JP7378367B2 - Fault diagnosis device and fault diagnosis method - Google Patents

Description

The present invention relates to a fault diagnosis device, a fault diagnosis method, and a learned model used for fault diagnosis for diagnosing the faults of various devices.

Equipment such as work vehicles may malfunction, and when a malfunction occurs, it is necessary to quickly confirm the exact symptoms of the malfunction, investigate the cause of the malfunction, and deal with the malfunction. For example, a worker who receives a malfunction report (warranty application, complaint, etc.) considers the structure of the work vehicle, imagines the range of influence and phenomenon expected from the report, and estimates the location and cause of the malfunction. and perform fault diagnosis. Furthermore, various auxiliary tools may be used when performing such failure diagnosis.

For example, a work machine such as a combine harvester is equipped with various actuators and sensors, and is equipped with a controller that receives detection signals from the sensors and controls the operation of the actuators. The checker device disclosed in Patent Document 1 connects to a controller, operates an actuator, and checks a detection signal from a sensor to perform a failure diagnosis to identify a failure location.

In addition, as shown in Patent Document 2, control data is acquired from a vehicle, calculations are performed to reveal data fluctuations due to the occurrence of mechanical or control abnormalities (failures), and abnormalities are diagnosed from the calculation results. Sometimes it is done.

Furthermore, as shown in Patent Document 3, a plurality of pieces of event information including various elements are accumulated in a server, and elements related to a failure are input, and based on event information having a predetermined number or more of elements that match the input elements, In some cases, analysis information is created and the cause of failure is predicted from the analysis information.

In addition, as shown in Patent Document 4, failure information is obtained based on current operation data and historical operation data of the acquired railway vehicle, a maintenance solution corresponding to the failure information is acquired, and the railway is operated based on the maintenance solution. Vehicle maintenance may also be performed.

In addition, as shown in Patent Document 5, failure events are organized in a tree shape that is easy for general workers to understand while taking into account the cause-and-effect relationship of failures from MFM information, MFM ancillary information, and influence ripple rules. In some cases, FTA information and FMEA information are generated, and fault diagnosis is performed using the FTA information and FMEA information.

In addition, as shown in Patent Document 6, a target ontology is selected, target ontology data and diagnostic ontology data in which the target ontology is mutually associated are extracted, and an FTA case having an appropriate diagnostic range and diagnosis content is created. In some cases, it is displayed and used for fault diagnosis.

Japanese Patent Application Publication No. 2004-199500 Japanese Patent Application Publication No. 2012-198144 JP 2017-117193 Publication Special table 2019-531523 publication Patent No. 3808893 Japanese Patent Application Publication No. 2000-322125

However, in order to understand the structure of a device such as a work vehicle and perform a failure diagnosis taking this into consideration, a considerable degree of understanding of the structure of the device is required. Furthermore, as the functions of devices such as work vehicles continue to improve, their structures become increasingly complex, and even if the auxiliary tools shown in Patent Documents 1 to 4 mentioned above are used, it is easy to use them without a certain degree of understanding of the structure of work vehicles. The situation has become such that failure diagnosis cannot be performed. Furthermore, in the auxiliary tools shown in Patent Documents 5 and 6 mentioned above, information in a tree structure organized in consideration of the structure of devices such as work vehicles is provided, and failure diagnosis is performed based on this information. When the problem becomes complex, even if organized information is provided, it is often not easy to diagnose the problem while checking it. Therefore, there is a need for a fault diagnosis that does not require trial and error and can be easily performed without any knowledge of fault diagnosis.

The failure diagnosis device according to one embodiment of the present invention is characterized in that the components constituting the target device, the connection relationships between the components, and the relationships between the components are described, thereby making it possible to mutually interact the plurality of components constituting the target device. Machine learning is performed using an ontology model showing the relationship and a failure mechanism database having a plurality of failure report texts including failure causes, failure parts, and failure events, which are failure information about failures that have occurred in the target device, as input data. A learning device that generates a trained model for performing fault diagnosis; and a learning device that inputs a fault report including at least one of the fault information into the trained model, and uses the fault report as the fault diagnosis. The present invention includes a failure diagnosis section that extracts the failure report text with a matching degree higher than a predetermined degree from the failure mechanism database, and a display section that displays the failure report text extracted by the failure diagnosis section.

Further, in the fault diagnosis method according to one embodiment of the present invention, parts constituting the target device, connection relationships between the components, and association relationships between the components are described, whereby a plurality of the components constituting the target device are described. The machine uses as input data an ontology model that shows the interrelationship between a step of generating a trained model for performing fault diagnosis by training, and inputting a fault report including at least one of the fault information to the trained model, and inputting the fault report as the fault diagnosis. The present invention includes a step of extracting the failure report text having a degree of matching with the failure mechanism database from the failure mechanism database, and a step of displaying the extracted failure report text.

Furthermore, the learned model according to an embodiment of the present invention includes an ontology model showing the mutual relationship between a plurality of parts constituting the target device, and failure information regarding the failure that has occurred in the target device, such as the cause of the failure and the location of the failure. , and a failure mechanism database having a plurality of failure report texts containing failure events as input data, are generated by machine learning, and a failure report including at least one of the failure information is input and used as a failure diagnosis. A computer is made to function to extract from the failure mechanism database the failure report text that has a degree of coincidence with the failure report that is higher than a predetermined degree.

As described above, a trained model is generated using the ontology model of the target device and the failure report texts regarding past failures that occurred in the target device as input data.

Therefore, failure diagnosis using this learned model takes into account the interrelationships of the plurality of components that constitute the target device. As a result, the operator does not have to go through trial and error when diagnosing a fault, and can directly obtain fault diagnosis results that take into account the configuration of the target device, even if he or she is not familiar with the configuration of the target device. can.

Further, the failure diagnosis result can be output as a failure report text consisting of failure information including failure cause, failure location, failure event, and the like. Therefore, the failure diagnosis results are output in the form of a failure report text as candidates for appropriate information corresponding to the failure, and workers can easily understand the failure diagnosis results even if they have no knowledge of failure diagnosis. , it is possible to perform accurate failure diagnosis simply by determining whether the failure report text is appropriate as a text for the failure.

Further, the failure report text may further include a failure mode that causes the failure event as the failure information.

With such a configuration, the failure mode that causes the failure event occurring at the failure part is recorded in the failure report text registered in the failure mechanism database. Therefore, a trained model is generated taking the failure mode into consideration, and the failure mode is described in the failure report text output by failure diagnosis. As a result, failure diagnosis is performed that takes the failure mode into consideration, and a more detailed failure report text is output, allowing workers to more easily understand the failure diagnosis results and perform failure diagnosis with greater accuracy. I can do it.

Further, the failure report text may further include countermeasures against the failure.

With such a configuration, a countermeasure policy for a failure is described in the failure report text outputted by the failure diagnosis, and an operator can easily take measures against the failure such as repair or design change.

Further, the failure report may be a word associated with each piece of the failure information.

The failure report input at the time of failure diagnosis is a word related to the failure, and the failure diagnosis can be easily performed without requiring a deep understanding of the structure of the target device. Furthermore, since the input words are associated with failure information, the relationship with the actually occurring failure becomes clear, and highly accurate failure diagnosis can be expected. From the above, failure diagnosis can be easily and accurately performed.

Furthermore, the apparatus may further include a failure report input section that causes the display section to display a failure report input screen that allows the failure information of the failure report to be input individually.

By inputting a failure report using the failure report input screen, the failure report can be easily input.

Further, in the failure diagnosis, a plurality of the failure report sentences may be extracted in order of the degree of agreement with the failure report, and the extracted failure report sentences may be displayed in order of degree of agreement with the failure report. .

With this configuration, as a result of fault diagnosis, multiple fault report texts that are likely to correspond to the fault that is actually occurring are displayed in the order of their likelihood of matching the fault. Diagnosis results can be determined by comparing the failure report sentences of , and by simply selecting the failure report sentence that is judged to be the most appropriate, failure diagnosis can be performed more easily and accurately.

Further, in the failure diagnosis, a probability of coincidence between the failure report and the failure report text may be calculated, and the extracted failure report text and the probability of coincidence may be displayed.

With this configuration, multiple failure report texts are displayed together with their respective matching probabilities, allowing the operator to judge the diagnostic results while referring to the matching probabilities, making it easier and more accurate. Fault diagnosis can be performed.

Further, the learned model may be generated by further performing machine learning using an arbitrary word related to a failure occurring in the target device as input data and a failure report text corresponding to the failure as training data.

The terms and expressions used in failure reports entered during failure diagnosis may vary depending on the operator, and even if the words have the same meaning, failure reports using words that are different from those described in the failure report text may be incorrect. If input, it may not be possible to accurately determine the degree of agreement between the failure report and the failure report text.

With the above configuration, machine learning is performed on the trained model using a wide range of words by a machine learning operator, and it becomes possible to determine synonyms of one word. Therefore, machine learning can also determine the degree of matching between a failure report and a failure report text while determining the similarity of words used in the failure report and the failure report text. As a result, even if a wide range of words are used in the failure report, it is possible to extract a failure report sentence that has a high degree of agreement with the failure report, and it is possible to perform failure diagnosis more easily and with higher accuracy.

Further, in addition to the machine learning for fault diagnosis, machine learning for natural language processing may be performed, and the learned model may have information related to the natural language processing.

With the above configuration, the trained model is compatible with natural language processing, so even if a wide range of words are used in a fault report, fault diagnosis can be performed by comparing synonyms of the words. It is possible to extract a failure report text that has a high degree of agreement with the failure report, and it is possible to perform failure diagnosis more easily and accurately.

Furthermore, the method may further include a step of inputting the failure report into the learned model to newly generate a failure report text.

By using a trained model such as the one described above, it is possible to not only judge the consistency between a failure report and a failure report text through failure diagnosis, but also take into account the content of the failure report and determine the failure report text to be registered in the failure mechanism database. It is possible to generate a text that conforms to the structure of the new failure report text. Therefore, by using the trained model as described above, it is possible not only to perform fault diagnosis with high accuracy but also to create a fault report text regarding a new fault.

Further, the apparatus may further include a failure diagnosis determination section that causes the display section to display a failure diagnosis determination screen for determining an appropriate failure report text from the failure report sentences extracted by the failure diagnosis section.

Thereby, an appropriate failure report text can be easily determined from a plurality of failure report texts.

Further, the target device is a work vehicle that performs a predetermined work at a work site, and the failure factor may include at least one of the work content, the work site environment, and the manner in which the work vehicle is used.

With such a configuration, failure diagnosis can be easily and accurately performed even for work vehicles. Further, information specific to the work vehicle that performs work at the work site, such as the content of the work to be performed, the environment of the work site where the work is performed, and the manner in which the work vehicle is used, can be input as the cause of the failure. Therefore, a trained model can be generated using a failure mechanism database containing information specific to a work vehicle as input data, and a failure report containing information specific to a work vehicle can be input to perform failure diagnosis. As a result, failure diagnosis can be performed with particularly high accuracy in the work vehicle.

FIG. 1 is a diagram showing a schematic configuration of a failure diagnosis device. FIG. 3 is a diagram showing a schematic structure of a failure report text. FIG. 3 is a diagram illustrating a failure report text. FIG. 3 is a diagram showing an example of a model configuration of a trained model. FIG. 2 is a block diagram illustrating the configuration of a failure diagnosis device. FIG. 3 is a diagram illustrating a failure report input screen. FIG. 3 is a diagram illustrating a processing flow of a failure diagnosis method. FIG. 3 is a diagram illustrating a failure diagnosis determination screen.

In addition to work vehicles such as agricultural machines and construction machines, various devices are composed of mechanical parts, electronic parts, and the like. Even if these devices are manufactured with sufficient durability and reliability, failures may occur. When a failure occurs, the failure location is identified, a failure diagnosis is performed to analyze the failure mechanism, and then repairs are made to the failure location and design changes are considered to prevent future failures.

Hereinafter, a failure diagnosis device for diagnosing a failure occurring in a device such as a work vehicle (hereinafter referred to as a "target device") will be described with reference to the drawings.

[Outline of failure diagnosis]
First, a schematic configuration in which a failure diagnosis device performs failure diagnosis will be described using FIG. 1.

As shown in FIG. 1, the fault diagnosis device 1 has a configuration in which a learning device 2 performs machine learning to generate a learned model 3, and a fault diagnosis section 4 performs fault diagnosis using the learned model 3.

The learning device 2 is equipped with AI (artificial intelligence) 5. Machine learning is performed by inputting the ontology model 6 of the target device and the failure mechanism database 7 as input data to the AI 5 of the learning device 2, and a learned model 3 is generated.

The ontology model 6 is information indicating the interrelationships of a plurality of components that constitute the target device. The target device is composed of mechanical parts, electronic parts, etc., which are organically linked. The ontology model 6 is information indicating the components included in the target device, their connection relationships, connection modes, modes of change in other components when a certain component operates, influence range of each component and operation, and the like.

The failure mechanism database 7 is a database in which a plurality of failure report texts 8 are stored. The failure mechanism database 7 is a database in which failure conditions, which are analysis results of failures that occurred in the past in the target device, are summarized as failure information and a failure report text 8 containing the failure information. Further, the failure report text 8 may appropriately include failure information not only for the target device but also for devices similar to the target device or devices having a common configuration with the target device. Failure report text 8 will be detailed later.

When a failure occurs in the target device, a failure report 9 indicating the status of the failure is input to the failure diagnosis section 4. For the failure report 9, various pieces of information related to the failure report 9 may be input into the failure diagnosis unit 4 one by one through screen input or the like, but the failure report 9 created as one data may be input to the failure diagnosis unit 4. You can also enter it directly. The failure report 9 created as one piece of data is failure information that indicates the symptoms of failure when a failure occurs in the target device, and is information created by the user of the target device after confirming the failure situation. be. For example, the failure report 9 created as a single piece of data can be used as a warranty application, a claim report, etc. for requesting warranty from the manufacturer or retailer when a failure occurs in the target device. . Specifically, as shown in FIG. 2, the information included in the failure report 9 is a failure cause 13, a failure location 14, a failure event 15, etc., and includes at least one of these. The failure cause 13, failure location 14, failure event 15, etc. are failure information written in the failure report text 8, and details will be described later. The failure report 9 and the failure report text 8 may further include a failure mode 18. The failure diagnosis unit 4 performs failure diagnosis by inputting the failure report 9 to the generated trained model 3.

[Failure report text]
The failure report text 8 shows failure information regarding the mechanism that led to the occurrence of the failure, including the cause, symptoms, and location of the failure. A failure is when a cause of failure exists, this cause affects a certain part, and a defect occurs in that part.

As shown in FIGS. 2 and 3, the failure report text 8 includes failure information including a failure cause 13, a failure location 14, and a failure event 15, and further includes a failure summary 16. The failure report text 8 may further include a functional block 17, a failure mode 18, a countermeasure policy 19, and the like. In the failure report text 8, the cause of the failure corresponds to the failure factor 13, and the part where the failure occurs corresponds to the failure part 14. The phenomenon occurring in the component corresponds to failure event 15, and the effect on the component corresponds to failure mode 18.

As mentioned above, the failure factor 13 is the cause of failure, and can be divided into internal factors and external factors. Internal factors are stresses that occur or increase under certain conditions due to the structure of the target product. For example, the internal factor is engine vibration, etc. External factors are stress that occurs or increases under certain specific conditions due to the usage method, usage environment, manufacturing/transportation, etc. of the target product. For example, the external factor may be vibration of the vehicle during high-load operation. Failure occurs when this stress increases and occurs due to internal and external factors.

The failure site 14 is a site of a component or the like where a failure has occurred due to the failure factor 13. For example, the failure part 14 is a "welded part of a fuel filter bracket," a "fuel filter," a "fuel hose," or the like.

The failure event 15 is information indicating what happened to the failure part 14 due to the failure. For example, in the failure event 15, information such as "damage", "falling off", "fuel leakage", etc. is described. This information corresponds to cases such as when the welded part of the fuel filter bracket is damaged and the fuel filter falls off, or when fuel leaks from the fuel hose. Furthermore, failure events 15 can also include events that occur as a result of these events. For example, if an abnormal noise occurs in the engine room due to contact/interference between the fuel hose and the radiator fan, descriptions such as "radiator fan," "engine room," or "abnormal noise occurred" are also possible.

The functional block 17 is a functional block that includes the failure part 14, and also includes relationships between related parts. For example, the functional block 17 related to the failure part 14 illustrated above is an engine auxiliary device (fuel system), and is shown to include a fuel filter and a bracket.

The failure mode 18 is information indicating how the failure part 14 etc. broke. The failure mode 18 has a causal relationship with the failure cause 13 and occurs together with the failure cause 13. In other words, a failure event 15, which is a malfunction event, occurs in the failure part 14 due to the failure mode 18 caused by the failure cause 13. Here, the failure mode 18 is a failure of a component that requires repair or replacement, and is defined as destruction of a component/destruction of an interface between components/irreversible change in physical properties. For example, failure mode 18 is described as "fatigue failure," "contact/interference," and "cutting."

The failure summary 16 is information that describes the failure situation in text, corresponding to failure information such as the failure cause 13, the failure location 14, and the failure event 15. Only this failure summary 16 may be output as the failure report text 8 output as a result of the failure diagnosis. When the failure information is the failure cause 13, failure location 14, and failure event 15 as exemplified above, the failure summary 16 is, for example, ``Due to engine vibration and vehicle body rocking during high-load operation, the fuel filter bracket The welded part of the engine was damaged due to fatigue failure, and the fuel filter fell off.The fuel hose came into contact with and interfered with the radiator fan, causing an abnormal noise in the engine room.The fuel hose broke and fuel leaked. .

The countermeasure policy 19 describes the countermeasures taken against the failure when a failure occurs, the details of the repair of the failure, and the matters that were changed in design in response to the failure. For example, countermeasure policy 19 is, ``Design the fuel filter bracket so that the stress is below the standard value by shifting the resonance point of the fuel filter bracket from the engine rotation range or lowering the amplitude level during resonance.'' Further, the countermeasure policy 19 can include design to-dos, experiment to-dos, and the like, which are matters to be performed during product development. Furthermore, the countermeasure policy 19 may include progress statuses, deliverables, etc. of design to-dos, experiment to-dos, and the like.

[Trained model]
As mentioned above, the ontology model 6 describes the parts that make up the target device, the connection relationships between the parts, and the relationships between the parts such as the components that intervene in the propagation of operations between the parts, and the ontology model 6 describes the parts that make up the target device, the connection relationships between the parts, and the relationships between the parts such as the components that intervene in the propagation of operations between the parts. The configuration of the device is described. Further, in the failure report text 8, the failure situation is described as failure information and a text containing the failure information.

Therefore, the learned model 3, which is machine-learned using the failure mechanism database 7 including a plurality of failure report sentences 8 and the ontology model 6 as input data, uses the ontology model in addition to the failure situation acquired from the failure report sentences 8. The relationship of parts related to the failure obtained from 6 is modeled. Therefore, in the learned model 3, past failure situations associated with all parts included in the functional blocks of the target device are modeled based on failures that have occurred in the past.

An example of such a trained model 3 will be described using FIG. 4 while referring to FIG. 2. Here, a model in which a part related to a failure in the fuel system illustrated in FIG. 3 is modeled will be described as an example.

As shown in Fig. 4, the trained model 3 is a model in which the failure situation is expressed as a word corresponding to the failure information, etc., and a sentence including the failure information regarding the event that occurred, the parts involved, the configuration related to the propagation of the event, etc. has been made into Among these, the connection relationships of parts and the configuration related to event propagation are extracted from the ontology model 6, and the occurring events, related parts, failure details, etc. are extracted from the failure mechanism database 7.

Specifically, the failure factor 13 is engine vibration and vehicle shaking during high-load operation, and the function block 17 in which the failure occurred is the fuel system of the engine auxiliary equipment, so the trained model 3 is , the effects of engine vibration and vehicle shaking during high-load operation on the fuel system are modeled.

The trained model 3 includes a failure report text 8 related to the failure situation shown in FIG. 3 as one failure in the fuel system. In this case, failure report text 8 states, ``Due to engine vibration and vehicle rocking during high-load operation, the welded part of the fuel filter bracket suffered fatigue failure, the fuel filter fell off, and the fuel hose and radiator fan were damaged. contact and interference, cutting the fuel hose and causing fuel leakage.

In the trained model 3, the interrelationships of fuel system components are further modeled from the ontology model 6. As shown in FIG. 4, the structure of the fuel system modeled from ontology model 6 is such that a fuel filter is held in a fuel filter bracket, and fuel is supplied from the fuel filter through a fuel hose.

By incorporating such an ontology model 6, "fatigue fracture" and "damage" in the failure report text 8 are linked to the welded part of the fuel filter bracket, as shown by the arrow in Figure 4, and "falling off" is associated with the welded part of the fuel filter bracket. ” is associated with the fuel filter, “contact/interference” is associated with the fuel hose, and “cut” is associated with the fuel hose. Furthermore, "contact/interference" is associated with "radiator fan" and "abnormal noise generation," and "abnormal noise generation" is associated with "engine room." Therefore, a learned model 3 is generated in which the failure report text 8 indicating the failure status of the fuel system is added with the device configuration of the fuel system.

As a result, the trained model 3 related to one of the fuel system failures shown in FIG. 3 includes the fuel system structure in addition to the fuel system failure situation. Furthermore, the learned model 3 is subjected to machine learning by inputting the failure mechanism database 7 into the failure status of other fuel systems. Further, the learned model 3 is subjected to machine learning by inputting ontology models 6 of other functional blocks and a failure mechanism database 7 related to failure conditions. As a result, the learned model 3 models various failure situations and device configurations of the target device, and becomes a model that can handle a wide range of failures.

[Fault diagnosis device]
Next, a configuration example of the failure diagnosis device 1 will be described using FIGS. 5 and 6 while referring to FIGS. 1 and 2.

The fault diagnosis device 1 includes a control section 20, a learning device 2, a fault diagnosis section 4, an input section 22, a storage section 23, and a display section 24.

The control unit 20 controls the operation of the failure diagnosis device 1. The input unit 22 receives an ontology model 6 and a failure mechanism database 7 that serve as input data for machine learning. The inputted ontology model 6 and failure mechanism database 7 are stored in the storage unit 23 under the control of the control unit 20 .

The learning device 2 is equipped with an AI 5 and, under the control of the control unit 20, inputs the ontology model 6 and the failure mechanism database 7 stored in the storage unit 23 as input data to the AI 5 for machine learning. Generate model 3. The generated learned model 3 is stored in the storage unit 23.

Note that the failure mechanism database 7 includes a large number of failure report texts 8. If the words used in these failure report texts 8 are not unified, a comparison with the input failure report 9 cannot be performed appropriately when diagnosing a failure using the generated trained model 3. Therefore, the failure report text 8 included in the failure mechanism database 7, which serves as input data for machine learning, is created in advance by standardizing the words, expressions, and syntax used. Further, even if the failure report text 8 is created using standardized words, expressions, and syntax, there may be some variation. Even in such a case, by performing machine learning on a large number of failure report sentences 8, the learned model 3 can be expected to become a model that can determine a certain degree of similarity in the words, expressions, and syntax used.

When performing a failure diagnosis, the failure diagnosis apparatus 1 may include a failure report input section 27 that causes the display section 24 to display a failure report input screen 30. The failure report 9 is input to the failure diagnosis apparatus 1 in a manner that selects the items displayed on the failure report input screen 30. The display unit 24 is, for example, a touch panel, and is configured to allow information to be input from a displayed failure report input screen 30.

On the failure report input screen 30, the failure cause 13, failure location 14, failure event 15, etc. contained in the failure report 9 can be individually input as the failure report 9, and the failure mode 18 can also be input.

The failure report input screen 30 displays an item switch 31 that can selectively input internal factors and external factors among the failure causes 13, failure location 14, failure event 15, and failure mode 18 as input items. When the item switch 31 corresponding to the item to be input is selected by touching the screen or the like, a pull-down menu 32 corresponding to the item is displayed. The pull-down menu 32 lists candidates to be input to that item, and by selecting one or more of the displayed candidates, information for that item is input. In FIG. 6, the item switch 31 corresponding to the internal factor is selected, and "engine vibration" is selected from among the candidates displayed in the pull-down menu 32. Then, by sequentially selecting the item switches 31 corresponding to each of the internal and external causes, failure location 14, failure event 15, and failure mode 18, all information related to the failure report 9 can be input. . If there is no information to be entered for that item, select "None" from the candidates in the pull-down menu 32, and if there is no corresponding information among the candidates, select "Other" and enter any information. Information can be entered. When "Others" is selected and arbitrary information is entered, a keyboard may be displayed on the display unit 24 and the arbitrary information may be entered from the keyboard. Any information may be input from an input means.

The information input in this manner is stored in the storage unit 23 as a failure report 9 in association with the corresponding item.

Then, under the control of the control unit 20, the failure report 9 stored in the storage unit 23 is input to the learned model 3 stored in the storage unit 23, and a failure diagnosis is performed. As a fault diagnosis, the fault diagnosis section 4 compares the fault report text 8 modeled in the trained model 3 with the fault report 9, and outputs one or more fault report texts 8 with a high degree of coincidence. Specifically, the failure report text 8 having a degree of coincidence with the failure report 9 that is equal to or higher than a predetermined degree is selected. For example, the failure diagnosis unit 4 extracts and outputs a failure report text 8 that includes a predetermined number or more or a predetermined ratio or more of words included in the failure report 9 or words similar to the words. When comparing failure report text 8 and failure report 9, the words and context of the description of failure report 9 are grasped by natural language processing, and synonyms and similar expressions are taken into account and compared with failure report text 8. Processing may be performed. In addition, since the ontology model 6 is modeled in the trained model 3, the fault diagnosis unit 4 can perform fault diagnosis in consideration of parts corresponding to the words included in the fault report 9 and related parts. There is also.

The display unit 24 displays the failure report text 8 output by the failure diagnosis unit 4 under the control of the control unit 20 . The display unit 24 may display only the portion of the failure report text 8 that corresponds to the failure summary 16. When a plurality of failure report texts 8 are output, the display unit 24 displays the plurality of failure report texts 8 in descending order of degree of matching.

Note that the failure diagnosis unit 4 may calculate the probability of coincidence between the failure report text 8 and the failure report 9. In this case, the display unit 24 displays the probability of matching together with the failure report text 8 in association with the failure report text 8.

As described above, in order to perform machine learning using the failure mechanism database 7 consisting of the failure report text 8 as input data, the failure diagnosis results are directly converted into the failure report text 8 that includes at least the failure summary 16, which is a text indicating the status of the failure. It is possible to output Therefore, even if the operator is not familiar with the structure of the target device such as a work vehicle, there is no need for trial and error during failure diagnosis, so failure diagnosis can be easily performed.

In particular, the output failure report text 8 is output in the form of a sentence consisting of words included in the failure report 9 input during failure diagnosis. Therefore, even if the operator is not familiar with the structure of the target device or has sufficient knowledge of failure diagnosis, the operator can easily understand the failure diagnosis results and write the failure report text that he or she deems most appropriate. Just by making a selection, you can perform fault diagnosis with high accuracy.

Furthermore, as described above, the failure report text 8 can include countermeasures. In this case, the failure report text 8 output as the failure diagnosis result describes the failure situation and countermeasures for the failure. Therefore, by simply checking the fault diagnosis results, workers can quickly confirm how to deal with the fault, how to repair the faulty part, or what design changes should be considered in response to the fault. It is possible to respond to major failures. Furthermore, when developing a new product, problems with the current target device can be detected by referring to the failure report text 8, and this can be referenced in the development of a new product.

[Fault diagnosis method]
Next, a failure diagnosis method will be described using FIG. 7 while referring to FIGS. 1 and 2. In addition, in the following explanation, a method of performing fault diagnosis using the fault diagnosis device 1 may be used, but the fault diagnosis method is not limited to the case where it is performed using the fault diagnosis device 1, and may be performed using other device configurations. , software, or a mixture of these.

First, an ontology model 6 indicating the structure of a target device such as a work vehicle is generated, and a failure mechanism database 7 is generated that collects failure report texts 8 indicating failure conditions that have occurred in the past with respect to the target device.

Next, the AI 5 is subjected to machine learning using the generated ontology model 6 and failure mechanism database 7 as input data to generate a trained model 3 (step #1).

When a failure occurs in the target device (Step #2 Yes), the operator creates a failure report 9 using as much information as possible regarding failure information such as failure cause 13, failure location 14, failure event 15, etc. (Step #3). The failure report 9 may be a sentence including the failure cause 13, the failure part 14, the failure event 15, etc., or may be a word associated with the failure cause 13, the failure part 14, the failure event 15, etc.

Next, the created failure report 9 is input to the trained model 3 to perform failure diagnosis (step #4). As a result of the fault diagnosis, among the fault report texts 8 stored in the fault mechanism database 7, one or more fault report texts 8 having a high degree of coincidence with the fault report 9 are output.

Next, the outputted failure report text 8 is displayed (step #5). The displayed failure report text 8 includes at least a failure summary 16. When a plurality of failure report texts 8 are output, the failure report texts 8 are displayed in descending order of degree of matching. Furthermore, the probability of matching each failure report text 8 with the failure report 9 may be displayed in association with each other.

Next, the operator checks the displayed failure report text 8 and checks whether the content of the displayed failure report text 8 matches the actual failure, or which failure report text 8 is the most likely. By examining whether the failure report text 8 corresponds to the actually occurring failure, the failure report text 8 is selected (step #6).

The operator then confirms the detailed situation of the failure from the selected failure report text 8 and considers countermeasures for the failure. Alternatively, if the countermeasure policy 19 is indicated in the failure report text 8, countermeasures are taken in accordance with the countermeasure policy 19 (step #7).

Even with the above-mentioned fault diagnosis method, similar to the fault diagnosis device 1, even if the worker is not familiar with the structure of the target device such as a work vehicle or the fault diagnosis, the diagnosis result can be interpreted to include at least the fault summary 16. This can be directly confirmed as a failure report text 8, which eliminates the need for trial and error during failure diagnosis, making it easy to diagnose failures.

[Determination of failure diagnosis]
As a result of the fault diagnosis performed by the fault diagnosis device 1 and fault diagnosis method, one or more fault report texts 8 are displayed on the display section 24. From the displayed failure report texts 8, a fault diagnosis operator may judge which fault report text 8 is appropriate, but the fault diagnosis determination unit 28 in the fault diagnosis apparatus 1 assists in this judgment. may be provided. A specific configuration example for determining an appropriate failure report text 8 will be described below with reference to FIGS. 2, 4, and 8.

The failure diagnosis device 1 displays a failure diagnosis determination screen 35 on the display unit 24 under the control of the failure diagnosis determination unit 28 . The failure diagnosis determination screen 35 displays failure event candidates 38 and component connection relationships 36. In the failure event candidates 38, failure units 39, which are failure events related to specific parts in the ontology model 6, are extracted from the failure report text 8 that is selected as a candidate by the failure diagnosis and are listed. At this time, the failure units 39 may be arranged in descending order of the degree of coincidence with reference to the degree of coincidence of the corresponding failure report texts 8, or may be displayed together with the probability of coincidence. In the component connection relationship 36, a plurality of components corresponding to the candidates of the failure report text 8 are listed as component candidates 37 for this specific component. Furthermore, in the component connection relationship 36, for one specific component selected from the component candidates 37, components connected upstream in the ontology model 6 are separately displayed as component candidates 37. The component candidates 37 displayed at this time may be sequentially displayed not only for one component on the upstream side but also for multiple components on the upstream side.

When a plurality of failure report texts 8 are selected as candidates by the fault diagnosis, first, the most downstream component of the ontology model 6 in the plurality of fault report texts 8 selected as candidates by the fault diagnosis is selected as the component candidate 37, and the part candidate 37 is used for the fault diagnosis. They are listed and displayed in the component connection relationship 36 on the determination screen 35. At the same time, a failure unit 39 corresponding to the most downstream component candidate 37 is displayed as a failure event candidate 38 on the failure diagnosis determination screen 35.

Next, the operator selects a failure unit 39 that matches the actual failure situation occurring in the target device from among the failure units 39 displayed in the failure event candidates 38. Such a selection is made by touching the corresponding failure unit 39 on the screen. When one failure unit 39 is selected, the part corresponding to the selected failure unit 39 among the component candidates 37 corresponding to the most downstream part in the component connection relationship 36 changes text color, blinks, etc. This is emphasized by Then, in the component candidate 37 corresponding to the upstream side of the ontology model 6, components connected to this component are listed and displayed. At this stage, the failure report text 8 is narrowed down to include the selected unit of failure.

When the failure unit 39 corresponding to the most downstream component is selected, in the ontology model 6, the upstream side of the component included in the narrowed-down failure report text 8 and corresponding to the selected failure unit 39 is selected. Units 39 of failures related to parts are listed and displayed in failure event candidates 38 . The operator selects the failure unit 39 that matches the actual failure situation occurring in the target device from among the displayed failure units 39.

Such processing is repeated up to the most upstream component of the ontology model 6, and the failure report text 8 is narrowed down. If the failure report text 8 is narrowed down to one by the time the ontology model 6 reaches the most upstream component, it is determined that the failure report text 8 corresponds to the failure that is actually occurring. can do. If multiple failure report texts 8 remain even for the most upstream part of the ontology model 6, the operator should take these failure report texts 8 into consideration and proceed with further investigation to find the appropriate fault. Determine report text 8.

For example, when the failure report text 8 illustrated in FIG. The display includes a failure unit 39, ``The fuel hose was cut and fuel leaked''. When "Fuel hose was cut and fuel leaked" is selected and the next failure unit 39 is displayed, "An abnormal noise occurred in the engine room" corresponding to failure unit B is selected. Finally, the failure report sentence 8 that includes "the fuel filter has fallen off" corresponding to the failure unit C remains, and an appropriate failure report sentence 8 is thereby extracted.

By determining the appropriate failure report text 8, failure diagnosis can be easily performed even if a plurality of failure report texts 8 are displayed as candidates.

[Another embodiment]
(1) In the failure diagnosis device 1, a failure report text generation section 26 may be further provided. The failure report text generation unit 26 inputs the failure report 9 into the trained model 3 to newly generate a failure report text 8 corresponding to the failure report 9.

The learned model 3 is generated using the failure report text 8 (failure mechanism database 7) and ontology model 6 as input data. Therefore, in the trained model 3, in addition to the failure report text 8, the structure of the target device is modeled. By inputting the failure report 9 into such a learned model 3, the failure report 9 can be analyzed while referring to the past failure report text 8 and taking into account the structure of the target device. As a result, a new failure report text 8 corresponding to the input failure report 9 can be generated with reference to the written content and document structure of the past failure report text 8.

Note that generation of the failure report text 8 is performed not only by the failure diagnosis device 1 but also by the process of inputting the failure report 9 into the trained model 3 and generating the failure report text 8 in the fault diagnosis method described in the above embodiment. May be added.

The failure report text 8 is created by an operator or the like after performing a failure diagnosis when a failure occurs and based on the diagnosis results. At this time, the words and expressions used differ depending on the worker, and the knowledge of the structure of the target device and the knowledge of failure diagnosis differs depending on the worker, so the completeness and expression of the created failure report text 8 may vary. was there. As described above, a new failure report text 8 can be generated by inputting the failure report 9 to the learned model 3. Therefore, consistency with the words and expressions used in the existing failure report text 8 is achieved, and the degree of completeness of the failure report text 8 tends to be constant. As a result, even an operator with limited knowledge of the structure of the target device or failure diagnosis can easily generate a failure report text 8 of the same level as other failure report texts 8.

Furthermore, the learned model 3 may be updated by inputting the failure mechanism database 7 containing the failure report text 8 generated by inputting the failure report 9 to the trained model 3 and performing machine learning again. The failure report text 8 generated by inputting the failure report 9 into the trained model 3 has the words, expressions, and syntax used standardized according to the same criteria as the failure report text 8 that has already been subjected to machine learning. Therefore, the updated trained model 3 has improved standardization of the words, expressions, and syntax used, and can further improve the accuracy of fault diagnosis.

(2) In each of the above embodiments, machine learning is not only performed using the failure mechanism database 7 and ontology model 6 as input data, but also uses words corresponding to failure information regarding an arbitrary failure as input data, and learns how to respond to that failure. Machine learning may be further performed using the failure report text 8 as training data.

Existing failure report texts 8 stored in the failure mechanism database 7 are organized in advance, and the words, expressions, and syntax used are standardized. Failure information regarding a separately prepared arbitrary failure may contain words or expressions that are different from words or expressions used in the existing failure report text 8 even if they have the same meaning. By adding machine learning as described above, the trained model 3 has expanded its versatility in terms of words, expressions, and syntax, and even if different words, expressions, and syntax are used during fault diagnosis, it can still properly report faults. It becomes possible to compare and verify the degree of agreement between the error report text 8 and the failure report text 8. As a result, failure diagnosis can be performed more easily and accurately.

(3) In each of the above embodiments, machine learning related to natural language processing may be further added as machine learning. Thereby, the trained model 3 is modeled so that synonyms of words, expressions, etc. can be determined. As a result, it becomes possible to appropriately compare and verify the degree of coincidence between the failure report 9 and the failure report text 8, and it becomes possible to perform failure diagnosis more easily and with higher accuracy.

(4) The configuration is not limited to inputting the failure report 9 via the failure report input screen 30, but the failure report 9 converted into data may be input from the input unit 22 and stored in the storage unit 23. When diagnosing a failure, the failure report 9 stored in the storage unit 23 is read out, and the failure diagnosis unit 4 performs the failure diagnosis. Note that the failure report 9 may be a sentence that includes a word corresponding to at least one of the failure cause 13, the failure part 14, and the failure event 15. Moreover, the failure report 9 may be a word corresponding to at least one of the failure cause 13, the failure part 14, and the failure event 15, which are failure information, without being written. Furthermore, the failure report 9 may further include failure mode 18 as failure information. Note that the input failure report 9 does not necessarily have to be converted into data, and words corresponding to the failure report 9 may be input one by one. Even in such a case, words similar to words included in the input failure report 9 are taken into consideration during failure diagnosis. Therefore, failure diagnosis can be performed without strictly restricting the words used in the failure report 9, and from this reason as well, failure diagnosis can be performed easily.

(5) In each of the above embodiments, the failure diagnosis device 1 includes a processor such as a CPU in addition to the AI 5. The failure diagnosis device 1 may be configured with hardware, but at least some of its functions may be configured with software. In this case, the software is stored in the storage unit 23 and operated by the processor. Further, the configuration of the functional blocks in the failure diagnosis device 1 is not limited to the above configuration, and each functional block may be appropriately aggregated or subdivided.

Further, the AI 5 is not limited to being provided in the fault diagnosis device 1, but may be provided in a state where the fault diagnosis device 1 can communicate via a network or the like. In this case, the input data for machine learning and the failure report 9 for failure diagnosis are transmitted to the AI 5 via the network, and the failure report text 8 that is the failure diagnosis result is transmitted from the AI 5 to the failure diagnosis device 1 via the network.

INDUSTRIAL APPLICABILITY The present invention can be applied to failure diagnosis for failures occurring in target devices such as various work vehicles such as agricultural vehicles and construction vehicles.

1 Fault diagnosis device 2 Learning device 3 Learned model 4 Fault diagnosis unit 6 Ontology model 7 Failure mechanism database 8 Failure report text 9 Failure report 13 Failure cause 14 Failure location 15 Failure event 18 Failure mode 19 Countermeasure policy 24 Display unit 26 Failure report Sentence generator

Claims (22)

An ontology model that describes the mutual relationships among the plurality of parts that make up the target device by describing the parts that make up the target device, the connection relationships between the parts, and the relationships between the parts; A trained model for performing fault diagnosis by performing machine learning using as input data a failure mechanism database that has multiple failure report texts that include failure causes, failure parts, and failure events, which are failure information about failures that have occurred. a learning device that generates
A failure report including at least one of the failure information is input into the learned model, and the failure report text having a degree of agreement with the failure report that is higher than a predetermined degree is extracted from the failure mechanism database as the failure diagnosis. A fault diagnosis section to extract,
A failure diagnosis device comprising: a display unit that displays the failure report text extracted by the failure diagnosis unit. The failure diagnosis apparatus according to claim 1, wherein the failure report text further includes a failure mode that causes the failure event as the failure information. 3. The failure diagnosis apparatus according to claim 1, wherein the failure report text further includes a countermeasure plan for the failure. The failure diagnosis device according to any one of claims 1 to 3, wherein the failure report is a word associated with each piece of failure information. The failure diagnosis device according to any one of claims 1 to 4, further comprising a failure report input unit that causes the display unit to display a failure report input screen that allows the failure information of the failure report to be input individually. The failure diagnosis unit extracts a plurality of failure report texts in order of their degree of matching with the failure report,
The failure diagnosis device according to any one of claims 1 to 5, wherein the display unit displays the extracted failure report sentences in order of degree of coincidence with the failure report. The failure diagnosis unit calculates a probability of coincidence between the failure report and the failure report text,
The failure diagnosis device according to claim 6, wherein the display unit displays the extracted failure report text and the matching probability. The learning device generates the learned model by further performing machine learning using an arbitrary word related to a failure that has occurred in the target device as input data and a failure report text corresponding to the failure as training data. 8. The failure diagnosis device according to any one of 7 to 7. The learning device performs machine learning for natural language processing in addition to machine learning for fault diagnosis,
The fault diagnosis device according to any one of claims 1 to 8, wherein the learned model includes information related to the natural language processing. The failure diagnosis device according to any one of claims 1 to 9, further comprising a failure report text generation unit that inputs the failure report into the learned model and generates a new failure report text. Any one of claims 1 to 10, further comprising a failure diagnosis determination unit that causes the display unit to display a failure diagnosis determination screen for determining the appropriate failure report text from the failure report text extracted by the failure diagnosis unit. The failure diagnosis device described in section. The target device is a work vehicle that performs a predetermined work at a work site,
The failure diagnosis device according to any one of claims 1 to 11, wherein the failure factor includes at least one of work content, work site environment, and usage mode of the work vehicle. An ontology model that describes the mutual relationships among the plurality of parts that make up the target device by describing the parts that make up the target device, the connection relationships between the parts, and the relationships between the parts; A trained model for performing fault diagnosis by performing machine learning using as input data a failure mechanism database that has multiple failure report texts that include failure causes, failure parts, and failure events, which are failure information about failures that have occurred. a step of generating
A failure report including at least one of the failure information is input into the learned model, and the failure report text having a degree of agreement with the failure report that is higher than a predetermined degree is extracted from the failure mechanism database as the failure diagnosis. The process of extracting;
A fault diagnosis method comprising the step of displaying the extracted fault report text. 14. The failure diagnosis method according to claim 13, wherein the failure report text further includes a failure mode that causes the failure event as the failure information. The failure diagnosis method according to claim 13 or 14, wherein the failure report text further includes a countermeasure plan for the failure. 16. The failure diagnosis method according to claim 13, wherein the failure report is a word associated with each piece of failure information. In the failure diagnosis, a plurality of failure report sentences are extracted in order of a high degree of coincidence with the failure report,
17. The failure diagnosis method according to claim 13, wherein the extracted failure report texts are displayed in order of degree of coincidence with the failure report. In the failure diagnosis, a probability of coincidence between the failure report and the failure report text is calculated,
18. The fault diagnosis method according to claim 17, wherein the extracted fault report text and the matching probability are displayed. 19. The learned model is generated by further performing machine learning using an arbitrary word related to a failure occurring in the target device as input data and a failure report text corresponding to the failure as training data. The failure diagnosis method described in item (1) above. In addition to machine learning for fault diagnosis, machine learning for natural language processing is performed,
The fault diagnosis method according to any one of claims 13 to 19, wherein the trained model includes information related to the natural language processing. The fault diagnosis method according to any one of claims 13 to 20, further comprising the step of inputting the fault report into the learned model to newly generate a fault report text. The target device is a work vehicle that performs a predetermined work at a work site,
22. The failure diagnosis method according to claim 13, wherein the failure factor includes at least one of work content, work site environment, and usage mode of the work vehicle.

Images