JP2754760B2 - Tool damage detection device - Google Patents

Tool damage detection device

Info

Publication number
JP2754760B2
JP2754760B2 JP17905189A JP17905189A JP2754760B2 JP 2754760 B2 JP2754760 B2 JP 2754760B2 JP 17905189 A JP17905189 A JP 17905189A JP 17905189 A JP17905189 A JP 17905189A JP 2754760 B2 JP2754760 B2 JP 2754760B2
Authority
JP
Japan
Prior art keywords
signal
output
neural network
tool
detecting
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Lifetime
Application number
JP17905189A
Other languages
Japanese (ja)
Other versions
JPH0349850A (en
Inventor
正則 佐藤
孝一 辻野
Original Assignee
オムロン株式会社
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by オムロン株式会社 filed Critical オムロン株式会社
Priority to JP17905189A priority Critical patent/JP2754760B2/en
Publication of JPH0349850A publication Critical patent/JPH0349850A/en
Application granted granted Critical
Publication of JP2754760B2 publication Critical patent/JP2754760B2/en
Anticipated expiration legal-status Critical
Expired - Lifetime legal-status Critical Current

Links

Description

DETAILED DESCRIPTION OF THE INVENTION Summary of the Invention A plurality of frequency components of an AE signal obtained from an AE sensor are extracted and supplied to a neural network, or an AE signal obtained from a plurality of airborne AE sensors having different center frequencies is converted to a neural network. Tool damage is detected by giving it to the network and processing it. The accuracy of damage detection is high, and the detection algorithm can be created relatively easily by learning the neural network.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to acoustic emission (AE: Acousti
The present invention relates to a tool damage detection device that detects using c Emission).

Conventional technology and its problems As an equipment for detecting tool damage on machine tools, AE generated when the tool is damaged is detected by an AE sensor installed near the tool on the machine tool, and tool damage is detected based on this detection signal. Things have been suggested. However, such a conventional device uses a damage detection algorithm that compares the level of the signal component of a specific frequency band in the AE signal with a threshold value, so malfunctions due to cutting noise and other noises may occur. Or damage could not be detected. To solve the above problems, damage detection algorithms using AE signal components in different frequency bands are being studied. However, as the number of signal types increases, the relationship between damage and the state of the AE signal increases. There is a problem that it is difficult to grasp, and eventually it is difficult to create an appropriate detection algorithm.

SUMMARY OF THE INVENTION Object of the Invention It is an object of the present invention to provide a tool damage detection device capable of accurately detecting damage using a plurality of features of an AE signal.

A tool damage detecting device according to the present invention is arranged near a tool of a machine tool, detects AE signals generated from the machine tool, and a plurality of AE signals provided from the AE detecting means. AE signal processing means for extracting a plurality of features and outputting a plurality of signals representing the features. The AE signal processing means receives a plurality of output signals from the AE signal processing means as inputs and is formed so that a tool damage state can be identified from other states by learning. A neural network means, and output means for outputting a tool damage detection signal in response to a signal indicating a tool damage state from the neural network means.

When the AE detecting means includes one AE sensor, the AE signal processing means is configured to extract a plurality of frequency components of an output signal of the AE sensor.

If the AE detecting means includes a plurality of airborne AE sensors having different center frequencies, the AE signal processing means processes output signals of the plurality of airborne AE sensors and outputs a plurality of outputs representing the processing results. It is configured to output a signal.

The AE signal processing means and the neural network means can be realized by an analog circuit and a digital circuit, and also by a programmed computer.

According to the present invention, the following operations and effects can be obtained. That is, several different frequency band components of the AE signal and other AE
Since the characteristic signal of a signal or a plurality of AE signals can be simultaneously processed, the accuracy of damage detection is increased. Even if the characteristics of the AE signal at the time of damage are not completely understood, a detection algorithm can be created by the learning function in the neural network. In addition, since the detection algorithm can be modified according to the use environment by the learning function, it can be applied to various environments.

FIG. 1 shows the overall configuration of a tool damage detection device.

A drill 3 is mounted on a main shaft 2 of a machine tool 1, and a work 5 fixed to a table 4 is machined by the drill 3. AE (Acoustic Emission) generated during this processing
Is detected by the AE sensor 6 fixed to the table 4. The AE signal output from the AE sensor 6 is amplified in the AE signal processing unit 7, and the characteristics of the signal are extracted as described later. The signals representing the extracted features are then provided to a neural network 8 and processed as described in more detail below. Based on the signal output from the neural network 8, the output circuit 9 outputs a drill breakage detection signal when the drill breaks.

FIG. 2 shows the configuration of the AE signal processing unit 7 and the neural network 8.

The AE signal detected by the AE sensor 6 is amplified to an appropriate level by an amplifier 10, and the band-pass filter 1
1, 12, 13 and the detection circuit 14. Band pass
The filters 11, 12, and 13 are filters having different pass frequency bands. In this embodiment, the center frequencies of the respective pass bands are 10 KHz, 50 KHz, and 200 KHz. An AE signal generally has a plurality of characteristic frequencies.
13 extracts this characteristic frequency component. The characteristic frequency can be known by frequency analysis of the AE signal at the time of drill damage. The outputs of the band-pass filters 11, 12, and 13 are supplied to detection circuits 21, 22, and 23, respectively, where the average values are detected and converted to DC levels. The envelope of the AE signal input to the detection circuit 14 is detected. Detection circuit
The output of 14 is given to a differentiating circuit 24, where a rising change component of the signal is extracted. The outputs of the detection circuits 21, 22, 23 and the differentiation circuit 24 are provided to the input layer 31 of the neural network 8. The detection circuit 14 and the differentiation circuit 24 do not always need to be provided.

A neural network combines multiple operators (neurons) with each other using a weighting function (synapse), and transmits an operation pattern performed by a neuron to an input pattern through a synapse to transmit an output pattern that can be concluded. This is the calculation method to obtain. This neural network has the ability to determine the weight of each synapse by learning the input pattern and the output pattern to be obtained from this input pattern a finite number of times, so it can be used for processing where it is difficult to analyze the algorithm. it can.

In FIG. 2, the neural network 8 has a multilayer structure of an input layer 31, an intermediate layer 32 and an output layer 33.
May be one or more layers. Each layer (excluding the output layer 33) is composed of a plurality of units. Output layer 33
Is composed of one unit. Here, the unit includes the above-mentioned neuron and a synapse that connects the neuron. Further, the neural network 8 is provided with a teacher signal output device 34 for learning described later.
The neural network 8 is generally realized by software of a digital computer.

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

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

When a certain unit i receives an input O j (j = 1 to k) from a plurality of units 1,..., J,.
An output O i is generated based on t i . This sum net i Is represented by

The output O i of the unit is a converted value (a value between 0 and 1) obtained by applying the sum net i obtained by the above equation to the function f i (sigmoid function) shown in FIG. .

The neural network 8 performs the above calculations sequentially on the input layer 31, the intermediate layer 32, and the output layer 33 based on the output signal from the AE signal processing unit 7, and when the drill breaks, 1 or a value close to 1 is output to the output layer 33. Is formed so as to be output.

The output circuit 9 is an output layer of the neural network 8.
In response to the output from 33, a drill breakage detection signal or the like is output. The output circuit 9 outputs a breakage detection signal when, for example, the output of the output layer 33 is 0.9 or more,
It is configured to output an alarm signal when less than. When the output of the output layer 33 is less than 0.8, no output is generated.

The formation of the neural network as described above is performed by learning. Learning is based on back propagation. Data at the time of drill breakage or the like is collected in advance, and learning is performed using these data. That is, the output value generated from the output layer 33 is compared with the output value of the teacher signal output device 34 to which a desired output value is given in advance, and the weight between the units is changed so that the difference is reduced. Go.

In FIG. 2, the band-pass filters 11, 12, 13 of the AE signal processing unit 7, the detection circuits 14, 21, 22, 23 and the differentiation circuit 24
Can be realized by an analog circuit or a digital circuit.

FIG. 5 shows another example of the AE signal processing unit. AE
The output signal of the sensor 6 passes through an amplifier 10, an anti-aliasing filter (a kind of low-pass filter) 41, and an A / A signal.
The signal is input to the D converter 42 and converted into a digital signal. A / D
The output signal of the converter 42 is alternately applied to the data buffer memories 43 and 44 at regular intervals. The data read from the data buffer memory 43 or 44 is supplied to the switching circuit 45.
Is given to an FFT (Fast Fourier Transform) processor 47, and is converted into a frequency component as shown in FIG. 6 (a frequency spectrum is created). The FFT processing requires some time. To deal with this, two buffer memories 4
3,44 are provided. That is, buffer memory 4
Reference numerals 3 and 44 are used while being switched by the FFT processor 47 in order to perform FFT processing of the time-series AE signal data in real time. As shown in FIG. 6, the waveform area calculating device 48 calculates the area of the hatched portions around 10 KHz, 50 KHz and 200 KHz in the result of the FFT processing. This calculation result corresponds to the output of the detection circuits 21, 22, and 23 in FIG. 2, and is given to the input layer 31 of the neural network 8.

Next, a specific example of learning for forming a neural network for performing drill breakage detection will be described in detail. The circuit shown in FIG. 5 is used as the AE signal processing unit. However, the waveform area calculation device 48 is configured to calculate the area of eight frequency bands as shown below, instead of the area of three frequency bands.

FIG. 7 shows an example of the power spectrum of the AE signal obtained from the FFT processor 47. Such a spectrum signal is provided to the waveform area calculator 48. The waveform area calculator 48 calculates the area (power) of the sections T1 to T8 obtained by dividing the frequency range of 0 to 400 KHz into eight equal parts. The result of this operation is shown in FIG. In FIG. 8, T1 is frequency 0-5
0KHz section, T2 is 50KHz ~ 100KHz section, and so on.
T3 to T8 indicate sections of 50 KHz each. Each section T1-T8
Are represented by I1 to I8.

FIG. 9 shows an example of the configuration of a neural network used for learning. In this neural network, an input layer 31 includes eight units U1 to U8 and an intermediate layer 32.
Has four units U11 to U14, and the output layer has one unit U21.

For training, 40 patterns were used as AE data when the drill was broken.
40 patterns were used as other data, for example, noise data such as work clamp sound. Two examples of data at the time of drill breakage and data other than drill breakage
The patterns are shown below. Where the data is I1, I2,…, I8
, And in this order the units U1, U2,
…, Input to U8. In addition, these data are normalized with the maximum value of all data (80 patterns) being +0.5 and the minimum value being -0.5 (that is, in this specific example, the input / output value is between 0 and 1). Instead of the value of
+0.5).

Drill break data −0.451111, −0.143826, −0.180023, −0.031947, −0.262238, −0.426132, −0.438874, −0.449643 −0.412872, −0.142519, −0.266363, −0.196906, −0.346527, −0.426137, −0.446903, − 0.469452 Data other than drill breakage −0.468254, −0.255373, −0.426633, −0.426967, −0.470333, −0.495223, −0.498779, −0.499449 −0.420794, −0.194203, −0.321077, −0.361384, −0.474583, −0.487298, −0.494725, −0.497785 Learning is to output 0.5 as the output when the data at the time of drill breakage is given to the neural network.
When other data is given to the neural network, the weight between units is corrected by back propagation so that -0.5 is output. The values of 0.5 and -0.5 described above are provided by the teacher signal output device. The learning is continued until the difference between the output from the output layer and the output of the teacher signal output device for all data patterns (80 patterns) becomes equal to or smaller than a predetermined value.

FIG. 10 shows the sigmoid function used in this learning. The sigmoid function f i is given by the following equation.

FIG. 11a shows weights before learning, and FIG. 11b shows weights finally obtained by learning. The weights before learning are randomly assigned values in the range of ± 0.3.

By the above learning, the output value of 0.499995 or more is output from the output layer of the neural network at the time of drill breakage after learning, otherwise -0.49 from the output layer.
Output value of 9995 or less is output.

FIG. 12 and FIG. 13 show another embodiment of the present invention. In the above-described embodiment, the AE sensor 6 is fixed to the table 4 (see FIG. 1), but in this embodiment, airborne AE sensors 6A, 6B, and 6C that detect AE in a non-contact manner are used. Since the AE signal that propagates in the air is weak, it is necessary to increase the sensitivity of the airborne AE sensor.
The sensitivity characteristics of the AE sensor are not uniform with respect to the frequency of the AE signal, and are configured so that the AE signal component can be detected only in a specific relatively narrow frequency band. Detects signal components in multiple frequency bands with different AE signals and
A plurality of airborne AE sensors 6A and 6 to provide to the network 8
That is, B and 6C are provided. Aerial AE sensors 6A, 6B, 6C having different center frequencies of sensitivity are fixed to a support member (not shown) toward a processing point of the drill 3 and the workpiece 5. In the embodiment, the center frequencies of these airborne AE sensors 6A, 6B and 6C are 10 KHz, 50 KHz and 200 KHz, respectively.

AE signals a, b, AE detected by the airborne AE sensors 6A, 6B, 6C
c is amplified to an appropriate level by the amplifiers 10A, 10B, and 10C, respectively, and sent to the band-pass filters 11A, 11B, and 11C, respectively. The band pass filters 11A, 11B, and 11C are filters having the same pass band as the center frequency of the airborne AE sensors 6A, 6B, and 6C, respectively, and cut noise components. The outputs of the band-pass filters 11A, 11B, 11C are applied to detection circuits 21A, 21B, 21C, respectively, where they are average-detected and converted to DC levels. On the other hand, the output of the detection circuit 21C is given to the differentiating circuit 21D, where the rising change component of the signal is extracted. Detection circuit 21A, 21B, 21C, differentiation circuit
The 21D output signal is provided to the neural network 8. The output of the differentiating circuit 21D does not necessarily have to be input to the neural network 8.

The configurations and operations of the neural network 8 and the output circuit 9 are the same as those of the above-described embodiment. A drill breakage detection signal is output from an output circuit 9 through a neural network 8 formed so that drill breakage can be detected by learning.

[Brief description of the drawings]

1 to 4 show an embodiment of the present invention. FIG. 1 shows an entire configuration of a tool damage detecting device, and FIG. 2 shows details of an AE signal processing section and a neural network in FIG. FIG. 3 is a diagram showing the structure of the unit in the neural network, and FIG. 4 is a graph showing the input / output characteristics of the unit. FIG. 5 is a block diagram showing another example of the AE signal processing unit, and FIG.
The figure is a graph showing the processing operation of the circuit of FIG. 7 to 11b show specific examples by simulation, FIG. 7 is a graph showing a frequency spectrum of an AE signal, FIG. 8 is a graph showing a spectrum area for each frequency band, and FIG. 9 is a simulation. Diagram showing the configuration of the neural network used, FIG. 10 is a graph showing input / output characteristics of units in the neural network of FIG. 9, and FIGS. 11a and 11b are diagrams before learning in the neural network of FIG. And an example of weights after learning. FIGS. 12 and 13 show another embodiment. FIG. 12 shows the overall configuration of a tool damage detection device. FIG. 13 is a block diagram showing the AE signal processing section and neural network in FIG. FIG. 1 ... machine tool, 3 ... drill (tool), 6 ... AE sensor, 6A, 6B, 6C ... aerial AE sensor, 7, 7A ... AE signal processing unit, 8 ... neural network, 9 ... ... Output circuit.

Claims (3)

    (57) [Claims]
  1. An AE detecting means arranged near a tool of a machine tool for detecting an AE signal generated from the machine tool, extracting a plurality of features of an AE signal supplied from the AE detecting means, and extracting a plurality of features representing the features. AE signal processing means for outputting a signal of the AE signal, neural network means for receiving a plurality of output signals of the AE signal processing means and formed so that a tool damage state can be distinguished from other states by learning, and the neural network Output means for outputting a tool damage detection signal in response to a signal indicating a tool damage state from the means.
  2. 2. The AE detecting means includes one AE sensor,
    The tool damage detection device according to claim 1, wherein the AE signal processing means extracts a plurality of frequency components of an output signal of the AE sensor.
  3. 3. The AE detecting means includes a plurality of aerial AE sensors having different center frequencies, and the AE signal processing means processes output signals of the plurality of aerial AE sensors and indicates a plurality of processing results. The tool damage detecting device according to claim 1, which outputs the output signal of (1).
JP17905189A 1989-07-13 1989-07-13 Tool damage detection device Expired - Lifetime JP2754760B2 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
JP17905189A JP2754760B2 (en) 1989-07-13 1989-07-13 Tool damage detection device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
JP17905189A JP2754760B2 (en) 1989-07-13 1989-07-13 Tool damage detection device

Publications (2)

Publication Number Publication Date
JPH0349850A JPH0349850A (en) 1991-03-04
JP2754760B2 true JP2754760B2 (en) 1998-05-20

Family

ID=16059267

Family Applications (1)

Application Number Title Priority Date Filing Date
JP17905189A Expired - Lifetime JP2754760B2 (en) 1989-07-13 1989-07-13 Tool damage detection device

Country Status (1)

Country Link
JP (1) JP2754760B2 (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7357197B2 (en) * 2000-11-07 2008-04-15 Halliburton Energy Services, Inc. Method and apparatus for monitoring the condition of a downhole drill bit, and communicating the condition to the surface
JP4934504B2 (en) * 2007-05-29 2012-05-16 日本クラウンコルク株式会社 connector
JP5943578B2 (en) * 2011-10-11 2016-07-05 株式会社東京精密 Wafer chamfering apparatus, and method for detecting surface state of chamfering grindstone or processing state of wafer by chamfering grindstone
CN102825504B (en) * 2012-09-18 2014-12-24 重庆科技学院 State detection method for main shaft of numerically-controlled machine tool
CN106217130B (en) * 2016-08-15 2018-06-15 大连理工大学 Milling cutter state on_line monitoring and method for early warning during complex surface machining
JP6487475B2 (en) * 2017-02-24 2019-03-20 ファナック株式会社 Tool state estimation device and machine tool

Also Published As

Publication number Publication date
JPH0349850A (en) 1991-03-04

Similar Documents

Publication Publication Date Title
Rangwala et al. Sensor integration using neural networks for intelligent tool condition monitoring
Tabrizi et al. Early damage detection of roller bearings using wavelet packet decomposition, ensemble empirical mode decomposition and support vector machine
Sambath et al. Automatic detection of defects in ultrasonic testing using artificial neural network
US4811308A (en) Seismo-acoustic detection, identification, and tracking of stealth aircraft
EP1927830B1 (en) Device for overall machine tool monitoring and corresponding method therefor
US7203132B2 (en) Real time acoustic event location and classification system with camera display
Wuxing et al. Classification of gear faults using cumulants and the radial basis function network
US6065342A (en) Apparatus and a method of locating a source of acoustic emissions in an article
Abu-Mahfouz Drilling wear detection and classification using vibration signals and artificial neural network
Sreejith et al. Fault diagnosis of rolling element bearing using time-domain features and neural networks
KR100511205B1 (en) Method for dividing the sound fields of individual sources by acoustic holography
Al-Bugharbee et al. A fault diagnosis methodology for rolling element bearings based on advanced signal pretreatment and autoregressive modelling
US4459851A (en) Method and device for the localization and analysis of sound emissions
Subrahmanyam et al. Using neural networks for the diagnosis of localized defects in ball bearings
Liu et al. Structure damage diagnosis using neural network and feature fusion
JP4241818B2 (en) Internal inspection device
EP1273927B1 (en) Environmental noise monitoring
EP0770854B1 (en) Method, apparatus and system for determining specimen
EP2031386B1 (en) Acoustic emission detector and controller
Boczar et al. Diagnostic expert system of transformer insulation systems using the acoustic emission method
CN108620949B (en) Cutter wear monitoring and predicting method
Lalitha et al. Emotion detection using MFCC and cepstrum features
EP1600791A1 (en) Sound source localization based on binaural signals
Wang et al. Neural network detection of grinding burn from acoustic emission
US9091613B2 (en) Multi-spectral ultrasonic gas leak detector