart of a model learning guidance method according to an application example of the third embodiment of the present invention.

[0020] Figure 5 This is a flowchart of the model learning guidance method according to the fourth embodiment of the present invention.

[0021] Figure 6 This is a flowchart of a model learning guidance method according to the fifth embodiment of the present invention.

[0022] Figure 7 This is a flowchart of the model learning guidance method according to the sixth embodiment of the present invention.

[0023] Explanation of reference numerals in the attached figures:

[0024] 100: Electronic devices

[0025] 110: Processor

[0026] 120: Storage device

[0027] L1: Abnormal Replacement Event

[0028] L2: Planned Replacement Event

[0029] L3: Data Anomaly Events

[0030] L4: Maintenance Event

[0031] L5: No Abnormal Events

[0032] M1: Test object replacement detection model

[0033] M2: Health Detection Model

[0034] M3: Behavior Detection Model

[0035] S205~S225: Steps of the model learning guidance method in the first embodiment

[0036] S305~S340: Steps of the Model Learning Guidance Method in the Second Embodiment

[0037] S405~S440: Steps of the Model Learning Guidance Method in the Third Embodiment

[0038] S505~S560: Steps of the Model Learning Guidance Method in the Fourth Embodiment

[0039] S605~S660: Steps of the Model Learning Guidance Method in the Fifth Embodiment

[0040] S705~S730: Steps of the Model Learning Guidance Method in the Sixth Embodiment Detailed Implementation

[0041] Figure 1 This is a block diagram of an electronic device according to an embodiment of the present invention. Please refer to... Figure 1 The electronic device 100 includes a processor 110 and a storage device 120. The processor 110 is coupled to the storage device 120. In one embodiment, the electronic device 100 monitors the device under test (DUT) via wired or wireless transmission. For example, the electronic device 100 receives operating data from the DUT at regular sampling intervals to determine whether new behavioral events have occurred on the DUT. Alternatively, in other embodiments, the DUT's own processor (i.e., the electronic device 100 is the DUT) can directly monitor its own operating data.

[0042] The processor 110 may be, for example, a central processing unit (CPU), a physical processing unit (PPU), a programmable microprocessor, an embedded control chip, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a programmable logic controller (PLC), or other similar devices.

[0043] Storage device 120 is, for example, any form of fixed or removable random access memory (RAM), read-only memory (ROM), flash memory, hard disk, or other similar device or combination thereof. The storage device includes one or more code segments, which, after being installed, are executed by processor 110 to perform the steps of the model learning guidance method for machine anomaly detection described below.

[0044] Figure 2 This is a flowchart of the model learning guidance method according to the first embodiment of the present invention. Please refer to... Figure 2 In step S205, the operating data of the machine under test is received. For example, at regular sampling intervals, the processor 110 acquires the operating data of the machine under test within the sampling time. The operating data includes parameters of the object (part) under test, machine settings, environmental settings, etc. Machine settings include, for example, motor rotation speed, vibration frequency, and voltage; environmental settings include, for example, temperature and humidity. The above is only an example; the operating data will vary depending on the machine under test.

[0045] Next, in step S210, the new behavior detection model determines whether a new behavior event has occurred on the machine under test based on the running data. In one embodiment, the new behavior detection model employs a machine learning algorithm. Machine learning algorithms are used to inductively derive useful rules from large amounts of data. For example, the new behavior detection model can be implemented using a learning algorithm based on Principal Components Analysis (PCA), an autoencoder, or a one-class support vector machine (SVM). In one embodiment, the new behavior detection model is trained using only running data of a single category labeled with normal tags.

[0046] In the event of a new behavioral event, step S215 determines whether the new behavioral event belongs to a first type of event or a second type of event. In one embodiment, the new behavior detection model may further provide a user interface. The user interface includes multiple options corresponding to the first type of event and the second type of event. When a new behavioral event is detected, the user interface is displayed through the new behavior detection model. After one of the options is selected, a marking page is displayed through the user interface to receive the marking selection. Here, the marking selection is used to mark the running data with a normal label or an abnormal label. Additionally, the user interface may further send text or voice messages to remind the user to mark the new behavioral event.

[0047] In another embodiment, at least one anomaly detection model can be used to determine whether a new behavioral event belongs to a first type of event or a second type of event (for details, please refer to the second to fifth embodiments described below). The anomaly detection model is, for example, a model employing a machine learning algorithm.

[0048] In step S220, in response to determining that the new behavioral event belongs to the first category of events, the runtime data is transmitted to the anomaly detection model to retrain it. The retrained anomaly detection model can then be used to determine whether the runtime data belongs to the first or second category of events. The anomaly detection model can be a health detection model or a device-to-measure (DTM) replacement detection model. The health detection model is used to determine whether the runtime data is abnormal. The DTM replacement detection model is used to determine whether the DTM has been replaced based on the runtime data.

[0049] Both the health status detection model and the object replacement detection model employ machine learning algorithms. For example, the health status detection model can be implemented using PCA-based learning algorithms, autoencoders, unary classification support vector machines, XGBoost (extreme gradient boosting) algorithms, LightGBM (light gradient-boosting machine) algorithms, CatBoost (categorical boosting) algorithms, or random forest algorithms. Similarly, the object replacement detection model can be implemented using XGBoost, LightGBM, CatBoost, or random forest algorithms.

[0050] In step S225, in response to determining that the new behavioral event belongs to the second type of event, the running data is transmitted to the new behavior detection model to retrain the new behavior detection model. At the next sampling time point, the retrained new behavior detection model can be used to detect whether a new behavioral event has occurred.

[0051] On the other hand, before a new behavioral event is determined to have occurred (step S210: No), the processor 110 returns to step S205 and continues to sample at intervals, obtaining the operational data between the current sampling time point and the previous sampling time point, and triggering an anomaly detection model to determine whether the operational data belongs to the first category of events or the second category of events. Accordingly, even when a known behavior is detected (i.e., no new behavioral event has occurred), the anomaly detection model can continuously monitor whether the operational data is abnormal or whether the tested object has been replaced. For example, a health detection model can be used to determine whether the operational data is abnormal; or, a tested object replacement detection model can be used to determine whether the tested object has been replaced based on the operational data.

[0052] Figure 3 This is a flowchart of the model learning guidance method according to the second embodiment of the present invention. The second embodiment is an application example of the first embodiment, and further explains how to determine which models to retrain when a new behavioral event is detected, provided that a health detection model is used as the anomaly detection model. Please refer to... Figure 3 In step S305, a health detection model is triggered upon detecting a new behavioral event. Next, in step S310, the health detection model is used to determine whether the operational data is abnormal.

[0053] In response to the determination that the running data is abnormal, in step S315, the new behavioral event is determined to belong to the first type of event. Here, the first type of event is a data anomaly event. Next, in step S320, the running data determined to be abnormal is marked with an anomaly label. Afterwards, in step S325, the marked running data is transmitted to the health detection model to retrain the health detection model.

[0054] In response to the determination that the operating data is normal, in step S330, the new behavioral event is determined to belong to the second type of event. Here, the second type of event is a maintenance event or a no-abnormality event. For example, an option can be provided in the user interface for the user to select a maintenance event or a no-abnormality event. Next, in step S335, the operating data determined to be normal is labeled with a normal tag. Afterwards, in step S340, the labeled operating data is transmitted to the new behavior detection model to retrain the new behavior detection model.

[0055] In addition, if no new behavioral events are detected, the health detection model will be triggered to determine whether the current running data is abnormal.

[0056] The following third embodiment illustrates another application example of the first embodiment. Figure 4A This is a flowchart of the model learning guidance method according to the third embodiment of the present invention. In the third embodiment, under the premise of using a test object replacement detection model as the anomaly detection model, it further explains how to determine which models to retrain when a new behavioral event is detected. Please refer to... Figure 4A In step S405, the test object replacement detection model is triggered when a new behavioral event is detected. Next, in step S410, the test object replacement detection model determines whether the test object has been replaced based on the running data.

[0057] In response to the determination that the test object has been replaced, in step S415, the new behavioral event is determined to belong to the first type of event. Here, the first type of event is the test object replacement event. Next, in step S420, the operating data obtained after the test object is replaced (i.e., the operating data obtained at the current sampling time point) is labeled with a normal label. Afterwards, in step S425, the labeled operating data is transmitted to the test object replacement detection model to retrain the test object replacement detection model.

[0058] In response to the determination that the object under test has not been replaced, in step S430, the new behavioral event is determined to belong to the second type of event. Next, in step S435, the running data of the determined object under test has not been replaced is marked with a normal label. Afterwards, in step S440, the marked running data is transmitted to the new behavior detection model to retrain the new behavior detection model.

[0059] Furthermore, if it is determined that the tested object has been replaced, it can be further determined whether the replacement was abnormal based on user selection, such as... Figure 4B As shown. For example, the user interface provides options for the user to choose between an abnormal change and a planned change. If the object under test is determined to be an abnormal change, the new behavioral event is considered an abnormal change event. If the object under test is determined not to be an abnormal change, the new behavioral event is considered a planned change event.

[0060] Figure 4B This is a flowchart of a model learning guidance method according to an application example of the third embodiment of the present invention. Figure 4B In this embodiment, in step S450, the test object replacement detection model is triggered upon detecting a new behavioral event. Next, in step S452, the test object replacement detection model determines whether the test object has been replaced based on operational data. In response to the determination that the test object has been replaced (step S452: Yes), the new behavioral event is determined to belong to the first type of event, and in step S454, it is determined whether the test object has undergone an abnormal replacement.

[0061] If the tested object is determined to have been replaced abnormally (abnormal replacement event), in step S456, the operating data obtained before the replacement (e.g., each set of operating data and classification results are recorded in storage device 120) is labeled as abnormal, and the operating data obtained after the replacement (operating data obtained at the current sampling time point) is labeled as normal. Furthermore, in step S458, the operating data labeled as normal and abnormal are transmitted to the health detection model to retrain it. Additionally, in step S462, only the operating data labeled as normal is transmitted to the tested object replacement detection model to retrain it.

[0062] If it is determined that the replacement of the device under test is not an abnormal event (planned replacement event), in step S460, the operating data obtained after replacing the device under test is labeled as normal. In step S462, the operating data labeled as normal is transmitted to the device under test replacement detection model to retrain the device under test replacement detection model.

[0063] In response to the determination that the tested object has been replaced (step S452: No), the new behavioral event is determined to be a second type of event, and in step S454, it is determined whether the tested object has been abnormally replaced. In step S464, the running data determined that the tested object has not been replaced is marked with a normal label. Then, in step S466, the running data marked with the normal label is transmitted to the new behavior detection model to retrain the new behavior detection model.

[0064] Furthermore, even without detecting new behavioral events, the test object replacement detection model will be triggered to determine whether the test object has been replaced. In one embodiment, it can be further configured such that the test object replacement detection model can independently determine whether the test object has been replaced after it is launched online, rather than being used to determine whether the test object has been replaced only when a new behavioral event is detected. However, the decision to retrain the test object replacement detection model is based on the subsequent judgment result of the detected new behavioral event. In addition, whenever it is determined that the test object has been replaced, the health detection model needs to be retrained. For example, the user interface provides an option for the user to select whether the detected replacement is an abnormal replacement or a planned replacement. If an abnormal replacement is selected, the processor 110 will further mark the operating data obtained before the replacement of the test object with an abnormal label and the operating data obtained after the replacement of the test object with a normal label, and then send it to the health detection model to retrain the health detection model.

[0065] In another embodiment, two anomaly detection models (health detection model and object replacement detection model) can be used to determine whether a new behavioral event belongs to the first type of event or the second type of event.

[0066] Figure 5 This is a flowchart of the model learning guidance method according to the fourth embodiment of the present invention. The fourth embodiment is an application example of the first embodiment. It further explains how to determine which models to retrain when a new behavioral event is detected, given that both a health detection model and a test object replacement detection model are used as anomaly detection models. In the fourth embodiment, the health detection model is first triggered to determine whether the running data is abnormal, and then the test object replacement detection model is triggered to determine whether the test object has been replaced. Furthermore, the first type of event is a test object replacement event or a data anomaly event, and the second type of event is a data normal event.

[0067] Please refer to Figure 5 In step S505, a health detection model is triggered upon detecting a new behavioral event. Next, in step S510, the health detection model determines whether the operational data is abnormal.

[0068] If the operating data is determined to be abnormal, in step S515, the detection model for the tested object is replaced. In step S520, based on the abnormal operating data and the detection model for the tested object being replaced, it is determined whether the tested object has been replaced.

[0069] In response to the determination that the operational data is abnormal and that the tested object has not been replaced, in step S525, the new behavioral event is determined to be a data anomaly event, and the operational data is marked with an anomaly label. Then, the operational data marked with the anomaly label is transmitted to the health detection model for retraining in step S530.

[0070] In response to the determination that the operating data is abnormal and that the tested object has been replaced, in step S535, the new behavioral event is determined to be a tested object replacement event. The operating data obtained before the tested object was replaced is labeled as abnormal, and the operating data obtained after the tested object was replaced is labeled as normal. Then, the operating data labeled as normal and abnormal are transmitted to the health detection model for retraining in step S530. Alternatively, only the operating data labeled as normal is transmitted to the tested object replacement detection model for retraining in step S540.

[0071] On the other hand, if the operating data is determined to be normal, in step S545, the detection model of the tested object is triggered to be replaced. In step S550, based on the operating data determined to be normal, it is determined whether the tested object has been replaced by replacing the detection model of the tested object.

[0072] In response to the determination that the running data is normal and that the tested object has been replaced, steps S535 and S540 are executed. In response to the determination that the running data is normal and that the tested object has not been replaced, in step S555, the new behavioral event is determined to be a normal data event, and the running data is labeled as normal. Then, the labeled running data is transmitted to the new behavior detection model and the health detection model, so that the health detection model and the new behavior detection model can be retrained in steps S530 and S560, respectively.

[0073] Figure 6 This is a flowchart of the model learning guidance method according to the fifth embodiment of the present invention. The fifth embodiment is an application example of the first embodiment. It further explains how to determine which models to retrain when a new behavioral event is detected, given that both a health detection model and a test object replacement detection model are used as anomaly detection models. In the fifth embodiment, the test object replacement detection model is first triggered to determine whether the test object has been replaced, and then the health detection model is triggered to determine whether the running data is abnormal. Furthermore, the first type of event is a test object replacement event or a data anomaly event, and the second type of event is a data normal event.

[0074] Please refer to Figure 6 In step S605, a health detection model is triggered upon detecting a new behavioral event. Next, in step S510, the detection model for the object under test is used to determine, based on operational data, whether the object under test has been replaced.

[0075] If it is determined that the tested object has been replaced, the health detection model is triggered in step S615. In step S620, the health detection model is used to determine whether the operating data is abnormal.

[0076] In response to the determination that the tested object has been replaced and the determination that the operating data is normal, in step S625, the new behavioral event is determined to be a planned replacement event (a type of tested object replacement event), and the operating data is marked with a normal label. Then, the marked operating data is transmitted to the tested object replacement detection model and the health detection model, so that in steps S630 and S640 respectively, the tested object replacement detection model and the health detection model are retrained.

[0077] In response to the determination that the tested object has been replaced and the operational data is abnormal, in step S627, the new behavioral event is determined to be an abnormal replacement event (one type of tested object replacement event), and the operational data obtained before the tested object was replaced is labeled as abnormal, while the operational data obtained after the tested object was replaced is labeled as normal. Then, the operational data labeled as normal and abnormal are transmitted to the health detection model for retraining in step S640. Alternatively, only the operational data labeled as normal is transmitted to the tested object replacement detection model for retraining in step S630.

[0078] If it is determined that the tested object has not been replaced, the health detection model is triggered in step S645. In step S650, the health detection model is used to determine whether the operating data is abnormal.

[0079] In response to the determination that the tested object has not been replaced and the determination that the operating data is abnormal, steps S635 and S640 are executed. In step S635, the new behavioral event is determined to be a data abnormality event, and the operating data is marked with an abnormal label. Then, the marked operating data is transmitted to the health detection model for retraining in step S640. In response to the determination that the tested object has not been replaced and the determination that the operating data is normal, in step S655, the new behavioral event is determined to be a data normality event, and the operating data is marked with a normal label. Then, in step S660, the marked operating data is transmitted to the new behavior detection model for retraining.

[0080] Figure 7 This is a flowchart of the model learning guidance method according to the sixth embodiment of the present invention. The sixth embodiment is an application example of the first embodiment. In the sixth embodiment, it is assumed that the first type of events includes abnormal replacement event L1, planned replacement event L2, and data anomaly event L3. The second type of events includes maintenance event L4 and no abnormality event L5.

[0081] When the new behavior detection model M3 detects a behavior that has never been seen before (a new behavior event is detected), it triggers a marking mechanism to ask the user which category the detected new behavior belongs to. Here, the new behavior detection model M3 further provides a user interface, which includes five options (options one through five) corresponding to abnormal replacement events L1, planned replacement events L2, data anomaly events L3, maintenance events L4, and no-abnormality events L5, allowing the user to choose one. After selecting one of the five options, a marking page is displayed on the user interface to receive the marking selection. Here, the marking selection can be done manually by the user on the marking page, or automatically by the processor 110 through the marking page. If automatic marking is inappropriate, manual marking by the user will then be required.

[0082] Please refer to Figure 7 If the new behavioral event is determined to be an abnormal replacement event L1 and the first option is selected in the user interface, a labeling page is displayed in the user interface to receive label selections. After receiving the label selections, the running data obtained before the replacement of the tested object is labeled as abnormal (step S705), and the running data obtained after the replacement of the tested object is labeled as normal (step S710). Then, only the running data labeled as normal is transmitted to the tested object replacement detection model M1 to retrain the tested object replacement detection model M1. Additionally, the two sets of running data labeled as normal and abnormal are transmitted to the health detection model M2 to retrain the health detection model M2.

[0083] If the new behavioral event is determined to be a planned replacement event L2 and the second option is selected in the user interface, the running data (obtained after replacing the test object) is marked with a normal label through the labeling page (step S715), and the labeled running data is transmitted to the test object replacement detection model M1 and the health detection model M2 respectively, so as to retrain the test object replacement detection model M1 and the health detection model M2.

[0084] If the new behavioral event is determined to be a data anomaly event L3 and the third option is selected in the user interface, the running data will be marked with an anomaly label (step S720), and the marked running data will be transmitted to the health detection model M2 to retrain the health detection model M2.

[0085] If the new behavior event is determined to be maintenance event L4 and the fourth option is selected in the user interface, the running data is marked with a normal label (step S725), and the marked running data is transmitted to the new behavior detection model M3 to retrain the new behavior detection model M3.

[0086] If the new behavioral event is determined to be a normal event L5 and the fifth option is selected in the user interface, the running data is labeled as normal (step S730), and the labeled running data is transmitted to the new behavior detection model M3 to retrain the new behavior detection model M3. If the detected running data is determined to correspond to a maintenance event L4 or a normal event L5, the new behavior detection model is updated, so that subsequent running data with the same state will not be regarded as a new behavioral event.

[0087] In one embodiment, before the runtime data labeled with abnormalities is imported, a semi-supervised learning algorithm based on unary classification (e.g., a learning algorithm based on PCA (Principal Components Analysis), an autoencoder, or a support vector machine for unary classification) is used to train the health detection model M2 only on runtime data labeled with normalities. After the runtime data labeled with normalities is imported, a supervised learning algorithm based on binary classification (e.g., the XGBoost algorithm, the LightGBM algorithm, the CatBoost algorithm, or the Random Forest algorithm) is used to train the health detection model M2.

[0088] In one embodiment, a semi-supervised learning approach using unary classification is employed, training the new behavior detection model M3 only on running data labeled with normal parameters.

[0089] In addition, when the accuracy of the test object replacement detection model M1 reaches a specified threshold, the retraining mechanism of the test object detection model M1 can be turned off (the marking action will no longer be performed). Subsequently, after the test object replacement detection model M1 automatically judges the test object replacement event, it feeds back the normal data after the test object replacement and the abnormal data marked in history to the health detection model M2, so as to automatically update the health detection model M2 and achieve the purpose of automatically maintaining the health detection model M2.

[0090] In summary, in the above embodiments, operational data is collected during the operation of the machine under test. When a new behavioral event that has never occurred before is detected by the new behavior detection model, the operational data belonging to the first type of event is fed back to the anomaly detection model, and the operational data belonging to the second type of event is fed back to the new behavior detection model. This achieves the goal of automatically updating and maintaining the model. Furthermore, this disclosure further uses the new behavior detection model to assist in the labeling of abnormal data, ensuring that the labeling logic of the data used for model training is consistent, thereby improving the accuracy of the model.

[0091] Furthermore, if no new behavioral events are detected, a health detection model can be used to continuously monitor whether the operational data is normal. Additionally, if no new behavioral events are detected, a device replacement detection model can be used to determine if the device has been replaced, and if an abnormal replacement is determined, the health detection model is retrained. Therefore, when known behaviors are detected (i.e., no new behavioral events occur), an anomaly detection model can be used to continuously monitor whether operational data shows anomalies or that the device has been replaced. When unknown behaviors are detected (i.e., new behavioral events), the new behavioral events are marked, and a decision is made to retrain the new behavior detection model and / or the anomaly detection model.

Claims

1. A model learning guidance method for machine tool anomaly detection, executed using a processor, the model learning guidance method comprising: Receive operational data from a machine under test; A new behavior detection model is used to determine whether a new behavior event has occurred on the machine under test based on the operational data. If the test machine is found to have a new behavioral event, the test machine is determined to be either a first-class event or a second-class event based on the operating data. In response to determining that the new behavioral event belongs to the first type of event, the running data is transmitted to at least one anomaly detection model to retrain the at least one anomaly detection model; as well as In response to determining that the new behavioral event belongs to the second type of event, the runtime data is transmitted to the new behavior detection model to retrain the new behavior detection model. The criteria for determining that the new behavioral event has occurred on the machine under test also include: The new behavior detection model displays a user interface, which includes multiple options corresponding to the first type of event and the second type of event; and After one of these options is selected, a labeling page is displayed through the user interface to receive a label selection, which is used to label the running data with a normal label or an abnormal label. The at least one anomaly detection model includes a health detection model and a test object replacement detection model. The health detection model is used to determine whether the operating data is abnormal, and the test object replacement detection model is used to determine whether a test object has been replaced based on the operating data. The first category of events includes an abnormal replacement event, a planned replacement event, and a data anomaly event; the second category of events includes a maintenance event and a no-abnormality event. These options include a first option, a second option, a third option, a fourth option, and a fifth option, corresponding to the abnormal replacement event, the planned replacement event, the data anomaly event, the maintenance event, and the no-abnormality event, respectively. The first option is used to mark the operational data obtained before replacing the test object with the abnormal label, mark the operational data obtained after replacing the test object with the normal label, and transmit the operational data marked with the normal label and the abnormal label to the health detection model; and transmit the operational data marked with the normal label to the test object replacement detection model. The second option is used to mark the operational data obtained after replacing the test object with the normal label, and to transmit the operational data marked with the normal label to the test object replacement detection model. The third option is used to mark the running data with the abnormal label and then send the marked running data to the health detection model. The fourth and fifth options are used to label the running data with the normal label and then send the labeled running data to the new behavior detection model to retrain the new behavior detection model.

2. The model learning guidance method as described in claim 1, wherein, for the health detection model, when it is determined that the new behavioral event has occurred on the machine under test, the step of determining whether the new behavioral event belongs to the first type of event or the second type of event based on the operating data includes: The health detection model is used to determine whether the operational data is abnormal. In response to the determination that the running data is abnormal, the new behavioral event is determined to belong to the first type of event, and the running data that is determined to be abnormal is marked with an abnormal label and then sent to the health detection model for retraining; In response to the determination that the running data is normal, the new behavior event is determined to belong to the second type of event, and the running data that is determined to be normal is marked with a normal label and then sent to the new behavior detection model for retraining.

3. The model learning guidance method as described in claim 1, wherein, when the detection model is changed for the object under test, and the new behavioral event is determined to occur on the machine under test, the step of determining whether the new behavioral event belongs to the first type of event or the second type of event based on the operating data includes: The detection model uses the operational data to determine whether a test object has been replaced. In response to the determination that the object under test has been replaced, the new behavioral event is determined to belong to the first type of event, and the running data obtained after the object under test is replaced is marked with a normal label and transmitted to the object under test replacement detection model for retraining. as well as In response to the determination that the object under test has not been replaced, the new behavior event is determined to belong to the second type of event. The running data that determines that the object under test has not been replaced is marked with a normal label and then sent to the new behavior detection model for retraining.

4. The model learning guidance method as described in claim 3, wherein responding to the determination that the test object has been replaced includes: Based on a user's selection, determine whether the tested item has been replaced abnormally; If the tested item is determined to have been replaced abnormally. The operational data obtained before replacing the test object is marked with an abnormal label, and the operational data obtained after replacing the test object is marked with a normal label; The running data marked with the normal label and the running data marked with the abnormal label are transmitted to a health detection model, wherein the health detection model is used to determine whether the running data is abnormal; as well as The operational data marked as normal will be transmitted to the tested object to replace the detection model; as well as If it is determined that the test object is not replaced abnormally, the operating data obtained after replacing the test object is marked with the normal label, and the operating data marked with the normal label is transmitted to the test object replacement detection model.

5. The model learning guidance method as described in claim 1, wherein, when it is determined that the new behavioral event has occurred on the test machine, the step of determining whether the new behavioral event belongs to the first type of event or the second type of event based on the operating data includes: Based on the operational data, using the health status detection model and the object replacement detection model, it is determined whether the new behavioral event belongs to the first type of event or the second type of event. The runtime data identified as a replacement event of the tested object is labeled as normal and then sent to the replaced object detection model for retraining. The runtime data identified as an anomaly event is labeled with an anomaly tag and then sent to the health detection model for retraining. The running data that is determined to be a normal event is labeled with the normal tag and then sent to the new behavior detection model for retraining.

6. The model learning guidance method as described in claim 5, wherein in the case of first triggering the health detection model to determine whether the running data is abnormal, and then triggering the test object replacement detection model to determine whether the test object has been replaced, the method includes: In response to determining that the operating data is abnormal and that the tested object has been replaced, the new behavioral event is determined to be a tested object replacement event; In response to the determination that the operational data is abnormal and that the tested object has not been replaced, the new behavioral event is determined to be a data abnormality event; In response to determining that the operating data is normal and that the tested object has been replaced, the new behavioral event is determined to be a tested object replacement event; as well as In response to the determination that the operational data is normal and that the tested object has not been replaced, the new behavioral event is determined to be a normal event for the data.

7. The model learning guidance method as described in claim 5, wherein, in the case of first triggering the detection model to change the test object to determine whether the test object has been changed, and then triggering the health detection model to determine whether the running data is abnormal, the method includes: In response to the determination that the tested object has been replaced and the determination that the operating data is normal, the new behavior event is determined to be a tested object replacement event; In response to the determination that the tested object has been replaced and the determination that the operating data is abnormal, the new behavioral event is determined to be a data abnormality event; In response to the determination that the tested object has not been replaced and the determination that the operating data is abnormal, the new behavioral event is determined to be a data abnormality event; as well as In response to the determination that the tested object has not been replaced and that the operating data is normal, the new behavioral event is determined to be a normal data event.

8. The model learning guidance method as described in claim 1, wherein if it is determined that the new behavioral event has not occurred on the test machine, it further includes: When the at least one anomaly detection model includes a health detection model, the health detection model is used to determine whether the operational data is abnormal. as well as When the at least one anomaly detection model includes a test object replacement detection model, the test object replacement detection model determines whether a test object has been replaced based on the operational data.

Images