Estimation device, estimation method, program, and learning model generation device

Abstract

To allow for estimating deformation-related performance of a member before the member becomes defective.SOLUTION: An estimation device (1) provided herein is configured to use physical quantities of an object (2) that is linearly or non-linearly deformable, the physical quantities being indicative of a plurality of different types of member properties that vary over time according to deformation of the object, and a physical quantity indicative of the state of deformation-related performance of the member as learning data, and input the physical quantities (3, 4) indicative of the plurality of member properties of a member to be estimated into a learning model (51) trained to receive input of the physical quantities indicative of the plurality of member properties and output the physical quantity indicative of the state of deformation-related performance of the member so as to estimate the physical quantity (6) indicative of the state of deformation-related performance of the member to be estimated.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to an estimation device, an estimation method, a program, and a learning model generation device. [Background technology]

[0002] Conventionally, members including elastic bodies such as spring members and rubber members can be deformed by an applied force, and when a member including an elastic body is used in a device, if a fault occurs in the member including the elastic body, the device will have a malfunction. For this reason, the malfunction of the device is eliminated by replacing the member including the elastic body where the fault occurs. As a technique for identifying a fault in a member including an elastic body, a technique for analyzing faults occurring in elastic bodies such as rubber members is known (for example, see Non-Patent Document 1). This technique focuses on unknown parameters used in a non-linear model of a pneumatic artificial muscle (PAM) and analyzes the faults of the PAM. In this analysis, changes in dynamics such as the steady-state response to an increase in contractile motion are measured for a fault state in which a hole is drilled in part of the PAM and a normal state, and the state is classified into a fault state and a normal state in association with changes in the parameters of the non-linear model. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] https: / / ieeexplore.ieee.org / document / 8511330 Summary of the Invention [Problem to be solved by the invention]

[0004] However, the performance of a member including an elastic body gradually or suddenly decreases in relation to deformation depending on factors such as usage conditions and aging. Therefore, even if the failure state can be classified, it is difficult to specify how a failure such as a failure in a member including an elastic body affects the performance of the member including an elastic body. For example, when controlling a device using a member including an elastic body, the device is formed on the premise of the performance of the member including an elastic body in relation to deformation. The performance of a member including an elastic body gradually or suddenly decreases in relation to deformation depending on factors such as usage conditions and aging. Therefore, it is required to specify the performance of the member including an elastic body in relation to deformation. However, if a failure occurs in a member including an elastic body, and the member is replaced, the device will be stopped during operation, and there is room for improvement.

[0005] The present disclosure aims to provide an estimation device, an estimation method, a program, and a learning model generation device that can estimate the performance related to deformation of a component before a defect occurs in the component. [Means for solving the problem]

[0006] In order to achieve the above object, the first aspect is an estimation unit that inputs physical quantities indicating a plurality of component characteristics of an estimation target component, the physical quantities being different in type and changing over time in response to the deformation of a component that deforms linearly or nonlinearly, and a physical quantity indicating a performance state related to the deformation of the component, as learning data, into a learning model that has been trained to input the physical quantities indicating the plurality of component characteristics and output the physical quantities indicating the performance state related to the deformation of the component, and estimates the physical quantities indicating the performance state related to the deformation of the estimation target component; The estimation device includes:

[0007] A second aspect is the estimation device of the first aspect, the electrical characteristics of the member change in response to the deformation; the physical quantities indicative of the plurality of member characteristics include a first physical quantity indicative of a pressure characteristic that deforms the member, and a second physical quantity indicative of an electrical characteristic that changes in response to the deformation of the member; the physical quantity indicating the performance state related to the deformation of the member includes a third physical quantity indicating a frequency at which the member is deformed; The learning model is trained to receive the first physical quantity and the second physical quantity as inputs and to output the third physical quantity.

[0008] A third aspect is the estimation device according to the second aspect, The member includes an elastic body having a hollow interior and generating a contraction force in a predetermined direction when a pressurized fluid is supplied to the hollow interior, the first physical quantity is a pressure characteristic indicating a plurality of time-series pressure values ​​in a case where the pressurized fluid is supplied and a case where the supply of the pressurized fluid is stopped, the second physical quantity is an electrical characteristic indicating a plurality of time-series electrical resistance values ​​of the elastic body that change in response to the first physical quantity; The third physical quantity is a performance index that indicates a performance state for each of a plurality of frequency groups obtained by dividing a predetermined frequency, which is a physical quantity that indicates a performance state of the component that can deform while maintaining a predetermined performance, into a plurality of stages.

[0009] A fourth aspect is the estimation device according to the second aspect, The member includes an elastic body having a hollow interior and generating a contraction force in a predetermined direction when a pressurized fluid is supplied to the hollow interior, the first physical quantity is a pressure characteristic indicating a plurality of time-series pressure values ​​in a case where the pressurized fluid is supplied and a case where the supply of the pressurized fluid is stopped, the second physical quantity is an electrical characteristic indicating a plurality of time-series electrical resistance values ​​of the elastic body that change in response to the first physical quantity; The third physical quantity is a part life index that indicates a performance state from a performance state at a frequency at which the component is deformed to a predetermined frequency that is predetermined as a performance state of the component that can be deformed while maintaining a predetermined performance.

[0010] A fifth aspect is the estimation device according to any one of the first to fourth aspects, The learning model is a model generated by learning using a recurrent neural network.

[0011] A sixth aspect is the estimation device according to any one of the first to fifth aspects, The learning model is a model generated by learning using a network based on reservoir computing.

[0012] A seventh aspect is the estimation device according to any one of the first to fifth aspects, The learning model is a model generated by training a network using physical reservoir computing that uses a reservoir that accumulates multiple physical quantities indicating the operating state of the component, physical quantities indicating the component characteristics that change in response to the deformation, and physical quantities indicating the performance of the component.

[0013] The eighth aspect is Computer A learning model is trained to input a plurality of physical quantities indicating a plurality of component characteristics of an estimation target component, the physical quantities being different in type and changing over time in response to the deformation of a component that deforms linearly or nonlinearly, and a physical quantity indicating a performance state related to the deformation of the component, as learning data, and the learning model is trained to input the physical quantities indicating the plurality of component characteristics and output the physical quantities indicating the performance state related to the deformation of the component, and estimates the physical quantities indicating the performance state related to the deformation of the estimation target component. This is an estimation method.

[0014] The ninth aspect is Computer an estimation unit that inputs physical quantities indicating a plurality of component characteristics of an estimation target component, the physical quantities being different in type and changing over time in response to the deformation of a component that deforms linearly or nonlinearly, and a physical quantity indicating a performance state related to the deformation of the component, as learning data, into a learning model that has been trained to input the physical quantities indicating the plurality of component characteristics and output the physical quantities indicating the performance state related to the deformation of the component, and estimates the physical quantities indicating the performance state related to the deformation of the estimation target component; This is a program that functions as a

[0015] The tenth aspect is an acquisition unit that acquires a plurality of physical quantities indicating a plurality of component characteristics of different types that change over time in response to a linearly or nonlinearly deforming component, and a plurality of physical quantities indicating a performance state related to the deformation of the component; a learning model generation unit that generates a learning model trained to input physical quantities indicating the plurality of component characteristics based on the results acquired by the acquisition unit and output physical quantities indicating a performance state related to the deformation of the component; The learning model generating device includes: Effect of the Invention

[0016] The present disclosure has an effect of making it possible to estimate the deformation-related performance of a component before failure or defects occur in the component. [Brief description of the drawings]

[0017] [Figure 1] 1 is a block diagram showing an example of a functional configuration of an elastic body performance estimation device according to a first embodiment. FIG. [Diagram 2] FIG. 2 is an explanatory diagram of a member according to the first embodiment. [Diagram 3] 3A to 3C are diagrams illustrating an example of a structure of a member according to the first embodiment. [Figure 4] FIG. 2 is an explanatory diagram of a learning process for learning a learning model according to the first embodiment. [Diagram 5] 1 is a block diagram showing an example of a measurement device according to a first embodiment. [Figure 6] FIG. 4 is a diagram showing an example of member properties according to the first embodiment. [Figure 7] FIG. 4 is a diagram showing electrical characteristics in an ideal environment according to the first embodiment. [Figure 8] 5 is a flowchart showing an example of a learning data collection process according to the first embodiment. [Figure 9A] 5A to 5C are diagrams illustrating an example of electrical characteristics collected by the measurement device according to the first embodiment. [Figure 9B]FIG. 4 is a diagram showing an example of pressure characteristics collected by the measurement device according to the first embodiment. [Figure 10] FIG. 4 is a diagram showing an example of electrical characteristics in an operation cycle of a performance index according to the first embodiment. [Figure 11] FIG. 4 is an explanatory diagram of a learning process in a learning processing unit according to the first embodiment. [Figure 12] 5 is a flowchart showing an example of the flow of a learning process according to the first embodiment. [Figure 13] FIG. 1 is a block diagram showing an example of a device that realizes various functions of the elastic body performance estimation device according to the first embodiment. [Figure 14] 5 is a flowchart showing an example of the flow of an estimation process according to the first embodiment. [Figure 15] FIG. 13 is a diagram showing the results of verifying the performance state of a rubber actuator. [Figure 16] FIG. 13 is a block diagram showing a functional configuration of a modified example of an elastic body performance estimation device. [Figure 17] 13 is an explanatory diagram of a learning process in a learning processing unit according to the second embodiment. FIG. [Figure 18] 10 is a flowchart showing an example of the flow of a learning process according to the second embodiment. [Figure 19] 13 is an explanatory diagram of a learning process in a learning processing unit according to the third embodiment. FIG. [Figure 20] 13 is a flowchart showing an example of the flow of a learning process according to the third embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] Hereinafter, embodiments for implementing the technology of the present disclosure will be described in detail with reference to the drawings. In addition, components and processes that perform the same actions and functions are given the same reference numerals throughout the drawings, and duplicated explanations may be omitted as appropriate. In addition, the present disclosure is not limited to the following embodiments, and can be implemented with appropriate modifications within the scope of the object of the present disclosure. In addition, the present disclosure mainly describes the estimation of physical quantities for members that deform nonlinearly, but it goes without saying that the present disclosure can also be applied to the estimation of physical quantities for members that deform linearly.

[0019] In this disclosure, the term "member" refers to a material that deforms nonlinearly and whose electrical properties change in response to the deformation. The term "elastic body" refers to an example of a member, and refers to a soft material such as rubber, foam, and resin. The term "elastic contraction body" refers to an example of an elastic body, and refers to a member that generates a contraction force in a predetermined direction according to a given physical quantity. The predetermined direction in which the contraction force is generated may be a linear direction that indicates expansion and contraction expressed in two dimensions, or a curved direction that indicates bending expressed in three dimensions. The elastic contraction body includes a member that is formed with a hollow interior and generates a contraction force in a predetermined direction when a pressurized fluid is supplied to the hollow interior.

[0020] Soft elastic bodies such as rubber materials behave nonlinearly with respect to a given force. For example, when looking at a two-dimensional change in shape, the distance that the body expands or contracts in a certain direction (e.g., a linear direction) changes nonlinearly in response to a given force (i.e., physical quantity or energy) (see Figure 2). For example, in an elastic body whose length changes in response to the magnitude of a given pressure, the change in length differs depending on the direction in which the pressure increases or decreases (for example, the length L1 is the same for pressures P1 and P2).

[0021] Furthermore, a member including a soft elastic body such as a rubber member may experience a decrease in performance from the initial performance (for example, performance indicating the expansion and contraction force) due to repeated deformation such as expansion and contraction (stretching and contraction in a two-dimensional shape change).The estimation device of the present disclosure uses a learning model that has been learned in advance to estimate a physical quantity indicating a performance state related to the deformation of a member that deforms nonlinearly before a failure or defect occurs in the member.

[0022] Specifically, the estimation device of the present disclosure includes a learning model. The learning model is trained using a plurality of learning data in which physical quantities indicating a plurality of component characteristics and physical quantities indicating a performance state of the component are associated with each other. The physical quantities indicating the plurality of component characteristics indicate physical quantities indicating a plurality of component characteristics of different types that change over time in response to deformation. In this training, the physical quantities indicating the plurality of component characteristics are input, and the estimation device is trained to output physical quantities indicating the performance state of the component. Using this trained learning model, physical quantities indicating a plurality of component characteristics of an estimation target component are input, and the output is estimated as a physical quantity indicating the performance state of the estimation target component.

[0023] (First embodiment) FIG. 1 shows an example of the configuration of an elastic body performance estimation device 1 as an estimation device according to the present disclosure.

[0024] In this embodiment, in order to grasp the physical quantity of an elastic body, as an example, a case will be described in which the technology of the present disclosure is applied to an elastic contraction body that includes a soft elastic body such as a rubber member among elastic bodies that deform nonlinearly and generates a contraction force in a predetermined direction due to an applied physical quantity. That is, in this embodiment, a case will be described in which the technology of the present disclosure is applied to an estimation process that estimates the performance state of an elastic contraction body.

[0025] In the following, as an example of an elastic contractile body, a case where an airbag type elastic contractile body according to a known technique is applied will be described (for example, see Japanese Patent Publication No. 52-40378). An example of an airbag-type elastic contraction body (hereinafter referred to as rubber actuator 2) has a main body 21 in which the outer periphery of a tubular body made of a soft elastic body such as a rubber member is covered with a braided reinforcement structure of organic or inorganic high-tensile fibers, for example, aromatic polyamide fibers, and both end openings 22 are sealed with a blocking member 23. The rubber actuator 2 is expanded and deformed by supplying a pressurized fluid to the internal cavity through a connection port 24 provided in the blocking member 23, and a contraction force is generated along the axial direction. The length of the rubber actuator 2 changes due to the expansion and deformation of the rubber actuator. However, the rubber actuator 2 is merely an example of the application target, and the estimation device of the present disclosure can be applied to elastic contraction bodies other than the rubber actuator 2 or members including an elastic body. The estimation device of the present disclosure can also be applied to members that deform linearly or nonlinearly by applying and releasing pressure.

[0026] As shown in FIG. 3, the rubber actuator 2 has a main body 21 in which the outer circumference of a tubular body 212 made of a rubber member or the like having an internal cavity 211 is covered with a reinforcing structure 214. When a pressurized fluid is supplied to the internal cavity 211, the rubber actuator 2 is expanded and deformed, and a contraction force is generated along the axial direction. When the supply of the pressurized fluid is released (or the pressurized fluid is discharged), the shape is restored. The contraction force changes according to the operation cycle of the supply and release of the pressurized fluid, that is, the expansion and contraction of the main body 21. The length of the rubber actuator 2 changes nonlinearly according to this contraction force (FIG. 2). This rubber actuator 2 may have a defect in the main body 21 due to repeated deformation due to expansion and contraction in the main body 21 or due to usage conditions such as aging. For example, a defective part 216 such as a crack may occur in a part of the tubular body 212. Due to the occurrence of the defective part 216, the performance of the rubber actuator 2 decreases from the performance state (e.g., contraction force) at the time of manufacture. In other words, the performance state of the rubber actuator 2 changes.

[0027] It is difficult to identify the defective part inside the rubber actuator 2 caused by such a change in the performance state of the rubber actuator 2 by measuring the structure. Although it is possible to inspect the inside of the rubber actuator 2 using a large-scale device capable of computer tomography (CT scan) to identify the defective part, this is not realistic. Therefore, in this embodiment, a learning model that has been learned in advance about the performance state of the rubber actuator 2 is used to estimate the performance state regarding the deformation of the rubber actuator 2. More specifically, the performance state of the rubber actuator 2 is estimated without using a large-scale device before a situation occurs in which a failure occurs due to a defect in the tubular body 212 and before a situation in which use becomes difficult due to the defect.

[0028] In this embodiment, an example in which a physical quantity indicating pressure characteristics and a physical quantity indicating electrical characteristics are applied to a physical quantity indicating a plurality of different types of material characteristics will be described. In the following description, the physical quantity indicating the material characteristics will be referred to as material characteristics. Also, the physical quantity indicating the pressure characteristics will be referred to as pressure characteristics, and the physical quantity indicating electrical characteristics will be referred to as electrical characteristics. For the pressure characteristics, a pressure characteristic indicating a plurality of pressure values ​​in a time series is applied as a first physical quantity that deforms the rubber actuator 2. Specifically, a pressure value that changes in a time series, that is, a pressure characteristic indicating a plurality of pressure values ​​in a time series that are applied to the rubber actuator 2 for deformation by pressure application and release (operation cycle), is applied. Also, for the electrical characteristics, an electrical characteristic of an electrical resistance value is applied as a second physical quantity that changes according to the deformation of the rubber actuator. Specifically, an electrical resistance value that changes in a time series, that is, an electrical characteristic of a plurality of electrical resistance values ​​that change in a time series by an operation cycle, is applied.

[0029] The performance index of the rubber actuator 2 is applied as a third physical amount indicating the frequency of deformation of the rubber actuator 2 to the physical amount indicating the performance state regarding the deformation of the member. Specifically, a performance index indicating the performance state for each of a plurality of frequency groups obtained by dividing a predetermined frequency (for example, a frequency of a malfunction state or a state where a malfunction occurs) at which the rubber actuator 2 can deform while maintaining a predetermined performance, which is a performance state in which the rubber actuator 2 can contract in a predetermined direction and return to its original state by releasing the contraction, into a plurality of stages is applied. Note that, although details will be described later, in this embodiment, the performance index indicates any one of an initial state, a middle state, and a later state. This performance index can be regarded as a lifespan, such as a period during which the performance of the member can be maintained. In this case, a part lifespan index indicating the performance state from a performance state at a certain frequency to a predetermined frequency that is a performance state of the rubber actuator 2 capable of deforming while maintaining a predetermined performance can be applied.

[0030] The estimation process in the elastic body performance estimation device 1 uses a trained learning model that performs machine learning on the pressure data labeled with performance data and the electrical resistance data of the rubber actuator 2 as learning data. The performance data is a physical quantity (performance index) that indicates the performance state of the rubber actuator 2 described above (see FIG. 7). The pressure data is a pressure characteristic that indicates a time series of pressure values ​​according to the above-mentioned operation cycle (see FIG. 6). The electrical resistance data is an electrical characteristic that indicates a time series of multiple electrical resistance values ​​according to the operation cycle (see FIG. 6). Using this trained learning model, the performance data of the rubber actuator 2 corresponding to the pressure data and electrical resistance data of the unknown rubber actuator 2 is estimated. In other words, the elastic body performance estimation device 1 estimates the performance state of the rubber actuator 2 that changes nonlinearly depending on the operation cycle of the rubber actuator 2. This makes it possible to estimate the performance state of the rubber actuator 2, even if the inside is invisible and deforms nonlinearly, without using a large-scale device.

[0031] As shown in FIG. 1, the elastic body performance estimation device 1 includes an estimation unit 5. To the estimation unit 5, first input data 3 indicating the magnitude of pressure on the rubber actuator 2 (pressure characteristics due to operation cycles) is input as pressure data. In addition, second input data 4 indicating the magnitude of the electrical characteristics in the rubber actuator 2 (electrical characteristics due to operation cycles) is input as electrical resistance data. In addition, the estimation unit 5 outputs output data 6 indicating the performance state (performance index) of the rubber actuator 2 as an estimated result as performance data. The estimation unit 5 includes a learning model 51 that has been trained.

[0032] The learning model 51 is a model that has completed learning to derive the performance state (output data 6) of the rubber actuator 2 from the time-series pressure characteristics (first input data 3) of the rubber actuator 2 accompanying deformation and the electrical characteristics (second input data 4) of the rubber actuator 2. The learning model 51 is, for example, a model that defines a trained neural network, and is expressed as a collection of information on the weights (strengths) of connections between nodes (neurons) that constitute the neural network.

[0033] The learning model 51 is generated by the learning process of the learning processing unit 52 (FIG. 4). The learning processing unit 52 performs the learning process using measured physical quantities as time-series physical quantities in the rubber actuator 2. That is, a large amount of data obtained by measuring the component characteristics of the rubber actuator 2 in a time series is used as learning data. The learning data includes a large amount of sets of input data including pressure characteristics (first input data 3) and electrical characteristics (second input data 4) and performance states (output data 6) corresponding to the input data. Note that the time-series physical quantities may be associated with time-series information by adding information indicating the measurement time to, for example, pressure values ​​and electrical resistance values. Also, the above-mentioned sets may be associated with time-series information by adding information indicating the measurement time.

[0034] Next, the learning process performed by the learning processing unit 52 will be described.

[0035] First, the learning data used in the learning process will be described. FIG. 5 shows an example of a measuring device 7 for measuring a physical quantity of the rubber actuator 2.

[0036] In the measuring device 7, one closing member 23 of the rubber actuator 2 is attached to a mounting plate 72 fixed to a base 71, and the other closing member 23 is attached to a movable plate 73. The connection port 24 of the rubber actuator 2 includes a pressure sensor that detects pressure (a first physical quantity that deforms the rubber actuator 2), and is connected to a supply unit 75 that supplies pressurized fluid to the rubber actuator 2. An electrical characteristic detection unit 76 including a sensor that detects the electrical resistance value (a second physical quantity that indicates an electrical characteristic) of the rubber actuator 2 is attached to the closing members 23 at both ends of the rubber actuator 2. In addition, a fixed plate 74 is fixed to the base 71, to which a distance sensor 77 such as a laser sensor that detects the distance from the movable plate 73 is attached. The distance sensor 77 is connected to a length identification unit 78. The length identification unit 78 identifies the length of the rubber actuator 2 from the distance detected by the distance sensor 77. For example, the length identification unit 78 stores the length L of the rubber actuator 2 in the initial state where no pressurized fluid is supplied (shown as initial state 200 in FIG. 5) and the distance (La) detected by the distance sensor 77 as initial values. The length identification unit 78 can then identify the length of the rubber actuator 2 (L=Lb-La) by subtracting the difference in distance in the state where pressurized fluid is supplied (shown as supply state 210 in FIG. 5) from the length in the initial state. The measuring device 7 can be equipped with an air pressure detection unit 79 consisting of a load cell and an air pressure cylinder.

[0037] In addition, the above-mentioned measuring device 7 has been described as including a configuration for identifying the length of the rubber actuator 2, but since identifying the length of the rubber actuator 2 is not essential in this embodiment, the configuration for identifying the length can be omitted.

[0038] A counter 80 is also connected to the supply unit 75. The counter 80 counts the number of operation cycles resulting from the supply of pressurized fluid to the rubber actuator 2 and the release of the supply by the supply unit 75. The configuration for counting the number of operation cycles is not limited to connecting the counter 80 to the supply unit 75. For example, the number of instructions to the supply unit 75 may be counted, or the number of operation cycles may be counted by other methods.

[0039] The measuring device 7 includes a controller 70 connected to a supply unit 75, an electrical characteristic detection unit 76, and a counter 80. The controller 70 controls the supply unit 75, acquires the time-series pressure values ​​and electrical resistance values ​​of the rubber actuator 2 during the operation cycle of the rubber actuator 2, and the number of operation cycles (cumulative number of cycles), and stores them in association with each other. In other words, the measuring device 7 can acquire multiple data sets of the pressure characteristics, electrical characteristics, and number of cycles of the rubber actuator 2, which deforms nonlinearly, in a time-series manner during the supply control of the pressurized fluid.

[0040] Specifically, the measurement device 7 can acquire pressure values ​​and electrical resistance values ​​in time series for each pressure characteristic and electrical characteristic for each operation cycle, for each minute time (for example, 0.01 seconds) within the operation cycle.

[0041] Figure 6 shows an example of pressure characteristics and electrical characteristics (material characteristics) for the rubber actuator 2. Figure 6 shows material characteristics from measurement start time t0, time t1 when the supply of pressurized fluid to the rubber actuator 2 starts, time t2 when the supply is released, and time t3 when the measurement of the next operation cycle starts. The time from t0 to t3 is also shown as an operation cycle with one period Ta.

[0042] As shown in FIG. 6, the pressure characteristic of the rubber actuator 2 is that the pressure value rises suddenly when the supply of pressurized fluid starts, and drops suddenly when the supply is stopped. The electrical characteristics change suddenly when the supply of pressurized fluid starts and when the supply is stopped. The pressure value and electrical resistance value are measured in time series for each of the pressure characteristics and electrical characteristics in this operating cycle at very short time intervals (for example, 0.01 seconds). The pressure characteristics and electrical characteristics (component characteristics) measured in time series are profiles of the pressure characteristics and electrical characteristics of the operating cycle at the number of cycles (frequency) during measurement, and can also be regarded as patterns showing the shape of each waveform.

[0043] It is considered that the rubber actuator 2 tends to have an increasing number and size of defective parts 216 (FIG. 3), such as cracks, that occur in the tubular body 212 as the number of operation cycles (cumulative number of cycles) increases. As the number and size of the defective parts 216 (FIG. 3) increase, the performance of the rubber actuator 2 decreases from its initial performance (e.g., performance showing contraction force). In a situation where the number and size of such defective parts 216 increase, the electrical characteristics of the rubber actuator 2, for example, the maximum value of electrical resistance within an operation cycle, tend to increase.

[0044] FIG. 7 shows the expected electrical characteristics when a durability test of the rubber actuator 2 is carried out under ideal conditions.

[0045] As shown in FIG. 7, the electrical characteristics of the rubber actuator 2, for example, the maximum value of the electrical resistance in the operation cycle, tend to increase as the cumulative number of cycles increases. Then, at the final cycle number Cte, the rubber actuator 2 cannot perform its performance, for example, it is in a failure state. If the range of cycle numbers from the unused state cycle number "0" (Ct0) to the final cycle number Cte is the operable range CtX of the rubber actuator 2, the performance maintenance range CtY of the number of cycles in which the rubber actuator 2 can be deformed while maintaining its performance is the same or narrower (CtX≧CtY). In this embodiment, the performance maintenance range CtY is divided into a plurality of operation cycle groups (for example, three stages), and a performance index indicating a performance state is associated with each of the plurality of operation cycle groups. Specifically, the performance maintenance range CtY is classified into three stages, and a performance index indicating an initial state, a middle state, and a later state is associated with each of the classified cycle number groups. The performance index indicating the initial state is determined for the number of cycles in the range from the number of cycles Ct0 to the number of cycles Ct1, the performance index indicating the middle state is determined for the number of cycles in the range from the number of cycles Ct1 to the number of cycles Ct2, and the performance index indicating the later state is determined for the number of cycles in the range from the number of cycles Ct2 to the number of cycles Ct3 (≦Cte). This makes it possible to confirm the performance state of the rubber actuator 2 by specifying the number of cycles.

[0046] The performance index of the rubber actuator 2 indicated by the initial state, middle state, and later state can be regarded as an index indicating the lifespan of the rubber actuator 2. That is, the performance maintenance range of the rubber actuator 2 becomes smaller in the order of the initial state, middle state, and later state. In other words, the cumulative number of cycles increases in the order of the initial state, middle state, and later state. Therefore, the number of cycles until the final number of cycles Cte at which a malfunction occurs in the rubber actuator 2 decreases. Therefore, if the total number of cycles in the operable range CtX (or performance maintenance range CtY) is the lifespan of the rubber actuator, this corresponds to the lifespan becoming shorter in the order of the initial state, middle state, and later state.

[0047] However, the degree of performance degradation of the rubber actuator 2 varies depending on the usage conditions and individual differences. Therefore, even if the maximum value of the electrical resistance and the number of cycles are specified, it is difficult to uniformly associate a performance index indicating any one of the initial state, middle state, and late state as the performance state of the rubber actuator 2. Therefore, in this embodiment, a learning model 51 is generated so that the performance index for each rubber actuator 2 can be estimated.

[0048] The controller 70 shown in FIG. 5 can be configured to include a computer including a CPU (not shown), and is adapted to execute a learning data collection process.

[0049] FIG. 8 shows an example of the learning data collection process executed by the controller 70.

[0050] As shown in FIG. 8, in step S100, the controller 70 instructs the supply unit 75 to perform an operation cycle of supplying and releasing the pressurized fluid. In the next step S102, the controller 70 acquires a set of physical quantities, which are the time series pressure value (pressure characteristic) and the time series electric resistance value (electric characteristic) of the rubber actuator 2 during the operation cycle, and the number of cycles. In the next step S104, the controller 70 temporarily stores the set of physical quantities (the set of physical quantities in the current operation cycle). In step S106, the controller 70 judges whether or not a predetermined judgment condition is met, and if the judgment is negative, the process returns to step S100, and if the judgment is positive, the process proceeds to step S108. As the judgment condition, any of the termination conditions is applied, such as when a malfunction such as a breakdown is detected in the rubber actuator 2, or when a predetermined number of cycles is reached as a durability test.

[0051] When the controller 70 judges in step S106 that the result is positive, the controller 70 associates the temporarily stored physical quantity set with a performance index in step S108, and ends this processing routine. Specifically, a performance index is associated with each set of the time-series pressure value (pressure characteristic) and electrical resistance value (electrical characteristic) for each operating cycle in the operable range CtX or performance maintaining range CtY of the rubber actuator 2, and the number of cycles.

[0052] FIG. 9A shows an example of the electrical characteristics of the rubber actuator 2 collected by the controller 70 of the measuring device 7 over time, and FIG. 9B shows an example of the pressure characteristics of the rubber actuator 2. FIG. 9A and FIG. 9B show the measurement results in each state of the initial state, the middle state, and the later state over time when a durability test was performed on an unused rubber actuator 2 in which the operation cycle was repeated until a failure was detected. In FIG. 9A and FIG. 9B, the electrical resistance value (electrical characteristic) and the pressure value (pressure characteristic) are standardized (i.e., normalized) so that the maximum values ​​in each operation cycle are the same. In the example shown in the figure, the pressure characteristics of the rubber actuator 2 have a pattern that is approximately the same from the initial state to the later state (FIG. 9B). On the other hand, the electrical characteristics have a pattern in which the electrical characteristics (electrical resistance value) gradually differ relatively over time. In particular, a difference appears in the electrical characteristics (electrical resistance value) in the latter half of the operation cycle (a state in which the supply of pressurized fluid is released) (FIG. 9A).

[0053] The controller 70 divides the operable range CtX (or CtY) from the number of cycles Ct0 at the start of the measurement of the rubber actuator 2 to the number of cycles Cte at the end into three (FIG. 7), and associates each of the performance indexes of the three stages of the initial state, middle state, and late state. This division may be made so that the number of cycles is equal, or may be made with a characteristic electrical resistance value, such as a stable electrical resistance value, as a boundary. In the operation cycles within the ranges associated with the performance indexes of the initial state, middle state, and late state, the electrical characteristics have characteristic portions.

[0054] FIG. 10 shows an example of electrical characteristics in the range of operation cycles associated with the performance indexes of the initial state, the middle state, and the later state. FIG. 10 shows the electrical characteristics at cycle number Ct01 among the operation cycles associated with the performance index of the initial state shown in FIG. 9A. Similarly, the electrical characteristics at cycle number Ct11 associated with the performance index of the middle state, and the electrical characteristics at cycle number Ct21 associated with the performance index of the later state are also shown. Note that the electrical characteristics shown in FIG. 10 are standardized (i.e., normalized) so that the maximum electrical resistance value is the same for each operation cycle. FIG. 10 shows the electrical resistance value Rx before standardization as an example in the diagram of each electrical characteristic.

[0055] As shown in FIG. 10, the electrical characteristics have one characteristic that, as the performance state transitions from the initial state to the later state, the maximum value of the electrical resistance value increases, but the pattern, which is the profile of the electrical characteristics when the supply of pressurized fluid is released, moves toward a stable state. The stable state indicates a characteristic in which the fluctuation of the electrical resistance value is suppressed. In another aspect, the minimum value of the electrical resistance value decreases as the state transitions, but the waveform shape of the standardized electrical characteristics profile pattern gradually transitions to a flat shape on the release side of the pressurized fluid (indicated by the length of the arrow in FIG. 10). In another aspect, the fluctuation range of the maximum and minimum values ​​of the electrical resistance value increases as the state transitions from the initial state to the later state. In addition, another characteristic portion indicates a characteristic including a characteristic portion such as the degree of the decrease rate of the electrical characteristics on the release side of the pressurized fluid, for example, the rate of change of the electrical resistance value decreases as the state transitions from the initial state to the later state. These characteristics are considered to affect the performance state of the rubber actuator 2, and function effectively in the learning process.

[0056] Therefore, the controller 70 can obtain the pressure characteristics, electrical characteristics and number of cycles of the rubber actuator 2 in chronological order and store them in association with the performance index by controlling the supply of pressurized fluid to the rubber actuator 2. A set of physical quantities indicating the performance state including the pressure characteristics, electrical characteristics and number of cycles of the rubber actuator 2 stored in chronological order in the controller 70 and the performance index becomes the learning data.

[0057] It is preferable to execute the above-mentioned processing routine shown in FIG. 8 for a plurality of rubber actuators 2 to collect a larger amount of learning data.

[0058] Next, the learning processing unit 52 will be described with reference to FIG. The learning processing unit 52 includes a generator 54 and a calculator 56. The generator 54 has a function of generating an output in consideration of the context of a time-series input.

[0059] In addition, the learning processing unit 52 holds a large number of sets of first input data 3 (pressure characteristics) and second input data 4 (electrical characteristics) for each operating cycle measured by the measuring device 7, and output data 6 (performance status) as learning data.

[0060] 11, the generator 54 includes an input layer 540, an intermediate layer 542, and an output layer 544 to configure a known recurrent neural network, for example, a recurrent neural network (RNN). The recurrent neural network (for example, RNN) itself is a known technology, so a detailed description will be omitted, but the intermediate layer 542 includes a large number of node groups (neuron groups) having inter-node connections and feedback connections. Data from the input layer 540 is input to the intermediate layer 542, and data of the calculation result of the intermediate layer 542 is output to the output layer 544.

[0061] Specifically, the generator 54 is a neural network that generates generated output data 6A indicating the performance state from the input first input data 3 (pressure characteristics) and second input data 4 (electrical characteristics). Here, the above-mentioned performance index is applied as a physical quantity indicating the performance state. This performance state may be the number of cycles. The generated output data 6A is data that estimates the performance state (performance index) of the rubber actuator 2 from the first input data 3 (pressure characteristics) and the second input data 4 (electrical characteristics). The generator 54 generates generated output data indicating a performance state (performance index) close to the performance state (performance index) obtained by measuring the performance state (performance index) of the rubber actuator 2 due to nonlinear deformation from the first input data 3 (pressure characteristics) and the second input data 4 (electrical characteristics) input in a time series. The generator 54 learns using a large number of first input data 3 (pressure characteristics) and second input data 4 (electrical characteristics), and is thereby able to generate generated output data 6A that is closer to the measurement result of the performance state (performance index) of the rubber actuator.

[0062] The calculator 56 is a calculator that compares the generated output data 6A with the output data 6 of the learning data, and calculates the error of the comparison result. The learning processing unit 52 inputs the generated output data 6A and the output data 6 of the learning data to the calculator 56. In response to this, the calculator 56 calculates the error between the generated output data 6A and the output data 6 of the learning data, and outputs a signal indicating the calculation result.

[0063] The learning processing unit 52 performs learning of the generator 54 by tuning the weight parameters of the connections between the nodes based on the error calculated by the calculator 56. Specifically, the weight parameters of the connections between the nodes of the input layer 540 and the intermediate layer 542 in the generator 54, the weight parameters of the connections between the nodes in the intermediate layer 542, and the weight parameters of the connections between the nodes of the intermediate layer 542 and the output layer 544 are fed back to the generator 54 using a method such as the gradient descent method or the error backpropagation method. That is, with the output data 6 of the learning data as the target, the connections between all the nodes are optimized so as to minimize the error between the generated output data 6A and the output data 6 of the learning data.

[0064] The learning model 51 is generated by a learning process of the learning processing unit 52. The learning model 51 is expressed as a collection of information on weight parameters (weights or strengths) of connections between nodes as a result of learning by the learning processing unit 52.

[0065] The learning processing unit 52 is configured to include a computer including a CPU (not shown) and is capable of executing the learning process. For example, as shown in FIG. 12, the learning processing unit 52 acquires the first input data 3 (pressure characteristics), the second input data 4 (electrical characteristics), and the output data 6 (performance index) for each operation cycle, which are learning data obtained by measuring the rubber actuator 2 in a time series, in step S110. In step S112, the learning processing unit 52 generates a learning model 51 using the learning data obtained by measuring the rubber actuator 2 in a time series. That is, the learning processing unit 52 obtains a set of information on the weight parameters (weights or strengths) of the connections between the nodes of the learning result learned using a large number of learning data as described above. Then, in step S114, the data expressed as a set of information on the weight parameters (weights or strengths) of the connections between the nodes of the learning result is stored as the learning model 51.

[0066] The generator 54 has a function of generating an output taking into account the context of the time series input, and although an example using a recurrent neural network has been described above, the technology of the present disclosure is not limited to using a recurrent neural network. In other words, the technology of the present disclosure may be any technology that has a function of generating an output taking into account the context of the time series input, and other methods may be used.

[0067] The elastic body performance estimation device 1 uses the trained generator 54 (i.e., data expressed as a set of information on weight parameters of bonds between nodes as a result of learning) generated by the above-mentioned method as the learning model 51. If a sufficiently trained learning model 51 is used, it is not impossible to estimate the performance state (performance index) of a rubber actuator that deforms nonlinearly from the pressure characteristics based on the time-series pressure values ​​in operation cycle units and the electrical characteristics based on the electrical resistance values. In other words, the elastic body performance estimation device 1 can estimate the performance index of any of the initial state, middle state, and late state as an output from the pressure characteristics and electrical characteristics of the operation cycle of the rubber actuator 2 to be estimated.

[0068] The process by the learning processing unit 52 is an example of the process of the learning model generating device of the present disclosure. Also, the elastic body performance estimation device 1 is an example of the estimation unit and estimation device of the present disclosure.

[0069] The above-mentioned elastic body performance estimation device 1 can be realized, for example, by causing a computer to execute a program representing each of the above-mentioned functions.

[0070] FIG. 13 shows an example of a configuration including a computer as an execution device that executes processes to realize various functions of the elastic body performance estimation device 1.

[0071] The computer functioning as the elastic body performance estimation device 1 shown in FIG. 13 includes a computer main body 100. The computer main body 100 includes a CPU 102, a RAM 104 such as a volatile memory, a ROM 106, an auxiliary storage device 108 such as a hard disk drive (HDD), and an input / output interface (I / O) 110. The CPU 102, the RAM 104, the ROM 106, the auxiliary storage device 108, and the input / output I / O 110 are connected via a bus 112 so as to be able to exchange data and commands with each other. In addition, a communication interface (I / F) 114 and an operation display unit 116 such as a display and a keyboard are connected to the input / output I / O 110. The communication I / F 114 functions as an input / output unit that inputs and outputs the first input data 3, the second input data 4, and the output data 6 to and from an external device.

[0072] A control program 108P for causing the computer main body 100 to function as an elastic body performance estimation device 1 as an example of an estimation device of the present disclosure is stored in the auxiliary storage device 108. The CPU 102 reads out the control program 108P from the auxiliary storage device 108, loads it into the RAM 104, and executes processing. As a result, the computer main body 100 that has executed the control program 108P operates as an elastic body performance estimation device 1 as an example of an estimation device of the present disclosure.

[0073] The auxiliary storage device 108 stores a learning model 108M including the learning model 51, and data 108D including various data. The control program 108P may be provided by a recording medium such as a CD-ROM.

[0074] Next, the estimation process in the elastic body performance estimation device realized by a computer will be described.

[0075] FIG. 14 shows an example of the flow of the estimation process by the control program 108P executed in the computer main body 100. In FIG.

[0076] 14 is executed by the CPU 102 when the computer main body 100 is powered on. That is, the CPU 102 reads out the control program 108P from the auxiliary storage device 108, loads it into the RAM 104, and executes the process.

[0077] First, in step S200, the CPU 102 reads the learning model 51 from the learning model 108M in the auxiliary storage device 108 and expands it in the RAM 104, thereby acquiring the learning model 51. Specifically, a network model in which the connections between nodes are determined by weight parameters expressed as the learning model 51 is expanded in the RAM 104. Thus, the learning model 51 in which the connections between nodes are determined by weight parameters is constructed.

[0078] Next, in step S202, the CPU 102 acquires the unknown first input data 3 (pressure characteristics) and the unknown second input data 4 (electrical characteristics) to be used to estimate the performance state of the rubber actuator 2 via the communication I / F 114. That is, the CPU 102 acquires the pressure characteristics and the electrical characteristics due to the operation cycle of the rubber actuator 2 whose performance state is unknown.

[0079] Next, in step S204, the CPU 102 uses the acquired learning model 51 to estimate output data 6 (performance index indicating the performance state of the rubber actuator 2) corresponding to the acquired first input data 3 (pressure characteristics) and second input data 4 (electrical characteristics).

[0080] Then, in the next step S206, output data 6 (performance index) of the estimation result is output via the communication I / F 114, and this processing routine ends.

[0081] The estimation process shown in FIG. 14 is an example of a process executed by the estimation method of the present disclosure.

[0082] As described above, according to this embodiment, it is possible to estimate the performance state of a rubber actuator 2 whose performance state is unknown from the first input data 3 (pressure characteristics) and the second input data 4 (electrical characteristics) according to the operation cycle of the rubber actuator 2. That is, by inputting the pressure characteristics and electrical characteristics according to the operation cycle of the rubber actuator 2 whose performance state is unknown, it is possible to estimate the performance index of any of the initial state, the middle state, and the later state as the performance state of the rubber actuator 2. Therefore, by detecting the pressure value and the electrical resistance value of the rubber actuator 2 in the operation cycle in a time series, the performance state of the rubber actuator 2 can be estimated, and there is no need to measure the performance state using a large-scale device. As a result, according to this embodiment, it is possible to specify the performance state of the rubber actuator 2, which was difficult to specify by directly measuring the structure of the rubber actuator 2.

[0083] Next, the results of estimating the performance state of the rubber actuator 2 were verified.

[0084] FIG. 15 shows the results of verifying the estimation results of the performance indexes estimated from the pressure characteristics and electrical characteristics over the operation cycles for a plurality of rubber actuators 2.

[0085] In this verification, the number and size of defective parts 216 (Fig. 3), such as cracks, that occurred in the tubular body 212 were measured, and each of the rubber actuators 2 classified into performance indexes of the initial state, middle state, or late state based on the measurement results was used as the estimation target. Then, for each of the rubber actuators 2 to be estimated, the pressure characteristics and electrical characteristics of the operating cycle at the time of the above-mentioned classification were obtained, and each performance index was estimated. The estimation results in Fig. 15 show the correspondence between the classification results based on the measurement and the estimation results based on the pressure characteristics and electrical characteristics of the operating cycle.

[0086] As shown in FIG. 15, with the above-described elastic body performance estimation device 1, it is possible to confirm that there is an appropriate correspondence between the classification results based on measurements of the rubber actuator 2 and the estimation results.

[0087] (First Modification) Number of Cycles In the above, the case where the performance index is estimated as the physical quantity indicating the performance state has been described, but the present disclosure is not limited to this. For example, the number of cycles in the operable range CtX or performance maintaining range CtY of a predetermined standard rubber actuator 2 may be estimated as the physical quantity indicating the performance state.

[0088] Specifically, the performance index of the above-mentioned output data 6 may be replaced with the number of cycles (cumulative number of cycles) of a predetermined standard rubber actuator 2, and the learning model 51 may be generated by the learning process of the learning processor 52. The number of cycles (cumulative number of cycles) of the standard rubber actuator 2 may be derived by calculating the average value or standard deviation of the operable range CtX or the performance maintaining range CtY of a plurality of rubber actuators.

[0089] In this way, by estimating the number of cycles of a standard rubber actuator 2, it is possible to quantitatively confirm the performance state of the rubber actuator 2 being estimated at the time of estimation based on the number of operating cycles (cumulative number of cycles).

[0090] (Second Modification) Life Estimation In the above, the case where the performance index or the number of cycles is estimated as the physical quantity indicating the performance state has been described, but the present disclosure is not limited to this. For example, information indicating the life of the rubber actuator 2 may be estimated as the physical quantity indicating the performance state.

[0091] Specifically, the number of cycles that the rubber actuator 2 to be estimated can operate may be applied as a part life index using the estimation of the number of cycles for the above-mentioned standard rubber actuator 2. For example, the number of cycles remaining at the time of estimating the performance state of the rubber actuator 2 to be estimated may be applied as a part life index by storing the final number of cycles Cte for the above-mentioned standard rubber actuator 2 in a memory in advance and subtracting the estimated number of cycles from the final number of cycles Cte.

[0092] In this way, by estimating the lifespan of the rubber actuator 2, it is possible to quantitatively confirm the number of operating cycles remaining in the rubber actuator 2 being estimated, i.e., the lifespan. This makes it possible to set the timing for maintenance, such as replacement, before a malfunction occurs in the rubber actuator 2 being estimated.

[0093] In addition, the above-mentioned three-stage performance index can be applied to a multi-stage performance index of four or more stages, and can be used as information indicating the life of the rubber actuator 2. In other words, by identifying the stage at which the estimated performance index is by the final cycle number Cte, it is possible to use it as information indicating the life of the rubber actuator 2.

[0094] (Third variant) Estimation based on electrical characteristics only In the above, the performance state of a rubber actuator 2 whose performance state is unknown is estimated from the first input data 3 (pressure characteristics) and the second input data 4 (electrical characteristics) according to the operation cycle of the rubber actuator 2. The technology of the present disclosure is not limited to using both the pressure characteristics and the electrical characteristics as input data. For example, when the rubber actuator 2 to be estimated is operated in an operation cycle according to a predetermined steady pressure characteristic, the pressure characteristics can be stored in advance and the first input data 3 (pressure characteristics) can be omitted by using the stored pressure characteristics.

[0095] Specifically, as shown in Fig. 16, information 3M indicating the first input data 3 (pressure characteristics) may be stored in the estimation unit 5. With this configuration, it becomes possible to estimate the performance state by simply acquiring, as the second input data 4, the electrical characteristics when a pressure stimulus 3P is applied to the rubber actuator 2 to be estimated.

[0096] Second embodiment Next, a second embodiment will be described. The second embodiment is intended to improve the estimation speed when estimating the performance state of the rubber actuator 2. Since the second embodiment has a configuration substantially similar to that of the first embodiment, the same parts are denoted by the same reference numerals and detailed description thereof will be omitted.

[0097] In a typical recurrent neural network, weight parameter information is optimized for each of the connections from the input layer 540 to the intermediate layer 542, the connections and feedback connections between nodes in the intermediate layer 542, and the connections between nodes from the intermediate layer 542 to the output layer 544 (Fig. 11). However, learning using time-series data with a temporal correlation for the rubber actuator 2 that deforms nonlinearly requires an enormous learning time. Also, a huge memory is required to go back in time when learning using time-series learning data.

[0098] In order to reduce this enormous learning time, a well-known network model called reservoir computing can be applied to estimating the length of the nonlinearly deforming rubber actuator 2. The network model called reservoir computing (RC) (hereinafter referred to as RCN) itself is a well-known technology, so a detailed explanation will be omitted. One example of RCN is one in which a part of a recurrent neural network is fixed (replaced with a random network) and only the connections between nodes from the intermediate layer 542 to the output layer 544 are optimized.

[0099] The learning process in the learning processing unit 52A using RCN will be described with reference to Fig. 17. The learning processing unit 52A shown in Fig. 17 differs from the learning processing unit 52 shown in Fig. 11 in that the generator 54 shown in Fig. 11 is replaced with a generator 54A, and learning is performed by reflecting the error derived by the calculator 56 only on the output layer 544 side.

[0100] As shown in Fig. 17, the generator 54A of the learning processing unit 52 includes an input layer 540 similar to that of Fig. 11, a reservoir layer 543 having a similar configuration to that of Fig. 11 instead of the intermediate layer 542 of Fig. 11, and an output layer 544 similar to that of Fig. 11, forming a known RCN. In the RCN, fixed weight parameter information (hereinafter, referred to as weight coefficients) is set for the node connections from the input layer 540 to the reservoir layer 543, and for the connections between nodes and feedback connections in the reservoir layer 543. The connections between nodes from the reservoir layer 543 to the output layer 544 are, for example, linear connections, and each weight parameter is optimized by learning the learning data.

[0101] The fixed weight coefficient is set in advance. The fixed weight coefficient can be set to a coefficient determined as an initial value. The fixed weight coefficient may be set to a weight coefficient obtained by optimizing the connections between nodes, etc., for a predetermined number of times or for a predetermined time period that is insufficient to minimize the error, using the learning data and targeting the output data 6 of the learning data so as to minimize the error between the generated output data 6A and the output data 6 of the learning data.

[0102] On the other hand, the weight parameters that define the connections between the nodes from the reservoir layer 543 to the output layer 544 are derived by learning using a large amount of training data so as to minimize the error between the generated output data 6A and the output data 6 of the training data.

[0103] The learning processing unit 52A including the generator 54A is configured to include a computer including a CPU (not shown) and can execute the learning process. For example, as shown in FIG. 18, the learning processing unit 52A acquires the first input data 3 (pressure characteristics), the second input data 4 (electrical characteristics), and the output data 6 (performance index) as learning data (step S120), and constructs the input layer 540 and the reservoir layer 543 (step S122). Here, the learning process is performed using a part of the learning data, similar to the generator 54 shown in FIG. 11. Specifically, a set of information on weight parameters of the connections between nodes of the learning result learned using a part of the learning data is obtained, and the connections of the nodes from the input layer 540 to the reservoir layer 543 and the connections and feedback connections between the nodes in the reservoir layer 543 are derived as weight coefficients. The input layer 540 and the reservoir layer 543 are identified by the derived weight coefficients, and the input layer 540 and the reservoir layer 543 are constructed.

[0104] Next, the learning processing unit 52A generates the learning model 51 using a large amount of learning data obtained by measuring in time series (step S124). That is, the learning processing unit 52A learns only about the connections between nodes from the reservoir layer 543 to the output layer 544, and constructs an RCN by obtaining a set of information on weight parameters of the connections between nodes as a result of the learning. Then, the weight coefficients derived in step S122 and data expressed as a set of information on weight parameters of the connections between nodes as a result of the learning in step S124 are stored as the learning model 51 (step S126).

[0105] In the elastic body performance estimation device 1, the generated trained generator 54A is used as the learning model 51. That is, here, the learning model 51 corresponds to the node connections from the input layer 540 to the reservoir layer 543, the weighting coefficients representing the connections and feedback connections between the nodes in the reservoir layer 543, and the weighting parameters representing the connections between the nodes from the reservoir layer 543 to the output layer 544. If a sufficiently trained learning model 51 is used, it is not impossible to estimate a performance index from the pressure characteristics and electrical characteristics in an operating cycle for a rubber actuator that deforms nonlinearly.

[0106] As described above, according to this embodiment, instead of a general recurrent neural network, a network is constructed by RCN to optimize the learning model 51. This makes it possible to reduce the learning time required compared to constructing a learning model by a general recurrent neural network. In addition, it is also possible to reduce the huge memory (memory capacity) required for time retrospection in a general recurrent neural network.

[0107] Third embodiment Next, a third embodiment will be described. The third embodiment is designed to improve the learning effect of the learning model 51 for estimating the performance state of the rubber actuator 2. Since the third embodiment has a configuration similar to that of the first and second embodiments, the same parts are denoted by the same reference numerals and detailed description thereof will be omitted.

[0108] As described above, by using an RCN in which a part of the recurrent neural network is fixed, the learning time can be suppressed. However, when fixed weight parameters are used for the node connections from the input layer 540 to the reservoir layer 543 and for the connections between nodes and the feedback connections in the reservoir layer 543, the learning effect may be insufficient. This is because even if the weight parameters of the connections between nodes from the reservoir layer 543 to the output layer 544 are learned, the output from the reservoir layer 543 may not be sufficient for optimization in the reservoir layer 543 in which a limited number of nodes are set by fixed weight parameters. For this reason, it is expected that the structure of the recurrent neural network used in the reservoir layer 543 will be complicated, but this is not preferable because it requires time to set up the reservoir layer 543.

[0109] By the way, as is well known, reservoir computing (RCN) is a method of projecting an input into a high-dimensional feature space by nonlinear transformation into a high-dimensional space. In this regard, a network model called physical reservoir computing (PRC) (hereinafter referred to as PRCN) is known, which uses a nonlinear dynamic system in the reservoir layer 543 instead of a recurrent neural network. Since PRC and PRCN themselves are well-known technologies, detailed explanations will be omitted, but they can be used to store data on the deformation of the rubber actuator 2 that deforms nonlinearly in the reservoir layer 543. In other words, PRCN can be suitably applied to estimating the performance state of the rubber actuator 2 that deforms nonlinearly.

[0110] Here, the learning process in the learning processing unit 52B will be described with reference to Fig. 19. The learning processing unit 52B shown in Fig. 19 differs from the learning processing unit 52A shown in Fig. 17 in that the generator 54A shown in Fig. 17 is replaced with a generator 54B, and learning is performed by reflecting the error derived by the calculator 56 only on the output layer 544 side.

[0111] As shown in FIG. 19, the generator 54B of the learning processing unit 52B includes an input layer 540 similar to that of FIG. 11, a physical reservoir layer 545 instead of the reservoir layer 543 of FIG. 17, and an output layer 544 similar to that of FIG. 11, forming a known PRCN. Since the PRCN itself is a known technology, detailed description will be omitted. In the PRCN, fixed weight coefficients are set for the connection of nodes from the input layer 540 to the reservoir layer 543 as described above. The physical reservoir layer 545 is configured to store features for a large number of time-series correlations and output a large number of features close to the input. The connection between nodes from the physical reservoir layer 545 to the output layer 544 is, for example, a linear combination, and each weight parameter is optimized by learning that uses a large number of learning data to minimize the error between the generated output data 6A and the output data 6 of the learning data.

[0112] As described above, the fixed weighting coefficient may be set to a coefficient determined as an initial value, or a weighting coefficient obtained by optimizing a predetermined number of times or for a predetermined period of time using learning data may be set.

[0113] The physical reservoir layer 545 stores a large number of physical correlations in the time series of the rubber actuator 2, extracts performance states (performance indexes) corresponding to input data (pressure characteristics and electrical characteristics) that are close to unknown input data (pressure characteristics and electrical characteristics) from the input layer 540, and outputs them as multiple feature quantities to the output layer 544. Conceptually, as the behavior of the rubber actuator 2, a large number of correlations between the pressure characteristics and electrical characteristics that change in time series in the operating cycle and the performance indexes that are performance states are stored, and each of the multiple rubber actuator 2 performance states (performance indexes) that are close to the unknown input (pressure characteristics and electrical characteristics) is selected and output as a feature quantity. This makes it possible to suppress the execution of complex calculations.

[0114] The learning processing unit 52B including the generator 54B is configured to include a computer including a CPU (not shown) and can execute the learning process. For example, as shown in FIG. 20, the learning processing unit 52B acquires the first input data 3 (pressure characteristics), the second input data 4 (electrical characteristics), and the output data 6 (performance state), which are learning data obtained by measuring in a time series (step S130), and constructs the input layer 540 and the physical reservoir layer 545 (step S132). The input layer 540 is set to a predetermined weighting factor. Therefore, the input layer 540 is constructed by identifying the connection of the nodes from the input layer 540 to the physical reservoir layer 545 by the predetermined weighting factor. On the other hand, the physical reservoir layer 545 is stored with a large number of correlations between the performance state and each of the learning data, that is, the pressure characteristics and electrical characteristics of the operating cycle. Therefore, the physical reservoir layer 545 is constructed by storing the correlation between the length that exhibits nonlinear behavior with respect to the pressure value and electrical resistance value that change over time based on the learning data as features, and configuring it to output multiple features that are close to the input from among them.

[0115] Next, in step S134, the learning processing unit 52B generates the learning model 51 using a large amount of learning data obtained as a result of time-series measurement. That is, the learning processing unit 52B learns only about the connections between nodes from the physical reservoir layer 545 to the output layer 544, and constructs a PRCN by obtaining a set of information on weight parameters of the connections between nodes as a result of the learning. Next, the weight coefficients derived in step S132 and data expressed as a set of information on weight parameters of the connections between nodes as a result of the learning in step S134 are stored as the learning model 51 (step S136).

[0116] In the elastic body performance estimation device 1, the generated trained generator 54B is used as the learning model 51. That is, here, the learning model 51 corresponds to weight parameters representing node connections from the input layer 540 to the physical reservoir layer 545, and connections between the nodes of the physical reservoir layer 545 and from the physical reservoir layer 545 to the output layer 544.

[0117] The elastic body performance estimation device 1 extracts a performance state (performance index) corresponding to input data (pressure characteristics and electrical characteristics) close to the unknown input data (pressure characteristics and electrical characteristics) from the input layer 540 from the time-series physical correlation of the rubber actuator 2 stored in the physical reservoir layer 545 for the pressure characteristics and electrical characteristics of the rubber actuator 2 for the operation cycle, which is unknown input data, and outputs it to the output layer 544 as a plurality of feature amounts. Then, the output layer 544 estimates the performance state (performance index) of the rubber actuator 2 by, for example, linearly combining the plurality of feature amounts from the physical reservoir layer 545 using learned weight parameters. If a sufficiently learned learning model 51 is used, it is not impossible to estimate the performance state (performance index) of the rubber actuator 2 from the pressure characteristics and electrical characteristics.

[0118] As described above, according to the present disclosure, a network is constructed using PRCN instead of RCN to optimize the learning model 51. This improves the learning effect of the learning model 51 compared to the case where a learning model is constructed using RCN.

[0119] In this way, by constructing a network using PRCN instead of RCN and optimizing the learning model 51, it is possible to estimate the length of the rubber actuator 2 that is extremely close to the actual length of the rubber actuator 2.

[0120] As described above, in this disclosure, a case where a rubber actuator is applied as a member has been described, but the member is not limited to a rubber actuator. Also, as at least three different types of physical quantities that change according to the deformation of the member, a pressure value that represents the magnitude of pressure is set as a first physical quantity, an electrical resistance value that represents the magnitude of electrical characteristics is set as a second physical quantity, and a length that represents the magnitude of deformation is set as a target physical quantity, and a case where the length is estimated from the pressure value and the electrical resistance value has been described. However, each of the first physical quantity, the second physical quantity, and the target physical quantity is not limited to this, and the target physical quantity may be set to a pressure value or an electrical resistance value.

[0121] Although the embodiments have been described above, the technical scope of the present disclosure is not limited to the scope described in the above embodiments. Various modifications or improvements can be made to the above embodiments without departing from the gist of the present disclosure, and the modifications or improvements are also included in the technical scope of the present disclosure.

[0122] In addition, in the above embodiment, the inspection process is described as being realized by a software configuration using a flowchart, but this is not limited to this, and for example, each process may be realized by a hardware configuration.

[0123] Furthermore, a part of the elastic body performance estimation device, for example a neural network such as a learning model, may be configured as a hardware circuit. [Explanation of symbols]

[0124] 1. Elastic body performance estimation device 2 Rubber Actuator 3. First input data 4. Second input data 5 Estimation part 6 Output Data 6A Generated Output Data 7. Measuring Equipment 51 Learning Model 52, 52A, 52B Learning processing unit 54, 54A, 54B generator 56 Arithmetic unit 70 Controller 75 Supply section 76 Electrical characteristic detection section 100 Computer main unit 108 Auxiliary storage 108D Data 108M Learning Model 108P control program 540 Input Layer 542 Middle Class 543 Reservoir Layer 544 Output Layer 545 Physical Reservoir Layer

Claims

1. an estimation unit that inputs physical quantities indicating a plurality of component characteristics of an estimation target component and estimates the physical quantities indicating the performance state related to the deformation of the estimation target component to a learning model that has been trained to input the physical quantities indicating the plurality of component characteristics and output the physical quantities indicating the performance state related to the deformation of the component, using a plurality of physical quantities indicating a plurality of component characteristics of a component that deforms linearly or nonlinearly and that change over time in response to the deformation of the component as learning data; the electrical characteristics of the member change in response to the deformation; the physical quantities indicative of the plurality of member characteristics include a first physical quantity indicative of a pressure characteristic that deforms the member, and a second physical quantity indicative of an electrical characteristic that changes in response to the deformation of the member; the physical quantity indicating the performance state related to the deformation of the member includes a third physical quantity indicating a frequency at which the member is deformed; The learning model is trained to receive the first physical quantity and the second physical quantity as inputs and to output the third physical quantity. Estimation device.

2. The member includes an elastic body having a hollow interior and generating a contraction force in a predetermined direction when a pressurized fluid is supplied to the hollow interior, the first physical quantity is a pressure characteristic indicating a plurality of pressure values ​​in a time series in a case where the supply of the pressurized fluid is started and a case where the supply of the pressurized fluid is stopped, the second physical quantity is an electrical characteristic indicating a plurality of time-series electrical resistance values ​​of the elastic body that change in response to the first physical quantity; The third physical quantity is a performance index that indicates a performance state for each of a plurality of frequency groups obtained by dividing a predetermined frequency, which is a physical quantity that indicates a performance state of the member that can be deformed while maintaining a predetermined performance, into a plurality of stages. The estimation device according to claim 1 .

3. The member includes an elastic body having a hollow interior and generating a contraction force in a predetermined direction when a pressurized fluid is supplied to the hollow interior, the first physical quantity is a pressure characteristic indicating a plurality of pressure values ​​in a time series in a case where the supply of the pressurized fluid is started and a case where the supply of the pressurized fluid is stopped, the second physical quantity is an electrical characteristic indicating a plurality of time-series electrical resistance values ​​of the elastic body that change in response to the first physical quantity; The third physical quantity is a part life index that indicates a performance state from a performance state at a frequency at which the member is deformed to a predetermined frequency that is predetermined as a performance state of the member that can be deformed while maintaining a predetermined performance. The estimation device according to claim 1 .

4. The learning model is a model generated by learning using a recurrent neural network. The estimation device according to any one of claims 1 to 3.

5. The learning model is a model generated by learning using a network based on reservoir computing. The estimation device according to any one of claims 1 to 4.

6. The learning model is a model generated by learning using a network based on physical reservoir computing using a reservoir that stores a plurality of physical quantities, each of which includes a physical quantity indicating the operating state of the component, a physical quantity indicating the component characteristics that change in response to the deformation, and a physical quantity indicating the performance of the component. The estimation device according to any one of claims 1 to 4.

7. Computer An estimation method comprising: using, as learning data, physical quantities indicating a plurality of component characteristics of a component that deforms linearly or nonlinearly and different types of physical quantities that change over time in response to the deformation of the component, and a physical quantity indicating a performance state related to the deformation of the component; inputting the physical quantities indicating the plurality of component characteristics of an estimation target component into a learning model that has been trained to output physical quantities indicating the performance state related to the deformation of the component, and estimating the physical quantities indicating the performance state related to the deformation of the estimation target component, the electrical characteristics of the member change in response to the deformation; the physical quantities indicative of the plurality of member characteristics include a first physical quantity indicative of a pressure characteristic that deforms the member, and a second physical quantity indicative of an electrical characteristic that changes in response to the deformation of the member; the physical quantity indicating the performance state related to the deformation of the member includes a third physical quantity indicating a frequency at which the member is deformed; The learning model is trained to receive the first physical quantity and the second physical quantity as inputs and to output the third physical quantity. Estimation method.

8. Computer a learning model that is trained to input a plurality of physical quantities indicating a plurality of component characteristics of an estimation target component, the physical quantities being different in type and changing over time in response to a linearly or nonlinearly deforming component, and a physical quantity indicating a performance state related to the deformation of the component, as learning data, and output a physical quantity indicating the performance state related to the deformation of the component; the electrical characteristics of the member change in response to the deformation; the physical quantities indicative of the plurality of member characteristics include a first physical quantity indicative of a pressure characteristic that deforms the member, and a second physical quantity indicative of an electrical characteristic that changes in response to the deformation of the member; the physical quantity indicating the performance state related to the deformation of the member includes a third physical quantity indicating a frequency at which the member is deformed; The learning model is trained to receive the first physical quantity and the second physical quantity as inputs and to output the third physical quantity. program.

9. an acquisition unit that acquires a plurality of physical quantities indicating a plurality of component characteristics of different types that change over time in response to a linearly or nonlinearly deforming component, and a plurality of physical quantities indicating a performance state related to the deformation of the component; a learning model generation unit that generates a learning model trained to input physical quantities indicating the plurality of component characteristics based on the results acquired by the acquisition unit and output physical quantities indicating a performance state related to the deformation of the component; Equipped with the electrical characteristics of the member change in response to the deformation; the physical quantities indicative of the plurality of member characteristics include a first physical quantity indicative of a pressure characteristic that deforms the member, and a second physical quantity indicative of an electrical characteristic that changes in response to the deformation of the member; the physical quantity indicating the performance state related to the deformation of the member includes a third physical quantity indicating a frequency at which the member is deformed; The learning model is trained to receive the first physical quantity and the second physical quantity as inputs and to output the third physical quantity. Learning model generation device.

Images