Maintenance work assistance device, maintenance work assistance method, and maintenance work assistance program

Abstract

Provided are a maintenance work assistance device, a maintenance work assistance method, and a maintenance work assistance program that improve operation efficiency. An information collection unit (12) collects operation information of a bill processing device (20). A feature amount calculation unit (13) classifies the operation information collected by the information collection unit (12) and calculates a feature amount that follows a normal distribution. An inference unit (14) inputs the feature amount calculated by the feature amount calculation unit (13) into a machine learning model (16) and performs an anomaly determination for the bill processing device (20).

Description

Maintenance work assistance device, maintenance work assistance method, and maintenance work assistance program

[0001] The present invention relates to a maintenance work assist device, a maintenance work assist method, and a maintenance work assist program.

[0002] Conventionally, when a banknote transport device has an abnormality such as a stoppage in its operating environment, maintenance work has been performed in the following procedure. First, the operation side worker, such as a store clerk who actually uses the banknote transport device, performs the maintenance work. In such cases, the operation side worker typically combines the error code displayed by the banknote transport device with their personal know-how to identify the cause of the device condition in the operating environment and take corrective measures.

[0003] Next, if the problem is too difficult for the operation staff to handle, a request is sent to a maintenance worker who specializes in maintenance work on banknote transport devices. The maintenance worker who receives the withdrawal request combines their professional knowledge and personal know-how to identify the cause of the equipment condition in the operation environment and take measures to improve it. A more specific procedure would be for the maintenance worker to refer to a maintenance manual or other source to isolate the problem one by one and take measures to resolve the problem.

[0004] As a technology related to the maintenance work of banknote transport devices, a technology has been proposed that uses a machine learning model to determine abnormalities based on the control state of a paper sheet processing mechanism that transports and processes paper sheets and on sounds or vibrations detected by sensors. Another technology has been proposed that detects abnormalities that occur inside a banknote validator and notifies users of the abnormality by coding the details based on the detection results. Another technology has been proposed for determining abnormal products, which extracts the features of an object to be determined and uses a machine learning model to determine whether the distribution of the features follows a normal distribution to determine whether an abnormality has occurred.

[0005] Japanese Patent Application Laid-Open No. 2023-124668 Japanese Patent Application Laid-Open No. 11-203529 Japanese Patent Application Laid-Open No. 2020-173496

[0006] However, when an abnormality occurs in a banknote transport device in an operational environment, the error codes and potential causes displayed vary widely. As a result, maintenance work often relies on individual specialized knowledge and know-how, as described above. This has resulted in significant variations in the accuracy of cause identification and the labor hours required for remedial measures, not only among operational personnel without specialized knowledge but also among maintenance personnel with specialized knowledge. If the quality of maintenance work varies from worker to worker, there is a risk of opportunity loss due to inappropriate responses. Furthermore, to reduce the variation in the quality of maintenance work, appropriate personnel training for maintenance is required, but this is time-consuming and potentially costly, making it difficult. As such, conventional maintenance methods have made it difficult to improve the operational efficiency of banknote transport devices.

[0007] Furthermore, with a technology that uses a machine learning model to determine an abnormality based on detected sound or vibration, identifying the cause of the abnormality depends on the ability of the operator, making it difficult to improve the operational efficiency of the banknote transport device. Furthermore, with a technology that encodes and notifies the content of an abnormality based on the detection result of an abnormality occurring inside a banknote validation device, analyzing the content of the notification depends on the ability of the operator, making it difficult to improve the operational efficiency of the banknote transport device. Furthermore, a technology that determines an abnormal product based on whether the feature distribution extracted from the target object follows a normal distribution is difficult to apply to a banknote validation device, making it difficult to improve the operational efficiency of the banknote transport device.

[0008] The disclosed technology has been made in view of the above, and aims to provide a maintenance work assist device, a maintenance work assist method, and a maintenance work assist program that improve operational efficiency.

[0009] In one aspect of the maintenance work assist device, maintenance work assist method, and maintenance work assist program disclosed herein, an information collection unit collects operation information of a banknote processing device. A feature calculation unit classifies the operation information collected by the information collection unit and calculates feature amounts according to a normal distribution. An inference unit inputs the feature amounts calculated by the feature calculation unit into a machine learning model and determines whether or not there is an abnormality in the banknote processing device.

[0010] According to one aspect of the maintenance work assist device, the maintenance work assist method, and the maintenance work assist program disclosed in the present application, it is possible to achieve an effect of improving operational efficiency.

[0011] FIG. 1 is a block diagram of a maintenance work assist device according to an embodiment. FIG. 2 is a schematic diagram showing an example of a banknote transaction device. FIG. 3 is a diagram showing an example of actuator drive information and firmware control information. FIG. 4 is a diagram showing an example of a command. FIG. 5 is a diagram showing an example of a contribution calculated by explainable AI. FIG. 6 is a diagram showing an example of correspondence information. FIG. 7 is a flowchart of a feature calculation process performed by the maintenance work assist device according to an embodiment. FIG. 8 is a flowchart of a feature calculation process for a cash box module. FIG. 9 is a flowchart of a feature calculation process for a recycling stacker module. FIG. 10 is a flowchart of a learning process performed by the maintenance work assist device according to an embodiment. FIG. 11 is a flowchart of an abnormal state determination process performed by the maintenance work assist device according to an embodiment. FIG. 12 is a diagram showing an example of a module that is the target of abnormality determination. FIG. 13 is a hardware configuration diagram of the maintenance work assist device.

[0012] Hereinafter, embodiments of the maintenance work assist device, the maintenance work assist method, and the maintenance work assist program disclosed in the present application will be described in detail with reference to the drawings. Note that the maintenance work assist device, the maintenance work assist method, and the maintenance work assist program disclosed in the present application are not limited to the following embodiments.

[0013] 1 is a block diagram of a maintenance work assist device according to an embodiment. The maintenance work assist device 1 is connected to an automated teller machine 2, an operation company system 3, and a maintenance company system 4. Here, the operation company is a company that operates the banknote transport device, such as store staff, and the like who actually use the banknote transport device. The maintenance company is a company that specializes in providing services for maintaining banknote transport devices.

[0014] The automated teller machine 2 is a device that executes transactions such as depositing and withdrawing cash with users. The automated teller machine 2 has a banknote handling device 20. The banknote handling device 20 executes depositing and withdrawing of banknotes. Fig. 2 is a schematic diagram showing an example of a banknote handling device. As shown in Fig. 2, the banknote handling device 20 has, for example, a top module 201, a recycling stacker module 202, and a cash box module 203.

[0015] The top module 201 is a mechanism for receiving and delivering banknotes to and from a user. The top module 201 has a banknote deposit / withdrawal port 211. The banknote deposit / withdrawal port 211 receives cash inserted by a user and sends it into the top module 201. The banknote deposit / withdrawal port 211 also ejects in-machine banknotes sent to the top module 201 so that the user can receive them.

[0016] The recycling stacker module 202 is a mechanism for storing and dispensing banknotes. The recycling stacker module 202 has a recycling stacker 212. The recycling stacker 212 stores banknotes and delivers the stored banknotes when dispensing. For example, the recycling stacker 212 can store banknotes using a film winding method in which banknotes are sandwiched in order between films and wound into a roll.

[0017] The cash box module 203 is a mechanism for collecting banknotes. The cash box module 203 has a cash box 213. The cash box 213 stores collected banknotes such as banknotes collected during settlement of banknotes stored in the recycling stacker 212 and damaged banknotes.

[0018] Banknotes are transported using the transport path 210. For example, banknotes inserted through the banknote deposit / withdrawal slot 211 are transported via the transport path 210 to the recycling stacker 212 or the cash box 213. In FIG. 2 , a path to one recycling stacker 212 is shown as an example, but paths also extend to other recycling stackers 212. The transport path 210 is also used when sending banknotes from the recycling stacker 212 to the banknote deposit / withdrawal slot 211 or the cash box 213. Although not shown, the banknote processing device 20 also has other mechanical modules such as a validation unit that validates banknotes. Here, a module is a component that has a predetermined function and is made up of multiple parts. A mechanical module is a module that includes a mechanism that performs a mechanical operation. A mechanical module may include other mechanical modules.

[0019] Continuing the explanation, returning to Fig. 1, the operation company system 3 is an information system used by a company such as a bank that actually operates the automated teller machine 2. The operation company system 3 notifies the administrator of the operation status of the automated teller machine 2, etc.

[0020] The maintenance company system 4 is an information system used by a maintenance company that is in charge of maintenance work on the automated teller machine 2. The maintenance company system 4 notifies maintenance workers and the like of the operational status of the automated teller machine 2.

[0021] The maintenance work assist device 1 performs an abnormality determination to detect an abnormality from data such as the operation history of the banknote processing device 20, predicts an abnormal state indicating what type of abnormality has occurred, and notifies the operation company system 3 and the maintenance company system 4. The maintenance work assist device 1 will be described in detail below. As shown in FIG. 1 , the maintenance work assist device 1 has an operation information database 11, an information collection unit 12, a feature calculation unit 13, an inference unit 14, a learning execution unit 15, a machine learning model 16, and an explainable AI (artificial intelligence) 17. The maintenance work assist device 1 has, as operation phases, a learning phase in which the machine learning model 16 is made to learn, and an inference phase in which an abnormality determination is made and an abnormal state is predicted in response to input operation information for which the correct answer is unknown.

[0022] The operation information database 11 is a database that stores operation information of the banknote processing device 20, such as drive information of the actuators of the banknote processing device 20 and control information of firmware. Fig. 3 is a diagram showing an example of the drive information of the actuators and the control information of firmware.

[0023] Table 301 in Fig. 3 shows actuator drive information. For example, the actuator drive information for the cash box module 203 includes the drive time of the arm that pushes banknotes into the cash box 213. The actuator drive information for the recycling stacker module 202 includes the surface speed of the recycling roll that winds up banknotes. The actuator drive information for the top module 201 includes the rotation speed of the pick roller that pulls banknotes from the banknote deposit / withdrawal slot 211 into the banknote processing device 20. The actuator drive information for the entire banknote processing device 20 also includes the transport speed of banknotes in each transport section of the transport path 210. The actuator drive information for the mechanical module in which a sensor is installed also includes the sensor level (light intensity) of the media detection sensor.

[0024] Table 302 in Fig. 3 shows firmware control information. For example, the firmware control information for the cash box module 203 includes a speed setting set by the firmware for the arm that pushes banknotes into the cash box 213. The firmware control information for the recycling stacker module 202 includes a speed set by the firmware for the motor that rotates the recycling roll, a setting value for the direction of rotation, and a flag indicating whether the rotation speed is stable. The firmware control information for the entire banknote processing device 20 also includes control commands set by the firmware and a history of errors that have occurred within the device.

[0025] Continuing the explanation, returning to Fig. 1 , the information collection unit 12 acquires firmware log files containing operation information of the banknote processing device 20, such as actuator drive information and firmware control information, from the banknote processing device 20. The information collection unit 12 then accumulates the firmware log files obtained hourly in the operation information database 11.

[0026] The feature amount calculation unit 13 acquires operation information of the banknote processing device 20 from the firmware log file stored in the operation information database 11. Next, the feature amount calculation unit 13 acquires operation information for determination that indicates the operation state of each mechanical module of the banknote processing device 20, performs data cleansing, and then classifies the information using predetermined indices.

[0027] The operation information for determination is information used to detect abnormalities in each mechanical module and is information that represents the operating state of each mechanical module. For example, the operation information for determination can be drive information of the actuators in each mechanical module. This operation information for determination is an example of "first information."

[0028] The predetermined index is an index for classifying the judgment-use operation information so that it can be determined whether or not there is an abnormality. The information used as the predetermined index can be information indicating the control status of each mechanical module. For example, the predetermined index can be an index such that the statistics of the judgment-use operation information after classification when it is normal are normally distributed, and anything that deviates from this normal distribution can be determined to be abnormal. The information used as the predetermined index is an example of "second information."

[0029] The feature calculation unit 13 calculates statistics for one command of the classified operation information for determination, which information follows a normal distribution, and sets the information as the feature of the mechanical module to be subjected to abnormality determination. A command is an instruction to cause each module to perform a predetermined process. In response to one command, each module that executes the predetermined process specified by the ancestor command performs one or more operations. In this embodiment, one command is treated as one piece of data, and the feature calculation unit 13 calculates statistics within the same command. The feature calculation unit 13 calculates, for example, the average, variance, or number of times as the statistical quantity. Then, if the system is in the learning phase, the feature calculation unit 13 outputs the feature for abnormality determination of the module to be subjected to abnormality determination to the learning execution unit 15. If the system is in the inference phase, the feature calculation unit 13 outputs the feature for abnormality determination of the module to be subjected to abnormality determination to the inference unit 14.

[0030] A specific example of classification will be described below. For example, the operation will be described when the target of abnormality determination is the cash box module 203. The feature amount calculation unit 13 acquires, as the operation information for determination, the drive time of the arm that pushes banknotes into the cash box 213, which is drive information of the actuator in the cash box module 203. Next, the feature amount calculation unit 13 classifies the acquired actuator drive information using a control command obtained from the firmware control information as a predetermined index to be used for classification.

[0031] 4 is a diagram showing an example of a command. For example, commands include a reset command, a deposit command, a return command, and a collection command, as shown in Table 303. For each command, the actuator in each mechanical module of the banknote processing device 20 performs an operation corresponding to the command.

[0032] Next, the feature calculation unit 13 further classifies the actuator drive information using the control speed mode obtained from the firmware control information as a predetermined index used for classification.

[0033] Thereafter, the feature calculation unit 13 calculates, for each classification, statistics of the drive time of the arm that pushes banknotes into the cash box 213. Here, the statistics obtained by classifying the drive time of the arm that pushes banknotes into the cash box 213 by command and control speed mode follow a normal distribution. Therefore, if the statistics of the arm drive time acquired for abnormality determination deviates from this normal distribution, the statistics are likely to be abnormal values, and there is a high possibility that an abnormality has occurred in the cash box 213.

[0034] As another example, the operation when the target of abnormality judgment is the recycling stacker module 202 will be described. The feature calculation unit 13 acquires the surface speed of the recycling roll that winds up banknotes, which is drive information of the actuator in the recycling stacker module 202, as operation information for judgment. Next, the feature calculation unit 13 classifies the acquired actuator drive information using control commands obtained from the firmware control information as a predetermined index to be used for classification. Furthermore, the feature calculation unit 13 classifies the actuator drive information using control speed modes obtained from the firmware control information as a predetermined index to be used for classification.

[0035] The feature calculation unit 13 then calculates statistics of the surface speed of the recycling roll that winds up the banknotes for each classification. Here, the statistics obtained by classifying the surface speed of the recycling roll that winds up the banknotes for each classification using the command and control speed mode follow a normal distribution. Therefore, if the statistics of the surface speed of the recycling roll that winds up the banknotes obtained for abnormality determination deviates from this normal distribution, it is likely to be an abnormal value, and there is a high possibility that an abnormality has occurred in the recycling stacker 212.

[0036] The machine learning model 16 is a model that receives the feature amount obtained by the feature amount calculation unit 13 as an input and outputs a determination result of the occurrence of an abnormality in the mechanical module that is the target of the abnormality determination.

[0037] In the learning phase, the learning execution unit 15 causes the machine learning model 16 to perform the following learning: The learning execution unit 15 receives input of the feature amounts of each mechanical module from the feature calculation unit 13. The learning execution unit 15 also receives input of information as to whether the mechanical module whose feature amount has been calculated by the feature calculation unit 13 is actually abnormal or normal from an input device or the like (not shown).

[0038] For example, the operation information of an abnormal banknote processing device 20 can be obtained as data obtained when an abnormality occurs in an abnormality reproduction experiment of the relevant mechanical module or in an actually operating banknote processing device 20. Furthermore, the operation information of a normal banknote processing device 20 can be obtained as data obtained when an operation experiment of a normal banknote processing device 20 or a banknote processing device 20 in actual operation is operating normally.

[0039] Next, the learning execution unit 15 generates learning data by labeling the acquired feature quantities with classification labels indicating whether the mechanical module from which the feature quantities were acquired is normal or abnormal. The learning execution unit 15 then uses the learning data to train and learn the machine learning model 16. In this way, the learning execution unit 15 generates a trained machine learning model 16. The trained machine learning model 16 can, for example, determine that a module having feature quantities that deviate from a normal distribution is abnormal.

[0040] In the inference phase, the inference unit 14 predicts an abnormal state of the mechanical module that is the target of abnormality determination, using the trained machine learning model 16 and the explainable AI 17, for the feature amounts of the mechanical module that is the target of abnormality determination calculated by the feature amount calculation unit 13. The operation of the inference unit 14 will be described in detail below. The inference unit 14 has an abnormality determination unit 141 and an abnormal state analysis unit 142.

[0041] The abnormality determination unit 141 receives an input of feature amounts of a mechanical module to be subjected to abnormality determination, for which the correct answer as to whether the module is normal or abnormal is unknown, from the feature amount calculation unit 13. Next, the abnormality determination unit 141 inputs the acquired feature amounts to the trained machine learning model 16. Then, the abnormality determination unit 141 obtains the abnormality determination result output from the machine learning model 16. The abnormality determination unit 141 obtains an abnormality determination result when the mechanical module to be subjected to abnormality determination has feature amounts that deviate from a normal distribution.

[0042] If the determination result indicates no abnormality and the banknote processing device 20 is normal, the abnormality determination unit 141 ends the abnormal state determination process. On the other hand, if the determination result indicates an abnormality, the abnormality determination unit 141 outputs the determination result of the abnormality determination to the abnormal state analysis unit 142. The determination result of the abnormality determination may also include information on the inference process, etc.

[0043] The explainable AI 17 is an AI that calculates the contribution of each feature value to an abnormality in an abnormality judgment from the judgment result of the abnormality judgment.

[0044] The abnormal state analysis unit 142 receives an input of the abnormality judgment result of the mechanical module for which an abnormality judgment has been obtained from the abnormality judgment unit 141. Then, the abnormal state analysis unit 142 inputs the abnormality judgment result to the explainable AI 17, and acquires as its output the contribution of each feature amount of the mechanical module judged to be abnormal to the abnormality.

[0045] Here, the abnormal state analysis unit 142 has correspondence information that maps each abnormal state to a pattern of important feature values ​​in that abnormal state. An abnormal state is information that indicates the specific type of abnormality occurring in the corresponding mechanical module. The abnormal state analysis unit 142 uses the correspondence information to determine the abnormal state corresponding to the abnormality determination result based on the contribution of each feature value obtained from the explainable AI 17. Thereafter, the abnormal state analysis unit 142 transmits information on the determined abnormal state to the operation company system 3 and the maintenance company system 4.

[0046] FIG. 5 is a diagram showing an example of contribution degrees calculated by the explainable AI. When the explainable AI 17 receives an input of the result of the abnormality determination, it outputs, for example, the contribution degrees for the feature quantities F1 to F4 shown in FIG. 5. In FIG. 5, the vertical axis represents the type of feature quantity, and the horizontal axis represents the contribution degree. In FIG. 5, a positive contribution degree indicates that the contribution degree contributes to the determination of an abnormality, and a negative contribution degree indicates that the contribution degree contributes to the determination of a normality. Furthermore, the sum of the contribution degrees of the feature quantities F1 to F4 corresponds to the abnormality determination result output from the machine learning model 16 for the data from which the abnormality determination was made. That is, in this case, the abnormality determination result output from the machine learning model 16 is an abnormality determination with a confidence level of +5.5.

[0047] Here, the contribution 101 of the feature F1 is +3.5. The contribution 102 of the feature F3 is +3.0. The contribution 103 of the feature F2 is +0.5. The contribution 104 of the feature F4 is −3.0. That is, by obtaining this information, the abnormal state analysis unit 142 determines that an abnormality has occurred based on the feature values ​​F1 to F3, and that the feature values ​​F1 and F3 have made the greatest contributions.

[0048] Fig. 6 is a diagram showing an example of correspondence information. The abnormal state analysis unit 142 has correspondence information indicating abnormal states and patterns of important feature quantities in those abnormal states, as shown in table 110 in Fig. 6. When the contribution levels shown in Fig. 5 are obtained, the abnormal state analysis unit 142 determines that the abnormal state is storage disorder B1 in the cash box 213, which corresponds to the pattern of feature quantities including the feature quantities F1 and F3 in the correspondence information, because the contribution levels of the feature quantities F1 and F3 to the abnormality determination are large.

[0049] 7 is a flowchart of the feature calculation process performed by the maintenance work assisting device 1 according to the embodiment. Next, the flow of the feature calculation process performed by the maintenance work assisting device 1 according to the embodiment will be described with reference to FIG.

[0050] The information collection unit 12 collects firmware log files including operation information of the banknote processing device 20 such as actuator drive information and firmware control information from the banknote processing device 20 and stores the collected files in the operation information database 11 (step S1).

[0051] The feature calculation unit 13 acquires operation information for determination to determine abnormalities in each mechanical module from the operation information of the banknote processing device 20 included in the firmware log file stored in the operation information database 11 (step S2).

[0052] Next, the feature amount calculation unit 13 classifies the determination operation information using any information among the operation information of the banknote processing device 20 included in the firmware log file as a predetermined index (step S3).

[0053] Thereafter, the feature amount calculation unit 13 calculates the statistics of the determination-use operation information for each category within one command, and sets the calculated statistics as feature amounts (step S4). These feature amounts follow a normal distribution.

[0054] 8 is a flowchart of the feature calculation process for the cashbox module. Next, the flow of the feature calculation process for the cashbox module 203 will be described with reference to FIG. 8 as a specific example of the feature calculation process shown in FIG.

[0055] The feature amount calculation unit 13 acquires operation information of the banknote processing device 20 contained in a firmware log file stored in the operation information database 11 (step S101).

[0056] Next, the feature calculation unit 13 acquires drive information of the actuator of the cash box module 203 as operation information for determining whether there is an abnormality in the cash box module 203 (step S102). The drive information of the actuator of the cash box module 203 is, for example, the drive time of the arm that pushes banknotes into the cash box 213.

[0057] Next, the feature amount calculation unit 13 performs data cleansing on the acquired actuator drive information of the cash box module 203 (step S103).

[0058] Next, the feature calculation unit 13 classifies the actuator drive information of the cash box module 203 using the control command as a predetermined index. Here, the control command will be described assuming that commands c1 to cn exist. The feature calculation unit 13 determines whether the control command is command c1 (step S104).

[0059] If the control command is command c1 (step S104: Yes), the feature calculation unit 13 uses the control speed mode as a predetermined index to classify the drive information of the actuator of the cash box module 203 when the control command is command c1. Here, a case where speeds v1 to vx are present as the control speed mode will be described. For example, the feature calculation unit 13 determines whether the control speed mode is speed v1 (step S105). If the control speed mode is not speed v1 (step S105: No), the feature calculation unit 13 determines whether the control speed mode is speed v2 (step S108). While FIG. 8 illustrates the determination when the control speed mode is speed v1 and the determination when the control speed mode is speed v2, the feature calculation unit 13 repeats the same determination for speeds v3 to v(x-1) in order to classify the drive information into speeds v1 to vx.

[0060] If the control speed mode is speed v1 (step S105: Yes), the feature calculation unit 13 calculates a statistic within one command of the actuator drive information when the control command is command c1 and the control speed mode is speed v1. Then, the feature calculation unit 13 sets the calculated feature as feature F1 (step S106).

[0061] If the control speed mode is speed v2 (step S108: Yes), the feature calculation unit 13 calculates a statistic within one command of the actuator drive information when the control command is command c1 and the control speed mode is speed v2. Then, the feature calculation unit 13 sets the calculated feature as feature F2 (step S109).

[0062] If the control speed mode is not speed v2 (step S108: No), the feature calculation unit 13 performs the same process on the actuator drive information when the control command is command c1 and the control mode is each of speeds v3 to vx. For example, the feature calculation unit 13 calculates statistics within one command of the actuator drive information when the control command is command c1 and the control speed mode is speed vx. Then, the feature calculation unit 13 sets the calculated feature as feature Fx (step S110). The feature calculation unit 13 calculates feature F1 to Fx.

[0063] On the other hand, if the control command is not command c1 (step S104: No), the feature calculation unit 13 repeats the determination for commands c2 to c(n-1) in order and classifies them into commands c1 to cn. While FIG. 8 shows an example of the determination for the case where the control command is command c1, the feature calculation unit 13 performs a similar determination for commands c2 to c(n-1). Then, for each of the control commands c2 to c(n-1), the feature calculation is performed in the same manner as in steps S105 to S110 for command c1.

[0064] For example, when the control command is command cn, the feature calculation unit 13 uses the control speed mode as a predetermined index to classify the drive information of the actuator of the cash box module 203 when the control command is command cn. Then, the feature calculation unit 13 determines whether the control speed mode is speed v1 (step S111). If the control speed mode is not speed v1 (step S111: No), the feature calculation unit 13 determines whether the control speed mode is speed v2 (step S113). The feature calculation unit 13 repeats the same determination for speeds v3 to v(x-1) in order, and classifies them into speeds v1 to vx.

[0065] If the control speed mode is speed v1 (step S111: Yes), the feature calculation unit 13 calculates a statistic within one command of the actuator drive information when the control command is command cn and the control speed mode is speed v1. Then, the feature calculation unit 13 sets the calculated feature as feature F((n-1)x+1) (step S112).

[0066] If the control speed mode is speed v2 (step S113: Yes), the feature calculation unit 13 calculates a statistic within one command of the actuator drive information when the control command is command cn and the control speed mode is speed v2. Then, the feature calculation unit 13 sets the calculated feature as feature F((n-1)x+2) (step S114).

[0067] If the control speed mode is not speed v2 (step S113: No), the feature calculation unit 13 performs the same process on the actuator drive information when the control command is command cn and the control mode is each of speeds v3 to vx. For example, the feature calculation unit 13 calculates statistics within one command of the actuator drive information when the control command is command cn and the control speed mode is speed vx. Then, the feature calculation unit 13 sets the calculated feature as feature F(nx) (step S115). The feature calculation unit 13 calculates feature quantities F((n-1)x+1) to F(nx). In summary, the feature calculation unit 13 calculates feature quantities F1 to F(nx).

[0068] 9 is a flowchart of the feature calculation process for the recycling stacker module. Next, the flow of the feature calculation process for the recycling stacker module 202 will be described with reference to FIG. 9 as a specific example of the feature calculation process shown in FIG.

[0069] The feature amount calculation unit 13 acquires operation information of the banknote processing device 20 included in the firmware log file stored in the operation information database 11 (step S201).

[0070] Next, the feature calculation unit 13 acquires driving information of the actuator of the recycling stacker module 202 as operation information for determining whether there is an abnormality in the recycling stacker module 202 (step S202). The driving information of the actuator of the recycling stacker module 202 is, for example, the surface speed of the recycling roll that winds up the banknotes.

[0071] Next, the feature amount calculation unit 13 performs data cleansing on the acquired driving information of the actuator of the recycling stacker module 202 (step S203).

[0072] Next, the feature amount calculation unit 13 classifies the driving information of the actuators of the recycling stacker module 202 using the control command as a predetermined index (step S204).

[0073] Next, the feature calculation unit 13 classifies the drive information of the actuator of the recycling stacker module 202 classified for each control command using the rotation direction of the recycling motor as a predetermined index (step S205).

[0074] Next, the feature calculation unit 13 calculates the statistics within one command of the actuator drive information of the recycling stacker module 202, which is classified using the control command and the rotation direction of the recycling motor as predetermined indicators, and sets the calculated statistics as the feature (step S206).

[0075] 10 is a flowchart of the learning process performed by the maintenance work assisting device 1 according to the embodiment. Next, the flow of the learning process performed by the maintenance work assisting device 1 according to the embodiment will be described with reference to FIG.

[0076] The feature amount calculation unit 13 executes the feature amount calculation process illustrated in FIG. 6 (step S11).

[0077] Next, the learning execution unit 15 generates learning data by labeling each of the feature amounts calculated by the feature amount calculation unit 13 with a classification label of abnormal or normal. Then, the learning execution unit 15 uses the learning data to train the machine learning model 16 (step S12).

[0078] Then, the learning execution unit 15 acquires the trained machine learning model 16 (step S13).

[0079] 11 is a flowchart of the abnormal state determination process performed by the maintenance work assisting device 1 according to the embodiment. Next, the flow of the abnormal state determination process performed by the maintenance work assisting device 1 according to the embodiment will be described with reference to FIG.

[0080] The feature amount calculation unit 13 executes the feature amount calculation process illustrated in FIG. 6 (step S21).

[0081] The abnormality determination unit 141 performs abnormality determination on the feature calculated by the feature calculation unit 13 using the trained machine learning model 16 (step S22).

[0082] The abnormality determination unit 141 determines whether or not the target mechanical module is abnormal based on the output result from the trained machine learning model 16 (step S23).

[0083] If the abnormality determination unit 141 determines that the target mechanical module is normal (step S23: No), the abnormal state determination process ends.

[0084] On the other hand, if the abnormality determination unit 141 determines that the target mechanical module is abnormal (step S23: Yes), the abnormal state analysis unit 142 analyzes the abnormality determination result using the explainable AI 17 (step S24). For example, the explainable AI 17 analyzes the determination result and calculates the contribution of each feature amount.

[0085] Then, the abnormal state analysis unit 142 determines an abnormal state using the output from the explainable AI 17 (step S25). For example, the abnormal state analysis unit 142 determines an abnormal state by using the contribution of each feature calculated by the explainable AI 17 as correspondence information that maps each abnormal state and a pattern of important feature values ​​in that abnormal state (step S25).

[0086] Thereafter, the abnormal state analysis unit 142 transmits information on the determined abnormal state to the operation company system 3 and the maintenance company system 4, and notifies them of the abnormal state of the banknote processing device 20 (step S26).

[0087] In the above description, the recycling stacker module 202 and the cash box module 203 are the targets of abnormality determination, but the targets of abnormality determination are not limited to these. FIG. 12 is a diagram showing an example of a mechanical module that is the target of abnormality determination. For example, as shown in FIG. 12, the top module 201, the banknote deposit / withdrawal slot 211, etc. can also be the targets of abnormality determination. In this case, the predetermined index used for classification can be information included in the actuator drive information shown in Table 301 of FIG. 3 or the firmware control information shown in Table 302.

[0088] As described above, the maintenance work assist device according to this embodiment acquires determination operation information for determining an abnormality in a module from operation information of a banknote processing device, and classifies the determination operation information according to a predetermined index to obtain statistical information according to a normal distribution.The maintenance work assist device then trains a machine learning model using the statistical information according to the normal distribution of the classified determination operation information as feature quantities.Furthermore, the maintenance work assist device performs an abnormality determination using the trained machine learning model using the statistical information according to the normal distribution of the classified determination operation information as feature quantities, and determines an abnormal state using explanatory AI for the determination result.

[0089] By identifying causes using operational information about the banknote processing device, it is possible to estimate and present causes based on the actual device status. This reduces the dependency on individual skills in identifying causes, increasing the opportunities for even personnel at the operating company who actually operate the banknote processing device, who do not have specialized knowledge or know-how, to appropriately identify causes and take corrective measures. Furthermore, maintenance companies can detect high risks of machine shutdowns in advance and propose proactive measures to the operating company. This increases the success rate of operational companies' corrective measures, thereby reducing opportunity losses due to machine shutdowns. Furthermore, by optimizing the number of maintenance personnel dispatches, maintenance costs are reduced. For maintenance companies, reducing the dependency on individual skills in maintenance reduces the variability in the labor required for cause identification and corrective measures due to the level of expertise of maintenance personnel, thereby reducing the time and cost required for training maintenance personnel. This makes it possible to improve the operational efficiency of banknote processing devices.

[0090] (Hardware Configuration) Fig. 13 is a hardware configuration diagram of the maintenance work assist device 1. Next, an example of a hardware configuration for realizing each function of the maintenance work assist device 1 will be described with reference to Fig. 13 .

[0091] 13, the maintenance work assist device 1 includes, for example, a CPU (Central Processing Unit) 91, a memory 92, a hard disk 93, and a network interface 94. The CPU 91 is connected to the memory 92, the hard disk 93, and the network interface 94 via a bus.

[0092] The network interface 94 is an interface for communication between the maintenance work assist device 1 and an external device. The network interface 94 relays communication between the operation company system 3, the maintenance company system 4, the banknote processing device 20, and the CPU 91, for example.

[0093] The hard disk 93 is an auxiliary storage device. The hard disk 93 stores the operation information database 11, the machine learning model 16, and the explainable AI 17 illustrated in Fig. 1. The hard disk 93 also stores various programs, including programs for realizing the functions of the information collection unit 12, the feature calculation unit 13, the inference unit 14, and the learning execution unit 15 illustrated in Fig. 1.

[0094] The memory 92 is a main storage device, and may be, for example, a dynamic random access memory (DRAM).

[0095] The CPU 91 reads various programs from the hard disk 93, expands them into the memory 92, and executes them. As a result, the CPU 91 realizes the functions of the information collection unit 12, the feature calculation unit 13, the inference unit 14, and the learning execution unit 15 illustrated in FIG.

[0096] REFERENCE SIGNS LIST 1 Maintenance work assistance device 2 Automatic teller machine 3 Operation company system 4 Maintenance company system 11 Operation information database 12 Information collection unit 13 Feature calculation unit 14 Inference unit 15 Learning execution unit 16 Machine learning model 17 Explainable AI 20 Banknote processing device 141 Abnormality determination unit 142 Abnormal state analysis unit

Claims

1. An information collection unit that collects operation information of a banknote processing device, a feature amount calculation unit that classifies the operation information collected by the information collection unit and calculates a feature amount that follows a normal distribution, and an inference unit that inputs the feature amount calculated by the feature amount calculation unit into a machine learning model to perform an abnormality determination of the banknote processing device. A maintenance work assistance device characterized by comprising.

2. The feature amount calculation unit acquires first information indicating an operation state included in the operation information, and classifies the first information using second information indicating a control state included in the operation information as an index, and calculates a feature amount that follows a normal distribution. The maintenance work assistance device according to claim 1, characterized by doing so.

3. The feature amount calculation unit uses drive information of an actuator as the first information and uses control information of firmware as the second information. The maintenance work assistance device according to claim 2, characterized by doing so.

4. The feature amount calculation unit calculates statistical information of the classification result within one command and uses it as the feature amount. The maintenance work assistance device according to claim 2, characterized by doing so.

5. The feature amount calculation unit calculates an average, a variance, or a frequency as the statistical information. The maintenance work assistance device according to claim 4, characterized by doing so.

6. When the inference unit determines that the banknote processing device is abnormal, it analyzes the abnormal state of the banknote processing device using explainable AI with respect to the result of the abnormality determination by the machine learning model. The maintenance work assistance device according to claim 1, characterized by doing so.

7. A maintenance work assistance method, characterized in that a maintenance work assistance device collects operation information of a banknote processing device, classifies the collected operation information, calculates a feature amount that follows a normal distribution, and inputs the calculated feature amount into a machine learning model to perform an abnormality determination of the banknote processing device.

8. A maintenance work assistance program, characterized in that a computer is caused to execute a process of collecting operation information of a banknote processing device, classifying the collected operation information, calculating a feature amount that follows a normal distribution, and inputting the calculated feature amount into a machine learning model to perform an abnormality determination of the banknote processing device.

Images