Tooth contact position setting quantity estimation device, machine learning device and robot system

Abstract

Tooth contact position setting quantity estimation device (1, 1') which performs the processing with reference to the estimation of a tooth contact position setting quantity for dimensional data of parts that represent a power transfer mechanism, comprising: a machine learning device (100) that performs the processing with respect to the estimation of the tooth contact position setting quantity for the dimensional data of parts that represent the power transmission mechanism, wherein the machine learning device (100) observes part dimension data, which are dimensional data of parts that observe the power transmission mechanism, as a state variable indicating a current state of an environment, and performs the processing relating to the estimation of the tooth contact position setting quantity for dimensional data of parts that represent the power transmission mechanism during the assembly of the power transmission mechanism by performing the processing relating to machine learning based on the observed state variable.

Description

BACKGROUND OF THE INVENTION 1. Field of the invention

[0001] This invention relates to a tooth contact position adjustment amount estimation device, a machine learning device, and a robot system. Specifically, this invention relates to a tooth contact position adjustment amount estimation device, a machine learning device, and a robot system for estimating a tooth contact position adjustment amount of gears used in a structure for transmitting power via shafts whose axial lines intersect. 2. Description of the state of the art

[0002] A bevel gear is used in a structure for transmitting power through shafts whose axial lines intersect. A worm gear is used in a structure for transmitting power through shafts whose axial lines are spaced apart. If the tooth contact of the gears mounted on each shaft and meshing with each other is not adequate, abnormal noise, vibration, and abnormal tooth wear may occur. It is important to optimize the shape of the gears for appropriate tooth contact. The positions in which the gears are mounted may differ due to a dimensional error of the parts mounted with the gears.

[0003] Fig. Figure 6 is a diagram illustrating a transmission structure using bevel gears. Fig. 6, the power on the input side is transmitted to the output side through the bevel gears, whose axial lines are arranged at a right angle. The bevel gears are mounted on a reducer or a holder. The reducer or holder is mounted on a housing. Thus, the tooth contact of the bevel gears changes due to a dimensional error of the mounted parts, such as the bevel gears or the holder. For example, if the shaft mounted on the bevel gear on the input side is in Fig. 6 is short, the meshing position of the input-side bevel gear and the output-side bevel gear differs from each other. This causes problems such as unusual noise, vibration, or abnormal tooth wear. In such a case, the tooth contact must be adjusted. A tooth contact position adjusting shim is inserted between parts such as the gears or the holder to adjust the tooth contact (JP 2016-142379 A, ​​JP 2013-160361 A).

[0004] JP 3586997 B2 discloses a technique for determining tooth contact adjustment in image processing, for example, as a conventional technique for adjusting tooth contact between gears.

[0005] However, dimensional variation exists in parts such as gears or the holder. Therefore, the optimal mounting position is different for each part. In the tooth contact adjustment process, the optimal tooth contact position adjustment quantity (such as the thickness of the disk or the number of disks) is finally determined by an experienced assembly worker. One problem is that it takes a long time for an inexperienced worker to determine the optimal tooth contact position adjustment quantity.

[0006] Tooth contact between gears is evaluated using an image in JP 2016-142379 A, ​​but the tooth contact must be evaluated while the gears are actually assembled. If the tooth contact position is found to be inappropriate, a step of removing the assembled parts and inserting a washer is performed. Thus, productivity or work efficiency decreases in the technique disclosed in JP 2016-142379 A.

[0007] It is therefore an object of this invention to provide a tooth contact position adjustment amount estimating device capable of determining the thickness or number of washers to be used before assembling a power transmission mechanism, a machine learning device, and a robot system. DE 20 2017 106 132 U1 discloses a neural network system, wherein the neural network system is configured to receive an observation characterizing a current state of a real environment with which a robot agent interacts to perform a robot task, and to process the observation to generate a control output defining an action to be performed by the robot agent in response to the observation. DE 10 2017 009 822 A1 describes a machine learning device that creates a learning model, thereby predicting the machine delay time caused in a machine with high accuracy. US 9 976 631 B1 relates to a planetary gear that can be adapted for different gear ratios. JP H10-252 872A discloses a bevel gear mechanism.JPH03-163 327A describes a method in which a tapered roller bearing is loaded to select suitable washers, and CN 1 05 954 021 A relates to a method for detecting a tooth flank contact area of ​​a spiral bevel gear in an automotive rear axle differential mechanism by means of a camera. SUMMARY OF THE INVENTION

[0008] A tooth contact position adjustment amount estimation device, a machine learning device, and a robot system according to this invention solve at least some of the above problems. According to one embodiment of this invention, machine learning is performed using input data and output data corresponding to the input data as teaching data. The input data is dimensional data of parts such as gears, which are a structure for transmitting power via shafts whose axial lines intersect. The output data indicates the optimal tooth contact position adjustment amount when the parts are assembled. A learned model acquired as a result of machine learning is used to estimate the tooth contact position adjustment amount from the part dimensional data. The thickness or number of washers is determined based on the estimated tooth contact position adjustment amount.This makes the mounting position adjustment step easier and the assembly is carried out efficiently.

[0009] According to one aspect of this invention, there is provided a tooth contact position adjustment amount estimation device that performs processing related to estimation of a tooth contact position adjustment amount for dimensional data of parts constituting a power transmission mechanism, comprising: a machine learning device that performs processing related to estimation of the tooth contact position adjustment amount for dimensional data of parts constituting a power transmission mechanism, wherein the machine learning device observes part dimensional data, which is dimensional data of parts observing the power transmission mechanism, as a state variable indicating a current state of an environment, and the processing related to estimation of the tooth contact position adjustment amount for dimensional data of parts constituting the power transmission mechanism,when assembling the power transmission mechanism by executing machine learning processing based on the observed state variable.

[0010] According to another aspect of this invention, there is provided a machine learning apparatus that performs processing related to estimation of a tooth contact position adjustment amount for dimensional data of parts constituting a power transmission mechanism, wherein the machine learning apparatus observes part dimensional data, which is dimensional data of parts constituting the power transmission mechanism, as a state variable indicating a current state of an environment, and performs processing related to estimation of the tooth contact position adjustment amount for dimensional data of parts constituting the power transmission mechanism in assembling the power transmission mechanism by executing the processing related to machine learning based on the observed state variable.

[0011] According to another aspect of this invention, there is provided a robot system including a controller that determines the number of washers used to assemble a power transmission mechanism and instructs a robot to mount a predetermined number of washers on the parts based on the tooth contact position adjustment amount estimated by the tooth contact position adjustment amount estimating device.

[0012] According to another aspect of this invention, there is provided a tooth contact position adjustment amount estimation system in which a plurality of devices are connected to each other via a network, the plurality of devices including at least one tooth contact position adjustment amount estimation device equipped with a learning unit.

[0013] According to the present invention, the thickness or number of shims is determined based on the estimated tooth contact position adjustment amount. This reduces the worker's workload during the assembly position adjustment step and performs the assembly efficiently. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] The above and other objects and features of the present invention will become clear from the following descriptions of the embodiments with reference to the accompanying drawings, in which: Fig. 1 is a schematic hardware configuration diagram of a tooth contact position adjustment amount estimation device according to a first embodiment; Fig. 2 is a diagram including a schematic configuration of a robot system; Fig. 3 is a schematic functional block diagram of the tooth contact position adjustment amount estimation device according to the first embodiment; Fig. 4 is a diagram illustrating exemplary part dimension data; Fig. 5 is a schematic functional block diagram illustrating a form of the tooth contact position adjustment amount estimation device according to the second embodiment; Fig. 6 is a diagram illustrating an exemplary power transmission mechanism; Fig. 7 is a diagram illustrating an exemplary system of a three-tier structure including a cloud server, a fog computer, and an edge computer; Fig. 8 is a schematic hardware configuration diagram of a tooth contact position adjustment amount estimation device implemented on a computer; Fig. 9 is a schematic configuration diagram of a tooth contact position adjustment amount estimating device according to a third embodiment; Fig. 10 is a schematic configuration diagram of a tooth contact position adjustment amount estimation device according to a fourth embodiment; and Fig. 11 is a schematic configuration diagram of a tooth contact position adjustment amount estimating device according to a fifth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Embodiments of this invention will be described below with reference to the drawings.

[0016] Fig. 1 is a schematic hardware configuration diagram illustrating a tooth contact position adjustment amount estimation device 1 according to the first embodiment. The tooth contact position adjustment amount estimation device 1 is implemented on a computer such as a personal computer provided for a robot system for manufacturing a machine, or on a cell computer, a host computer, an edge server, and a cloud server connected to the robot system via a wired / wireless network. According to this embodiment, the tooth contact position adjustment amount estimation device 1 is implemented on a personal computer provided for a robot system 2.

[0017] The tooth contact position adjustment amount estimation device 1 is connected to the robot system 2, which Fig. 2, via an interface 19. The robot system 2 includes a controller 2A and a robot 2B controlled by the controller 2A. The robot 2B assembles the parts that have been previously fed and arranged by a parts feeder or the like, thereby assembling a power transmission mechanism. When assembling the parts, the robot system 2 mounts only the required gear adjustment disks on the parts. The tooth contact position adjustment amount estimation device 1 acquires dimensional data of the respective parts that have been previously input to the controller 2A via the interface 19. The tooth contact position adjustment amount estimation device 1 outputs the tooth contact position adjustment amount as an estimation result to the controller 2A via the interface 19.The controller 2A determines the number of disks to be mounted on the parts or the thickness of one disk based on the tooth contact position adjustment amount input from the tooth contact position adjustment amount estimator 1. A detailed configuration of the robot system 2 for assembling a power transmission mechanism is known, and therefore, a detailed description thereof is omitted from this specification.

[0018] A central processing unit (CPU) 11 provided for the tooth contact position adjustment amount estimation device 1 is a processor for fully controlling the tooth contact position adjustment amount estimation device 1. The CPU 11 reads a system program stored in a read-only memory (ROM) 12 via a bus 20. The CPU 11 controls the entire tooth contact position adjustment amount estimation device 1 according to the system program. Temporary calculation data or display data, or various data items, input by a worker via an input device 71 are temporarily stored in a random access memory (RAM) 13.

[0019] A non-volatile memory 14 is backed up by, for example, a battery (not shown). The non-volatile memory 14 is a memory in which stored states are retained even when the tooth contact position adjustment amount estimation device 1 is turned off. The non-volatile memory 14 has a setting area in which the setting information for the functions of the tooth contact position adjustment amount estimation device 1 is stored. The non-volatile memory 14 stores the program input from the input device 71, the programs read from the external storage device (not shown), or dimensional data of the respective parts acquired by the robot system 2 via the interface 19. The programs or the various data items of the data stored in the non-volatile memory 14 can be loaded into the RAM 13 upon execution / use.Furthermore, various system programs such as a known analysis program (including a system program for controlling communication with a machine learning device 100, which will be described later) are previously written into the ROM 12.

[0020] A display device 70 displays each data point read into the memory or all data acquired due to an executed program or the like. Data or the like output from the machine learning device 100 described below is input to the display device 70 via an interface 17. The display device 70 displays the input data or the like thereon. The input device 71 is configured from a keyboard or a pointing device. The input device 71 receives commands based on worker operations, data, and the like. The input device 71 forwards the received commands, data, and the like to the CPU 11 via an interface 18.

[0021] An interface 21 connects each part of the tooth contact position adjustment amount estimation device 1 to the machine learning device 100 in the tooth contact position adjustment amount estimation device 1. The machine learning device 100 includes a processor 101, a ROM 102, a RAM 103, and a non-volatile memory 104. The processor 101 controls the entire machine learning device 100. The ROM 102 controls a system program and the like. The RAM 103 stores temporary data in each machine learning processing. The non-volatile memory 104 stores a learning model and the like. The machine learning device 100 observes each piece of information (such as dimensional data of each part acquired by the robot system 2) capable of being acquired by the tooth contact position adjustment amount estimation device 1 via the interface 21.Further, the tooth contact position adjustment amount estimation device 1 causes the display device 70 to display the information on the tooth contact position adjustment amount estimated by the machine learning device 100. The tooth contact position adjustment amount estimation device 1 outputs the information on the tooth contact position adjustment amount estimated by the machine learning device 100 to the robot system 2.

[0022] Fig. 3 is a schematic functional block diagram of the tooth contact position adjustment amount estimation device 1 according to the first embodiment. The CPU 11 for the tooth contact position adjustment amount estimation device 1 and the processor 101 of the machine learning device 100 shown in Fig. 1, executes the system programs, and thus controls the function of each part, so that each function of a function block as shown in Fig. 3 illustrated, was implemented.

[0023] The tooth contact position adjustment amount estimation device 1 includes an output unit 34 for outputting the information on the tooth contact position adjustment amount estimated by the machine learning device 100 to the outside. The output unit 34 can output the estimation result of the tooth contact position adjustment amount output, for example, from the machine learning device 100 to the display device 70. This configuration is used when the worker needs to adjust, for example, the number of pulleys used to assemble a power transmission mechanism in the robot system 2. Furthermore, the output unit 34 can directly output the estimation result of the tooth contact position adjustment amount output, for example, from the machine learning device 100 to the robot system 2.The configuration is used when the tooth contact adjustment amount estimation device 1 adjusts the number of disks required for assembling a power transmission mechanism in the robot system 2 via a communication line.

[0024] On the other hand, the machine learning device 100 provided for the tooth contact position adjustment amount estimation device 1 performs machine learning to estimate the tooth contact position adjustment amount for the dimensional data of the parts constituting a power transmission mechanism assembled by the robot system 2. The machine learning device 100 includes software (such as a learning algorithm) and hardware (such as the processor 101) for executing machine learning. The machine learning device 100 learns a model structure indicating a correlation between the dimensional data of the parts constituting the power transmission mechanism and the tooth contact position adjustment amount.

[0025] As in Fig. As shown in Figure 3, the machine learning device 100 includes a state observation unit 106, a label data acquisition unit 108, a learning unit 110, and an estimation result output unit 122. The state observation unit 106 observes part dimension data S1 indicating the dimensions of parts constituting a power transmission mechanism assembled by the robot system 2 as a state variable S indicating a current state of an environment. The label data acquisition unit 108 acquires label data L including tooth contact position adjustment amount data L1. The learning unit 110 associates and learns the dimension data of the parts constituting the power transmission mechanism and the estimated tooth contact position adjustment amount by using the state variable S and the label data L.

[0026] The part dimension data S1 in the state variable S observed by the state observation unit 106 is dimension data of the parts constituting the power transmission mechanism detected by the robot system 2. As shown in Fig. As illustrated in Figure 4, the dimensional data of the parts constituting the power transmission mechanism can be defined as a matrix, where the dimensional data of the respective parts are elements. The dimensions of the parts are observed using a 3D measuring device or the like before the parts are supplied to the robot system 2. The dimensions of the parts constituting the power transmission mechanism can be measured by an image of each part previously acquired by an imaging device. The measured dimensional data of the parts are stored in the memory of the robot system 2 via a network or an external device such as a USB device. When the robot system 2 is provided with a sensor such as a 3D measuring device or an imaging device, the dimensional data of each part can be measured when the part is supplied to the robot system 2.

[0027] The label data acquisition unit 108 acquires, as adjustment amount data L1, the data on the tooth contact position adjustment amount acquired by the robot system 2 at the time of assembling a power transmission mechanism during learning. The label data acquisition unit 108 may acquire, as adjustment amount data L1, the data on the tooth contact position adjustment amount input by the worker via the input device 71. The adjustment amount data L1 may indicate the distance between a reference position for adjusting the tooth contact position and a predetermined part of the gear. Furthermore, the adjustment amount data L1 may indicate the number of disks inserted. Multiple points of the adjustment amount data L1 may be acquired from one power transmission mechanism. For example, in the case of the power transmission mechanism shown in Fig. As illustrated in Figure 6, the label data acquisition unit 108 acquires two adjustment sets, including the input-side adjustment set y1 and the output-side adjustment set y2, as the adjustment set data L1. The label data acquisition unit 108 is used for learning. Thus, the label data acquisition unit 108 may not be critical to the machine learning device 100 after the learning is completed.

[0028] The learning unit 110 learns a correlation between the state variable S (the part dimension data S1 indicating the dimensions of the parts constituting the power transmission mechanism) and the label data L (the tooth contact position adjustment amount data L1) according to a learning algorithm. For example, the learning unit 110 learns a correlation between the part dimension data S1 included in the state variable S and the adjustment amount data L1 included in the label data L. The learning unit 110 repeatedly performs learning based on a collection of data including the state variable S and the label data L. The learning unit 110 performs machine learning by using the adjustment amount data L1 acquired by the label data acquisition unit 108 and the part dimension data S1 of the parts constituting the power transmission mechanism corresponding to the adjustment amount data L1.

[0029] The learning unit 110 performs machine learning online with the robot system 2. Furthermore, the learning unit 110 can acquire accumulated data when a power transmission mechanism is mounted on the robot system 2 via an external device such as a USB device (not shown). That is, the learning unit can perform learning offline. In this case, when the data acquired by the external device (not shown) is stored in, for example, the non-volatile memory 14, the state observation unit 106 and the label data acquisition unit 108 acquire the state variable S and the label data L, respectively, from the stored data. The learning unit 110 performs machine learning based on the acquired data.

[0030] The learning cycle is repeated, so that the learning unit 110 automatically interprets a correlation between the dimensional data (the part dimension data S1) of the parts constituting the power transmission mechanism and the tooth contact position adjustment amount (the adjustment amount data L1). The correlation between the part dimension data S1 and the adjustment amount data L1 is essentially unknown at the beginning of the learning algorithm. The learning unit 110 gradually interprets the correlation between the part dimension data S1 and the adjustment amount data L1 as learning progresses. As a result, the learning unit 110 acquires a learned model that captures the correlation between the part dimension data S1 and the adjustment amount data L1.

[0031] The estimation result output unit 122 estimates the tooth contact position adjustment amount based on the dimensional data of the parts constituting the power transmission mechanism by using the result (learned model) learned by the learning unit 110. The estimation result output unit 122 outputs the estimated tooth contact position adjustment amount. The estimation result of the tooth contact position adjustment amount output by the estimation result output unit 122 is output to the display device 70 by the output unit 34 for display thereon. The worker operates the estimation result displayed on the display device 70 and adjusts the tooth contact position adjustment amount. Further, the output unit 34 can output the estimation result of the tooth contact position adjustment amount output by the estimation result output unit 122 to the robot system 2.In this case, the controller 2A of the robot system 2 determines the number of disks required for assembling the power transmission mechanism based on the input estimation result of the tooth contact position adjustment amount. The controller 2A of the robot system 2 outputs a control command to assemble a specific number of disks to the robot 2B. The robot 2B assembles the power transmission mechanism in response to the control command. The estimation result output unit 122 is used to estimate the tooth contact position adjustment amount. Thus, the estimation result output unit 122 may not be essential to the machine learning device 100 if only the learning unit 110 performs learning.

[0032] The learning algorithm executed by the learning unit 110 is not specifically limited, and a known learning algorithm is used for machine learning. Fig. Fig. 5 illustrates a second embodiment of the tooth contact position adjustment amount estimation device 1 shown in Fig. 3. The learning unit 110 performs learning by using a supervised learning algorithm. In supervised learning, a known data set (referred to as teaching data) including an input and an output corresponding to the input is output to the learning unit 110. The teaching data is, for example, a data set including dimensional data (input) of each of the parts acquired in the past and data (output) indicating the tooth contact position adjustment amount adjusted for each item of the dimensional data. The learning unit 110 identifies the characteristics indicating a correlation between the input and output of the teaching data. Thereby, the learning unit 110 derives a correlation model to estimate an output for a new input.

[0033] The learning unit 110 includes an error calculation unit 112 and a model update unit 114. The error calculation unit 112 calculates an error E between a correlation model M for estimating the tooth contact position adjustment amount and a correlation characteristic identified by the teaching data T. The model update unit 114 updates the correlation model M to reduce the error E. The model update unit 114 in the learning unit 110 thus repeatedly updates the correlation model M to learn the estimation of the tooth contact position adjustment amount. That is, the learning unit 110 thus performs machine learning to derive the correlation model M to indicate the correlation between the part dimension data and the optimal tooth contact position adjustment amount for the part dimension data.

[0034] For example, the correlation between the state variable S and the label data L is simplified (e.g., by a linear function) at the beginning of learning and expressed as a correlation model M. The simplified correlation model M is given to the learning unit 110 before supervised learning begins. According to this embodiment, the part dimension data of the power transmission mechanism acquired in the past and the data on the tooth contact position adjustment amount are used for the teaching data T. When the tooth contact position adjustment amount estimation device 1 is operated, the teaching data T is regularly given to the learning unit 110. The error calculation unit 112 identifies the correlation characteristics indicating the correlation between the part dimension data of the power transmission mechanism and the tooth contact position adjustment amount from the teaching data T given to the learning unit 110.The error calculation unit 112 finds the error E between the identified correlation property and the correlation model M corresponding to the state variable S and the label data L in the current state. The model update unit 114 updates the correlation model M to reduce the error E, for example, according to a predetermined update rule.

[0035] In the next learning cycle, the error calculation unit 112 estimates the tooth contact position adjustment amount using the state variable S according to the updated correlation model M. The error calculation unit 112 finds the error E between the estimation result of the tooth contact position adjustment amount and the actually acquired label data L. The model update unit 114 updates the correlation model M to reduce the error again. Thus, the unknown correlation between the current state of the environment and the estimation thereof gradually becomes clear. That is, the correlation between the dimensional data of the parts constituting the power transmission mechanism and the tooth contact position adjustment amount data for the dimensional data gradually becomes clear.

[0036] The learning unit 110 can use a neural network to perform supervised learning. For example, a neural network comprising three layers (input layer, middle layer, and output layer) can be used as a neural network. Furthermore, the learning unit 110 can perform learning using a neural network with three or more layers or deep learning. The learning unit 110 can perform more effective learning and estimation by using deep learning.

[0037] A configuration of the machine learning device 100 can be described as a machine learning method (or software) executed by the processor 101.The machine learning method is to learn the estimation of the tooth contact position adjustment amount from the dimensional data of the parts constituting the power transmission mechanism assembled by the robot system 2, and the processor 101 performs, in the machine learning method, a step of observing the dimensional data (part dimensional data S1) of the parts constituting the power transmission mechanism assembled by the robot system 2 as a state variable S indicating a current state, a step of acquiring the tooth contact position adjustment amount (adjustment amount data L1) as label data L, and a step of associating and learning the dimensional data of the parts constituting the power transmission mechanism assembled by the robot system 2 and the tooth contact position adjustment amount by using the state variable S and the label data L.

[0038] In the third to fifth embodiments described below, the case will be described where the tooth contact position adjustment amount estimation device 1 according to the first and second embodiments is implemented as part of a system in which the tooth contact position adjustment amount estimation device 1 is connected to a plurality of devices, such as a cloud server, a host computer, a front computer, and an edge computer (robot controller, controller, etc.), via a wired / wireless network. As exemplified in Fig. As shown in Figure 7, in the following third to fifth embodiments, each of a plurality of devices is configured to be logically divided into three layers, that is, a layer including a cloud server 6, a layer including fog computers 7, and so on, and a layer including edge computers 8 (robot controllers included in a cell 9, controllers, and so on), and so on, with each of the plurality of devices being connected to the network. In the system described above, the tooth contact position adjustment amount estimation device 1 can be implemented on each of the cloud server 6, the fog computers 7, and the edge computers 8. The tooth contact position adjustment amount estimation device 1 can share learning data used in machine learning processing among the plurality of devices via the network, and perform, for example, distributed learning.The tooth contact position adjustment amount estimation device 1 can collect the generated learning model in the fog computers 7 or the cloud server 6 for large-scale analysis. Furthermore, the tooth contact position adjustment amount estimation device 1 can mutually reuse the generated learning model. In a system exemplified in FIG. Fig. As shown in Figure 7, multiple cells 9 are deployed for each factory at numerous locations. The fog computers 7, which represent the upper layer of the cell 9, manage the respective cells 9 in predetermined units (such as per factory or by multiple factories of the same manufacturer). The cloud server 6, which represents the upper layer of the fog computer 7, then collects and analyzes the data collected and analyzed by these fog computers 7. The information resulting from the analysis can be used to control each edge computer, and so on.

[0039] Fig. 8 illustrates a schematic hardware configuration diagram of a tooth contact position adjustment amount estimation device 1 implemented on a computer such as a cloud server, a fog computer, and so on.

[0040] A central processing unit (CPU) 311 provided for the tooth contact position adjustment amount estimation device 1' implemented on a computer according to the embodiment is a processor for completely controlling the tooth contact position adjustment amount estimation device 1'. The CPU 311 reads a system program stored in a read-only memory (ROM) 312 through a bus 320. The CPU 311 controls the entire tooth contact position adjustment amount estimation device 1 according to the system program. Temporary calculation data or display data, or various data items, input by a worker via an input device 371 are temporarily stored in a random access memory (RAM) 313.

[0041] A non-volatile memory 314 is backed up by, for example, a battery (not shown). The non-volatile memory 314 is a memory in which stored states are retained even when the tooth contact position adjustment amount estimation device 1 is turned off. The non-volatile memory 314 stores the programs input via the input device 371. The non-volatile memory 314 stores various data items acquired by the robot system 2 and the like about each part of the tooth contact position adjustment amount estimation device 1' or a network 5. The programs or the various data items of the data stored in the non-volatile memory 314 can be loaded into the RAM 313 upon execution / use.Furthermore, various system programs such as a known analysis program (including a system program for controlling communication with a machine learning device 100, which will be described later) are previously written into the ROM 312.

[0042] The tooth contact position adjustment amount estimation device 1' is connected to a wired / wireless network 5 via an interface 319. At least one robot system 2, another tooth contact position adjustment amount estimation device 1, an edge computer 8, a front computer 7, a cloud server 6, and so on are connected to the network 5. These devices, connected to the network 5, exchange mutual data with the tooth contact position adjustment amount estimation device 1'.

[0043] A display device 370 displays each data point read into memory or all data acquired as a result of an executed program or the like, which is output via an interface 317. The input device 371 is configured as a keyboard or a pointing device. The input device 371 transmits commands based on the worker's operations, data, and the like to the CPU 311 via an interface 318.

[0044] An interface 321 is an interface that connects each part of the tooth contact position adjustment amount estimation device 1' to the machine learning device 100. The machine learning device 100 has the same configuration as in Fig. 1 described.

[0045] As described above, in a case where the tooth contact position adjustment amount estimation device 1' is implemented on a computer such as a cloud server and a fog computer, functions provided with the tooth contact position adjustment amount estimation device 1' are the same as the functions described in the first and second embodiments, except that information from the robot system 2 and commands for setting a tooth contact position adjustment amount are exchanged to the robot system via a network 5.

[0046] Fig. 9 is a schematic hardware configuration diagram according to the third embodiment, illustrating a tooth contact position adjustment amount estimation system provided with the tooth contact position adjustment amount estimation device 1'. A tooth contact position adjustment amount estimation system 500 is provided with a plurality of tooth contact position adjustment amount estimation devices 1, 1', a plurality of robot systems 2, and a network 5 connecting the plurality of tooth contact position adjustment amount estimation devices 1, 1' and the plurality of robot systems 2.

[0047] In the tooth contact position adjustment amount estimation system 500, the tooth contact position adjustment amount estimation device 1' equipped with a machine learning device 100 estimates the tooth contact position adjustment amount in the assembly of a power transmission mechanism by the robot system 2 using a learning result by the learning unit 110. At least one tooth contact position adjustment amount estimation device 1' learns the tooth contact position adjustment amount common to all the tooth contact position adjustment amount estimation devices 1, 1' in the assembly of a power transmission mechanism by the robot system 2 based on a state variable and label data acquired by each of the plurality of other tooth contact position adjustment amount estimation devices 1, 1'.The tooth contact position adjustment amount estimation device 1' is configured to share the learning result among all of the tooth contact position adjustment amount estimation devices 1, 1'. Therefore, according to the tooth contact position adjustment amount estimation system 500, the speed and reliability of the learning system can be improved by using various acquisition data (including the state variables S and label data L) as input.

[0048] Fig. 10 is a schematic configuration diagram according to the fourth embodiment, illustrating a system in which a machine learning device and a tooth contact position adjustment amount estimation device are implemented on different devices. A tooth contact position adjustment amount estimation system 500' includes at least one machine learning device 100 implemented on a part of a computer, such as a cloud server, a host computer, a fog computer, and so on ( Fig. 10 illustrates an example of a machine learning device 100 implemented as a part of a fog computer 7) and a plurality of tooth contact position adjustment amount estimation devices 1". Further, the tooth contact position adjustment amount estimation system 500' is provided with a network 5 connecting the plurality of tooth contact position adjustment amount estimation devices 1" and a computer. The hardware configuration of the computer is the same as the schematic hardware configuration of the tooth contact position adjustment amount estimation devices 1' as shown in Fig. 8, in which hardware such as a CPU 311, a RAM 313, and a non-volatile memory 314 provided for a general computer are connected to each other via a bus 320.

[0049] In the tooth contact position adjustment amount estimation system 500' structured above, the machine learning device 100 of the tooth contact position adjustment amount estimation system 500' learns, based on the state variable S and the label data L acquired for each of the plurality of tooth contact position adjustment amount estimation devices 1", the tooth contact position adjustment amount common to all the tooth contact position adjustment amount estimation devices 1" when assembling a power transmission mechanism by the robot system 2. The tooth contact position adjustment amount estimation system 500' can adjust the tooth contact position adjustment amount when assembling a power transmission mechanism by the robot system 2 using the learning result.According to the tooth contact position adjustment amount estimation system 500', with respect to the embodiment, any of the required number of tooth contact position adjustment amount estimation devices 1" can be connected to the machine learning device 100 as needed, regardless of an existing location or time.

[0050] Fig. 11 is a schematic configuration diagram according to the fifth embodiment, illustrating a tooth contact position adjustment amount estimation system 500" provided with a machine learning device 100' and a tooth contact position adjustment amount estimation device 1'. A tooth contact position adjustment amount estimation system 500" is provided with at least one machine learning device 100 connected to a computer such as an edge computer, a fog computer, a host computer, and a cloud server ( Fig.11 illustrates an example of a machine learning device 100 implemented as part of a fog computer 7 and provided with a plurality of tooth contact position adjustment amount estimation devices 1. Furthermore, the tooth contact position adjustment amount estimation system 500' is provided with a wired / wireless network 5 connecting the tooth contact position adjustment amount estimation devices 1 and the computer. In the tooth contact position adjustment amount estimation system 500' structured above, the fog computer 7 provided with the machine learning device 100' acquires, from each of the tooth contact position adjustment amount estimation devices 1, learning models acquired as a result of machine learning by the machine learning device 100' included in the tooth contact position adjustment amount estimation devices 1.The machine learning device 100' included in the fog computer 7 then optimizes the knowledge and performs efficiency-improving processing based on the plurality of these learning models to generate new, optimized, efficiency-improved learning models. The machine learning device 100' distributes the generated learning models to each of the tooth contact position adjustment amount estimation devices 1.

[0051] One example of optimizing or improving the efficiency of the learning models performed by the machine learning device 100' is the generation of a distilled model based on multiple learning models acquired by each of the tooth contact position adjustment amount estimation devices 1. The machine learning device 100' according to the present embodiment learns input data to be inputted into the learning models and then learns by using an output acquired as a result of inputting the input data to each learning model, thereby generating a new learning model (distilled model). As described above, the distilled model thus generated is distributed to an external storage medium or distributed and used via a network to the tooth contact position adjustment amount estimation devices 1 and other computers.

[0052] Another example of optimizing or improving the efficiency of the learning models performed by the machine learning device 100' may be as follows. First, in the process of performing distillation on a plurality of learning models acquired by each of the tooth contact position adjustment amount estimation devices 1, an output data distribution of each learning model for input data is analyzed by a general statistical plan. Next, outliers are extracted from sets of input data and output data. Next, distillation is performed by using sets of input data and output data excluding the outliers. By adopting this approach, exceptional estimation results are excluded from the sets of input data and output data acquired from each learning model.Specifically, a distilled model can be generated by using sets of input data and output data while excluding exceptional estimation results. A distilled model generated in this way can be used more generally than a learning model generated in multiple tooth contact position adjustment amount estimation devices 1.

[0053] It should be noted that any of the other general plans of optimization or efficiency improvement of the learning models (such as analyzing each learning model and optimizing a hyperparameter of the learning models based on the analysis result) can be introduced as appropriate.

[0054] In the tooth contact position adjustment amount estimation system 500" according to the present embodiment, the machine learning device 100' is arranged on a fog computer 7 installed as an edge computer for a plurality of the tooth contact position adjustment amount estimation devices 1. Learning models generated by the tooth contact position adjustment amount estimation devices 1 are collectively stored on the fog computer 7. Optimization and efficiency improvement are performed based on the plurality of stored learning models. Then, the optimized, efficiency-improved learning models can be redistributed to each of the tooth contact position adjustment amount estimation devices 1 as needed.

[0055] Furthermore, in the tooth contact position adjustment amount estimation system 500" according to this embodiment, for example, learning models collectively stored on the fog computer 7 and learning models optimized or improved in efficiency on the fog computer 7 can be acquired to another upper host computer or cloud server. In this case, by using these learning models, application to intellectual tasks in factories or at the manufacturer of the tooth contact position adjustment amount estimation devices 1 can be performed (such as building and redistributing other versatile learning models on an upper server, supporting maintenance based on the result of the learning model analysis, analyzing the performance of each of the tooth contact position adjustment amount estimation devices 1, and others, and applying them to new machine development).

[0056] The embodiments of this invention have been described above, but this invention is not limited to the above embodiments and can be changed as needed to be implemented in various aspects.

[0057] For example, the learning algorithm or calculation algorithm executed by the machine learning device 100, the control algorithm executed by the tooth contact position adjustment amount estimation device 1, and the like are not limited to the above, and various algorithms can be employed.

[0058] Furthermore, in the above embodiment, the tooth contact position adjustment amount estimation device 1 and the machine learning device 100 each have different CPUs, but the machine learning device 100 can also be implemented by the CPU 11 provided for the tooth contact position adjustment amount estimation device 1 and the system program stored in the ROM 12.

Claims

[1] A tooth contact position adjustment amount estimation device (1, 1') that performs processing with reference to estimation of a tooth contact position adjustment amount for dimensional data of parts constituting a power transmission mechanism, comprising: a machine learning device (100) that performs processing related to estimation of the tooth contact position adjustment amount for the dimensional data of parts constituting the power transmission mechanism, wherein the machine learning device (100) observes part dimension data, which is dimension data of parts constituting the power transmission mechanism, as a state variable indicating a current state of an environment, and executes processing related to estimation of the tooth contact position adjustment amount for dimension data of parts constituting the power transmission mechanism in assembling the power transmission mechanism by executing processing related to machine learning based on the observed state variable. [2] Tooth contact position adjustment amount estimation device (1, 1') according to claim 1, wherein the machine learning device (100) comprises: a state observation unit (106) that observes the part dimension data, which is dimension data of parts constituting the power transmission mechanism, as the state variable indicating the current state of the environment; a label data acquisition unit (108) that acquires tooth contact position adjustment amount data adjusted during assembly of the power transmission mechanism as label data; and a learning unit (110) that learns the tooth contact position adjustment amount in assembling the power transmission mechanism for the dimensional data of parts constituting the power transmission mechanism by using the state variable and the label data. [3] Tooth contact position adjustment amount estimation device (1, 1') according to claim 2, wherein the learning unit (110) comprises: an error calculation unit (122) that calculates an error between a correlation model for estimating the tooth contact position adjustment amount in assembling the power transmission mechanism from the state variable and a correlation characteristic identified by previously prepared teaching data; and a model update unit (114) that updates the correlation model to reduce the error. [4] The tooth contact position adjustment amount estimation device (1, 1') according to claim 2 or 3, wherein the learning unit (110) calculates the state variable and the label data in a multi-layer structure. [5] Tooth contact position adjustment amount estimation device (1, 1') according to claim 1, wherein the machine learning device (100) comprises: a state observation unit (106) that observes the part dimension data, which is dimension data of parts constituting a power transmission mechanism, as a state variable indicating the current state of an environment; a learning unit (110) that has learned a tooth contact position adjustment amount in assembling a power transmission mechanism for dimensional data of parts constituting the power transmission mechanism; and an estimation result output unit (122) that estimates and outputs a tooth contact position adjustment amount in assembling a power transmission mechanism based on a state variable observed by the state observation unit (106) and a learning result by the learning unit (110). [6] A machine learning device (100) that performs processing related to estimation of a tooth contact position adjustment amount for dimensional data of parts constituting a power transmission mechanism, wherein the machine learning device (100) observes part dimensional data that is dimensional data of parts observing the power transmission mechanism as a state variable indicating a current state of an environment, and performs processing related to estimation of the tooth contact position adjustment amount for dimensional data of parts constituting the power transmission mechanism in assembling the power transmission mechanism by performing processing related to machine learning based on the observed state variable. [7] Machine learning device (100) according to claim 6, further comprising: a state observation unit (106) that observes the part dimension data, which is dimension data of parts constituting the power transmission mechanism, as the state variable indicating the current state of the environment; a label data acquisition unit (108) that acquires the tooth contact position adjustment amount data adjusted during assembly of the power transmission mechanism as the label data; and a learning unit (110) that learns the tooth contact position adjustment amount in assembling the power transmission mechanism for the dimensional data of parts constituting the power transmission mechanism by using the state variable and the label data. [8] Machine learning device (100) according to claim 6, further comprising: a state observation unit (106) that observes the part dimension data, which is dimension data of parts constituting the power transmission mechanism, as the state variable indicating the current state of the environment; a learning unit (110) that has learned the tooth contact position adjustment amount in assembling a power transmission mechanism for the dimensional data of parts constituting the power transmission mechanism; and an estimation result output unit (122) that estimates and outputs the tooth contact position adjustment amount in assembling the power transmission mechanism based on the state variable observed by the state observation unit (106) and a learning result by the learning unit (110). [9] A robot system (2) comprising a controller (2A) that determines the number of disks used for assembling a power transmission mechanism and instructs a robot (2B) to assemble a predetermined number of disks onto the parts based on the tooth contact position adjustment amount estimated by the tooth contact position adjustment amount estimating device (1, 1') according to claim 5. [10] A tooth contact position adjustment amount estimation system (500, 500') in which a plurality of devices (1, 1', 2, 7) are connected to each other via a network, wherein the plurality of devices (1, 1', 2, 7) comprise at least one tooth contact position adjustment amount estimation device (1, 1') which is the tooth contact position adjustment amount estimation device (1, 1') according to any one of claims 2 to 4. [11] Tooth contact position adjustment amount estimation system (500, 500') according to claim 10, wherein the plurality of devices (1, 1', 2, 7) comprise a computer (7) provided with a machine learning device (100), and the computer (7) acquires a learning model as at least one of the results of learning by the first tooth contact position adjustment amount estimation device (1, 1'), and the machine learning device (100) performs the optimization or efficiency improvement based on the learning model. [12] The tooth contact position adjustment amount estimation system (500, 500') according to claim 10, wherein the plurality of devices (1, 1', 2, 7) includes a second tooth contact position adjustment amount estimation device (1, 1') different from the first tooth contact position adjustment amount estimation device (1, 1'), and the result of learning by the first tooth contact position adjustment amount estimation device (1, 1') is shared with the second tooth contact position adjustment amount estimation device (1, 1'). [13] The tooth contact position adjustment amount estimation system (500, 500') according to claim 10, wherein the plurality of devices (1, 1', 2, 7) includes a second tooth contact position adjustment amount estimation device (1, 1') different from the first tooth contact position adjustment amount estimation device (1, 1'), and data observed by the second tooth contact position adjustment amount estimation device (1, 1') can be used to learn the first tooth contact position adjustment amount estimation device (1, 1') via a network (5).

Images