JPH0349845A - Control device adapted to cutting work - Google Patents

Abstract

PURPOSE:To obtain a control device, adapted to cutting work with reliability of a suitable control enhanced, by obtaining a plurality of machining condition signals for showing a condition of the cutting work and synthetically deciding these signals by a neural network. CONSTITUTION:A cutting dynamometer 11, which detects cutting resistance applied to a tool 8, and an AE sensor 10, which detects AE (acoustic emission) generated during machining, are mounted to a tool rest 7 respectively with an output of the cutting dynamometer 11 given to a cutting resistance signal processing part 21 and an AE signal, detected by the AE sensor 10, given to an AE signal processing part 22, here the character of each signal is extracted and fed to a neural network 23, and an optimum condition of rotation in a spindle motor 14 and a feed shaft motor 6 is calculated. A result of the calculation is given to a control command output circuit 24 and converted into an override amount of main and feed shafts given to an NC unit 25, and a speed of the spindle motor 14 and the feed shaft motor 6 is controlled.

Description

DETAILED DESCRIPTION OF THE INVENTION Summary of the Invention A cutting process adaptive control device that processes output signals of a plurality of sensors that detect the state of cutting processes using a neural network and determines process conditions. It becomes possible to optimally control machining in order to improve machining efficiency, machining accuracy, and reduce production costs. Also.

Adaptive control algorithms can be easily created by learning neural networks.

Background of the invention The present invention relates to a cutting process adaptive control device.

In conventional machine tools that perform cutting.

In order to improve machining efficiency and precision and reduce production costs, some machines are equipped with an adaptive control system that controls machining conditions to obtain predetermined results. In conventional systems, any one of machining conditions such as cutting power, cutting torque, or cutting resistance is measured, and the machining conditions are controlled so that these values match predetermined constraint conditions. However, in conventional systems, there is only a single type of measurement value that represents the machining state, and with this single measurement value, it is not possible to accurately grasp the machining conditions so as to meet predetermined constraint conditions. Therefore, there is a problem in that the reliability of adaptive control becomes low depending on the conditions.

Summary of the Invention An object of the present invention is to obtain a plurality of machining status signals representing the machining status, and to comprehensively judge these status signals using a neural network, thereby improving the reliability of adaptive control. An object of the present invention is to provide an adaptive control device.

The cutting processing adaptive control device according to the present invention is obtained from a plurality of processing state measuring means each detecting a plurality of types of cutting processing states in a machine tool that performs cutting according to set processing conditions, and a plurality of processing state measuring means. Neural network means configured to input a plurality of machining status signals and calculate machining condition control commands through learning, and control for outputting machining condition control commands based on output signals from the neural network means. It is characterized by comprising a command output means.

The plurality of machining state measuring means include, for example, an AE detecting means for detecting an AE multiplied signal generated during cutting, a cutting resistance detecting means for measuring cutting resistance applied to a tool, and a measuring means for measuring current or voltage of a motor of a machine tool. power consumption measurement means,
and a surface roughness measuring means for measuring the state of surface roughness of the work being processed.

According to the present invention, it is possible to comprehensively monitor cutting conditions, and it is possible to optimally control machining in order to improve machining efficiency, machining accuracy, and reduce production costs. Furthermore, even if the relationship between the plurality of signals from the machining state measuring means and the actual machining state is not known, an algorithm for determining machining conditions can be created using the learning function of the neural network. The learning function of the binary network allows the algorithm for determining machining conditions to be modified depending on the usage environment.

DESCRIPTION OF EMBODIMENTS FIG. 1 shows the overall configuration of a cutting process adaptive control system.

A frame 2 of the NC lathe 1 is rotatably equipped with a main shaft 3, and the main shaft 3 is rotationally driven by a main shaft motor 14. A workpiece 4, which is a workpiece, is attached to the main shaft 3. A movable table 5 is provided on a horizontal base 2A fixed to the frame 2 so as to be movable in two directions, and this movable table 5 is driven to move by a feed shaft motor 6. A cutting tool (tool) 8 is mounted on the movable table 5 via a tool mount 7. Cutting control of the cutting tool 8 is performed by a motor 9 provided on the movable table 5. Main shaft 3 by motor 14
is rotationally driven, and the moving table 5 (tool mounting table 7) is transferred in the Z direction by the motor 6, thereby moving the workpiece 4.
Cutting is performed using the cutting tool 8.

The tool mount 7 is equipped with a cutting dynamometer 11 that detects the cutting resistance applied to the tool (bite 8), and a cutting dynamometer 11 that detects the cutting resistance that is generated during machining.
An AE sensor 10 for detecting (Acoustic Emission) is attached. Cutting dynamometer 11
is, for example, a piezoelectric strain detector, which detects strain in the tool mount 7 caused by cutting resistance during cutting. The output of this cutting dynamometer 11 is given to a cutting resistance signal processing section 21, where features of the cutting resistance signal are extracted as will be detailed later, and signals representing the features are sent to a neural network 23.

The AE multiplied signal detected by the AE sensor 10 is given to an AE signal processing section 22, where features of the AE multiplied signal are extracted, and a signal representing the extracted features is sent to a neural network 23.

Based on the input signal, the neural network 23 calculates the optimal state of rotation of the main shaft motor 14 and the rotation of the feed shaft motor 6, as will be described in detail later, and the results are given to the control command output circuit 24. This circuit 24 converts it into override amounts for the main axis and feed axis.

The override amount of the main axis and feed axis is given to the NC control device 25.

Based on this, the NC control device 25 controls the rotational speed of the main shaft motor 14 and the feed shaft motor 6.

FIG. 2 shows the cutting resistance signal processing section 21. AE signal processing section 22
and shows the configuration of the neural network 23.

The output of cutting dynamometer 11 is amplified by amplifier 30 and provided to band pass filters 31 and 32, respectively. Band pass filters 31 and 32 are each
For example, it is a band pass filter whose pass frequency band is set at 20 Hz and IKHz, respectively, and has the function of extracting characteristic fluctuation components of the cutting force applied to the cutting tool 8 and removing noise components. The output signals of the band pass filters 31 and 32 are each converted to a corresponding DC level by a detection circuit 85 and 36 and sent to the input layer 51 of the neural network 23.

The output of the AE sensor IO is amplified by an amplifier 40 and sent to band pass filters 41 and 42. band·
Pass filters 41 and 42 each have, for example, 5
A band pass filter with passband center frequencies of 0KHz and 200KHz, which eliminates the bite (
Tools) 8. Extraction of characteristic AE frequency components generated from the work 4 and removal of noise components are performed. The outputs of the band pass filters 41 and 42 are each converted to a DC level by a detection circuit 45 and 46, and sent to the input layer 51 of the neural network 23.

A neural network connects multiple operators (neurons) to each other using weighted functions (synapses), and transmits the calculated values performed by the neurons on input patterns through the synapses to produce a conclusion. This is a calculation method for obtaining cover turns. This neural network.

Since it has the ability to independently determine the weighting of each synapse by learning the input pattern and the output pattern to be obtained from the input pattern a finite number of times, it can be used for processes that are difficult to analyze using algorithms.

In FIG. 2, neural network 23 has input layers 51 . The multilayer structure includes intermediate layers 52 and output layers 53, and the number of intermediate layers 52 may be one or more and any number. Each layer (excluding the output layer 53) is composed of a plurality of units.

The output layer 53 is composed of one unit. Here, the unit includes the above-mentioned neurons and synapses that connect them. Also, neural network 2
3 is provided with a teacher signal output device 54 for learning, which will be described later. Neural network 23 is generally realized by digital or computer software.

FIG. 3 shows the structure of the unit, and FIG. 4 shows the input/output characteristics of the unit.

A unit consists of a part that receives input from other units, a part that converts the input according to certain rules, and a part that outputs the result. A variable weight Wlj (the above-mentioned synapse) is attached to each connection portion with other units, and by changing the value of this weight through learning or the like, the structure of the network changes. Changing the structure of the network means changing the signal processing performed there.

A certain unit i is a plurality of units 1. ... j.

..., when input Oj (j-1 to k) is received from k, the weighting is based on the sum net 1 and the output oj
occurs. This total net j is.

netl-E wljOj It is expressed as

The output O9 of the unit is the sum n e obtained from the above formula.
t f is applied to the function f, (sigmoid function) shown in FIG. 4, resulting in a converted value (which is a value between 0 and 1).

The neural network 23 inputs input layers 51 . Middle layer 52. The above-mentioned calculations are sequentially performed on the output layer 53, and the optimum control amount of the rotational speed of the main shaft motor 4 and feed shaft motor 6 of the NC lathe 1 is output as a calculation result from the output layer 53.

5 and 6 show the output A from the output layer 53.

It shows the relationship between B and the output of the control command output circuit 24.

As shown in FIG. 5, in response to the spindle rotation calculation value output A output from the neural network 23, the control command output circuit 24 outputs a spindle rotation override (Over r
ide) it is output between -50% and 50%.

The override amount is a percentage of the rotational speed to be increased or decreased, and when the override amount is positive and large, the rotational speed is controlled to become faster, and when it is negative, the rotational speed is slowed down. When the spindle rotation calculation value output A is 0.9 or more, cutting is not being performed normally, and it is difficult to return to normal cutting using override control.

A cutting abnormality signal is output from the control command output circuit 24.

Allows you to perform operations such as canceling processing.

FIG. 6 shows the control command output circuit 2 in response to the feed axis rotation calculation value output B output from the neural network 23.
4 shows the feed axis override amount output from 4.

The signal processing algorithm of the neural network 23 is formed by learning. Learning is by back propagation. Cutting resistance signal processing section 21 and A
The value output from the neural network 23 when a certain input is given from the E signal processing section 22, and the desired a value given from the teacher signal output device 54 for that input (this is set in advance to the teacher signal output device 54). Learning is performed by comparing the given values and changing the weight between each unit so that the difference between them is reduced.

By controlling the main spindle and feed axis as described above, it is possible to perform high-speed cutting without damaging the tool, and it is possible to reduce production costs and improve processing efficiency.

In FIG. 2, the band pass and
Filter 31.32.41.42. Detection circuit 35.36
45 and 48 can be realized by analog circuits or digital circuits.

In addition, the AE signal processing unit converts the AE multiplied signal into digital data and then performs fast Fourier transform to create a frequency spectrum, and calculates the area for a predetermined frequency band on this spectrum to calculate the AE multiplied signal characteristics. It can also be configured to extract. The cutting resistance signal processing section can also be configured in the same way.

Next, a specific example of learning for forming a neural network that performs cutting process adaptive control will be explained. Here, only the control of the feed axis will be described.

AE signal processing section 22. A more specific configuration example of the cutting resistance signal processing unit 21 and the neural network 23 is shown in FIG. In these figures, the same components as those shown in FIG. 2 are given the same reference numerals.

The signal from the AE sensor IO is amplified by an amplifier 40,
The signal is sent to band pass filters 41A to 44A. Band/bass filters 4LA, 42A.

The passbands of 43A and 44A are 0 to 100, respectively.
KHz, too ~200 KHz, 200~30
0 KHz.

and 300-4QQKHz. The outputs of the band pass filters 41A to 44A are sent to the detection circuits 45A to 411.
1A, and the average amplitude of the AE multiplied signal in each frequency band is detected. The outputs of the detection circuits 45A-48A are sent to normalization circuits 61-64, respectively. Here, the signal level is normalized in order to process the signal in the neural network 23 at the next stage. Normalization is performed so that the average of the data input to the neural network 23 becomes zero.

The cutting resistance signal processing section 21 has the same configuration as the AE signal processing section 22. The pass bands of band pass filters 31 and 32 are 0 to 250 Hz and 250 to 50 Hz, respectively.
0 Hz.

The neural network 23 has input layer 6 units U
1-U6. It has a three-layer structure including two intermediate layer units Ull and U12 and one output layer unit U21. The output layer 53 (unit U21) outputs the feed shaft rotation calculation value output B.

The output function (input/output characteristic) of each unit is shown in FIG. 8, and the relationship between the output value B of the unit of the output layer 53 and the override amount is shown in FIG.

Learning is performed by inputting the neural network 23 to the outputs of the normalization circuits 61, 82, 83, 64, 37 and 38)
This is done by giving the desired feed axis rotation calculation value at that time to the neural network 23 from the teacher signal output device 54. In this example, 15 pair patterns of input signals and teacher signals were used. These data are shown below. The "input signal pattern" is the units Ul, U2, U3, etc. of the input layer 51. U4. The outputs of normalization circuits 61, 62, 63, 64, 37 and 3B (i.e. neural network 23) to be applied to U5 and U6
) are arranged in this order.

Input signal pattern 1 1.191429. 1.991429゜4.6914
29.3.891429 Teacher signal output 1 0.5 4J91429. 5.391429° input signal pattern 2 -0.408571.

-0,808571° -1,308571° -1,108571° 1.391429° -1,108571 Teacher signal output 2 Input signal pattern 3 -2.608571.

-3°108571.

-2,408571° -1,608571.

-1,108571° −1,6([571 Teacher signal output 3 -0.3 Input signal pattern 4 -0.608571° -0,808571° -1,408571° -1,008571.

1.591429°-1,208571 Teacher signal output 4 Input signal pattern 5 -2.708571°-3,208571°-2,508571°-1,508571°-1,108571,-1,508571 Teacher signal output 5 - 〇, 3 Input signal pattern 6 -2.408571. -3.208571° -1,30
8571, -1, 808571 Teacher signal output 6 -0,3 -2.408571° -1,608571° Input signal pattern 7 0.891429. 1.891429゜4.3914
29.3.891429 Teacher signal output cuff 0.5 4.391429° 5.391429° Input signal pattern 8 -0.608571. -1.108571゜1.391
429. -1.108571 IJ signal output 8 -1.608571.

-1,108571° Input signal pattern 9 1.291429. 1.791429゜4.6914
29.3.791429 Teacher signal output 9 0.5 4.591429° 5.191429° Input signal pattern l0 -0.308571. -1.408571゜1.591
429. -1.008571 Teacher signal output 10 -1.308571° -1,108571° Input signal pattern 11 -2.608571. -3.208571° -1,40
8571, -1, 908571 Teacher signal output 11 -0.3 -2.508571° -1,508571° Input signal pattern 12 1.191429° 1.791429° 4.591429° 5.591429° 4.591429° 3. 791429 Teacher signal output 12 0.5 Input signal pattern 13 -0.308571°-1,108571°-1,108571°-1,108571°1.591429°-1,2 (18571 Teacher signal output 13 Input signal pattern 14 0.891429゜1.891429゜4.891429゜5.491429゜4.391429゜3J91429 Teacher signal output 14 0.5 Input signal pattern 15 -2.908571゜-3,208571゜-2,808571゜-1, 8H571°-1, 208571, -1, 708571 Teacher signal output 15 -0.3 Set the weight between each unit so that the output of the neural network has a value close to the teacher signal output for all input signal patterns. Correction is performed by back propagation. Learning is performed for all 15 sets of patterns until the difference between the output of the neural network and the output of the teacher signal becomes less than a certain value.

The weight before learning and the weight after learning between each unit are shown in FIG. 10a and FIG. 10b, respectively. The weight before learning is -0,
It was given as a random number between 3 and 0.3.

FIG. 11 shows another example of the configuration of the cutting process adaptive control system.

In the system shown in FIG. 1, the cutting dynamometer 11 is expensive and requires modification of the tool mount 7 for installation. Furthermore, when the cutting dynamometer 11 is attached, there is a problem that the rigidity of the tool system decreases, and in some cases, accurate cutting cannot be performed.

FIG. 11 shows a system that solves the above problems. A power consumption signal processing section 2B is provided in place of the cutting dynamometer 11 and the cutting resistance signal processing section 21, and the load applied to the spindle motor is detected by observing the fluctuation in power consumption.

FIG. 12 shows the configuration of the power consumption signal processing section 2B.

The power consumption meter 71 is connected to the power supply 70 of the spindle motor I4, and detects the power consumption of the spindle motor 14. The output of the power consumption meter 71 is processed by a low-pass filter 72 for removing noise components and a detection circuit 73 for converting the output of the filter 72 into a corresponding DC level, and then sent to the input layer 51 of the neural network 23. given to. Other configurations and processing are similar to the cutting process adaptive control system shown in FIGS. 1 to 6.

In the manner described above, it is possible to more easily reduce production costs and improve processing efficiency.

FIG. 13 shows still another example of the cutting process adaptive control system.

In this example, an optical fiber probe 12 for detecting the roughness state of the surface of the workpiece 4 is installed on the tool mount 7. Based on the optical signal obtained from the optical fiber probe 12, the characteristics of the roughness state of the workpiece surface are extracted by the surface roughness signal processing section 27, and the output thereof is given to the neural network 23.

FIG. 14 shows a specific configuration of the surface roughness signal processing section 27. The optical fiber probe 12 is made up of a plurality of optical fibers, some of which are used for transmitting light and others for receiving light. The photonic sensor 81 is equipped with a white light source, and the work surface is irradiated with white light through a light-emitting fiber, and the reflected light is detected by a phototransistor in the photonic ψ sensor 81 through a light-receiving fiber. . Dynamic components and high frequency noise are removed from the output by a low pass filter 82 to obtain a signal representing the surface roughness state of the workpiece, which is applied to the neural network 23.

The neural network 23 includes the AE signal processing unit 21
．． Power consumption signal processing section 26 and surface roughness signal processing section 2
Processing is performed based on the output signal of 7.

The output from the output layer 53 of the neural network 23 is the spindle rotation calculation value and the feed axis rotation calculation value, as in the above embodiment. In the cutting process adaptive control system of this embodiment, since a signal representing the surface condition of the workpiece is also added as an input to the neural network 23, it is possible to control the cutting process in consideration of the processing quality of the workpiece. In other words, it is possible to cut as quickly as possible without damaging the tool, and to create the desired product.

[Brief explanation of drawings]

1 to 6 show embodiments of the present invention, FIG. 1 shows the overall configuration of the cutting process adaptive control system, and FIG. 2 shows the cutting resistance signal processing section and AE signal processing section in FIG. 1. Figure 3 is a diagram showing the structure of the unit in the neural network, Figure 4 is a graph showing the input/output characteristics of the unit, Figures 5 and 6 are the main axes, respectively. 2 is a graph showing the relationship between the output of one neural network and the output of the control command output circuit for rotation calculation values and feed axis rotation calculation values. Figures 7 to 10b show concrete examples based on simulations. Figure 7 is a block diagram showing details of the cutting resistance signal processing unit, AE signal processing unit, and neural network, and Figure 8 is a block diagram showing the details of the cutting resistance signal processing unit, AE signal processing unit, and neural network. A graph showing the input/output characteristics of the neural network unit used, FIG. 9 is a graph showing the relationship between the output of the neural network and the output of the control command output circuit, and FIGS. It shows an example of weights before and after learning in a neural network. 11 and 12 show other embodiments, FIG. 11 is a block diagram showing the overall configuration of the cutting process adaptive control system, and FIG. 12 is a block diagram showing the configuration of the power consumption signal processing section. be. Figures 13 and 14 show still other embodiments, with Figure 13 being a block diagram showing the overall configuration of the cutting process adaptive control system, and Figure 14 being a block diagram showing the configuration of the surface roughness signal processing section. It is a diagram. 1...NC lathe. 6...Feed axis motor. 8...Bite (tool). lO...AE sensor. 11...Cutting dynamometer. 12... Optical fiber 14... Main shaft motor. 21... Cutting resistance signal processing section. 22...AE signal processing section. 23... Neural network. 24...Control command output circuit. 25...NC control device. 26...Power consumption signal processing section. 27...Surface roughness signal processing section. 71... Power consumption meter. probe. 81...Photoniraku sensor. that's all

Claims (2)

[Claims] (1) Input a plurality of machining state measurement means that respectively detect a plurality of types of cutting states in a machine tool that performs cutting according to set machining conditions, and a plurality of machining state signals obtained from the plurality of machining state measurement means. neural network means formed to calculate control commands for the machining conditions through learning; and control command output means for outputting control commands for the machining conditions based on output signals from the neural network means. Equipped with a cutting processing adaptive control device. (2) The plurality of machining state measuring means include an AE detecting means for detecting an AE signal generated during cutting, a cutting resistance detecting means for measuring the cutting resistance applied to the tool, and a current or voltage of the motor of the machine tool. The cutting adaptive control device according to claim 1, wherein the cutting processing adaptive control device is at least two of a power consumption measuring means for measuring the state of surface roughness of the workpiece being machined, and a surface roughness measuring means for measuring the state of surface roughness of the workpiece being machined. .