Patent Application
20240219268
The present disclosure relates to an estimation device, an estimation method, a program, and a learning model generation device.
Hitherto, when a member that includes an elastic body, such as a spring member, a rubber member, or the like, and that is deformable by force being imparted thereto is utilized in a device, then a flaw occurs in the device when a defect occurs in the elastic body containing member. The flaw in the device is accordingly eliminated by replacing the elastic body containing member in which a defect has occurred. As technology to identify defects in members including elastic bodies, technology to analyze defects occurring in elastic bodies such as rubber members is known (see, for example, https://ieeexplore.ieee.org/document/8511330). Nonlinear models of Pneumatic Artificial Muscles (PAMs) are employed in this technology that focuses on unknown parameters to analyze defects in PAMs. In such analysis, changes in dynamics, such as the steady-state response to increasing contraction movements, are measured for both a defective state in which a hole has been opened in part of the PAM and a normal state thereof, these changes in dynamics are then associated with changes to a parameter of the nonlinear model, and classification is performed into defective states or normal states.
A first aspect of the present disclosure is an estimation device including an estimation section. The estimation section inputs a physical quantity indicating a plurality of member characteristics for an estimation target member into a learning model that: employs, as training data, a plurality of sets of a physical quantity indicating a plurality of member characteristics of different types which change in time series according to deformation of a linearly- or nonlinearly-deforming member and a plurality of sets of a physical quantity indicating a performance state related to deformation of the member, and has been trained by being input with physical quantities indicating the plurality of member characteristics so as to output physical quantities indicating the performance state related to deformation of the member; and estimates a physical quantity indicating a performance state related to deformation of the estimation target member.
An elastic body containing member suffers a gradual or rapid drop in performance related to deformation due to factors such as the conditions of use, ageing, and the like. This means that even when defective states can be classified, it is still difficult to identify how damage to an elastic body containing member, such as a defect, is going to affect the performance related to deformation. For example, in cases in which control is performed on a device employing an elastic body containing member, a device is formed under an assumption of a performance related to deformation of the elastic body containing member. The elastic body containing member suffers a gradual or rapid drop in the performance related to deformation due to factors such as the conditions of use, ageing, and the like. This means that there is a demand to identify the performance related to deformation of the elastic body containing member. However, measures taken after a defect has occurred in an elastic body containing member, such as replacing the member, lead to halting the device during operation thereof, and so there is room for improvement therein.
The present disclosure is able to estimate performance related to deformation of a member prior to a flaw occurring in the member.
A first aspect of technology disclosed herein is an estimation device including an estimation section. The estimation section inputs a physical quantity indicating plural member characteristics for an estimation target member into a learning model that employs as training data plural sets of a physical quantity indicating plural member characteristics of different types which change in time series according to deformation of a linearly- or nonlinearly-deforming member and a physical quantity indicating a performance state related to deformation of the member, and that has been trained by being input with physical quantities indicating the plural member characteristics so as to output physical quantities indicating the performance state related to deformation of the member, and the estimation section estimates a physical quantity indicating a performance state related to deformation of the estimation target member.
A second aspect is the estimation device of the first aspect, wherein the member has an electrical characteristic that changes according to the deformation, the physical quantity indicating the plural member characteristics includes a first physical quantity that indicates a pressure characteristic for deforming the member and a second physical quantity that indicates the electrical characteristic that changes according to the deformation of the member, the physical quantity indicating the performance state related to deformation of the member includes a third physical quantity that indicates a number of repetitions of deformation of the member, and the learning model is trained so as to output the third physical quantity with the first physical quantity and the second physical quantity as inputs.
A third aspect is the estimation device of the second aspect, wherein the member includes an elastic body that is formed with a hollow inside and that generates a contraction force in a specific direction when a pressurized fluid is supplied inside the hollow, the first physical quantity is a pressure characteristic indicating plural pressure values in a time series when the pressurized fluid is supplied and the pressurized fluid supply is cancelled, the second physical quantity is an electrical characteristic indicating plural electrical resistance values in a time series of the elastic body that change according to the first physical quantity, and the third physical quantity is a performance indicator indicating a performance state for plural respective repetition number groups that result from dividing a predetermined specific number of repetitions, as a physical quantity indicating a performance state of the member when deformable while sustaining a specific performance, into plural levels.
A fourth aspect is the estimation device of the second aspect, the member includes an elastic body that is formed with a hollow inside and that generates a contraction force in a specific direction when a pressurized fluid is supplied inside the hollow, the first physical quantity is a pressure characteristic indicating plural pressure values in a time series when the pressurized fluid is supplied and the pressurized fluid supply is cancelled, the second physical quantity is an electrical characteristic indicating plural electrical resistance values in a time series of the elastic body that change according to the first physical quantity, and the third physical quantity is a product lifespan indicator that indicates a performance state from a performance state at the number of repetitions of deformation of the member until a predetermined specific number of repetitions as a performance state of the member when deformable while sustaining a specific performance.
A fifth aspect is the estimation device of any one of the first aspect to the fourth aspect, wherein the learning model is a model generated by training using a recurrent neural network.
A sixth aspect is the estimation device of any one of the first aspect to the fifth aspect, wherein the learning model is a model generated by training using a network by reservoir computing.
A seventh aspect is the estimation device of any one of the first aspect to the fifth aspect, wherein the learning model is a model generated by training using a network by physical reservoir computing employing a reservoir accumulated with plural sets of a physical quantity indicating an operation state of the member, a physical quantity indicating a member characteristic that changes according to the deformation, and a physical quantity indicating a performance of the member.
An eighth aspect is an estimation method including a computer inputting a physical quantity indicating plural member characteristics for an estimation target member into a learning model that employs as training data plural sets of a physical quantity indicating plural member characteristics of different types which change in time series according to deformation of a linearly- or nonlinearly-deforming member and a physical quantity indicating a performance state related to deformation of the member, and that has been trained by being input with physical quantities indicating the plural member characteristics so as to output physical quantities indicating the performance state related to deformation of the member, and the computer estimating a physical quantity indicating a performance state related to deformation of the estimation target member.
A ninth aspect is program that causes a computer to function as an estimation section. The estimation section inputs a physical quantity indicating plural member characteristics for an estimation target member into a learning model that employs as training data plural sets of a physical quantity indicating plural member characteristics of different types which change in time series according to deformation of a linearly- or nonlinearly-deforming member and a physical quantity indicating a performance state related to deformation of the member, and that has been trained by being input with physical quantities indicating the plural member characteristics so as to output physical quantities indicating the performance state related to deformation of the member, and the estimation section estimates a physical quantity indicating a performance state related to deformation of the estimation target member.
A tenth aspect is a learning model generation device including an acquisition section that acquires plural sets of a physical quantity indicating plural member characteristics of different types which change in time series according to deformation of a linearly- or nonlinearly-deforming member and a physical quantity indicating a performance state related to deformation of the member, and a learning model generation section that based on results of the acquisition by the acquisition section generates a learning model that is input with physical quantities indicating the plural member characteristics, and that is trained so as to output physical quantities indicating the performance state related to deformation of the member.
The present disclosure enables a performance related to deformation of a member to be estimated prior to occurrence of a defect or flaw in the member.
Detailed description follows regarding an exemplary embodiment to implement technology disclosed herein, with reference to the drawings.
Note that the same reference numerals will be appended throughout the drawings to configuration elements and processing performing the same operation/function, and duplicate explanation thereof will sometimes be omitted as appropriate. Moreover, the present disclosure is not limited to any of the exemplary embodiments described below, and appropriate modifications may be implemented thereto within a range of the object of the present disclosure. Moreover, although in the present disclosure description follows regarding estimation of a physical quantity related to a member that mainly undergoes nonlinear deformation, obviously the present disclosure is applicable to estimation of a physical quantity related to a member that undergoes linear deformation.
Reference in the present disclosure to a “member” encompasses materials that deform nonlinearly, and that have an electrical characteristic that changes according to deformation. An “elastic body” is an example of a member, and encompasses soft materials such as rubber, foamed materials, resin materials, and the like. Moreover, an “elastic contraction body” is an example of an elastic body, and encompasses a member that generates a contraction force in a specific direction due to being imparted with a physical quantity. The specific direction a contraction force is generated in may be a straight line direction indicating elongation-compression as expressed in two dimensions, and may be a curvilinear direction indicating flexing as expressed in three dimensions. Moreover, the elastic contraction body includes a member that is formed with a hollow portion inside, and that generates a contraction force in a specific direction by a pressurized fluid being supplied inside the hollow portion.
A soft elastic body such as a rubber member exhibits nonlinear behavior in response to imparted force. For example from the perspective of deformation change, when looking at two dimensional deformation change, a distance of elongation-compression in a given direction (for example a straight line direction) in response to imparted force (namely, a physical quantity or energy) changes in a nonlinear manner (see
Moreover, sometimes performance of a member containing a soft elastic body, such as a rubber member, deteriorates from the performance when initially manufactured (for example, a performance indicating an expansion-contraction force) due to repeated deformation such as expansion-contraction (elongation-compression for two dimensional deformation change). However, for a member that undergoes nonlinear deformation, the estimation device of the present disclosure employs a pre-trained learning model to estimate a physical quantity indicating a performance state related to deformation of the member, prior to a defect or flaw occurring in the member.
More specifically, the estimation device of the present disclosure includes a learning model. The learning model is trained using plural training data in which physical quantities indicating plural member characteristics and physical quantities indicating performance states of the member have been associated with each other. The physical quantities indicating the plural member characteristics indicate physical quantities that indicate plural member characteristics of different types that change in a time series according to deformation. This training is training so as to output a physical quantity indicating a performance state of the member by inputting the physical quantities indicating the plural member characteristics. This trained learning model is then employed, the physical quantities indicating the plural member characteristics of an estimation target member are input thereto, and the output therefrom is estimated as the physical quantity indicating the performance state of the estimation target member.
Note that in the present exemplary embodiment, description follows regarding a case in which, in order to ascertain a physical quantity of an elastic body, as an example, the technology disclosed herein is applied to an elastic contraction body that includes a soft elastic body such as a rubber member from out of elastic bodies that deform nonlinearly, and that generates contraction force in a specific direction by being imparted with a physical quantity. Namely, in the present exemplary embodiment description follows regarding application of the technology disclosed herein to estimation processing for estimating a performance state of an elastic contraction body.
Description follows regarding a case in which known technology of an airbag type of elastic contraction body is employed as an example of an elastic contraction body (see, for example, Japanese Patent Application Publication No. S52-40378).
An example of the airbag type of elastic contraction body (hereafter referred to as rubber actuator 2) is configured including a main body 21 including a tube shaped body configured from a soft elastic body such as a rubber member, with the outer periphery of the tube shaped body covered with a braided reinforcement structure made from organic or inorganic high tensile strength fibers, such as aromatic polyamide fibers for example, and with openings 22 at both ends thereof sealed by closure members 23. The rubber actuator 2 is configured to perform diameter enlargement deformation when a pressurized fluid is supplied into an internal hollow therein through a connection port 24 provided in the closure member 23, so as to thereby generate a contraction force along an axial direction. In such a rubber actuator, the length of the rubber actuator 2 changes due to diameter enlargement deformation. However, using the rubber actuator 2 as an application target is merely an example thereof, and the estimation device of the present disclosure is applicable to members including elastic contraction bodies or elastic body containing members other than the rubber actuator 2. Moreover, the estimation device of the present disclosure is applicable to members that deform either linearly or nonlinearly due to pressure being imparted and canceled.
As illustrated in
A flaw site that has occurred inside the rubber actuator 2 and causes a change to the performance state of the rubber actuator 2 is difficult identify by measuring the structure. Note that although it would be possible to identify the flaw site by scanning the inside of the rubber actuator 2 using a bulky apparatus capable of performing computer tomographic imaging (a CT scan), this is not realistic in practice. The present exemplary embodiment accordingly employs a learning model pre-trained on the performance state of the rubber actuator 2 to estimate the performance state relating to deformation of the rubber actuator 2. More specifically, the performance state of the rubber actuator 2 is estimated without employing a bulky apparatus prior to a situation in which a defect caused by a flaw arising in the tube shaped body 212 occurs and prior to a situation in which use becomes difficult due to the flaw.
In the present exemplary embodiment description follows regarding an example in which a physical quantity indicating a pressure characteristic and a physical quantity indicating an electrical characteristic are applied as the physical quantity indicating plural member characteristics of different types. In the following description the physical quantities indicating the member characteristics will be referred to as member characteristics. Moreover, the physical quantity indicating the pressure characteristic will be referred to as a pressure characteristic, and the physical quantity indicating the electrical characteristic will be referred to as an electrical characteristic. Regarding the pressure characteristic, a pressure characteristic indicating a time series of plural pressure values is applied as a first physical quantity to deform the rubber actuator 2. More specifically, a pressure value that changes in a time series, namely a time series of plural pressure values imparted to the rubber actuator 2 to cause deformation by pressure being imparted and canceled (operation cycling), is applied as the pressure characteristic. Moreover, as the electrical characteristic, application is made to an electrical characteristic of electrical resistance values as a second physical quantity that changes according to deformation of the rubber actuator. More specifically, an electrical resistance value that change in a time series, namely plural electrical resistance values that change in a time series by operation cycling, is applied as the electrical characteristic.
A performance indicator of the rubber actuator 2, serving as a third physical quantity indicating a number of repetitions of deformation of the rubber actuator 2, is applied as the physical quantity indicating the performance state related to member deformation. More specifically, a predetermined specific number of repetitions at which the rubber actuator 2 is deformable while sustaining a specific performance, this being a performance state where recovery is possible under contraction in a specific direction and cancellation of contraction (for example, a number of repetitions when an operation malfunction state or a flaw state has occurred) is applied as a performance indicator to indicate a performance state with respect to each group of plural repetitions arrived at by dividing into plural levels. Note that although described in detail later, in the present exemplary embodiment the performance indicator indicates one or other state from out of an initial state, an intermediate state, or a later state. The performance indicator may be understood as being a lifespan, such as a duration of sustainable performance of a member. Such cases are applicable to a product lifespan indicator expressing a lifespan of the rubber actuator 2, namely indicating a performance state from given number of repetitions up to a performance state at a predetermined specific number of repetitions as the performance state when the rubber actuator 2 is deformable while sustaining a specific performance.
In the estimation processing of the elastic body performance estimation device 1, a learning model is employed that has been trained by performing machine learning using pressure data and electrical resistance data of the rubber actuator 2 to label performance data as the training data for the rubber actuator 2. The performance data is a physical quantity indicating the performance state of the rubber actuator 2 (i.e. a performance indicator) as described above (see
As illustrated in
The learning model 51 is a trained model for deriving a performance state (the output data 6) of the rubber actuator 2 from time series accompanying deformation of the pressure characteristic (first input data 3) of the rubber actuator 2 and the electrical characteristic (the 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 set of information about weights (strengths) of connections between respective nodes (neurons) configuring the neural network.
The learning model 51 is generated by training processing of a training processing section 52 (
Next, description follows regarding training processing performed by the training processing section 52.
First description follows regarding training data employed in the training processing.
In the measurement device 7, one of the closure members 23 of the rubber actuator 2 is attached to an attachment plate 72 that is fixed to a base 71, and the other of the closure members 23 is attached to a movable plate 73 that is capable of moving. A pressure sensor to detect pressure (the first physical quantity that deforms the rubber actuator 2) is included at the connection port 24 of the rubber actuator 2, and a supply section 75 that supplies a pressurized fluid to the rubber actuator 2 is in communication with the connection port 24. An electrical characteristic detection section 76 that includes a sensor to detect electrical resistance values (the second physical quantity indicating an electrical characteristic) of the rubber actuator 2 is attached to the closure members 23 at both ends of the rubber actuator 2. A fixed plate 74 is fixed to the base 71, and a distance sensor 77 such as a laser sensor that detects a distance to the movable plate 73 is attached to the fixed plate 74. The distance sensor 77 is connected to a length identification section 78. The length identification section 78 identifies a length of the rubber actuator 2 from the distance detected by the distance sensor 77. For example, the length identification section 78 stores, as initial values, a length L of the rubber actuator 2 in an initial state (indicated by initial state 200 in
Note that although the measurement device 7 described above has been described for a case including configuration to identify the length of the rubber actuator 2, identifying the length of the rubber actuator 2 is not essential in the present exemplary embodiment, and configuration for length identification may be omitted.
Moreover, a counter 80 is connected to the supply section 75. The counter 80 counts a number of cycles of operation cycling performed by supplying pressurized fluid to the rubber actuator 2 using the supply section 75 and canceling supply thereof.
Configuration to count the number of cycles of operation cycling is not limited to the counter 80 connected to the supply section 75. For example, a number of times of instruction to the supply section 75 may be counted, or the number of cycles of operation cycling may be counted by another method.
The measurement device 7 includes the supply section 75, the electrical characteristic detection section 76, and a controller 70 connected to the counter 80. The controller 70 performs control of the supply section 75, acquires time series of the pressure values and the electrical resistance values of the rubber actuator 2 during operation cycling of the rubber actuator 2 and the cycle number (cumulative cycle number) of operation cycling, and stores these associated with each other. Namely, by controlling the supply of pressurized fluid, the measurement device 7 is able to acquire plural time series of data sets, including the pressure characteristic, the electrical characteristic, and the cycle number, for the rubber actuator 2 that deforms in a nonlinear manner.
More specifically, the measurement device 7 is able to acquire time series respectively for the pressure characteristic and the electrical characteristic in operation cycle units, as pressure values and electrical resistance values for each minute period of time (for example, 0.01 second) in each of these operation cycles.
As illustrated in
It may be supposed that the rubber actuator 2 has an increasing trend in the number and size of flaw sites 216 (see
As illustrated in
Note that the performance indicator of the rubber actuator 2 expressed by the initial state, intermediate state, and later state may be taken as being an indicator expressing the lifespan of the rubber actuator 2. Namely, the sustainable performance range of the rubber actuator 2 is shorter in the sequence of the initial state, then the intermediate state, and then the later state. In other words, the cumulative cycle number increases in the sequence of the initial state, then the intermediate state, and then the later state. A quantity of the cycle number until reaching the end cycle number Cte when a flaw occurs in the rubber actuator 2 is reduced accordingly. This is equivalent to saying that, taking a total of the cycle number in the operable range CtX (or in the sustainable performance range CtY) as being the lifespan of the rubber actuator, then the lifespan becomes shorter in the sequence of the initial state, then the intermediate state, and then the later state.
However, in the rubber actuator 2, the way in which performance falls changes according to the conditions of use and individual differences. This means that even if the maximum value of the electrical resistance value and the cycle number has been identified, the performance indicator indicating one of the performance states from out of the initial state, intermediate state, and later state is difficult to associate as the performance state across the board for all rubber actuators 2. Thus in the present exemplary embodiment, the learning model 51 is generated so as to be capable of estimating the performance indicator corresponding to individual rubber actuators 2.
The controller 70 illustrated in
As illustrated in
When affirmative determination is made at step S106, the controller 70 associates a performance indicator with the physical quantity set temporarily stored at step S108 and then ends the current processing routine. More specifically, a performance indicator is able to be associated with each set, which includes the time series of pressure values (the pressure characteristic) and electrical resistance values (the electrical characteristic) for each operation cycle in the operable range CtX or in the sustainable performance range CtY of the rubber actuator 2, together with the cycle number.
The controller 70 proportionally divides the operable range CtX (or the CtY) from the cycle number Ct0 at the start of measuring the rubber actuator 2 to the end cycle number Cte into three (see
As illustrated in
The controller 70 is accordingly able to acquire time series of the pressure characteristic and the electrical characteristic of the rubber actuator 2 and each cycle number by performing control of the supply of pressurized fluid to the rubber actuator 2, and is able to associate and store the performance indicators therewith. Sets including the physical quantities indicating the performance states, including the pressure characteristic, the electrical characteristic, and the cycle number of the rubber actuator 2, and including the performance indicators thereof, are stored as time series in the controller 70 for use as training data.
Note that the processing routine illustrated in
Next, description follows regarding the training processing section 52, with reference to
The training processing section 52 includes a generator 54 and a computation unit 56. The generator 54 includes a function to generate an output that considers precedence relationships of the time series input.
Moreover, as training data, the training processing section 52 holds multiple sets that include the first input data 3 (the pressure characteristic) and the second input data 4 (the electrical characteristic) for each operation cycle as measured by the measurement device 7, and include the output data 6 (performance state).
In the example illustrated in
More specifically, the generator 54 is a neural network that generates, from the input first input data 3 (the pressure characteristic) and second input data 4 (the electrical characteristic), generated output data 6A that indicates a performance state. The performance indicator described above is applied as the physical quantity indicating the performance state. This performance state may be a cycle number. The generated output data 6A is data of a performance state (performance indicator) of the rubber actuator 2 as estimated from the first input data 3 (the pressure characteristic) and the second input data 4 (the electrical characteristic). The generator 54 generates, from the first input data 3 (pressure) and the second input data 4 (the electrical characteristic) that have been input as time series, the generated output data indicating a performance state (performance indicator) close to a performance state (performance indicator) obtained by measurement for the performance state (performance indicator) of the rubber actuator 2 that deforms in a nonlinear manner. The generator 54 performs training using multiple first input data 3 (the pressure characteristic) and multiple second input data 4 (the electrical characteristic) so as to be able to generate the generated output data 6A closer to the measurement result of the performance state (performance indicator) of the rubber actuator.
The computation unit 56 is a computation unit that compares the generated output data 6A and the training-data output data 6, and computes an error of the comparison result. The training processing section 52 inputs the generated output data 6A and the training-data output data 6 to the computation unit 56. In response thereto, the computation unit 56 computes an error between the generated output data 6A and the training-data output data 6, and outputs a signal representing this computation result.
Based on the error computed in the computation unit 56, the training processing section 52 tunes weight parameters of the inter-node connections so as to perform training of the generator 54. More specifically, a method such as, for example, gradient descent, error backpropagation, or the like is employed to feedback the respective weight parameters of inter-node connections between the input layer 540 and the middle layer 542 of the generator 54, weight parameters of inter-node connections within the middle layer 542, and weight parameters of inter-node connections between the middle layer 542 and the output layer 544, to the generator 54. Namely, using the training-data output data 6 as the target, all the inter-node connections are optimized so as to minimize error between the generated output data 6A and the training-data output data 6.
The learning model 51 is generated by the training processing of the training processing section 52. The learning model 51 is represented by a set of information of the weight parameters (weights or strengths) of inter-node connections resulting from training by the training processing section 52.
The training processing section 52 may be configured including a computer that contains a non-illustrated CPU and executes the training processing. For example, as illustrated as an example of training processing in
Note that the generator 54 includes a function to generate an output that considers precedence relationships of time series input, and although a description has been given above of an example in which a recurrent neural network is employed, the technology disclosed herein is not limited to employing a recurrent neural network. Namely, the technology disclosed herein may employ another method that includes a function to generate an output that considers precedence relationships of time series input.
The elastic body performance estimation device 1 employs as the learning model 51 the trained generator 54 generated by the method given as an example above (namely, the data represented by a set of information of the weight parameters of inter-node connections of the training result). Employing the sufficiently trained learning model 51 means that, for a rubber actuator that deforms in a nonlinear manner, it is not impossible to estimate a performance state (performance indicator) from the pressure characteristic configured by a time series of the pressure values and the electrical characteristic configured by a time series of electrical resistance values for operation cycle units. Namely, as a performance state derived from the pressure characteristic and the electrical characteristic of operation cycling of the estimation target rubber actuator 2, the elastic body performance estimation device 1 is able to estimate as an output the performance indicator for one or other out of the initial state, the intermediate state, and the later state.
Note that the processing by the training processing section 52 is an example of processing of a learning model generation device of the present disclosure. Moreover, the elastic body performance estimation device 1 is an example of an estimation section and estimation device of the present disclosure.
The elastic body performance estimation device 1 described above is, for example, able to be implemented by causing a program representing each of the above functions to be executed in a computer.
The computer that functions as the elastic body performance estimation device 1 illustrated in
A control program 108P to cause the computer main body 100 to function as the elastic body performance estimation device 1, serving as an example of the estimation device of the present disclosure, is stored in the auxiliary storage device 108. The CPU 102 reads the control program 108P from the auxiliary storage device 108, and expands and executes the processing of the control program 108P in the RAM 104. The computer main body 100 that has executed the control program 108P thereby operates as the elastic body performance estimation device 1 serving as an example of the estimation device of the present disclosure.
Note that a learning model 108M including the learning model 51, and data 108D including various data, is stored on the auxiliary storage device 108. The control program 108P may be configured so as to be provided by a recording medium such as a CD-ROM or the like.
Next, description follows regarding estimation processing in the elastic body performance estimation device implemented by a computer.
The estimation processing illustrated in
First at step S200, the CPU 102 acquires the learning model 51 by reading the learning model 51 from the learning model 108M of the auxiliary storage device 108 and expanding the learning model 51 in the RAM 104. More specifically, a network model configured by inter-node connections expressed by weight parameters is expanded as the learning model 51 in the RAM 104. The learning model 51 implemented by the inter-node connections with weight parameters is thereby built.
Next at step S202, the CPU 102 acquires, through the communication I/F 114, unknown first input data 3 (the pressure characteristic) and unknown second input data 4 (the electrical characteristic) to be subjected to estimation of a performance state of the rubber actuator 2. Namely, the pressure characteristic and the electrical characteristic under operation cycling on the rubber actuator 2 of unknown performance state are acquired.
Next as step S204. the CPU 102 uses the acquired learning model 51 to estimate the output data 6 (the performance indicator indicating the performance state of the rubber actuator 2) corresponding to the acquired first input data 3 (the pressure characteristic) and the acquired second input data 4 (the electrical characteristic).
Then at the next step S206, the output data 6 (performance indicator) of the estimation result is output through the communication I/F 114, and the current processing routine is ended.
Note that the estimation processing illustrated in
As described above, for a rubber actuator 2 of unknown performance state, the present exemplary embodiment is able to estimate the performance state of the rubber actuator 2 from the first input data 3 (the pressure characteristic) and the second input data 4 (the electrical characteristic) of the rubber actuator 2 by operation cycling. Namely, inputting the pressure characteristic and the electrical characteristic of the rubber actuator 2 of unknown performance state by operation cycling enables a performance indicator for one out of the initial state, the intermediate state, and the later state to be estimated as the performance state of the rubber actuator 2. Thus by detecting the pressure values and the electrical resistance values of the rubber actuator 2 by operation cycling as a time series, the performance state of the rubber actuator 2 can be estimated, without the need to measure the performance state using a bulky apparatus. The present exemplary embodiment accordingly enables identification of the performance state of the rubber actuator 2 that has hitherto been difficult to identify by directly measuring the structure of the rubber actuator 2.
The estimation result of the estimated performance state of the rubber actuator 2 was then validated.
In this validation, the number and size of flaw sites 216, such as cracks, (
As illustrated in
Although description has been given above regarding a case in which the performance indicator is estimated as a physical quantity indicating a performance state, the present disclosure is not limited thereto. For example, a cycle number in the operable range CtX or in the sustainable performance range CtY of a predetermined standard rubber actuator 2 may be estimated as the physical quantity indicating a performance state.
More specifically, the cycle number (cumulative cycle number) for the predetermined standard rubber actuator 2 may be substituted for the performance indicator of the output data 6 described above, and the learning model 51 may be generated by the training processing of the training processing section 52. Note that the cycle number (cumulative cycle number) of the standard rubber actuator 2 may be derived by computing an average value and standard deviation etc. of the operable range CtX or the sustainable performance range CtY of plural rubber actuators.
Adopting such an approach enables, by estimating the cycle number of the standard rubber actuator 2, quantitative confirmation of the performance state when the performance state of the estimation target rubber actuator 2 is estimated from the cycle number (cumulative cycle number) of the operation cycles.
Although description above is of a case in which the performance indicator or the cycle number is estimated as the physical quantity indicating the performance state, the present disclosure is not limited thereto. For example, information indicating the lifespan of the rubber actuator 2 may be estimated as the physical quantity indicating the performance state.
More specifically, an estimation of the cycle number of the standard rubber actuator 2 described above may be employed, so as to apply a number of operable cycles of the estimation target rubber actuator 2 as a product lifespan indicator. For example, the end cycle number Cte of the standard rubber actuator 2 described above may be pre-stored in memory, and a remaining number of cycles at the time of estimating the performance state for the estimation target rubber actuator 2, which results from subtracting the estimated cycle number from the end cycle number Cte, may be applied as the product lifespan indicator.
Estimating the lifespan of the rubber actuator 2 in this manner enables the number of cycles of operation cycling remaining, namely the lifespan, to be quantitatively confirmed for the estimation target rubber actuator 2. This thereby enables, for example, a time prior to a flaw occurring in the estimation target rubber actuator 2 to be set for performing maintenance, such as replacement.
Note that when a multi-level performance indicator having 4 or more levels is applied for the performance indicator of 3 levels in the example described, this may still be applied as information to indicate the lifespan of the rubber actuator 2. Namely, an estimated performance indicator identifying a level until the end cycle number Cte may be employed as information indicating the lifespan of the rubber actuator 2.
In the above description, the performance state of the rubber actuator 2 was estimated for the rubber actuator 2 of unknown performance state from the first input data 3 (the pressure characteristic) and the second input data 4 (the electrical characteristic) by performing operation cycling on the rubber actuator 2. The technology disclosed herein is not limited to employing both the pressure characteristic and the electrical characteristic as the input data. For example, in cases in which the estimation target rubber actuator 2 is operated by operation cycling under a predetermined routine pressure characteristic, this pressure characteristic may be pre-stored, and the stored pressure characteristic then employed so as to enable the first input data 3 (the pressure characteristic) to be omitted.
More specifically, as illustrated in
Next, description follows regarding a second exemplary embodiment. The second exemplary embodiment considers an improvement in the estimation speed when estimating the performance state of the rubber actuator 2. Note that due to the second exemplary embodiment being configured substantially the same as the first exemplary embodiment, the same reference numerals will be appended to the same portions and detailed explanation thereof will be omitted.
In a general recurrent neural network, information of the weight parameters is optimized for respective connections of the inter-node connections from the input layer 540 to the middle layer 542, the inter-node connections and feedback connections in the middle layer 542, and the inter-node connections from the middle layer 542 to the output layer 544 (
With an object of suppressing such extensive time being needed for training time, a well-known network model called reservoir computing may be employed to estimate a length of the rubber actuator 2 that deforms in a nonlinear manner. Network models called reservoir computing (RC) (hereafter referred to as RCNs) are themselves known technology, and so detailed explanation thereof will be omitted, however an example of an RCN is a network in which part of a recurrent neural network is fixed (switched to a random network), and only the inter-node connections from the middle layer 542 to the output layer 544 are optimized.
Description follows regarding training processing performed in a training processing section 52A employing an RCN, with reference to
As illustrated in
Note that fixed weight coefficients are coefficients that are pre-set. For these fixed weight coefficients, coefficients determined as initial values may be set. The fixed weight coefficients may be set as weight coefficients for cases in which training data is employed, and the inter-node connections and the like are optimized using the training-data output data 6 as the target so as to minimize error between the generated output data 6A and the training-data output data 6 by optimization performed either a specific number of times or for a specific duration that is insufficient for the errors to be minimized completely.
The weight parameters that define the inter-node connections from the reservoir layer 543 to the output layer 544 are derived by using multiple training data so as to train to minimize errors between the generated output data 6A and the training-data output data 6.
The training processing section 52A including the generator 54A may be configured by a computer including a non-illustrated CPU to execute the training processing. For example, as in the example of the training processing of the
Next, the training processing section 52A employs multiple training data of results measured in time series to generate the learning model 51 (step S124). Namely, the RCN is built by training only the inter-node connections from the reservoir layer 543 to the output layer 544, and then obtaining a set of information about the weight parameters of the inter-node connections of the training result. Data expressed as a set of the weight coefficients derived at step S122 and information about the weight parameters of the inter-node connections of the training result of step S124 is stored as the learning model 51 (step S126).
In the elastic body performance estimation device 1 described above, the generator 54A that has been generated and trained is employed as the learning model 51. Namely, the weight coefficients expressing the inter-node connections from the input layer 540 to the reservoir layer 543 and the inter-node connections and feedback connections in the reservoir layer 543, and the weight parameters expressing the inter-node connections from the reservoir layer 543 to the output layer 544, correspond to the learning model 51. Employing the sufficiently trained learning model 51 means that it is not impossible to estimate a performance indicator from the pressure characteristic and the electrical characteristic by operation cycling for a rubber actuator that deforms in a nonlinear manner.
As described above, the present exemplary embodiment optimizes the learning model 51 built with a network using an RCN instead of a general recurrent neural network. This thereby enables the required training time to be suppressed compared to building a learning model using a general recurrent neural network. Moreover, this also enables a need for an extensive memory (memory capacity) that would be needed to perform backpropagation through time as in a general recurrent neural network to be suppressed.
Next, description follows regarding a third exemplary embodiment. The third exemplary embodiment considers improving the training effect of the learning model 51 for estimating the performance state of the rubber actuator 2. Note that due to the third exemplary embodiment being configured substantially the same as the first exemplary embodiment and the second exemplary embodiment, the same reference numerals will be appended to the same portions and detailed explanation thereof will be omitted.
As described above, it is possible to suppress training time by employing an RCN in which part of a recurrent neural network is fixed. However, sometimes the effect of training is insufficient in cases in which fixed weight parameters are employed for the inter-node connections from the input layer 540 to the reservoir layer 543 and for the inter-node connections and feedback connections in the reservoir layer 543. This is because even when the weight parameters of the inter-node connections from the reservoir layer 543 to the output layer 544 are trained, the limited number of individual nodes in the reservoir layer 543 are set with the fixed weight parameters, and sometimes output from the reservoir layer 543 is not an output sufficient for optimization. Adopting a more complex structure for the recurrent neural network using the reservoir layer 543 might be envisaged, however this is not preferable due to the time needed to set the reservoir layer 543.
However, as is well known, reservoir computing (RCN) projects an input onto high dimension feature space by nonlinear conversion into high dimension space. In regard to this point, a known approach is to employ a nonlinear dynamic system in the reservoir layer 543 instead of a recurrent neural network in a network model called physical reservoir computing (PCR) (hereafter referred to as PRCN). Due to PRC and PRCN both themselves being known technology, detailed explanation thereof will be omitted, however they may be utilized to retain data related to deformation of the rubber actuator 2 that deforms in a nonlinear manner in the reservoir layer 543. Namely, the PRCN is appropriately applicable to estimating a performance state of the rubber actuator 2 that deforms in a nonlinear manner.
Explanation follows regarding training processing in a training processing section 52B, with reference to
As illustrated in
Note that the fixed weight coefficients, as described above, may be setting of fixed coefficients as initial values, and may be setting of weight coefficients in cases in which training data is employed for optimization performed only a specific number of times or for a specific duration.
The physical reservoir layer 545 retains multiple physical correlations in time series of the rubber actuator 2, extracts from the input layer 540 a performance state (performance indicator) corresponding to input data (the pressure characteristic and the electrical characteristic) close to the unknown input data (the pressure characteristic and the electrical characteristic), and outputs these as plural feature values to the output layer 544. Roughly speaking, multiple correlations of the performance indicator, which is the performance state, against the pressure characteristic and the electrical characteristic that change in time series by operation cycling, are stored as the behavior of the rubber actuator 2, and respective performance states (performance indicators) of the plural rubber actuators 2 close to the unknown input (the pressure characteristic and the electrical characteristic) are selected and output as feature values. This thereby enables execution of complex calculations to be suppressed.
The training processing section 52B including the generator 54B may be configured including a computer including a non-illustrated CPU that executes the training processing. For example, as illustrated in an example of the training processing in
Next at step S134, the training processing section 52B generates the learning model 51 employing the multiple training data of the results measured in time series. Namely, a PRCN is built by training only the inter-node connections from the physical reservoir layer 545 to the output layer 544, and obtaining a set of information of weight parameters of the inter-node connections of the training result. Next data expressed by the weight coefficients derived at step S132 and sets of information of weight parameters of the inter-node connections of the training result of step S134 are stored as the learning model 51 (step S136).
Then in the elastic body performance estimation device 1 described above the generated and trained generator 54B is employed as the learning model 51. Namely, weight parameters expressing the connections of nodes from the input layer 540 to the physical reservoir layer 545, the physical reservoir layer 545, and the inter-node connections from the physical reservoir layer 545 to the output layer 544 correspond to the learning model 51.
In response to the pressure characteristic and the electrical characteristic of the rubber actuator 2 under operation cycling, which is unknown input data, the elastic body performance estimation device 1 extracts, from out of the physical correlations to time series of the rubber actuator 2 retained in the physical reservoir layer 545, a performance state (performance indicator) corresponding to input data (the pressure characteristic and the electrical characteristic) close to the unknown input data (the pressure characteristic and the electrical characteristic) from the input layer 540, and outputs these as plural feature values to the output layer 544. The output layer 544 then uses the plural feature values from the physical reservoir layer 545 to estimate the performance state (performance indicator) of the rubber actuator 2 using the trained weight parameters, such as by linear combination for example. Using the sufficiently trained learning model 51 means that it is not impossible to estimate a performance state (performance indicator) from the pressure characteristic and the electrical characteristic for the rubber actuator 2.
As described above, the present disclosure builds a network using a PRCN instead of an RCN, and optimizes the learning model 51. An improvement is accordingly achieved in the training effect of the learning model 51 compared to a case in which the learning model is built with an RCN.
This thereby enables extremely close estimation of the length of the rubber actuator 2 to the length of the actual rubber actuator 2 by optimizing the learning model 51 built with a network by PRCN instead of RCN.
As described above, in the present disclosure a case has been described of application of the present disclosure to a rubber actuator as a member, however obviously the member is not limited to being a rubber actuator. Moreover, a case has been described in which there are at least 3 different types of physical quantity that change according to deformation of the member, with the pressure values expressing the magnitude of pressure as the first physical quantity, the electrical resistance values expressing the magnitude of the electrical characteristic as the second physical quantity, and the length expressing the magnitude of deformation as the target physical quantity, and with the length being estimated from the pressure value and electrical resistance value. However, each of the first physical quantity, the second physical quantity, and the target physical quantity are not limited thereto, and the pressure values or the electrical resistance values may be set as the target physical quantity.
Although a description has been given in terms of the exemplary embodiments, the technical scope of the present disclosure is not limited to the ranges of the exemplary embodiments described above. Various modifications and improvements may be made to the above exemplary embodiments within a scope not departing from the spirit of the present disclosure, and embodiments including such modifications and improvements are contained in the technical scope of the present disclosure.
Moreover, although in the above exemplary embodiments, processing for a case implemented by processing of a software configuration has been described using a flowchart, there is no limitation thereto, and, for example, an embodiment may be adopted in which each processing is implemented by a hardware configuration.
Moreover, part of the elastic body performance estimation device, for example a neural network such as a learning model or the like, may be configured as a hardware circuit.
The entire content of the disclosure of Japanese Patent Application No. 2021-080622 filed on May 11, 2021 is incorporated by reference in the present specification. All publications, patent applications and technical standards mentioned in the present specification are incorporated by reference in the present specification to the same extent as if each individual publication, patent application, or technical standard was specifically and individually indicated to be incorporated by reference.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2021-080622 | May 2011 | JP | national |
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2022/019726 | 5/9/2022 | WO |
