JPH06309017A - Numerical controller - Google Patents

Abstract

PURPOSE:To provide the numerical controller which can automatically adjust fuzzy inference part control rule parameters for the slow-motion correcting process of an object machine to be controlled. CONSTITUTION:The numerical controller which performs position control by estimating a slow-motion quantity by fuzzy inference from measurement data in machine operation is equipped with a fuzzy parameter learning part 40 which learns fuzzy parameters so that the square of the difference between slow-motion measurement data and an estimated value becomes minimum, and corrected fuzzy parameters are used for actual fuzzy inference. The fuzzy parameters which are learnt and updated are used as the center value, width, and consequent constant of a triangular membership function of simple fuzzy inference and slow-motion data measured at any time in the machine operation are difference data between a semiclosed loop current position and a full-closed loop current position right after semiclosed loop current position data are inverted.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a numerical control device for performing servo control, and more particularly to a numerical control device suitable for application to a control object such as a machine tool that requires an appropriate lost motion correction.

[0002]

2. Description of the Related Art In a machine tool to which a numerical control device is applied, it is always required to improve machining accuracy and productivity. Lost motion is generated in the feed drive mechanism mounted on the machine tool, which adversely affects the quality of the machined work. Therefore, measures for avoiding this have been considered. As shown in FIG. 9, a drive load system of a machine tool or the like to be controlled by the numerical controller includes system play, play, loose bearing (not shown) or bearing support, play at ball screw end, ball screw 8a and nut. Backlash in a broad sense such as play between the balls, expansion and contraction and twist of the ball screw 8a depending on the frictional force and rigidity of the system, axial elastic displacement of the nut, bearing displacement supporting the ball screw 8a, sliding between the bed 8c and the table 8b. There are so-called lost motion factors such as dynamic resistance. Among these, the elastic deformation of the ball screw 8a or the like changes depending on the position of the table 8b and the weight of the workpiece mounted on the table 8b, and the bed 8c and the table 8c.
The sliding resistance between b and the friction coefficient of the sliding surface greatly depend on the feed rate or the lubricating state of the sliding surface. FIG. 5 is a diagram showing the relationship between the lost motion amount and the feed speed Vel. Lost motion tends to decrease as the feed speed increases. Fig. 6 shows the relationship between the length Pos from the motor shaft end to the machine movable part (nut part) and the lost motion amount.
Shown in. It is known that elastic deformation such as "twist" or "deflection" of the ball screw 8a or the like is a function of the length from the drive portion to the nut portion position, and FIG.
It is a graph of the following equation. Further, FIG. 7 shows an example of the relationship between the lubrication state of the sliding surface and the amount of lost motion. Oil lubrication is applied to the sliding surfaces of machine tools to prevent seizure and to guide the movable parts stably. For this lubrication, parameters such as the lubrication interval time Tlint, which indicates the time interval of lubrication operation, and the lubrication time Tlon, which determines the time to actually eject the lubricating oil to the sliding surface by operating the shaft lubrication motor, are set in the numerical controller. By doing so, the timer is operated. As shown in Fig. 7, when the shaft lubrication motor operates and the lubricating oil is output to the sliding surface, the lost motion amount decreases, and after the lubricating oil output (lubrication) is stopped, the lost motion amount gradually increases. You can see it going. An example of a medium-sized NC lathe is shown in FIG.
Maximum value Lmax and minimum value L of the lost motion amount in
The difference Lw from min may become 10 μm or more. FIG. 7 shows that the lost motion amount depends on the above-mentioned lubrication interval time, lubrication time, elapsed time from the stop of lubrication, and the like.
It shows that it changes intricately.

In order to cope with the dynamic change of the lost motion amount as described above, fuzzy calculation is executed from the measured data of speed, position and shaft lubrication state which are measured at any time during machine operation, and the lost shaft is lost. A numerical control device has been devised which estimates a motion amount and calculates a lost motion correction amount from the estimated amount. In fuzzy control, if-th
Express with the en rule.

[Equation 1] In the i-th fuzzy inference process, the input X i = (x i
1, ..., Xir), i = 1, 2, ..., N, the estimated value output yi * by fuzzy inference is calculated as in Equation 2.

[Equation 2] In Expression 2, μ s (Xi), s = 1, ... R is the membership value of the antecedent part, and is given by Expression 3.

[Equation 3] Here, the simple fuzzy inference is performed by using f1, ..., Fr in the formulas 1 and 3 as constants.

FIGS. 8 to 14 show an example of a conventional numerical controller having a lost motion correction function using simplified fuzzy inference, and FIG. 9 shows a lost motion correction function using simplified fuzzy inference. It is a block diagram of a configuration of a numerical control device having. The operation of the conventional technique will be described below with reference to FIG. NC not shown in the figure
Based on the NC program 1 read and interpreted by the program reading unit and the NC program interpreting unit,
The function generator 2 calculates movement information (target position, speed, etc.) of the axis and transfers the position command to the axis controller 3. The axis control unit 3 includes a main control unit 31, a control program storage unit 32, an output unit 33, an input unit 34, a lost motion correction amount calculation unit 35, a control rule storage unit 36, a membership function storage unit 37, and a consequent part constant storage. It is composed of a section 38. The main control unit 31 uses the position command and the servo control program stored in the control program storage unit 32.
And a desired position, speed, and current control loops are calculated based on the position detection value of the motor position detector 7 that detects the rotor position of the servo motor 5 input via the input unit 34, and finally Is to transfer the current command value to the power amplifier 4 via the output unit 33. Power amplifier 4
Then, the current command value (P
From the WM command), PWM processing is performed by a known technique,
Each phase voltage to be applied to the servo motor 5 is generated. A drive torque is generated in the servomotor 5 by applying this voltage, and the drive load system including the ball screw 8a, the table 8b and the bed 8c is relayed through the coupling 6 to a desired position.
Drive at speed.

In the following description, each variable and mathematical expression will be described by omitting the subscript i indicating the i-th fuzzy inference process. In the embodiment of the prior art shown here, the following five variables are adopted as the input variable X of the fuzzy control rule. That is, X = (x1, x2, x3, x4, x5). Where x1: Feed speed Vel [mm / min] x2: Distance from motor shaft end to machine moving part Pos [m
m] x3: Lubrication interval time that represents the time interval of lubrication operation
Tlint [minutes] x4: Lubrication time that determines the time when the lubricating oil is actually jetted to the sliding surface Tlon [minutes] x5: Elapsed time from refueling stop Tidle [minutes] The control rule storage unit 36 stores R1 to A control rule for lost motion correction processing from the viewpoint of feed rate as shown in R3, a control rule for lost motion correction processing from the viewpoint of machine movable part position from the motor shaft end as shown at R4 to R6 in FIG. The control rules of the lost motion correction process from the viewpoint of the shaft lubrication state as indicated by R7 to Rr in FIG. 8 are stored. These rules are set empirically, and the antecedent part is a qualitative expression. As described above, there are five input variables in the control rule as a whole, but the number of input variables is three or less. That is, the input variable in the rules R1 to R3 is one variable X = (x1 = Vel), and the input variable in the rules R4 to R6 is one variable X = (x2 = Pos), the rule R
The input variables in 7 to Rr are X = (x3 = Tlint, x
4 = Tlon, x5 = Tidle). Accordingly, the antecedent part fuzzy set F in each rule
sj, s = 1, ... r, j = 1, ... 5 and antecedent part membership value μs (Xi), s = 1, ...
Specifically, r is given by Equation 4.

[Equation 4] In addition, the nonlinear equation f in the consequent part of each rule in FIG.
s, s = 1, ... r is a constant f in the simplified fuzzy inference
s, s = 1, ..., R, which are stored in the consequent part constant storage unit 38 as a constant table.

FIG. 10 shows the membership function storage 37.
~ Each membership function shown in Fig. 14 is stored. FIG. 10 shows membership functions PB (NB), PM (NM), PS (NS) for the feed rate Vel, and the membership function P shown in FIG. 10 (a).
B (NB) is a high-speed membership function, FIG.
The membership function PM (NM) shown in (b) is a medium speed membership function, and the membership function PS (NS) shown in FIG. 10 (c) is a low speed membership function. Next, FIG. 11 shows the membership functions LG, MD, ST with respect to the position of the machine moving part from the motor shaft end (the distance from the motor shaft end to the machine moving part) Pos, and FIG. The membership function LG shown is the membership function at a position far from the motor shaft end.
The membership function MD shown in (b) is a membership function at a middle position from the motor shaft end, and the membership function ST shown in FIG. 11 (c) is a membership function at a position close to the motor shaft end. FIG. 12 (a) shows the membership functions SI, MI, BI for the lubrication interval time Tlint representing the time interval of the lubrication operation, where the membership function SI is the membership function with a short lubrication interval time, The membership function MI is a membership function with a medium lubrication interval time, and the membership function BI is a membership function with a long lubrication interval time. Figure 1
3 (a) shows the membership functions SO, MO, BO for the lubrication time Tlon that determines the time for actually injecting lubricating oil to the sliding surface by operating the shaft lubrication motor. SO is a membership function with a short lubrication time, membership function MO is a membership function with a medium lubrication time, and membership function BO is a membership function with a long lubrication time. Further, FIG. 14A shows each membership function SD, MD, for the elapsed time Tidle after refueling stop.
The membership function SD is a membership function with a short elapsed time, the membership function MD is a membership function with a medium elapsed time, and the membership function BD is with a long elapsed time. Indicates the membership function.

Next, the lost motion correction processing operation of the apparatus configured as described above will be described. Now, it is assumed that the movement command required for the lost motion correction process is transferred from the function generation unit 2 to the main control unit 31 in the axis control unit 3. At this time, the main control unit 31 obtains the current moving direction and the current feed speed from the previous moving command and the current moving command, and transfers these information to the lost motion correction amount calculating unit 35. Further, the current position data of the movable portion obtained by the motor position detector 7 is inputted via the input unit 34 to the above-mentioned lubrication interval time / lubrication time and the elapsed time from the suspension of the lubrication oil, which is not shown in the figure. The data sequentially updated by the timer is transferred to the lost motion correction amount calculation unit 35. Upon receiving these data, the lost motion correction amount calculation unit 35 first obtains a membership value from each membership function corresponding to each control rule antecedent unit in FIG. That is, the control rules R1 to R3 in FIG.
Membership values corresponding to the antecedent part of μ 1, μ 2, μ 3
From the membership functions shown in FIGS. 10A, 10B, and 10C, the membership values μ4, μ5, and μ6 corresponding to the antecedent parts of the control rules R4 to R6 in FIG.
It is calculated from the membership functions shown in (a), (b) and (c) according to each input variable. Further, the membership values μ7 to μr corresponding to the antecedents of the control rules R7 to Rr in FIG. 8 are μs = μs3 × μs4 × μs5, s = 7, ...
, R, it is necessary to find each membership value μs3, μs4, μs5 from each membership function. For example, the membership value μ7 corresponding to the antecedent part of the control rule R7 in FIG. 8 is the membership value μ73 for the lubrication interval time from the membership function shown in FIG. 12 (b), and the membership value μ7 shown in FIG. 13 (b). Membership value from ship function to lubrication time μ74
From the membership function shown in FIG. 14 (b), the membership value μ75 for the elapsed time from the suspension of lubrication is determined, and the product of these is calculated. When each membership value is calculated in this way, the consequent part constant fs s = of each control rule shown in FIG.
1, ..., R are read from the consequent part constant storage unit 38, and the lost motion amount estimated value yi * in the i-th process is obtained based on the mathematical expression 2. When the estimated lost motion amount yi * is obtained as described above, the lost motion correction amount calculation unit 35 multiplies the correction coefficient as necessary to determine the final lost motion correction amount Lca. The lost motion correction amount calculation unit 35 determines this correction amount data Lca.
Is transmitted to the main control unit 31, and the main control unit 31 adds the lost motion correction amount Lca to the position command data at a predetermined timing in the position control loop to correct the moving amount of the movable part.

[0008]

There is a tuning problem in fuzzy control. Tuning is the work of modifying and reconfiguring fuzzy inference rules so that the difference between the inference value and the output data is minimized after the production rules are constructed. Specifically, it is the work of redefining the shape of the membership function in the control rule, the non-linear equation of the consequent part (a constant in this embodiment), etc. according to the control target. The absolute value of the lost motion amount of a drive control target such as a machine tool largely and complicatedly depends on various factors such as the physical size of the machine drive unit, the guide system, the feed rate, and the lubrication condition of the guide surface. Of course, if the specifications of the drive control target machine change, even if the machines of the same specifications change the machine base, or even if the machine environment changes the installation environment and drive conditions, the same machine Even if the installation environment and driving conditions are the same on the table, the amount of lost motion changes intricately with time. Specifically, the amount of lost motion varies, for example, in the range of about 15 to 35 μm in a medium-sized lathe, and in the range of about 40 to 100 μm in a slightly large horizontal machining center. Therefore, at the stage of shipping the machine equipped with the numerical control device according to the related art, the central value and width of the membership function shown in FIGS. 10 to 14 are specifically determined based on the measured data, and the membership function storage unit is stored. Must be set to 37. Further, similarly, the consequent part constant in FIG. 8 also needs to be specifically determined based on the actual measurement data and set in the consequent part constant storage unit 38. The analysis of the actual measurement data and the determination of each fuzzy parameter based on the analysis are manual work of the designer, and there is a problem that considerable man-hours must be generated and dealt with each time the machine specifications change. . Furthermore, even if the machines have the same specifications as described above, even if the machine is different, the amount of lost motion will be different, and even if the end user thinks about changing the fuzzy parameters in order to respond to changes over time, means are available. There is a problem that the lost motion correction process may not work effectively because it cannot be dealt with practically because it is not present. Even if the designer of the manufacturer tries to do this, it is practically impossible because the lost motion measurement data that is the basis of the fuzzy parameter determination of the controlled machine is not stored.

[0009]

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and in the numerical controller according to claim 1, the lost of the shaft concerned is obtained from the data measured at any time during machine operation. In a numerical control device that estimates a motion amount using fuzzy reasoning and calculates a lost motion correction amount from the lost motion amount estimated value, the lost motion data and the lost motion data that are measured at any time during machine operation are used. Equipped with a fuzzy parameter learning unit that learns the fuzzy parameter in the fuzzy control rule based on the difference from the motion amount estimated value, and uses the fuzzy parameter modified by the fuzzy parameter learning unit in actual fuzzy reasoning. It is characterized by doing so. Further, in the numerical controller according to claim 2, the fuzzy parameters to be learned / updated are the central value of the triangular membership function in the simplified fuzzy inference, its width and the consequent constant, and the lost motion measured at any time during machine operation. The data is the difference data between the semi-closed loop current position and the full closed loop current position immediately after the inversion of the semi-closed loop current position data, and the fuzzy parameters are learned by the error backpropagation learning model. is there.

As described above, in the numerical control device according to the first and second aspects, since the fuzzy parameters are learned and automatically updated based on the lost motion measurement data, it is necessary to generate the fuzzy parameters for each machine specification each time. Without
Optimal fuzzy parameters can be quickly determined.

[0010]

1 to 4 show an embodiment of a numerical control device according to the present invention, in which the functions and processing contents of the components indicated by the same numbers as those in FIG. Since they are the same, the description thereof will be omitted, and only the essential part of the present invention will be described below. In the present invention, the previously described lost motion correction processing operation in the prior art is interpreted as the forward processing operation of the neural network as follows. That is, the lost motion correction processing function of the conventional technique is as shown in FIG. 3, which is an input layer 1 layer (the number of neurons = 5), an intermediate layer 1 layer (the number of neurons = r), an output layer 1 layer (the number of neurons = 1). It is considered to be a neural network with the configuration of. Each neuron in the input layer has an input variable vector X = (x1, x2, x3, x4, x5;
x1 = feed speed Vel [mm / min], x2 = distance Pos [mm] from motor shaft end to machine moving part, x3 = lubrication interval time Tlint [min], which represents the time interval of lubrication operation,
x4 = Lubrication time Tlon [minutes] that determines the time when the lubricating oil is actually jetted to the sliding surface, x5 = Elapsed time T from the stop of lubrication
Each element variable of idle [minute]) is input, and the fuzzy control rule antecedent part fuzzy set matching degree in FIG. 8 for the input variable is transferred as an output to each neuron in the intermediate layer. In each neuron of the intermediate layer, the membership value μs related to the antecedent part of the fuzzy control rule in FIG.
(Xi), s = 1, ..., R are calculated according to Equation 4, and output to the output layer neuron together with the fuzzy control rule consequent part fs, s = 1 ,. Transfer to neurons. In the output layer neuron,
The lost motion amount estimated value yi * in the i-th fuzzy inference process is obtained by substituting the membership value and the consequent part constant for the antecedent part of each fuzzy control rule in FIG. Is calculated. In other words, the lost motion correction process using fuzzy inference is the forward process of the neural network shown in FIG. 3, and the coupling coefficient between the input layer and the intermediate layer is the input variable and the fuzzy control rule antecedent part fuzzy. The degree of agreement with the set, that is, the membership function,
The coupling coefficient between the hidden layer and the output layer can be considered to be the consequent constant. Therefore, in order to perform the lost motion amount estimation processing so as to bring about an effect by adapting it to the mechanical specifications, operating conditions, secular changes, etc. of the machine base, the input layer and the intermediate layer and the intermediate layer and the output in FIG. It suffices to prepare means for tuning the coupling coefficient between layers according to the actual condition of the controlled machine.

FIG. 1 is a block diagram showing the configuration of a numerical controller according to the present invention. In FIG. 1, a component different from the prior art is a fuzzy parameter learning section 40, in which the input layer, the intermediate layer and the intermediate layer shown in FIG. The coupling coefficient between the intermediate layer and the output layer is learned. As parameters for specifically determining the triangular membership function of the antecedent part of the fuzzy control rule, attention is paid to the central value asj and its width bsj of the triangular membership function shown in FIG. In the fuzzy parameter learning unit 40, the central value asj and the width b
sj, s = 1, ..., r, j = 1, ..., 5 and the consequent constants fs, s = 1 ,.
The r is adjusted by the error backpropagation learning model using the steepest descent method. However, this adjustment is an estimated value yi * of fuzzy inference when the input data Xi in the i-th process is input.
And the corresponding lost motion measurement data yi, that is, the estimation error E is calculated as in Equation 5, and is minimized.

[Equation 5] That is, the teacher data for update calculation in the neural network of FIG. 3 is the lost motion measurement data yi. Here, lost motion data y to be measured at any time
For i, the lost motion correction amount Lca calculated based on the i-th estimated lost motion amount yi * is added to the position command data at a predetermined timing in the position control loop in the main control unit 31, and this is added to the speed control loop. Difference data between the current position of the semi-closed loop and the current position of the full-closed loop at the timing when the data of the current position of the semi-closed loop is first inverted after the time when it becomes valid as an input. The vector indicating the direction in which the estimation error E decreases most is expressed by Equation 6.

[Equation 6] FIG. 4 shows a procedure for learning and updating fuzzy parameters in the fuzzy parameter learning unit 40. The procedure for learning and updating the fuzzy parameters in the fuzzy parameter learning unit 40 will be described below with reference to FIG. First, i
Input data Xi = (x required for calculating the estimated lost motion amount yi * in the th fuzzy inference process
1, x2, x3, x4, x5; x1 = feed rate Vel [m
m / min], x2 = distance from motor shaft end to moving part of machine
Pos [mm], x3 = Lubrication interval time Tlint [minutes], which represents the time interval of lubrication operation, x4 = Lubrication time Tlon that determines the time when the lubricating oil is actually jetted to the sliding surface.
[Minutes], x5 = elapsed time since refueling stop Tidle [minutes])
Each element variable of is taken in (step S1). Next, the i-th fuzzy inference process is performed based on the information stored in the control rule storage unit 36, the membership function storage unit 37, and the consequent part constant storage unit 38 using the input variables fetched in step S1. Lost motion estimation value y at
i * is calculated (step S2). Next, the fuzzy control rule consequent part fs is calculated by using the formula 7 by the steepest descent method.
, S = 1, ..., R are updated, and the consequent part constant storage unit 3 is updated.
Set to 8.

[Equation 7] Gf in Equation 7 is a learning coefficient, and the superscript t is the t-th
This indicates that the data is obtained by the second update (step S3). Furthermore, the central value asj of each fuzzy control rule antecedent triangular membership function and its width bsj, s = 1, ...
= 1, ..., 5 are updated and set in the membership function storage unit 37.

[Equation 8] Ga and Gb in Equation 8 are learning coefficients, and the superscript t
Indicates that the data is obtained by the t-th update (step S4). When the fuzzy parameter learning update and the resetting to each storage unit are executed by the fuzzy parameter learning unit 40 as described above, in the lost motion amount estimation process thereafter, the lost motion of the actual controlled machine is lost. The estimation process is performed with the fuzzy parameters updated to suit the situation of occurrence. If the temporal factor that determines the lost motion amount of the controlled machine base is, for example, one that is related to the lubrication state of the guide surface, it changes in the order of minutes to hours as shown in FIG. The change in the learning factor of the fuzzy parameters and the resetting process to each storage unit may be executed in a cycle of several tens of seconds to several tens of minutes because the change of the dynamic factor is slow, and the CPU processing load per unit time is large. It does not affect.

The present invention is not limited to the embodiments of the present invention shown in FIGS. 1 to 4 described above, and various variations should be considered without departing from the scope of the invention. For example, the learning update of the fuzzy parameters executed by the fuzzy parameter learning unit 40 and the resetting process to each storage unit are not always executed every time, but before the execution of step S1 in FIG. It is checked whether the difference between the value of Equation 5 and the value of Equation t-1 time before is within a certain value ε as shown in Equation 9 below. If Equation 9 is satisfied, A numerical control device that does not perform learning and updating of fuzzy parameters and resetting processing to each storage unit is conceivable.

[Equation 9]

[0013]

As described above, in the numerical control device according to claims 1 and 2, analysis of measured data such as determination of the central value and width of the membership function or determination of the consequent constant of the fuzzy control rule. And, it is possible to solve the problem that a considerable number of man-hours must be generated and dealt with each time the machine specifications are changed in determining each fuzzy parameter based on the above. Furthermore, the fuzzy parameters are learned based on the lost motion measurement data that is the basis of the fuzzy parameter determination of the controlled machine base, and it is automatically updated when necessary.
Even if the machines have the same specifications, there is an excellent effect that it is possible to deal with the problem that the amount of lost motion is different if the machine bases are different, and to cope with changes over time quickly and without the need for man-hours for human adjustment.

[Brief description of drawings]

FIG. 1 is a configuration block diagram of a numerical controller according to an embodiment of the present invention.

FIG. 2 is a diagram showing a triangular membership function.

FIG. 3 is a diagram illustrating a lost motion amount estimation calculation and a fuzzy parameter learning calculation by a neural network.

FIG. 4 is a diagram showing a procedure for learning and updating a fuzzy control rule consequent part constant and a central value and a width of a triangular membership function by the steepest descent method.

FIG. 5 is a diagram showing a relationship between a feed speed and a lost motion amount.

FIG. 6 is a diagram showing a relationship between a position of a mechanical moving part from a motor shaft end and a lost motion amount.

FIG. 7 is a diagram showing a relationship between a lubrication state and a lost motion amount.

FIG. 8 is an explanatory diagram showing a fuzzy control rule.

FIG. 9 is a configuration block diagram of an embodiment of a numerical controller according to the prior art.

FIG. 10 is a diagram showing each membership function regarding a feed rate.

FIG. 11 is a diagram showing each membership function with respect to the position of the mechanical moving part from the motor shaft end.

FIG. 12 is a diagram showing each membership function with respect to a lubrication interval time.

FIG. 13 is a diagram showing each membership function with respect to lubrication time.

FIG. 14 is a diagram showing each membership function with respect to the elapsed time from the stop of refueling.

[Explanation of symbols]

 1 NC Program 2 Function Generation Unit 3 Axis Control Unit 4 Power Amplifier 5 Servo Motor 6 Coupling 7 Motor Position Detector 8a Ball Screw 8b Table 8c Bed 31 Main Control Unit 32 Control Program Storage Unit 33 Output Unit 34 Input Unit 35 Lost Motion Correction Quantity calculation unit 36 Control rule storage unit 37 Membership function storage unit 38 Consequent part constant storage unit 40 Fuzzy parameter learning unit

Claims (2)

[Claims] 1. A desired position control is performed by estimating the lost motion amount of the relevant axis using fuzzy reasoning from data measured at any time during machine operation and calculating the lost motion correction amount from the estimated lost motion amount value. In the numerical control device to be performed, a difference between the lost motion data measured at any time during machine operation and the estimated value of the lost motion is obtained, and a fuzzy parameter learning unit for learning a fuzzy parameter in a fuzzy control rule based on the difference is provided. , A numerical controller characterized in that the fuzzy parameters modified by the fuzzy parameter learning unit are used in actual fuzzy inference. 2. The learned and updated fuzzy parameters are the central value of the triangular membership function in simplified fuzzy inference, its width and the consequent constant, and the lost motion data measured at any time during machine operation is the semi-closed loop current position. 2. The numerical controller according to claim 1, wherein the fuzzy parameter is learned by an error backpropagation learning model by using difference data between the current position of the semi-closed loop and the current position of the full closed loop immediately after data inversion. .