JP7607934B2 - Trained model, control device, friction stir welding system, neural network system, and trained model generating method - Google Patents

Description

The present invention relates to a trained model for predicting joint characteristics in friction stir welding, a control device including the trained model, a friction stir welding system including the control device, a neural network system for generating the trained model, and a method for generating the trained model.

Friction stir welding is a method of joining materials by contacting them with a rod-shaped tool while rotating at high speed, utilizing the heat of friction between the materials (see, for example, Patent Document 1). Friction stir welding is a solid-phase joining method in which the maximum temperature reached during joining does not reach the melting point of the base material, and it is an innovative joining method that has higher joint efficiency than conventional fusion welding and, in some cases, the joint can be stronger than the base material (see, for example, Patent Document 1).

Patent No. 2712838

In friction stir welding, if the welding conditions such as welding speed and tool rotation speed are set, essentially the same joint can be obtained with good reproducibility. However, the appropriate welding conditions vary depending on the material, size, shape, etc. of the materials to be welded, and determining the welding conditions requires numerous preliminary experiments for each joint. In other words, it is difficult to accurately predict the joint characteristics from the welding conditions.

Therefore, the present invention aims to accurately predict joint characteristics from the joining conditions in friction stir welding. Note that in the present invention, friction stir welding includes butt joining, lap joining, line joining, spot joining, and combinations of these, as well as the friction stir process, which is a surface modification technology.

In order to solve the above-mentioned problems, the present invention provides a trained model for causing a computer to function so as to output a predicted value of joint characteristics based on input data indicating joining conditions of friction stir welding, the trained model being composed of a neural network including an input layer, an output layer, and a hidden layer having a plurality of hidden units interposed between the input layer and the output layer and represented by a nonlinear activation function, the neural network uses the variation in experimental data indicating the joint characteristics to learn weighting coefficients so as to minimize a function represented by the error between the experimental data and the predicted value, and performs a calculation based on the trained model on input data indicating the joining conditions of friction stir welding input to the input layer, and outputs a predicted value of the joint characteristics from the output layer and information indicating the reliability of the predicted value. A trained model for causing a computer to function is provided.
Here, the trained model may output the joining conditions. For example, the trained model may use the predicted values of the joint characteristics to calculate and output joining conditions for obtaining better joint characteristics.

In the trained model of the present invention having the above-mentioned configuration, it is preferable to cause the computer to function so that the neural network further learns the weighting coefficients using the variation in the weighting coefficients and further outputs information indicating the degree of influence of the joining conditions on the predicted value.

Furthermore, in the trained model of the present invention having the above-mentioned configuration, it is preferable that the neural network learns the weighting coefficients for each of a plurality of models having different numbers of hidden units, and selects, as the prediction model, a model among the plurality of models that minimizes the error between the predicted value and the experimental data.

In addition, in the trained model of the present invention having the above-mentioned configuration, it is preferable that the neural network learns the weighting coefficients for each of a plurality of models having different numbers of hidden units, ranks the plurality of models based on the error between the predicted value and experimental data, and among the plurality of new models generated by sequentially combining the highest-ranked models, the model that has the smallest error between the predicted value and experimental data is used as the prediction model.

The present invention also provides a control device including an input unit that acquires observation data indicating the joining conditions in friction stir welding, a memory unit that stores the above-mentioned trained model, a calculation unit that determines an operation amount of a friction stir welding device based on a result of applying the observation data to the trained model, and an output unit that instructs the friction stir welding device on the determined operation amount.

The present invention also provides a control device including an input unit that acquires observation data indicating the welding conditions and joint characteristics in friction stir welding, a memory unit that stores a predetermined learning algorithm, a calculation unit that determines an operation amount of a friction stir welding device based on a result of applying the observation data to the predetermined learning algorithm, and an output unit that instructs the friction stir welding device on the determined operation amount, wherein the predetermined learning algorithm is composed of a neural network including an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function, so as to cause the control device to function to output an appropriate operation amount of the friction stir welding device based on input data indicating the welding conditions of the friction stir welding device and the joint characteristics, the neural network learns weighting coefficients so as to minimize an error function represented by a value function including a relationship between the welding conditions and the joint characteristics, performs calculations based on the learned model for input data indicating the welding conditions of friction stir welding input to the input layer, and outputs the operation amount of the friction stir welding device from the output layer.

Furthermore, the present invention also provides a friction stir welding system including a friction stir welding apparatus for performing friction stir welding, a measuring instrument for measuring observation data indicating the welding conditions in the friction stir welding, and the above-mentioned control device.

The present invention also provides a neural network system that generates a trained model for causing a computer to function to output a predicted value of joint characteristics based on input data indicating the joining conditions of friction stir welding, the neural network including an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function, the neural network using the variation in experimental data indicating the joint characteristics to learn weighting coefficients so as to minimize a function represented by the error between the experimental data and the predicted value, and performing calculations based on the trained model on input data indicating the joining conditions of friction stir welding input to the input layer, to generate a trained model that outputs a predicted value of the joint characteristics from the output layer and outputs information indicating the reliability of the predicted value.

The present invention further provides a method for generating a learning model for causing a computer to function to output a predicted value of joint characteristics based on input data indicating the joining conditions of friction stir welding, the learning model for generating the trained model includes a neural network including an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function, the method including: learning weighting coefficients in the neural network using the variance in experimental data indicating the joint characteristics to minimize a function represented by the error between the experimental data and the predicted value; and performing a calculation based on the trained model on input data indicating the joining conditions of friction stir welding input to the input layer, to generate a trained model that outputs a predicted value of the joint characteristics from the output layer and outputs information indicating the reliability of the predicted value.

According to the present invention, in friction stir welding, joint characteristics can be accurately predicted from the welding conditions. Therefore, by applying the present invention to the control of an FSW device, various joints can be easily obtained using optimal welding conditions without the need for numerous preliminary experiments and regardless of the skill level of the worker. It is widely known that joints obtained by FSW have various excellent characteristics, and if it were possible to automatically grasp the appropriate welding conditions, it is expected that the number of users of friction stir welding will increase dramatically.

1 is a schematic diagram of a friction stir welding system 1 according to an embodiment of the present invention. FIG. 1 is a diagram showing an example of a rotating tool 11 used in friction stir welding. 2 is a cross-sectional view showing an example of a joint between materials 41 and 42. FIG. FIG. 1 is a conceptual diagram illustrating an example of a neural network. 1 is a flowchart illustrating an example of a procedure for generating a trained model. 13 is a flowchart illustrating an example of a procedure for determining the number of layers of a neural network. 11 is a flowchart illustrating an example of a procedure for combining a plurality of prediction models to generate a more appropriate model. 1 is a graph showing an example of predicted values and error bars output from a prediction model. 13 is a graph showing an example of the importance of bonding conditions output from a prediction model. 2 is a block diagram showing an example of a functional configuration of a control device 20. FIG. 4 is a flowchart showing an example of the operation of the control device 20. FIG. 2 is a block diagram showing an example of the functional configuration of a control device 120. 4 is a flowchart showing an example of the operation of the control device 120.

Representative embodiments of the present invention will be described in detail below with reference to the drawings. However, the present invention is not limited to these drawings. In addition, since the drawings are intended to conceptually explain the present invention, dimensions, ratios, or numbers may be exaggerated or simplified as necessary for ease of understanding.

Here, we will explain the overall picture of the friction stir welding system, and then explain the predictive model (trained model) of joint characteristics and its learning method, as well as the application of the model to control.

1. Overall Configuration of the Friction Stir Welding System The friction stir welding (FSW) system 1 according to this embodiment is for welding various materials 40 by friction stirring. Here, the material 40 to be welded is typically a metal material such as aluminum, magnesium, copper, titanium, zinc, lead, steel, nickel, cobalt, or an alloy or composite material of these materials, but may also be a resin material such as plastic or CFRP.

As shown in FIG. 1, the FSW system 1 includes a friction stir welding (FSW) apparatus 10 that performs friction stir welding, a control device 20 that controls the FSW apparatus 10, and a measuring device 30. The measuring device 30 is a collective term for one or more types of measuring equipment that measure the joining conditions and joint characteristics of FSW, and measures, for example, the state of the FSW apparatus 10 and the material 40 (e.g., the rotation speed, pressing load, angle, and moving speed of the rotating tool 11, and the temperature and width of the joint). The measuring device 30 includes an imaging device that photographs the joint. The measuring device 30 may be included in the FSW apparatus 10 and the control device 20.

1-1. Overview of Friction Stir Welding Apparatus An FSW apparatus 10 is equipped with a rotary tool 11 that rotates at a desired rotation speed to perform friction stir welding.
2, the rotary tool 11 includes a columnar shoulder 12 and a probe 13 that is disposed at the tip of the shoulder 12 and has a smaller diameter than the shoulder 12. As the rotary tool 11, either a type in which the probe 13 and the shoulder 12 are driven integrally, or a type in which the probe 13 and the shoulder 12 are driven independently can be used.

As shown in Figure 1, the probe 13 is pressed into the material 40 and moves along the butt surfaces to be joined. At this time, the material is restrained by the backing plate and is joined by the plastic flow caused by the rotating tool 11 while maintaining a solid state. Note that the rotating tool 11 pressed into the material 40 does not necessarily need to be moved laterally; spot joining can be achieved by simply pulling it out.

When the material 40 is an aluminum alloy, magnesium alloy, or other material that is easy to stir by friction, it is preferable that the probe 13 is formed with a screw 14. The length of the probe 13 is approximately equal to the plate thickness of the material 40, but it is preferable that it is slightly shorter (for example, about 0.05 mm to 0.2 mm) so as not to come into contact with the backing plate. The cross-sectional shape of the probe 13 is generally cylindrical, but other shapes such as elliptical may also be used.

The shoulder 12 prevents the softened material 40 from popping out and generates and maintains frictional heat. The shoulder 12 rotates while in contact with the material 40 and moves at a desired moving speed (joining speed) in the joining direction X.

The FSW device 10 also includes a tool holder that holds the rotating tool 11, a motor that drives the rotating tool 11, a pressing mechanism that presses the rotating tool 11 against the material 40, a moving mechanism that moves the rotating tool 11 in the joining direction (X direction) and in a direction perpendicular to the joining direction (Y direction), and a material holder that holds the material 40 (all not shown). The tool holder can hold the rotating tool 11 in a state in which the rotating tool 11 is tilted at a certain angle relative to the material. Therefore, during joining, the rotating tool 11 is tilted at a predetermined angle (for example, about 1 to 5 degrees) so that the tip of the probe 13 leads the shoulder 12.

As shown in Figure 3, the cross-sectional structure of the material 40 (41, 42) joined by FSW is divided into several regions. That is, in the center of the joint, there is a stir zone (joint) 43, which usually has a recrystallized structure consisting of equiaxed grains of several μm. Outside the stir zone 43, there is a thermal processing affected zone 44, which has a shape in which the crystal grains have been elongated due to plastic deformation, and further outside that, there is a heat affected zone (HAZ) 45, which has not been plastically deformed but has been affected by heat. The state of the stir zone 43, the thermal processing affected zone 44, and the HAZ 45 (e.g., width, degree of softening, etc.) affects the joint characteristics.

Therefore, in FSW, in addition to securely holding the materials 40 to be joined, it is important to properly adjust the joining conditions (FSW conditions) such as the shape of the rotating tool 11, pressing load, angle, rotation speed, and joining speed in order to obtain a good joint. However, the joining conditions differ depending on the type of material, plate thickness, and joining shape. In addition, the quality of the joint, that is, the joint characteristics, can be evaluated, for example, by the mechanical properties of the joint (tensile strength, bending strength, fatigue strength, corrosion resistance, etc.), the amount of burrs generated, the amount of energy input, the joining speed, etc.

1-2. Overview of the Control Device The control device 20 acquires measurement data from the measuring device 30 and controls the FSW device 10 based on this measurement data. The control device 20 can be configured as a computer including an arithmetic unit (CPU, GPU, etc.) and a storage device (RAM, ROM, etc.). The control device 20 may be a single computer or may be configured with multiple computers.

The control device 20 can control the FSW device 10 using a trained prediction model. The prediction model may be generated in the control device 20, or may be generated in an external computer, for example, by cloud computing.
Before describing the details of the control device 20, machine learning in this embodiment will be described below.

2. Building a predictive model
2-1. Learning by Neural Network A neural network can express complex relationships by combining nonlinear functions. In this embodiment, focusing on the flexibility of such a neural network, the neural network is applied to data processing of joint characteristics in which many elements are intricately intertwined.

However, it is not easy to immediately apply neural networks to predicting joint characteristics, and it is necessary to appropriately handle errors due to data variance and fitting. Therefore, the inventors decided to introduce the idea of Bayesian estimation into neural networks. This makes it possible to predict statistical error bars. The size of the error bar depends on the input conditions at the time, and if the data has a large variance and is unreliable, the error bar will be large.

More specifically, the structure of the neural network is shown in FIG. 4. As shown in the figure, the neural network includes an input layer and an output layer. When various joining conditions (FSW conditions) are given as input x _i , a predicted value y of the joint characteristics is obtained as output. The output y may be one or more, but for convenience of explanation, it is assumed here that the output y is one.

The input joining conditions include, for example, "pre-joining parameters" such as the chemical composition of the sample, the rotation speed of the rotating tool 11, the movement speed, the applied load (or the insertion amount of the probe 13), the advance angle, the shape of the sample, the thickness of the sample, the shape of the clamping jig, and the tool shape, as well as "joining parameters" such as the temperature of the joint, the tool torque during joining, the tool load during joining, and the tool position during joining. The output joint characteristics are numerical values that quantitatively represent the characteristics of the joint, and include, for example, the joint strength, the width of the stirring portion 43, the amount of burr generation, HAZ softening, hardness distribution, maximum hardness, and corrosion resistance.

Between the input layer and the output layer, a hidden layer having multiple hidden units is arranged, making it possible to express complex functions. Here, for the sake of convenience, the hidden units are assumed to constitute one hidden layer, but it goes without saying that the hidden units can constitute multiple hidden layers. Note that it is not necessary for all hidden units to be expressed by nonlinear functions, and they may include units expressed by linear functions.

In such a neural network, the relationship between the input x _j and the i-th hidden unit h _i can be expressed as follows using a nonlinear activation function:
where w _ij ⁽¹⁾ is the weight between x _j and h _i , and θ _i ⁽¹⁾ is a threshold value. The hyperbolic tangent function tanh is an example of a nonlinear function, and other nonlinear functions such as a sigmoid function or a rectified linear function (ReLU) may be used.

Moreover, the relationship between the hidden unit h _i and the output y can be expressed as a linear function as follows:
where w _i ⁽²⁾ is the weight between h _i and y, and θ ⁽²⁾ is a threshold value.

A network that can express complex relationships can be constructed by expressing the relationship between the input x and the hidden unit h as a nonlinear function as in the above formula 1. Hereinafter, the weights w _ij ⁽¹⁾ and w _i ⁽²⁾ in formulas 1 and 2 will be simply referred to as weight coefficients w, and the thresholds θ _i ⁽¹⁾ and θ ⁽²⁾ will be simply referred to as thresholds θ.

Here, the input variable x may be normalized to the range of ±0.5 using the following formula, and the output variable y may also be normalized using the same method.
Here, _xN is the normalized x, _xmax is the maximum value of the original data, and _xmin is the minimum value.

2-2. Energy function M(w)
In order to determine the above-mentioned weighting coefficient w and threshold value θ, it is considered to minimize the energy function M(w) expressed by the following equation.
Here, E _D is an error function, E _w(c) is a regularization term (term) and will be described in detail later. The parameter vector w includes a weighting coefficient w and a threshold θ. α _c and β are parameters that control the complexity of the model.

To minimize the energy function M(w) on a computer, for example, the variable metric method can be used, and the gradient of M(w) can be calculated using the backpropagation method, but we will not go into details here.

As shown in the above equation, the energy function M(w) is composed of an error function E _D and an appropriated term E _w(c) . The error function E _D can be expressed as the sum of squares of the difference between the predicted value y(x ^m ; w) of the prediction model and the experimental value t ^m as shown in the following equation.
Here, {x ^m , t ^m } is a data set, x ^m is an input variable, and t ^m is experimental data, i.e., a target, and m is a label for the combination of data and target.

From the above equation, when the prediction results of the prediction model match the experimental data well, ie, when y( ^xm ;w) is close to ^tm , the error function E _D is at a minimum.

Next, the fit term _Ew serves to make the model output y(x;w) a smooth function of the input x. In other words, the fit term encourages the weight coefficient w to be small, suppressing the overfitting of the predictive model to the variability of the data set, i.e., overfitting.

The appropriateness term _Ew may be expressed as the sum of a plurality of appropriateness terms _Ew(c) . For example, one class may be created based on the weight coefficient between the input x and the hidden unit h, one class based on the weight coefficient between the hidden unit h and the output y, and one class based on the threshold value of the hidden unit h, and each appropriateness term _Ew(c) may be calculated. In this case, each appropriateness term Ew _(c) is expressed as the sum of squares of the coefficients w _i belonging to each class, as shown in the following equation.

The parameters _αc and β then control the complexity of the model, along with the number of hidden units h. For example, if the data is distributed Gaussianly with a standard deviation of _σν , then β defines the variance of the data _σν2 = ¹ /β, and _αc defines the variance of the weighting coefficient w, σw _(c) ² = 1/ _αc .

From Equation 4, the parameter α _c has the effect of decreasing the weight coefficient w. Therefore, a large σ _w means that the corresponding input has a large change in the output. From this, σ _w can be used as an index showing the importance of each input. Also, σ _ν shows the variance of the data and can be used as an error bar as described later.

In this embodiment, as will be described in detail later, the parameters _αc and β can be calculated using the concept of Bayesian estimation. The initial values of each parameter may be appropriately determined by the user, and for example, the initial values may be set by allowing small variations in the weighting coefficient w.

2-3 Determination of parameters α _c and β Next, the parameters α _c and β will be described in detail.
If the parameter β is too large, the degree of freedom of the function increases, and overfitting becomes more likely. Conversely, if the parameter α _c (hereinafter simply referred to as α) is large, the function becomes too smooth and does not fit the data. Considering the importance of α and β, the inventors have decided to introduce the idea of Bayesian estimation to give α and β statistical meaning.

In general, the conditional probability p(w|D) that a certain combination w occurs for a weighting factor w and a threshold value θ, given that certain data D occurs, is expressed as follows:

In order to determine the most likely weighting coefficient w and threshold value θ, it is necessary to maximize p(w|D).
Since the following relationship exists, each probability is calculated assuming that the probabilities p(w|D) and p(w) included in the right-hand side vary according to a normal distribution.

First, the probability p(w|D) is calculated.
In general, a normal distribution f(x) is expressed by the following equation, where m is the mean and σ is the standard deviation.
Therefore, when the weighting coefficient w and the threshold value θ are expressed by a vector w, the variance of the data is expressed by the following equation.
In this case, x ^(m) is an input variable, t ^(m) is experimental data, that is, a target, _ZD is a normalization constant, and _σν is the variance of the data.

Here, by substituting the error function E _D (w) expressed by the formula 5 into the formula 10, the following formula is obtained.

Next, when the probability p(w) is calculated, p(w) also has variation. Taking into account Equation 5, the probability is as follows:
where Z _w is a normalization constant and σ _w is the deviation from the true w value.

Substituting Equations 11 and 12 into Equation 8, we get
It becomes.

Therefore, to maximize p(w|D),
should be minimized.

In comparison with minimizing the energy function M(w) expressed by Equation 4, the parameters α and β are
This has statistical significance, making it possible to provide precise training.

In this case, the error bar σ is expressed by the following equation.

2-4. Training Procedure As described above, training for generating a prediction model is performed according to the following procedure. The training may be performed via execution of a training program in the control device 20, or may be performed by an external computer capable of executing the training program.

More specifically, in step S11 of Fig. 5, training data ( ^xm , ^tm ) is acquired. Here, ^xm is a joining condition that is an input variable, and ^tm is a target joint characteristic (experimental data). In addition, in step S12, a parameter vector w is set. Here, the vector w may include a weighting coefficient w and a threshold value θ.

Next, in step S13, the variance σ _v of the training data (x ^m , t ^m ) and the variance σ _w of the parameter vector w are calculated.

Then, in step S14, a parameter vector w that minimizes the energy function M(w) is calculated. The variable metric method can be used for this calculation, and the gradient of M(w) can be calculated using, for example, the backpropagation method.

In step S15, the parameter vector w is updated in accordance with the calculation results of the previous step, and the variation σ _w is calculated and updated.

Then, steps S14 and S15 are repeated a predetermined number of times to end the series of procedures. Alternatively, the end condition may be that the parameter vector w converges within a preset range. The parameter vector w thus obtained is used as a parameter of the prediction model.

This process is repeated while adjusting the number of hidden units h, i.e., the number of elements in the parameter vector w. As a result, the difference between the predicted value of the predictive model and the experimental data monotonically decreases as the number of hidden units increases. Here, the more complex the relationship between the input and output, the more hidden units are required, but because experimental data always contains errors, increasing the number of hidden units h too much will result in overfitting, and the prediction accuracy, i.e., generalization performance, will actually decrease.

For example, half of the data can be randomly selected, the neural network can be trained using only that half of the data, and the remaining half can be used as test data to examine the generalization performance of the model. The test data is used to measure the error between the model's predicted value and the test data, i.e., the test error. If the number of data sets is small, most of the data can be used for training and the remaining data for testing. For example, if there are 30 data sets, it is preferable to use 27 of them for training and the remaining 3 for testing.

When the number of hidden units h reaches a certain number, the test error is minimized. If the model at this point is judged to be optimal and used as a trained model, the prediction accuracy will improve.

Therefore, the procedure for determining an appropriate number k ^* of hidden units h can be as follows.
First, in step S21 of FIG. 6, the above-mentioned training is repeated to prepare a plurality of prediction models having k (k=1, 2, . . . ) hidden units h ^k .

Next, in step S22, the test error is calculated for each prediction model, and in step S23, the number k ^* that minimizes the test error is determined, and the corresponding prediction model is set as the optimal model. Note that, for the combination of prediction models described below, the models may be ordered in ascending order of test error, such as the optimal model, the second optimal model, the third optimal model, and so on.

2-5. Optimization by combining predictive models Multiple models may be combined to further improve the prediction accuracy. In other words, with neural networks, it is possible to create many models that are not very different from the optimal model in terms of error but have completely different structures. By combining these models, the shortcomings of each individual model can be mutually compensated for, and the prediction accuracy can be further improved.

Specifically, in step S31 of FIG. 7, the error function (or the results of the comparison of the test errors described above) is used to rank the multiple predictive models as an optimal model, a second optimal model, a third optimal model, etc.

In step S32, the average value of the predicted value of the optimum model and the predicted value of the second optimum model is determined as the predicted value of the new model according to the following equation, and in step S33, the error between this predicted value and the experimental data is calculated.

Then, in step S34, it is determined whether the previously calculated error is a minimum. If it is determined that the error is not a minimum, in step S35, the next model is combined according to the above formula, and steps S33 and S34 are executed again.

If it is determined that the error is minimal, in step S36, the combination of models at that time is adopted as the prediction model. For example, if the error is reduced by combining the first four optimal models, but the error increases when the fifth optimal model and onward are added, the combination of the optimal model through the fourth optimal model can be used as the prediction model.

In this case, the error bar σ of the new prediction may be calculated using the following formula:
where N is the number of models, and y _i and σ _i are the predicted values and error bars of each model.

In the prediction model of this embodiment, it is possible to output predicted values and error bars of joint characteristics for certain joining conditions. For example, in the example of Fig. 8, the relationship between the joining speed and the predicted value of the joining strength is represented by a curve L, and is also represented by curves D indicating the upper and lower limits of the error bar on either side of the curve L. Alternatively, the error bar may be represented by a line segment E having a length corresponding to the variation σ _ν . Of course, the predicted value and the error bar may be represented by numerical values.

By obtaining such predicted values and error bars, the user can obtain a reasonable prediction of the joint characteristics and know the accuracy of the prediction. Since a large error bar means that the number of experimental data is insufficient or the accuracy of the experiment is insufficient, the user should increase the experimental data in the region with a large error bar or conduct a highly accurate experiment (conduct a reinforcement experiment), thereby improving the reliability of the model prediction and further deepening the research on the material. Alternatively, when the prediction model is used to control the FSW device 10, highly accurate and stable control can be expected by preferentially adopting the predicted value in the region with a small error bar.

Furthermore, the prediction model can indicate the importance (degree of influence) of each welding condition on the joint characteristics based on the variation σ _w of the parameter vector w, for example, as shown in Fig. 9. This allows the user to understand the welding conditions that have an influence on the relevant joint characteristics, and can use this as a guideline for research and development. In addition, in the control of the FSW device 10, it is also possible to quickly and efficiently approach the target value by, for example, preferentially adjusting the welding conditions with high importance.
Alternatively, the prediction model may output the joining conditions. For example, the prediction model may use the predicted values of the joint characteristics to calculate and output joining conditions for obtaining better joint characteristics. The prediction model may refer to the error bars and importance described above when selecting better joining conditions. This allows the user to obtain suggestions for obtaining a better joint.

For the prediction model obtained in this way, if a specific input among the multiple input variables (joining conditions) is changed while the remaining inputs are not changed, the output (predicted value) will show the response characteristics to that specific input. By examining the response characteristics of such an output for all input variables, it is possible to grasp the trend of the output with respect to the input, and therefore it is possible to grasp the joining conditions with excellent response characteristics together with or instead of the variation σ _w of the parameter vector w. This knowledge can be used for controlling the FSW device 10, and can also be used for research on the material in question.

3. Examples of control methods The above-mentioned model can also be applied to the control of an FSW device. Several application forms are possible, including control that utilizes reinforcement learning, so we will first provide an overview of reinforcement learning.

3-1. Reinforcement learning Reinforcement learning is a framework for learning control that adapts to the environment through trial and error.
The agent, which is the learning subject, makes a decision at time t according to the observed value s(t) of the state of the environment to be controlled, and outputs an action a(t). The agent's action causes the environment to transition to a state s(t+1), and a reward r(t) corresponding to the transition is given to the agent. This series of steps is repeated, and the agent learns a policy π from the state observation to the action output with the goal of maximizing the profit.

For example, in Q-learning, which is a type of reinforcement learning, an agent takes various actions a in a certain state s by trial and error, and learns the optimal action value Q(s, a) using the reward at that time. The update formula for the action value function Q(s, a) is expressed as follows:
where s _t and a _t are the state s of the environment and the action a at time t.

By the action a _t , the environment transitions to state s _t+1 , and the reward r _t+1 is calculated accordingly. In the above formula, the term maxQ(s _t+1 , a) is the Q value corresponding to the action a with the highest Q value in state s _t+1 , multiplied by the discount rate γ (γ is 0<γ≦1). In addition, α (0<α≦1) is the learning coefficient.

This update equation increases Q(s _t , a _{t )} if the evaluation value max _a Q(s t ₊₁ , a) of the best action in the next state s _{t+1 is} greater than the evaluation value Q(s _t , a _t ) of action a t in state s t, and decreases Q(s _t , a _t ) if it is smaller.

When applying reinforcement learning such as Q-learning to the control of an FSW device, action a can be made to correspond, for example, to the manipulated variable of the joining conditions, and state s can be made to correspond, for example, to various observed data. Furthermore, conditions related to reward include, for example, the joining quality, joining speed, and energy consumption, and a value function can be constructed that changes depending on these elements. For example, the reward increases when the joining strength approaches or matches the target value, and the reward decreases or decreases as it deviates from the target value. Furthermore, the reward increases if the amount of burrs generated is small, and decreases if the amount is large. Furthermore, the reward decreases if energy consumption is high, and increases if energy consumption is low.

Given that FSW involves complex phenomena, it is preferable that the action value function Q(s, a) be expressed by a function approximation rather than a table. To perform a function approximation of the action value function Q(s, a), first, the action value function Q(s, a) is modeled as a function Q'(s, a; w) represented by a parameter vector w. Then, during learning, the parameter vector w of Q' is updated rather than updating the Q value itself.

Supervised learning can then be applied to learn the parameter w. To deal with complex problems such as friction stir welding, for example, neural networks can be used. For example, a neural network can be prepared that, when a state s is input, outputs a value function Q for all possible actions a, and this can be used as the function approximation Q'(s, a; w) (Q network).

Here, the parameter w of the function approximation corresponds to the weight parameter of the neural network. The weight parameter can be adjusted so as to minimize the error function of the following equation:
Here, y _t (s, a) is the target value.
Therefore, the energy function M(w) of the above-mentioned equation 4 and the error function E _D of the above-mentioned equation 5 can be used.

However, in reinforcement learning, training data to be used as the target value y _t (s, a) is not given, so the target value y _t in the above equation is replaced with, for example, the state and reward actually sampled from the environment, and the following equation is used:
may also be used.

Then, the parameter w of the value function Q' is updated, for example, as follows:
This can be done by using the following:

The parameter w and the evaluation function Q' can be updated, for example, every time a connection is made. Various methods have been proposed for implementing the Q network, and these methods can be incorporated into this embodiment. In this case, the above formula is modified as appropriate before the execution process is performed.

4. Details of the Control Device In this embodiment, the above-mentioned trained model or learning model is used to control the FSW device 10. For example, the control device may determine the operation amount of the FSW device based on a predicted value output by the trained model (prediction model) (Configuration Example 1), or may cause a reinforcement learning learning model to determine the optimal operation amount (Configuration Example 2).

4-1. Control device functional configuration example 1
First, a description will be given of a control device 20 that determines the manipulated variable of the FSW device based on the predicted value of a prediction model. As shown in FIG. 10 , the control device 20 includes an input unit 21, a calculation unit 22, a storage unit 23, and an output unit 24.

The input unit 21 acquires observation quantities (observation data) related to FSW. The input unit may acquire the observation data by measuring it itself, or may acquire it from, for example, the FSW device 10 or another measuring device, or may even acquire it by user input. Furthermore, the observation data may be acquired during the execution of FSW, or may be acquired after execution. Furthermore, the observation data may be obtained by calculating multiple observation quantities, such as energy consumption.

Here, the observation data includes data related to the joining conditions of the FSW and data related to the joint characteristics. However, the observation data related to the joining conditions and the observation data related to the joint characteristics may overlap.

Specifically, the observed data on the joining conditions include, for example, the rotation speed of the rotating tool, the moving speed, the applied load (or the tool insertion amount), and the advance angle. The adjustment amount of each joining condition is determined by the control device 20.

Examples of observed data on joint characteristics include the joining temperature during FSW, tool rotational torque, tool Y-axis (joining direction) load, tool Z-axis (vertical) load, stir zone width, amount of burrs generated around the processing zone, joint strength, HAZ softening, hardness distribution, maximum hardness, corrosion resistance, etc. Many of these physical quantities are information related to the stir zone, and change depending on the joining conditions.

Next, the calculation unit 22 will be described. The calculation unit 22 loads a control program stored in the storage unit 23 into memory and executes it to predict joint characteristics based on observed data and to determine optimal joining conditions based on the prediction results. The calculation unit 22 includes a setting unit 25, a prediction unit 26, and an update unit 27.

The setting unit 25 performs initial setting of the joining conditions based on user input, etc. The conditions to be set include, for example, information on the materials to be joined, the target joint characteristics (strength, etc.), the working time, and the amount of energy consumption. The setting unit 25 may also use a trained prediction model to determine the joining conditions that satisfy the target joint characteristics, and set these conditions as the initial settings.

Next, the prediction unit 26 inputs the observation data into the prediction model to obtain a predicted value for the joint characteristics. The update unit 27 then compares this predicted value with a target value and updates the manipulated variable of the joining conditions based on the comparison result. For example, if the joint strength is less than the target, the update unit 27 inputs an amount obtained by adding a small amount Δ to the joining condition into the prediction model to grasp the change in the predicted value, and adjusts the joining condition so that the joint strength approaches the target value. At this time, the update unit 27 may take into account the degree of influence and reliability (variation) of the corresponding joining condition.

Then, the calculation unit 22 repeats the observation and adjustment of the joining conditions at a preset time interval, and instructs the FSW device 10 of the determined adjustment amount via the output unit 24.

Next, the memory unit 23 stores, for example, the learned prediction model, the control program of the FSW device 10, and various data. The various data include, for example, observation data, joining condition settings, and reward settings. The memory unit 23 may also store processing conditions such as joint shape, plate thickness, base material attitude, and gap amount. Note that all or a part of the memory unit 23 may be incorporated in the control device 20, or may be, for example, a memory area of an external computer.

The output unit 24 outputs instructions from the calculation unit 22 to the FSW device 10, and may also, for example, display the prediction results of the calculation unit 22 and the instructions to the FSW device 10 on a display not shown.

Next, an example of the operation of the control device 20 will be described with reference to FIG.
First, in step S41, the setting unit 25 sets the joining conditions in response to a user input. Next, in step S42, the input unit 21 acquires observed data and stores it in the storage unit 23 or an external storage device. Then, in step S43, the prediction unit 26 calculates a predicted value of the joint characteristics using the trained prediction model. Furthermore, in step S44, the update unit 27 refers to the prediction result of the prediction unit 25, adjusts and updates the joining conditions so as to bring the joint characteristics closer to the target values, and outputs the adjusted joining conditions.

The control device 20 then repeatedly executes steps S42 to S44 at set time intervals, allowing the control device 20 to automatically determine appropriate joining conditions according to the joint characteristics desired by the user.

In other words, because the prediction model is generated by a neural network including hidden units expressed by nonlinear functions, it can appropriately represent joining phenomena involving complex relationships. This improves the prediction accuracy of the joint characteristics, and at the same time, by incorporating this prediction model into the control device 20, it becomes possible to set more appropriate joining conditions according to the user's requests. The prediction accuracy of the prediction model further improves as learning opportunities increase, making it possible to set even more appropriate joining conditions. Therefore, various joints can be easily obtained using optimal joining conditions without the need for numerous preliminary experiments and regardless of the skill level of the worker.

In addition, the predictive model can provide error bars for the predicted values. Variation in the predicted values indicates a lack of data used and/or a lack of reliability, making it possible to clarify targets for which data should be actively collected. Furthermore, highly reliable control can be achieved by controlling based on these error bars.

4-2. Control device functional configuration example 2
Next, a control device 120 of a type that determines an optimal operation amount for a learning model of reinforcement learning will be described.

The control device 120 is programmed to calculate a reward based on the result of determining the connection condition (environment state transition), update an evaluation function based on the calculated reward, and determine a (more suitable) connection condition that will obtain the most reward by repeating the update of the evaluation function. As shown in FIG. 12, the control device 120 includes the functional units of an input unit 121, a calculation unit 122, a memory unit 123, and an output unit 124.

The input unit 121 acquires observation data related to FSW, similar to the input unit 21 described above. The memory unit 123 stores, for example, a learning model, a control program, observation data, setting information, etc. The output unit 124 indicates to the FSW device 10 the operation amount of the joining condition determined by the calculation unit 122.

The calculation unit 122 performs machine learning based on the observation data and determines optimal joining conditions. In other words, the calculation unit 122 plays a role equivalent to an agent in reinforcement learning. The calculation unit 122 includes a setting unit 125, a function update unit 126, a reward calculation unit 127, and a policy determination unit 128.

The setting unit 125 performs initial setting of the joining conditions based on user input, etc., and performs settings related to rewards (for example, initial values of rewards and value functions, conditions for granting rewards, etc.). In addition, when initially setting the joining conditions, the setting unit 125 may use a learned prediction model stored in the memory unit 123 or an external storage device to grasp joining conditions that satisfy the target joint characteristics, and set these conditions as the initial settings.

Function update unit 126 selects a maximum value function Q ^* for a possible action a by updating parameter vector w using, for example, the observation data and Equation 21, and updates value function Q'. Reward calculation unit 127 calculates a reward using the updated value function Q' based on the observed environmental state s. Policy determination unit 128 determines a policy π corresponding to the updated value function Q'.

The calculation unit 22 then repeats the observation and adjustment of the joining conditions at a preset time interval, and instructs the FSW device 10 on the determined measure, i.e., the amount of adjustment of the joining conditions, via the output unit 24.

Next, an example of the operation of the control device 120 will be described with reference to FIG.
First, in step S51, the setting unit 125 performs initial settings of joining conditions and the like in response to user input. Next, in step S52, the measure determination unit 128 determines the measure π based on the value function Q'. In addition, in step S53, the function update unit 126 updates the parameter vector w of the value function Q' using the learning model, thereby updating the value function Q'.

In step S54, the control unit 120 causes the FSW device 10 to execute the strategy π and obtains a new state s of the environment from the measuring instrument 30. In step S55, the reward calculation unit 127 calculates the reward using the observed new state s of the environment and the updated value function Q'.

The control device 20 then repeatedly executes steps S52 to S55 at set time intervals. This allows the connection conditions to be optimized according to the state of the environment. In other words, by repeating this process, the reliability of the action value function is increased, and by determining a strategy to obtain more rewards based on a highly reliable action value function, it becomes possible to more optimally determine the setting of the connection conditions.

In other words, the control device 20 can learn the relationship between various joining conditions and the joints obtained under those joining conditions through the concept of reward, and automatically determine appropriate joining conditions according to the joint characteristics desired by the user. Therefore, various joints can be easily obtained using optimal joining conditions without the need for numerous preliminary experiments and regardless of the skill level of the operator.

The above describes typical embodiments of the present invention, but the present invention is not limited to these, and various design modifications are possible, which are also included in the present invention.

1...Friction stir welding (FSW) system,
10...Friction stir welding (FSW) device,
20...Control device,
30...Measuring instrument.

Claims (9)

A trained model for causing a computer to function so as to output a predicted value of a joint characteristic based on input data indicating a joining condition of friction stir welding,
The neural network includes an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function;
the neural network learns weighting coefficients using information indicating a variance in experimental data indicating the joint characteristics so as to minimize a function represented by an error between the experimental data and the predicted value;
A trained model for causing a computer to function in such a way that, for input data indicating joining conditions for friction stir welding input to the input layer, calculations are performed based on the trained model, and a predicted value of joint characteristics is output from the output layer, along with information indicating the reliability of the predicted value. the neural network further learns the weighting coefficients using information indicating the variation of the weighting coefficients;
The predicted model according to claim 1 , which causes a computer to function to further output information indicating a degree of influence of the joining condition on the predicted value. The neural network comprises:
learning the weighting factors for each of a plurality of models having different numbers of hidden units;
The trained model according to claim 1 or 2, wherein a model among the plurality of models that has a minimum error between a predicted value and experimental data is set as a prediction model. The neural network comprises:
learning the weighting factors for each of a plurality of models having different numbers of hidden units;
Ranking the plurality of models based on the error between the predicted values and the experimental data;
3. The trained model according to claim 1 or 2, wherein, among a plurality of new models generated by sequentially combining models from a higher ranking, a model that has a minimum error between a predicted value and experimental data is set as a prediction model. An input unit for acquiring observation data indicating welding conditions in friction stir welding;
A storage unit that stores the trained model according to any one of claims 1 to 4;
A calculation unit that determines an operation amount of a friction stir welding apparatus based on a result of applying the observation data to the trained model; and
an output unit that instructs the friction stir welding apparatus on the determined operation amount;
A control device including: An input unit for acquiring observation data indicating welding conditions and joint characteristics in friction stir welding;
A storage unit that stores a predetermined learning algorithm;
a calculation unit that determines an operation amount of the friction stir welding apparatus based on a result of applying the observation data to the predetermined learning algorithm; and
an output unit that instructs the friction stir welding apparatus on the determined operation amount;
A control device comprising:
The predetermined learning algorithm is
The control device is caused to function so as to output an appropriate operation amount of the friction stir welding device based on input data indicating welding conditions of the friction stir welding device and the joint characteristics.
The neural network includes an input layer, an output layer, and a hidden layer that is interposed between the input layer and the output layer and is represented by a nonlinear activation function.
The neural network learns weighting coefficients so as to minimize an error function represented by a value function including a relationship between the joining conditions and the joint characteristics;
A control device that performs calculations based on a learned model on input data indicating welding conditions for friction stir welding input to the input layer, and outputs an operation amount of the friction stir welding device from the output layer. A friction stir welding apparatus for performing friction stir welding;
A measuring instrument for measuring observation data indicating welding conditions in the friction stir welding;
A control device according to claim 5 or 6;
A friction stir welding system comprising: A neural network system for generating a trained model for causing a computer to function to output a predicted value of a joint characteristic based on input data indicating a joining condition of friction stir welding,
The neural network includes an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function,
The neural network comprises:
learning a weighting factor so as to minimize a function represented by an error between the experimental data and the predicted value, using information indicating a variation in the experimental data indicating the joint characteristics;
A neural network system configured to generate a trained model that performs calculations based on the trained model on input data indicating welding conditions for friction stir welding input to the input layer, and outputs a predicted value of joint characteristics from the output layer as well as information indicating the reliability of the predicted value. A method for generating a trained model for causing a computer to function to output a predicted value of a joint characteristic based on input data indicating a joining condition of friction stir welding,
The learning model for generating the trained model includes a neural network including an input layer, an output layer, and a hidden layer interposed between the input layer and the output layer and having a plurality of hidden units represented by a nonlinear activation function;
a step of learning weighting coefficients in the neural network by using information indicating a variance in experimental data indicating the joint characteristics so as to minimize a function represented by an error between the experimental data and the predicted value;
A step of generating a trained model that performs a calculation based on the trained model on input data indicating welding conditions of friction stir welding input to the input layer, and outputs a predicted value of joint characteristics from the output layer and information indicating the reliability of the predicted value;
How to generate a trained model, including: