CN107891199A - Spark discharge condition checkout gear and recognition methods - Google Patents
Spark discharge condition checkout gear and recognition methods Download PDFInfo
- Publication number
- CN107891199A CN107891199A CN201711250142.3A CN201711250142A CN107891199A CN 107891199 A CN107891199 A CN 107891199A CN 201711250142 A CN201711250142 A CN 201711250142A CN 107891199 A CN107891199 A CN 107891199A
- Authority
- CN
- China
- Prior art keywords
- subprograms
- discharge
- pixels
- deep learning
- convolutional layer
- 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.)
- Granted
Links
Classifications
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23H—WORKING OF METAL BY THE ACTION OF A HIGH CONCENTRATION OF ELECTRIC CURRENT ON A WORKPIECE USING AN ELECTRODE WHICH TAKES THE PLACE OF A TOOL; SUCH WORKING COMBINED WITH OTHER FORMS OF WORKING OF METAL
- B23H1/00—Electrical discharge machining, i.e. removing metal with a series of rapidly recurring electrical discharges between an electrode and a workpiece in the presence of a fluid dielectric
-
- B—PERFORMING OPERATIONS; TRANSPORTING
- B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
- B23H—WORKING OF METAL BY THE ACTION OF A HIGH CONCENTRATION OF ELECTRIC CURRENT ON A WORKPIECE USING AN ELECTRODE WHICH TAKES THE PLACE OF A TOOL; SUCH WORKING COMBINED WITH OTHER FORMS OF WORKING OF METAL
- B23H11/00—Auxiliary apparatus or details, not otherwise provided for
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mechanical Engineering (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Electrical Discharge Machining, Electrochemical Machining, And Combined Machining (AREA)
Abstract
The invention discloses a kind of spark discharge condition checkout gear, including signal processing module, identification device and upper computer module.The invention also discloses the spark discharge state identification method using above-mentioned spark discharge condition checkout gear, the voltage signal of the interpolar of negative and positive two caused by a discharge pulse in electric spark machine tool electric discharge machining process is first input to optical coupling module by partial pressure circuit, analog-to-digital conversion is carried out again, digital filtering and cycle detection are carried out again, generate 256 sampled points, CPU element and GPU units are calculated simultaneously, comprehensive deep learning program and simulation manual features value extraction, voted by simple majority, draw final recognition result.The present invention is converted to the test problems of electric signal the identification problem of discharge waveform image, substantially reduces identification difficulty, improves recognition accuracy.Simple majority voting method enhances the fault freedom of the present invention, greatly improves the robustness of the present invention.
Description
Technical field
The present invention relates to the detection means of technical field and method, is that one kind is based on depth specifically
Practise the spark discharge condition checkout gear and method of assembled classifier.
Background technology
Electrical discharge machining is carried out by electric spark machine tool, specifically in media as well, utilizes tool-electrode and workpiece electricity
The energy of pulse feature spark discharge between pole is processed.Electrical discharge machining is in each field machine such as Aero-Space, automobile die
Tool processing in obtain extensive use, for solve it is various be difficult to, complex-shaped part.And electrical discharge machining is to use
Electric energy is processed, and process is easily achieved digital control, intelligent and Unmanned operation.From nineteen forty-three, Soviet Union's Moscow State University
Lazarenko Mr. and Mrs find spark discharge principle since, spark erosion technique is constantly progressive, but can not still meet
Modernize the demand of processing and manufacturing.Because the lifting of the limitation of spark discharge principle, machining accuracy encounters bottleneck.
Influenceing the factor of edm process mainly includes electrical parameter（Peak point current, crest voltage, pulse width, arteries and veins
Punching interval, processing polarity）With non-electrical parameter（Working fluid pressure, wash away mode）And its rapidoprint（Specific heat, density, thermal conductivity system
Number, fusing point）.The form of expression of these parametric synthesis can be described with spark discharge state.Further, accurately realize
Discharging headlamp, automatic adjustment servo feed direction and speed in real time, lifts processing stability, is to realize that Precision EDM adds
One of key technology of work.
The conventional method that spark discharge state differentiates（Average voltage detection, crest voltage detect, high fdrequency component detects,
Discharge breakdown delay etc.）It is heavily dependent on gap voltage.Except traditional pulsed discharge method, some intelligent methods
（Fuzzy logic, neutral net, wavelet transformation）Technology is also widely used.These detection methods or existence inspection
It is excessively broad to survey result classification range；Or accuracy of detection to be present inadequate due to the limitation in Cleaning Principle, structure, or detection
Method robustness it is bad.Especially since the sample size of sampling limits, the high frequency waves of discharge waveform can not be obtained completely
Dynamic composition, and this is a key index for distinguishing spark discharge and arc discharge.
Because discharge mechanism is sufficiently complex in edm process, especially aviation complex-shaped surface mould Precision Machining or
In fine electric spark forming, the high-frequency noise contained in discharge waveform increases sharply, and can make conventional spark discharge shape
State detection method is in half failure or entirely ineffective state.
Found by being retrieved to prior art literature, in periodical《The International Journal of
Advanced Manufacturing Technology》（International advanced manufacture periodical）November 2015, Volume 81,
ISSue 5-8, pp 1403-1418, " A new method for on-line monitoring diScharge pulSe
in WEDM-MS proceSS”（Towards a kind of new discharge condition online test method in middle wire linear cutter）In,
After discharge voltage signal is converted into image, discharge condition is identified using RF+SVM double classification method.But this article
Only with the feature of Traditional Man extraction algorithm acquisition discharge waveform, the electrical discharge machining to realize other materials, or
Processed in die cavity applied to complexity, after discharge waveform complexity increase, the robustness of its algorithm is with regard to pessimistic, it is necessary to again
Consider the extracting method of characteristic value.
The content of the invention
It is an object of the invention to provide a kind of spark discharge state-detection dress based on deep learning assembled classifier
Put, accuracy rate is high, and fault-tolerance is strong, and robustness is good.
To achieve the above object, spark discharge condition checkout gear of the invention includes signal processing module, identification dress
Put and upper computer module；
Signal processing module includes optical coupling module, analog-to-digital conversion module, arm processor and the first hi-speed USB interface；Optical coupling module
Input be connected by partial pressure circuit with the negative and positive two-stage of electric spark machine tool, described in the connection of the output end of optical coupling module
Analog-to-digital conversion module, the analog-to-digital conversion module connect the arm processor, and arm processor connects first high speed USB and connect
Mouthful；
Identification device includes embedded main board, the second hi-speed USB interface module, CPU element, GPU units, internal storage location, solid-state
Hard disk, Ethernet interfaces and netting twine；
Second hi-speed USB interface module connects first hi-speed USB interface by USB connecting lines；Second hi-speed USB interface mould
Block is connected by output line with embedded main board, the CPU element, GPU units, internal storage location, solid state hard disc and
Ethernet interfaces are connected by both-way communication circuit with the embedded main board respectively；Ethernet interfaces pass through the net
Line is connected with upper computer module.
Linux system is stored with solid state hard disc, three deep learning grader programs, spies are installed in Linux system
Levy extraction algorithm program, SVC/AdaBoost/DTC/RFC sorting algorithms program, assembled classifier program and for controlling entirety
The main program of flow；
The input of three deep learning grader programs is the discharge waveform image of complete cycle, and the resolution ratio of the image is 256*
256 pixels；
First deep learning grader program includes convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 journeys
Sequence, pond layer 2S2 subprograms, the first full connection level routine and the first soft recurrence subprogram；
The convolutional layer 1C1 subprograms of first deep learning grader are 256*256*8 pixels and its convolution kernel is 3*3 pixels；
The pond layer 1S1 subprograms of first deep learning grader are 128*128*8 pixels；The volume of first deep learning grader
Lamination 2C2 subprograms are 128*128*32 pixels and its convolution kernel size is 3*3 pixels；First deep learning grader pond
Layer 2S2 subprograms are 64*64*32 pixels；First full connection level routine is 4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of first deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, the first full connection level routine and the first soft sub- journey of recurrence
After the processing of sequence, final output discharge condition result vector [S1, S2, S3, S4, S5], wherein S1 is open-circuit condition, and S2 is fire
Flower electric discharge, S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；The value of vector is present discharge shape probability of state；
Second deep learning grader program includes convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 journeys
Sequence, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond layer 3S3 subprograms, the second full connection level routine and second are soft
Return subprogram；
The convolutional layer 1C1 subprograms of second deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of second deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of second deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of second deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of second deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of second deep learning grader are 32*32*64 pixels, and the second full connection level routine is
4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of second deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond layer 3S3 subprograms,
After two full connection level routines and the second soft processing for returning subprogram, final output discharge condition result vector [S1, S2, S3,
S4, S5], wherein S1 is open-circuit condition, and S2 is spark discharge, and S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；
The value of vector is present discharge shape probability of state；
3rd deep learning grader program includes global image processing sub-programme gather, topography's processing sub-programme gather, the
Three full connection level routines and the 3rd soft recurrence subprogram；
The discharge waveform image for the complete cycle that 3rd deep learning grader program is received while send at global image
Manage sub-programme gather and local image procossing sub-programme gather carries out parallel processing；
Global image processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms,
Pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms；
The convolutional layer 1C1 subprograms of 3rd deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of 3rd deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of 3rd deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of 3rd deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of 3rd deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of 3rd deep learning grader are 32*32*64 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of global image processing sub-programme gather, pond layer
At 1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms
After reason, 4096 pixel datas are conveyed to the 3rd full connection level routine；
Topography processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and
Pond layer 2S2 subprograms；The discharge waveform image of the complete cycle received is divided into 4 portions by topography's processing sub-programme gather
Point, respectively the upper left of image, upper right portion, bottom left section and lower right-most portion, 4 parts are 128*128 pixels；
Convolutional layer 1C1 subprograms are 128*128*8 pixels and its convolution kernel size is 3*3 pixels；Pond layer 1S1 subprograms are
64*64*8 pixels；Convolutional layer 2C2 subprograms are 64*64*32 pixels；Pond layer 2S2 subprograms are 32*32*32 pixels；
The discharge waveform image of complete cycle handles the convolutional layer 1C1 subprograms of sub-programme gather, pond layer 1S1 by topography
When program, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms are handled, topography handles convolutional layer 1C1 of sub-programme gather
Program, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms handle the upper left quarter of described image respectively
Point, a part in upper right portion, bottom left section and lower right-most portion, the data after each part of image is processed are
1024 pixels；Topography handles sub-programme gather and the data of obtained after processing 4 1024 pixels is conveyed into the 3rd full connection
Level routine；
1 4096 pixel data and 4 1024 pixel datas that 3rd full connection level routine is received merge into 8192
Pixel data, 8192 pixel datas are then conveyed to the 3rd soft recurrence subprogram；The 3rd soft subprogram that returns exports electric discharge shape
State result vector [S1, S2, S3, S4, S5], wherein S1 are open-circuit condition, and S2 is spark discharge, and S3 is arc discharge, and S4 is
Transient discharge, S5 are short-circuit condition；The value of vector is present discharge shape probability of state.
The analog-to-digital conversion module is more than the analog-to-digital conversion of 12 using switching rate in more than 10Msps and sampling precision
Module；
Three deep learning graders program passes through off-line training, and the Sample Storehouse of off-line training includes mould steel, titanium alloy
With aluminum-base silicon carbide material.
CPU in the CPU element uses CPU of the dominant frequency in more than 2.0GHz and physics kernel more than 4；CPU's
One physics kernel operation main program and assembled classifier program, are responsible for task scheduling and communication task, CPU another kernel
Operation characteristic extraction algorithm program, CPU most latter two kernel is by loading condition operation SVC/AdaBoost/DTC/RFC algorithms
Program.
The GPU units use core frequency for 1067MHz, have 2400 stream processing units and video memory capacity in 4GB
GPU above, GPU units are directed to each discharge waveform, complete the calculating task of 3 kinds of deep learning grader programs in real time.
, should the present invention also aims to provide a kind of spark discharge state identification method using above-mentioned detection device
Method is carried out according to the following steps：
First step is：After system starts, the Linux system of solid state hard disc storage inside, which is loaded into internal storage location, to be run, and
Start three deep learning grader programs, feature extraction algorithm program, SVC/AdaBoost/DTC/RFC sorting algorithms program,
Assembled classifier program and main program；
The voltage signal of the interpolar of negative and positive two caused by a discharge pulse in electric spark machine tool electric discharge machining process is passed through
Partial pressure circuit is input to optical coupling module, and after light-coupled isolation, access analog-to-digital conversion module is changed into data signal；
Described optical coupling module has more than 50M bandwidth, between electric spark machine tool region of discharge and identification device
Be electrically isolated completely to, and ensure the information of the voltage signal of spark discharge by not lost after analog-to-digital conversion module；
Second step is：The data signal of analog-to-digital conversion module output passes to arm processor；
The analog-to-digital conversion module is more than the analog-to-digital conversion module of 12 using switching rate in more than 10Msps and sampling precision,
Every 0.1 μ s are sampled once, no matter for long-pulse discharge waveform（≥100μs）Or short pulse discharge waveform（≤10μs）, all
It can guarantee that enough sampled datas, it is thus possible to accurately reflect discharge voltage waveform；
Third step is：The data that arm processor receives to it carry out digital filtering and cycle detection, then passing ratio conversion
By the data compression in discharge waveform cycle or it is stretched to 256 sampled points；
Arm processor by after processing complete cycle 256 sample point datas by the first hi-speed USB interface be transferred to second at a high speed
USB interface, the second hi-speed USB interface send the data to CPU element by embedded main board；
Four steps is：CPU element receives the data that the second hi-speed USB interface transmits, and the data are deposited into internal memory
In unit, Wave data complete cycle of history is stored by the way of circle queue；
5th step is：CPU element starts calculating task, after the current Wave data complete cycle composograph in internal storage location
It is sent in GPU units；The composograph is the gray level image of 256*256 pixels；
Then, CPU element and GPU units are calculated simultaneously；
GPU units are by three deep learnings grader program after off-line training is read and performed in internal storage location；Three
Recognition result after deep learning grader program calculates sends CPU element back to by embedded main board again, and is stored in internal storage location
In；
CPU element opens multithreading, is face respectively respectively by calculating the characteristic value of extraction Wave data composograph complete cycle
Color characteristic, patterned feature and SURF features, wherein color characteristic are 21 dimensions, and patterned feature is characterized as 13 dimensions, and SURF is characterized as 70
Dimension, 104 tie up altogether；
After the characteristic value for extracting image, then it is identified respectively by SVC/AdaBoost/DTC/RFC sorting algorithms program, its
The parameter E of middle AdaBoost algorithms is calculated three times using three different values, exports the recognition result of each time respectively；DTC algorithms
Parameter D calculated twice using two different values, export the recognition result of each time respectively；
6th step is：The result of CPU and GPU parallel computations is handled by assembled classifier program in 5th step, is adopted
Final identification state is determined with simple majority voting method, and final status data is passed through into netting twine through Ethernet interfaces
It is transmitted to upper computer module；
7th step is：Above-mentioned first to the 6th step is iteratively repeated progress 30 times, and CPU element counts this 30 discharge pulses
State status, provide electro-discharge machining qualitative data and by the qualitative data through Ethernet interfaces, host computer is transmitted to by netting twine
Module；
It is electrical spark working after real-time discharge condition data and 30 history discharge condition situations are fed back to host computer by CPU element
The control system operation of work lathe provides foundation.
In 7th step, 30 discharge pulses are calculated with ratio of the number of pulses of regular picture in overall pulse quantity respectively
Example, the ratio of the number of pulses of short-circuit condition in overall pulse quantity, so as to by this discharge condition vector data<Pulse ratio
Example, short-circuit ratio>Through Ethernet interfaces, upper computer module is transmitted to by netting twine.
Three deep learning grader programs after off-line training by being deposited into solid state hard disc, the sample of off-line training
For open-circuit condition, spark discharge, arc discharge, transient discharge and each 5000 of short-circuit condition, grown wherein every kind of state both included
Periodic discharging waveform, also comprising short periodic discharging waveform.
By netting twine by three deep learning grader download programs after host computer off-line training into solid state hard disc, it is right
Deep learning grader program is updated.
In 5th step, Wave data complete cycle is synthesized into method that gray level image uses as when discharge voltage is more than
During equal to 200V, gray value 0, when discharge voltage is less than or equal to 0V, gray value 1；Discharge voltage is less than 200V and is more than
Gray value during 0V is calculated using linear interpolation.
During work, the voltage signal of the caused interpolar of negative and positive two passes through partial pressure in electric spark machine tool electric discharge machining process
Circuit is input to optical coupling module, output signal of the analog voltage signal after optical coupling module is isolated as optical coupling module,
The output signal is converted to data signal by analog-to-digital conversion module, and passes to arm processor, and arm processor receives to it
Data carry out digital filtering and cycle detection, the second hi-speed USB interface is then sent to by the first hi-speed USB interface, then
Sent by embedded main board to CPU element.The data signal received is passed through embedded master by CPU element by waveform complete cycle
It is stored in after plate transfer in internal storage location.
Then, one side GPU units perform the deep learning calculating task after training, and the result of identification passes through embedded master
CPU element is returned after plate transfer；Another aspect CPU element carries out characteristics extraction to discharge waveform image, passes through SVC/ again
AdaBoost/DTC/RFC sorting algorithms are identified respectively.The result of above-mentioned CPU and GPU parallel computations passes through the group in CPU
Close grader program to be handled, by final spark discharge status data through Ethernet interfaces, be transmitted to by netting twine
Position machine module.
Arm processor is internally integrated 128K ROM and 16K RAM, is stored with ROM for Wave data collection, place
Reason and the program of transmission.Run into the RAM for the 16K being internally integrated from ROM loading procedures after electricity on arm processor, hold automatically
The collection of traveling wave graphic data, processing and transmission.
Solid state hard disc storage inside has Linux system, and deep learning grader program is provided with Linux system（It is with dividend right
Weight）, SVC/AdaBoost/DTC/RFC sorting algorithm program（Containing weight）, feature extraction algorithm program, SVC/AdaBoost/
DTC/RFC sorting algorithms program, assembled classifier program and main program.Main program is used to control overall flow, by process flow operation
The every other programs such as deep learning network, feature extraction algorithm program.It is people in the art according to flow establishment main program
The existing force of member.As needed, other auxiliary programs can also be stored on solid state hard disc by those skilled in the art.
CPU element starts main program after solid state hard disc loading Linux successes after upper electricity, and by other journeys of process flow operation
Sequence, load and recover three kinds of deep learning networks, the respective networks of SVC/AdaBoost/DTC/RFC that off-line training is crossed.
The analog-to-digital conversion module, it is preferred to use switching rate is more than the modulus of 12 in more than 10Msps and sampling precision
Modular converter, the discharge voltage data for making it into arm processor have higher sampling precision.
CPU in the CPU element preferably uses CPU of the dominant frequency in more than 2.0GHz and physics kernel more than 4；Make
CPU element real time execution discharge condition recognizer, CPU a physics kernel operation main program, assembled classifier program and
Three deep learning grader programs, it is responsible for task scheduling and communication task, CPU another kernel operation characteristic extraction algorithm
Program, CPU most latter two kernel is by loading condition operation SVC/AdaBoost/DTC/RFC algorithm routines.
In principle, the present invention is by the detection identification of discharge waveform, and discharge waveform is converted to by the test problems of electric signal
The problem of image recognition of image, so as to substantially reduce identification difficulty, improve recognition accuracy.
Integrated use image feature value of the present invention manually extracts（Color characteristic, Haralick features and SURF features）It is and deep
The automated characterization of learning network is spent, the mode for simulating manual identified discharge waveform is understood electric discharge image.Further, offline
Except including common mould steel in the Sample Storehouse of training（SKD11、718、H13）Outside, in addition to difficult-to-machine material titanium alloy, aluminium
Base silicon carbide material, the Covering domain of electric discharge sample is widened, so as to lift the accuracy rate of identification.
The present invention is on the basis of individual pulse discharge condition is identified, to present period（Nearest 30 pulsed discharges）
Discharge condition carry out statistical analysis, 30 discharge conditions comprising the present period including single discharge condition are all passed through
Ethernet interfaces are transferred to host computer.Accurate identification for individual pulse discharge condition, on the one hand provided for feeding control
Decision-making, the single pulse energy optimization that still further aspect is alternatively host computer provide reference frame.Know for the discharge condition of period
Not, foundation can be provided for the chip removal condition in analysis process, electrode carbon distribution.
Patent of the present invention can also fully meet the requirement handled in real time, by high speed analog-to-digital conversion, realize discharge wave figurate number
According to real-time collection, be forwarded in real time in identification device by arm processor.For than relatively time-consuming recognizer, this method
Comprehensive utilization of C PU and GPU parallel computation advantage, the small SVC/AdaBoost/DTC/RFC algorithms of amount of calculation are started by CPU element
Two thread parallels are handled, three depth networks（Deep learning grader program）Calculating transfer to special GPU units to enter
Row processing, by using the image of small size（256*256）, algorithm optimization and its lifting GPU runnability, it is possible to achieve
To a batch in 4 milliseconds（64 images）Discharge waveform identification, thus can also be completed for short pulse discharge waveform in real time
Identification.Further, by reducing the sample size of batch（32、8、2）, can also reduce GPU calculating time.Due to host computer number
Control system is 8 milliseconds and servo feed is adjusted, therefore by appropriate planning beat, disclosure satisfy that electric spark machine tool
Requirement of real-time control.
The present invention determines the final identification state of a discharge pulse using simple majority voting method, only in 7 kinds of meters
Calculate result（SVC/AdaBoost/DTC/RFC sorting algorithms program, which amounts to, calculates 4 kinds of state outcomes, three deep learning graders
Program, which amounts to, calculates three state result, altogether 7 kinds of state outcomes）In most of result of calculation same mistake is presented, can just make
The final identification state of discharge pulse produces mistake, than if any 4 correct results, 3 wrong results, then final output
State outcome be correct.In fact, discharge condition result vector has 5 kinds of states, wherein [S1, S2, S3, S4, S5], S1
For open-circuit condition, S2 is spark discharge, and S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition, by 7 kinds of state outcomes
Vector amount to together, correct state（Such as S2）Value be up to the event of maximum probability, 7 kinds of state knots calculated
In fruit, it is the very small event of probability largely to malfunction and misjudge for same state.So, it just greatly strengthen the present invention's
Fault freedom, greatly improve the robustness of apparatus and method of the present invention.
Brief description of the drawings
Fig. 1 is the structural representation of the present invention；
Fig. 2 is the principle schematic of the identification device of the present invention；
Fig. 3 is discharge waveform stretching/compressing schematic diagram；
Fig. 4 is circle queue Waveform storage schematic diagram；
Fig. 5 is structure discharge waveform schematic diagram；
Fig. 6 is that color feature value vector calculates schematic diagram；
Fig. 7 is current Single Pulse Discharge method for waveform identification flow chart；
Fig. 8 is the statistical flowsheet figure for the state status that 30 subpulses described in the 7th step discharge；
Fig. 9 is spark discharge state recognition mistake classification schematic diagram.
Embodiment
As shown in Figures 1 to 9, spark discharge condition checkout gear of the invention includes signal processing module, identification dress
Put and upper computer module；
Signal processing module includes optical coupling module, analog-to-digital conversion module, arm processor and the first hi-speed USB interface；Optical coupling module
Input be connected by partial pressure circuit with the negative and positive two-stage of electric spark machine tool, described in the connection of the output end of optical coupling module
Analog-to-digital conversion module, the analog-to-digital conversion module connect the arm processor, and arm processor connects first high speed USB and connect
Mouthful；
Identification device includes embedded main board, the second hi-speed USB interface module, CPU element, GPU units, internal storage location, solid-state
Hard disk, Ethernet interfaces and netting twine；
Second hi-speed USB interface module connects first hi-speed USB interface by USB connecting lines；Second hi-speed USB interface mould
Block is connected by output line with embedded main board, the CPU element, GPU units, internal storage location, solid state hard disc and
Ethernet interfaces are connected by both-way communication circuit with the embedded main board respectively；Ethernet interfaces pass through the net
Line is connected with upper computer module.
Linux system is stored with solid state hard disc, three deep learning grader programs, spies are installed in Linux system
Levy extraction algorithm program, SVC/AdaBoost/DTC/RFC sorting algorithms program, assembled classifier program and for controlling entirety
The main program of flow；
The input of three deep learning grader programs is the discharge waveform image of complete cycle, and the resolution ratio of the image is 256*
256 pixels；
The code name of first deep learning grader program is CNN2C, and first deep learning grader program includes convolutional layer
1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, the first full connection level routine
With the first soft recurrence subprogram（That is " Softmax regreSSion "）；
The convolutional layer 1C1 subprograms of first deep learning grader are 256*256*8 pixels and its convolution kernel is 3*3 pixels；
The pond layer 1S1 subprograms of first deep learning grader are 128*128*8 pixels；The volume of first deep learning grader
Lamination 2C2 subprograms are 128*128*32 pixels and its convolution kernel size is 3*3 pixels；First deep learning grader pond
Layer 2S2 subprograms are 64*64*32 pixels；First full connection level routine is 4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of first deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, the first full connection level routine and the first soft sub- journey of recurrence
After the processing of sequence, final output discharge condition result vector [S1, S2, S3, S4, S5], wherein S1 is open-circuit condition, and S2 is fire
Flower electric discharge, S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；The value of vector is present discharge shape probability of state；
The code name of second deep learning grader program is CNN3C, and second deep learning grader program includes convolutional layer
1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond
Change layer 3S3 subprograms, the second full connection level routine and the second soft recurrence subprogram；
The convolutional layer 1C1 subprograms of second deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of second deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of second deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of second deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of second deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of second deep learning grader are 32*32*64 pixels, and the second full connection level routine is
4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of second deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond layer 3S3 subprograms,
After two full connection level routines and the second soft processing for returning subprogram, final output discharge condition result vector [S1, S2, S3,
S4, S5], wherein S1 is open-circuit condition, and S2 is spark discharge, and S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；
The value of vector is present discharge shape probability of state；
The code name of 3rd deep learning grader program is CNNFC, and the 3rd deep learning grader program includes global figure
As processing sub-programme gather, topography's processing sub-programme gather, the 3rd full connection level routine and the 3rd soft recurrence subprogram；
The discharge waveform image for the complete cycle that 3rd deep learning grader program is received while send at global image
Manage sub-programme gather and local image procossing sub-programme gather carries out parallel processing；
Global image processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms,
Pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms；
The convolutional layer 1C1 subprograms of 3rd deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of 3rd deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of 3rd deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of 3rd deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of 3rd deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of 3rd deep learning grader are 32*32*64 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of global image processing sub-programme gather, pond layer
At 1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms
After reason, 4096 pixel datas are conveyed to the 3rd full connection level routine；
Topography processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and
Pond layer 2S2 subprograms；The discharge waveform image of the complete cycle received is divided into 4 portions by topography's processing sub-programme gather
Point, respectively the upper left of image, upper right portion, bottom left section and lower right-most portion, 4 parts are 128*128 pixels；
Convolutional layer 1C1 subprograms are 128*128*8 pixels and its convolution kernel size is 3*3 pixels；Pond layer 1S1 subprograms are
64*64*8 pixels；Convolutional layer 2C2 subprograms are 64*64*32 pixels；Pond layer 2S2 subprograms are 32*32*32 pixels；
The discharge waveform image of complete cycle handles the convolutional layer 1C1 subprograms of sub-programme gather, pond layer 1S1 by topography
When program, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms are handled, topography handles convolutional layer 1C1 of sub-programme gather
Program, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms handle the upper left quarter of described image respectively
Divide, a part in upper right portion, bottom left section and lower right-most portion（I.e. a subprogram handles a part）, image it is every
Data after individual part is processed are 1024 pixels；4 1024 pictures that topography's processing sub-programme gather will obtain after processing
The data of element are conveyed to the 3rd full connection level routine；
1 4096 pixel data and 4 1024 pixel datas that 3rd full connection level routine is received merge into 8192
Pixel data（1*4096+4*1024=8192）, 8192 pixel datas are then conveyed to the 3rd soft recurrence subprogram；3rd is soft
Subprogram output discharge condition result vector [S1, S2, S3, S4, S5] is returned, wherein S1 is open-circuit condition, and S2 is put for spark
Electricity, S3 are arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；The value of vector is present discharge shape probability of state.
The analog-to-digital conversion module is more than the analog-to-digital conversion of 12 using switching rate in more than 10Msps and sampling precision
Module；
Three deep learning graders program passes through off-line training, and the Sample Storehouse of off-line training includes mould steel（At least wrap
Containing SKD11,718 and H13 type mould steel）, titanium alloy and aluminum-base silicon carbide material.
CPU in the CPU element uses CPU of the dominant frequency in more than 2.0GHz and physics kernel more than 4；Make CPU
Unit real time execution discharge condition recognizer, CPU physics kernel operation main program and assembled classifier program, is responsible for
Task scheduling and communication task, CPU another kernel operation characteristic extraction algorithm program, CPU most latter two kernel is by negative
Load situation runs SVC/AdaBoost/DTC/RFC algorithm routines.
The GPU units use core frequency for 1067MHz, have 2400 stream processing units and video memory capacity in 4GB
GPU above, GPU units are directed to each discharge waveform, complete the calculating task of 3 kinds of deep learning grader programs in real time.
Wherein, partial pressure circuit is provided with least one resistor, and partial pressure circuit output voltage is its input electricity in the present invention
The 1% of pressure.
The invention also discloses using above-mentioned spark discharge condition checkout gear spark discharge state identification method,
Carry out according to the following steps successively：
First step is：After system starts, the Linux system of solid state hard disc storage inside, which is loaded into internal storage location, to be run, and
Start three deep learning grader programs, feature extraction algorithm program, SVC/AdaBoost/DTC/RFC sorting algorithms program,
Assembled classifier program and main program；
The voltage signal of the interpolar of negative and positive two caused by a discharge pulse in electric spark machine tool electric discharge machining process is passed through
Partial pressure circuit is input to optical coupling module, and after light-coupled isolation, access analog-to-digital conversion module is changed into data signal；
Described optical coupling module has more than 50M bandwidth, between electric spark machine tool region of discharge and identification device
Be electrically isolated completely to, and ensure the main information of the voltage signal of spark discharge by not lost after analog-to-digital conversion module；
Second step is：The data signal of analog-to-digital conversion module output passes to arm processor；
The analog-to-digital conversion module is more than the analog-to-digital conversion module of 12 using switching rate in more than 10Msps and sampling precision,
Every 0.1 μ s are sampled once, no matter for long-pulse discharge waveform（≥100μs）Or short pulse discharge waveform（≤10μs）, all
It can guarantee that enough sampled datas, it is thus possible to accurately reflect discharge voltage waveform；
Third step is：The data that arm processor receives to it carry out digital filtering and cycle detection, then passing ratio conversion
By the data compression in discharge waveform cycle or it is stretched to 256 sampled points；
Arm processor presses 256 sample point datas complete cycle after processing the discharge waveform transmission association positioned at network application layer
View is transferred to the second hi-speed USB interface by the first hi-speed USB interface, and the second hi-speed USB interface passes through the data embedded
Mainboard is sent to CPU element；
Because pulsed discharge waveform length differs and causes sampling number different in second step, therefore, it is necessary to passing ratio converts
By the data compression in discharge waveform cycle or 256 sampled points are stretched to, for long pulse waveform, the data of sampling surpass
256 points are crossed, transformation of scale is compressed transform, and for narrow pulse waveform, the data of sampling are less than 256 points, and transformation of scale is
Stretching conversion, its handling process is as shown in Figure 3.
Four steps is：CPU element receives the data that the second hi-speed USB interface transmits, and the data are deposited into internal memory
In unit, Wave data complete cycle of history is stored by the way of circle queue；Circle queue Waveform storage schematic diagram is for example attached
Shown in Fig. 4.
The size of the circle queue is no less than 1000, the prolonged discharge waveform data of available buffer.
5th step is：CPU element starts calculating task, by the current Wave data complete cycle composite diagram in internal storage location
It is sent to as after in GPU units；The composograph is the gray level image of 256*256 pixels, and magnitude of voltage is higher, and color is whiter；
Then, CPU element and GPU units are calculated simultaneously；
GPU units are by three deep learnings grader program after off-line training is read and performed in internal storage location；Three
Recognition result after deep learning grader program calculates sends CPU element back to by embedded main board again, and is stored in internal storage location
In；
CPU element opens multithreading, is face respectively respectively by calculating the characteristic value of extraction Wave data composograph complete cycle
Color characteristic, patterned feature（Haralick features）With SURF features, wherein color characteristic be 21 dimension, calculating process as shown in fig. 6,
Patterned feature is characterized as 13 dimensions, and SURF is characterized as 70 dimensions, altogether 104 dimension；
After the characteristic value for extracting image, then it is identified respectively by SVC/AdaBoost/DTC/RFC sorting algorithms program, its
The parameter E of middle AdaBoost algorithms uses three different values（E=30, E=50, E=100）Calculate three times, export respectively each time
Recognition result；The parameter D of DTC algorithms and RFC algorithms takes 6, each calculates once, exports respective recognition result respectively；
The parameter C of SVC algorithms takes 200, calculates once, exports recognition result.
Haralick features and SURF feature calculations are this area routine techniques, do not do excessive description here.AdaBoost
The parameter E and its value of algorithm, the parameter D of DTC algorithms and its value, the parameter C of SVC algorithms and its value and RFC algorithms
Parameter D and its value be this area routine techniques, do not do excessive description here.
6th step is：In 5th step the result of CPU and GPU parallel computations by assembled classifier program at
Reason, final identification state is determined using simple majority voting method, and final status data is passed through through Ethernet interfaces
Netting twine is transmitted to upper computer module；Current Single Pulse Discharge method for waveform identification flow chart is as shown in Figure 7.
Simple majority voting method refers to that SVC/AdaBoost/DTC/RFC sorting algorithms program, which amounts to, calculates 6 kinds of states
As a result (wherein：AdaBoost takes 3 kinds of different parameters [E=30, E=50, E=100]), three deep learning grader programs amount to
Three state result is calculated, altogether 9 kinds of state outcomes；In 9 kinds of state outcomes, the quantity of various state outcomes is calculated, quantity is most
More state outcomes is to be final when secondary discharge condition result.
7th step is：Above-mentioned first to the 6th step is iteratively repeated progress 30 times, and CPU element counts this 30 times electric discharge arteries and veins
The state status of punching, provide electro-discharge machining qualitative data and by the qualitative data through Ethernet interfaces, be transmitted to by netting twine
Position machine module；
It is electrical spark working after real-time discharge condition data and 30 history discharge condition situations are fed back to host computer by CPU element
The control system operation of work lathe provides foundation.
In 7th step, 30 discharge pulses are calculated with ratio of the number of pulses of regular picture in overall pulse quantity respectively
Example, the ratio of the number of pulses of short-circuit condition in overall pulse quantity, so as to by this discharge condition vector data<Pulse ratio
Example, short-circuit ratio>Through Ethernet interfaces, upper computer module is transmitted to by netting twine.The nearest 30 pulsed discharge situations system of history
Count flow chart as shown in Figure 8.
Three deep learning grader programs after off-line training by being deposited into solid state hard disc, the sample of off-line training
For open-circuit condition, spark discharge, arc discharge, transient discharge and each 5000 of short-circuit condition, grown wherein every kind of state both included
Periodic discharging waveform, also comprising short periodic discharging waveform.
By netting twine by three deep learning grader programs after host computer off-line training（Containing weighted value）Download to solid
In state hard disk, deep learning grader program is updated.
In 5th step, Wave data complete cycle is synthesized into method that gray level image uses as when discharge voltage is more than
During equal to 200V, gray value 0, when discharge voltage is less than or equal to 0V, gray value 1；Discharge voltage is less than 200V and is more than
Gray value during 0V is calculated using linear interpolation.The discharge waveform schematic diagram of structure is as shown in Figure 5.
In the present invention, the purpose of each Procedure Word is to distinguish and describe, and code name be able to can distinguished clearly in itself
On the premise of changed, therefore Procedure Word does not form limiting of its scope.
Implementation result：
In order to verify the validity of the inventive method, to three kinds of materials（SKD11,718 mould steel and titanium alloy ti6al4v）Put
Electrical waveform is identified, and the discharge waveform pulse width of electrical discharge machining is 10 ~ 400 milliseconds, and pulsewidth is 30 ~ 500 milliseconds.Test
Sample number be 1000, wherein SKD11 discharge waveforms be 330,718 mould steel discharge waveforms be 320, titanium alloy
Ti6Al4V discharge waveforms are 350.The state coverage of electric discharge five kinds of states to be identified, are respectively 200 waveforms.In order to avoid
Measurement error, experiment are repeated 3 times.
Table 1 compares for spark discharge state recognition result, and wherein EC with MC-CNN methods are patent institute of the present invention
It is proposed based on deep learning assembled classifier method, other method is the single method in assembled classifier method, in table
Data be recognition methods evaluation index, ideally, accuracy of identification andf- measure is 1.As shown in the data of table 1, use
After assembled classifier, accuracy of identification andf- measure is better than single method.
The spark discharge state recognition result of table 1 compares
Further, the accuracy of identification of the assembled classifier in table 1 andfData fluctuations of-the measure in testing three times are minimum,
It that is to say that the robustness of explanation this method is better than other method.Fig. 1 gives the electric spark of preceding 430 samples of single experiment
Discharge condition identification mistake classification schematic diagram.
It has been presented in Fig. 9 assembled classifier recognition methods（EC with MC-CNN）A kind of way of realization, by simple
Majority voting method exports final recognition result.As long as not most of method error, would not influence assembled classifier identification
The validity of method.Thus, Automatic Feature Extraction and simulation manual features extraction by spark discharge waveform image（Color
Feature, patterned feature and SURF features）, the accuracy of the invention that spark discharge state can not only be lifted, moreover it is possible to lift the inspection
Survey the robustness of device and recognition methods.
Above example is only to illustrative and not limiting technical scheme, although with reference to above-described embodiment to this hair
It is bright to be described in detail, it will be understood by those within the art that：Still the present invention can be modified or be waited
With replacing, any modification or partial replacement without departing from the spirit and scope of the present invention, the code name of each program is such as changed, its is equal
It should cover among scope of the presently claimed invention.
Claims (10)
1. spark discharge condition checkout gear, it is characterised in that：Including signal processing module, identification device and host computer mould
Block；
Signal processing module includes optical coupling module, analog-to-digital conversion module, arm processor and the first hi-speed USB interface；Optical coupling module
Input be connected by partial pressure circuit with the negative and positive two-stage of electric spark machine tool, described in the connection of the output end of optical coupling module
Analog-to-digital conversion module, the analog-to-digital conversion module connect the arm processor, and arm processor connects first high speed USB and connect
Mouthful；
Identification device includes embedded main board, the second hi-speed USB interface module, CPU element, GPU units, internal storage location, solid-state
Hard disk, Ethernet interfaces and netting twine；
Second hi-speed USB interface module connects first hi-speed USB interface by USB connecting lines；Second hi-speed USB interface mould
Block is connected by output line with embedded main board, the CPU element, GPU units, internal storage location, solid state hard disc and
Ethernet interfaces are connected by both-way communication circuit with the embedded main board respectively；Ethernet interfaces pass through the net
Line is connected with upper computer module.
2. spark discharge condition checkout gear according to claim 1, it is characterised in that：It is stored with solid state hard disc
Linux system, three deep learning grader programs, feature extraction algorithm program, SVC/ are installed in Linux system
AdaBoost/DTC/RFC sorting algorithms program, assembled classifier program and the main program for controlling overall flow；
The input of three deep learning grader programs is the discharge waveform image of complete cycle, and the resolution ratio of the image is 256*
256 pixels；
First deep learning grader program includes convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 journeys
Sequence, pond layer 2S2 subprograms, the first full connection level routine and the first soft recurrence subprogram；
The convolutional layer 1C1 subprograms of first deep learning grader are 256*256*8 pixels and its convolution kernel is 3*3 pixels；
The pond layer 1S1 subprograms of first deep learning grader are 128*128*8 pixels；The volume of first deep learning grader
Lamination 2C2 subprograms are 128*128*32 pixels and its convolution kernel size is 3*3 pixels；First deep learning grader pond
Layer 2S2 subprograms are 64*64*32 pixels；First full connection level routine is 4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of first deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, the first full connection level routine and the first soft sub- journey of recurrence
After the processing of sequence, final output discharge condition result vector [S1, S2, S3, S4, S5], wherein S1 is open-circuit condition, and S2 is fire
Flower electric discharge, S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；The value of vector is present discharge shape probability of state；
Second deep learning grader program includes convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 journeys
Sequence, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond layer 3S3 subprograms, the second full connection level routine and second are soft
Return subprogram；
The convolutional layer 1C1 subprograms of second deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of second deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of second deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of second deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of second deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of second deep learning grader are 32*32*64 pixels, and the second full connection level routine is
4096 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of second deep learning grader, pond layer
1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms, pond layer 3S3 subprograms,
After two full connection level routines and the second soft processing for returning subprogram, final output discharge condition result vector [S1, S2, S3,
S4, S5], wherein S1 is open-circuit condition, and S2 is spark discharge, and S3 is arc discharge, and S4 is transient discharge, and S5 is short-circuit condition；
The value of vector is present discharge shape probability of state；
3rd deep learning grader program includes global image processing sub-programme gather, topography's processing sub-programme gather, the
Three full connection level routines and the 3rd soft recurrence subprogram；
The discharge waveform image for the complete cycle that 3rd deep learning grader program is received while send at global image
Manage sub-programme gather and local image procossing sub-programme gather carries out parallel processing；
Global image processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms,
Pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms；
The convolutional layer 1C1 subprograms of 3rd deep learning grader are 256*256*8 pixels and its convolution kernel is 5*5 pixels；
The pond layer 1S1 subprograms of 3rd deep learning grader are 128*128*8 pixels；
The convolutional layer 2C2 subprograms of 3rd deep learning grader are 256*256*32 pixels and its convolution kernel is 5*5 pixels；
The pond layer 2S2 subprograms of 3rd deep learning grader are 64*64*32 pixels；
The convolutional layer 3C3 subprograms of 3rd deep learning grader are 64*64*64 pixels and its convolution kernel size is 3*3 pictures
Element；The pond layer 3S3 subprograms of 3rd deep learning grader are 32*32*64 pixels；
The discharge waveform image of complete cycle is successively by the convolutional layer 1C1 subprograms of global image processing sub-programme gather, pond layer
At 1S1 subprograms, convolutional layer 2C2 subprograms, pond layer 2S2 subprograms, convolutional layer 3C3 subprograms and pond layer 3S3 subprograms
After reason, 4096 pixel datas are conveyed to the 3rd full connection level routine；
Topography processing sub-programme gather include convolutional layer 1C1 subprograms, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and
Pond layer 2S2 subprograms；The discharge waveform image of the complete cycle received is divided into 4 portions by topography's processing sub-programme gather
Point, respectively the upper left of image, upper right portion, bottom left section and lower right-most portion, 4 parts are 128*128 pixels；
Convolutional layer 1C1 subprograms are 128*128*8 pixels and its convolution kernel size is 3*3 pixels；Pond layer 1S1 subprograms are
64*64*8 pixels；Convolutional layer 2C2 subprograms are 64*64*32 pixels；Pond layer 2S2 subprograms are 32*32*32 pixels；
The discharge waveform image of complete cycle handles the convolutional layer 1C1 subprograms of sub-programme gather, pond layer 1S1 by topography
When program, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms are handled, topography handles convolutional layer 1C1 of sub-programme gather
Program, pond layer 1S1 subprograms, convolutional layer 2C2 subprograms and pond layer 2S2 subprograms handle the upper left quarter of described image respectively
Point, a part in upper right portion, bottom left section and lower right-most portion, the data after each part of image is processed are
1024 pixels；Topography handles sub-programme gather and the data of obtained after processing 4 1024 pixels is conveyed into the 3rd full connection
Level routine；
1 4096 pixel data and 4 1024 pixel datas that 3rd full connection level routine is received merge into 8192
Pixel data, 8192 pixel datas are then conveyed to the 3rd soft recurrence subprogram；The 3rd soft subprogram that returns exports electric discharge shape
State result vector [S1, S2, S3, S4, S5], wherein S1 are open-circuit condition, and S2 is spark discharge, and S3 is arc discharge, and S4 is
Transient discharge, S5 are short-circuit condition；The value of vector is present discharge shape probability of state.
3. spark discharge condition checkout gear according to claim 1 or 2, it is characterised in that：The analog-to-digital conversion mould
Block is more than the analog-to-digital conversion module of 12 using switching rate in more than 10Msps and sampling precision；
Three deep learning graders program passes through off-line training, and the Sample Storehouse of off-line training includes mould steel, titanium alloy
With aluminum-base silicon carbide material.
4. spark discharge condition checkout gear according to claim 1 or 2, it is characterised in that：In the CPU element
CPU uses CPU of the dominant frequency in more than 2.0GHz and physics kernel more than 4；CPU physics kernel operation main program
With assembled classifier program, it is responsible for task scheduling and communication task, CPU another kernel operation characteristic extraction algorithm program,
CPU most latter two kernel is by loading condition operation SVC/AdaBoost/DTC/RFC algorithm routines.
5. spark discharge condition checkout gear according to claim 1 or 2, it is characterised in that：The GPU units use
Core frequency is 1067MHz, has the GPU of 2400 stream processing units and video memory capacity in more than 4GB, and GPU units are for every
Secondary discharge waveform, the calculating task of 3 kinds of deep learning grader programs is completed in real time.
6. the spark discharge state identification method of the spark discharge condition checkout gear according to claim 2, it is special
Sign is to carry out according to the following steps successively：
First step is：After system starts, the Linux system of solid state hard disc storage inside, which is loaded into internal storage location, to be run, and
Start three deep learning grader programs, feature extraction algorithm program, SVC/AdaBoost/DTC/RFC sorting algorithms program,
Assembled classifier program and main program；
The voltage signal of the interpolar of negative and positive two caused by a discharge pulse in electric spark machine tool electric discharge machining process is passed through
Partial pressure circuit is input to optical coupling module, and after light-coupled isolation, access analog-to-digital conversion module is changed into data signal；
Described optical coupling module has more than 50M bandwidth, between electric spark machine tool region of discharge and identification device
Be electrically isolated completely to, and ensure the information of the voltage signal of spark discharge by not lost after analog-to-digital conversion module；
Second step is：The data signal of analog-to-digital conversion module output passes to arm processor；
The analog-to-digital conversion module is more than the analog-to-digital conversion module of 12 using switching rate in more than 10Msps and sampling precision,
Every 0.1 μ s are sampled once, no matter for long-pulse discharge waveform（≥100μs）Or short pulse discharge waveform（≤10μs）, all
It can guarantee that enough sampled datas, it is thus possible to accurately reflect discharge voltage waveform；
Third step is：The data that arm processor receives to it carry out digital filtering and cycle detection, then passing ratio conversion
By the data compression in discharge waveform cycle or it is stretched to 256 sampled points；
Arm processor by after processing complete cycle 256 sample point datas by the first hi-speed USB interface be transferred to second at a high speed
USB interface, the second hi-speed USB interface send the data to CPU element by embedded main board；
Four steps is：CPU element receives the data that the second hi-speed USB interface transmits, and the data are deposited into internal memory
In unit, Wave data complete cycle of history is stored by the way of circle queue；
5th step is：CPU element starts calculating task, after the current Wave data complete cycle composograph in internal storage location
It is sent in GPU units；The composograph is the gray level image of 256*256 pixels；
Then, CPU element and GPU units are calculated simultaneously；
GPU units are by three deep learnings grader program after off-line training is read and performed in internal storage location；Three
Recognition result after deep learning grader program calculates sends CPU element back to by embedded main board again, and is stored in internal storage location
In；
CPU element opens multithreading, is face respectively respectively by calculating the characteristic value of extraction Wave data composograph complete cycle
Color characteristic, patterned feature and SURF features, wherein color characteristic are 21 dimensions, and patterned feature is characterized as 13 dimensions, and SURF is characterized as 70
Dimension, 104 tie up altogether；
After the characteristic value for extracting image, then it is identified respectively by SVC/AdaBoost/DTC/RFC sorting algorithms program, its
The parameter E of middle AdaBoost algorithms is calculated three times using three different values, exports the recognition result of each time respectively；DTC algorithms
Parameter D calculated twice using two different values, export the recognition result of each time respectively；
6th step is：The result of CPU and GPU parallel computations is handled by assembled classifier program in 5th step, is adopted
Final identification state is determined with simple majority voting method, and final status data is passed through into netting twine through Ethernet interfaces
It is transmitted to upper computer module；
7th step is：Above-mentioned first to the 6th step is iteratively repeated progress 30 times, and CPU element counts this 30 discharge pulses
State status, provide electro-discharge machining qualitative data and by the qualitative data through Ethernet interfaces, host computer is transmitted to by netting twine
Module；
It is electrical spark working after real-time discharge condition data and 30 history discharge condition situations are fed back to host computer by CPU element
The control system operation of work lathe provides foundation.
7. recognition methods according to claim 6, it is characterised in that：In 7th step, 30 discharge pulses are counted respectively
Calculate ratio of the number of pulses in overall pulse quantity of regular picture, the ratio of the number of pulses of short-circuit condition in overall pulse quantity
Example, so as to by this discharge condition vector data<Impulse ratio, short-circuit ratio>Through Ethernet interfaces, it is transmitted to by netting twine
Upper computer module.
8. recognition methods according to claim 6, it is characterised in that：Three deep learnings grader program is by offline
Be deposited into after training in solid state hard disc, the sample of off-line training is open-circuit condition, spark discharge, arc discharge, transient discharge and
Each 5000 of short-circuit condition, wherein every kind of state both includes long period discharge waveform, also comprising short periodic discharging waveform.
9. recognition methods according to claim 8, it is characterised in that：By netting twine by three after host computer off-line training
Deep learning grader download program is updated into solid state hard disc to deep learning grader program.
10. recognition methods according to claim 6, it is characterised in that：In 5th step, Wave data complete cycle is synthesized
The method used for gray level image for, when discharge voltage is more than or equal to 200V, gray value 0, when discharge voltage is less than or equal to
During 0V, gray value 1；Gray value when discharge voltage is less than 200V and is more than 0V is calculated using linear interpolation.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711250142.3A CN107891199B (en) | 2017-12-01 | 2017-12-01 | Spark discharge condition checkout gear and recognition methods |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711250142.3A CN107891199B (en) | 2017-12-01 | 2017-12-01 | Spark discharge condition checkout gear and recognition methods |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107891199A true CN107891199A (en) | 2018-04-10 |
| CN107891199B CN107891199B (en) | 2019-02-15 |
Family
ID=61807179
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711250142.3A Active CN107891199B (en) | 2017-12-01 | 2017-12-01 | Spark discharge condition checkout gear and recognition methods |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107891199B (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109396576A (en) * | 2018-09-29 | 2019-03-01 | 郑州轻工业学院 | Stability of EDM and power consumption state Optimal Decision-making system and decision-making technique based on deep learning |
| CN109580268A (en) * | 2018-12-22 | 2019-04-05 | 西安瑞联工业智能技术有限公司 | A kind of product abnormal sound, abnormal sound intelligent detecting method |
| CN109829408A (en) * | 2019-01-23 | 2019-05-31 | 中国科学技术大学 | Intelligent lightening recognition device based on convolutional neural networks |
| CN111558753A (en) * | 2020-05-11 | 2020-08-21 | 杭州台业机械设备有限公司 | Detection control method for slow-speed wire-feeding servo tracking voltage |
| CN111599025A (en) * | 2020-05-19 | 2020-08-28 | 湖北智旅云科技有限公司 | Touch inquiry ticketing system based on travel ticketing |
| CN114043024A (en) * | 2021-11-30 | 2022-02-15 | 大连工业大学 | Digital twin electric spark machining based cavity morphology online monitoring system and online monitoring method |
| CN114330489A (en) * | 2021-11-24 | 2022-04-12 | 江苏方天电力技术有限公司 | Fault diagnosis method and system for monitoring equipment |
| CN114367710A (en) * | 2022-02-22 | 2022-04-19 | 广州大学 | Electric spark machining control method based on deep learning and acoustic emission signals |
| WO2024002365A1 (en) * | 2022-07-01 | 2024-01-04 | 深圳市创客工场科技有限公司 | Flame recognition method, flame recognition apparatus and numerical control machine |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101474762A (en) * | 2009-01-08 | 2009-07-08 | 上海交通大学 | Electric spark clearance discharge condition detection apparatus and method based on wavelet transformation |
| JP2009166190A (en) * | 2008-01-17 | 2009-07-30 | Mitsubishi Electric Corp | Wire electric discharge machining apparatus |
| CN102248235A (en) * | 2011-05-23 | 2011-11-23 | 哈尔滨工业大学 | Device and method for detecting discharge state in electric spark linear cutting work gap |
-
2017
- 2017-12-01 CN CN201711250142.3A patent/CN107891199B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009166190A (en) * | 2008-01-17 | 2009-07-30 | Mitsubishi Electric Corp | Wire electric discharge machining apparatus |
| CN101474762A (en) * | 2009-01-08 | 2009-07-08 | 上海交通大学 | Electric spark clearance discharge condition detection apparatus and method based on wavelet transformation |
| CN102248235A (en) * | 2011-05-23 | 2011-11-23 | 哈尔滨工业大学 | Device and method for detecting discharge state in electric spark linear cutting work gap |
Non-Patent Citations (1)
| Title |
|---|
| 明五一: "精密线切割电加工建模仿真与工艺优化研究", 《中国博士学位论文全文数据库 工程科技Ⅰ辑》 * |
Cited By (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109396576A (en) * | 2018-09-29 | 2019-03-01 | 郑州轻工业学院 | Stability of EDM and power consumption state Optimal Decision-making system and decision-making technique based on deep learning |
| CN109580268A (en) * | 2018-12-22 | 2019-04-05 | 西安瑞联工业智能技术有限公司 | A kind of product abnormal sound, abnormal sound intelligent detecting method |
| CN109829408A (en) * | 2019-01-23 | 2019-05-31 | 中国科学技术大学 | Intelligent lightening recognition device based on convolutional neural networks |
| CN111558753A (en) * | 2020-05-11 | 2020-08-21 | 杭州台业机械设备有限公司 | Detection control method for slow-speed wire-feeding servo tracking voltage |
| CN111558753B (en) * | 2020-05-11 | 2021-08-03 | 杭州台业机械设备有限公司 | Detection control method for slow-speed wire-feeding servo tracking voltage |
| CN111599025A (en) * | 2020-05-19 | 2020-08-28 | 湖北智旅云科技有限公司 | Touch inquiry ticketing system based on travel ticketing |
| CN114330489A (en) * | 2021-11-24 | 2022-04-12 | 江苏方天电力技术有限公司 | Fault diagnosis method and system for monitoring equipment |
| CN114043024A (en) * | 2021-11-30 | 2022-02-15 | 大连工业大学 | Digital twin electric spark machining based cavity morphology online monitoring system and online monitoring method |
| CN114043024B (en) * | 2021-11-30 | 2022-08-30 | 大连工业大学 | Digital twin electric spark machining based cavity morphology online monitoring system and online monitoring method |
| CN114367710A (en) * | 2022-02-22 | 2022-04-19 | 广州大学 | Electric spark machining control method based on deep learning and acoustic emission signals |
| WO2024002365A1 (en) * | 2022-07-01 | 2024-01-04 | 深圳市创客工场科技有限公司 | Flame recognition method, flame recognition apparatus and numerical control machine |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107891199B (en) | 2019-02-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107891199A (en) | Spark discharge condition checkout gear and recognition methods | |
| CN109396953B (en) | Machine tool working state intelligent identification system based on signal fusion | |
| CN109917205B (en) | Electromagnetic valve fault diagnosis device and method based on feature extraction and multilayer perceptron | |
| CN114619292B (en) | Milling cutter wear monitoring method based on fusion of wavelet denoising and attention mechanism with GRU network | |
| CN109711682B (en) | Data-driven ship overall state evaluation method | |
| CN101067976B (en) | Pressurized water reactor nuclear power station control bar driving mechanism on-line monitoring and fault diagnosing method | |
| CN107553219A (en) | A kind of Tool Wear Monitoring method based on multiple types sensor composite signal | |
| CN110355608A (en) | Based on the tool abrasion prediction technique from attention mechanism and deep learning | |
| CN106404441B (en) | A kind of failure modes diagnostic method based on non-linear similarity index | |
| CN104794492B (en) | Machine tool processing running status ONLINE RECOGNITION method based on power features model | |
| CN111639461B (en) | Tool wear state detection method aiming at industrial unbalanced data | |
| CN111230740B (en) | Method and device for predicting grinding burn of aero-engine blade robot | |
| CN112192319A (en) | Tool wear monitoring method and system of unsupervised model | |
| CN109023429A (en) | A kind of aluminium cell intelligence crust breaking and intelligent feeding system and method | |
| CN110434676A (en) | A kind of boring monitoring chatter method of multisensor time-frequency characteristics fusion | |
| CN101229599B (en) | Digital pulse power source | |
| CN113554621A (en) | Tool wear state identification system and method based on wavelet scale map and depth migration learning | |
| CN109333159A (en) | The depth core extreme learning machine method and system of cutting-tool wear state on-line monitoring | |
| CN116787225A (en) | Rotary cutter abrasion monitoring method based on sound vibration signal characteristic fusion graph | |
| CN113762182A (en) | Milling cutter wear state prediction method and device based on depth network self-adaptation | |
| CN109739178A (en) | A kind of work pieces process deformation on-line prediction method of monitoring data driving | |
| CN114354184A (en) | Deep learning-based method and device for establishing health early warning model of main shaft of large-scale rotating equipment | |
| Wang et al. | Research on tool remaining life prediction method based on CNN-LSTM-PSO | |
| CN116372665A (en) | OPC UA-based aviation numerical control machine tool health monitoring method | |
| Hanjie et al. | EDM machining quality control and parameters optimization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |