CN110209131A - A kind of qualitative forecasting method based on error propagation network and promotion tree algorithm - Google Patents
A kind of qualitative forecasting method based on error propagation network and promotion tree algorithm Download PDFInfo
- Publication number
- CN110209131A CN110209131A CN201910375901.1A CN201910375901A CN110209131A CN 110209131 A CN110209131 A CN 110209131A CN 201910375901 A CN201910375901 A CN 201910375901A CN 110209131 A CN110209131 A CN 110209131A
- Authority
- CN
- China
- Prior art keywords
- feature
- quality
- network
- machining
- formula
- 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
- 238000013277 forecasting method Methods 0.000 title claims abstract description 20
- 238000003754 machining Methods 0.000 claims abstract description 101
- 238000000034 method Methods 0.000 claims abstract description 47
- 238000004519 manufacturing process Methods 0.000 claims abstract description 41
- 238000012545 processing Methods 0.000 claims abstract description 36
- 230000008569 process Effects 0.000 claims abstract description 19
- 238000012797 qualification Methods 0.000 claims abstract description 18
- 238000005457 optimization Methods 0.000 claims abstract description 11
- 238000010845 search algorithm Methods 0.000 claims abstract description 8
- 238000000342 Monte Carlo simulation Methods 0.000 claims abstract description 7
- 238000011156 evaluation Methods 0.000 claims abstract description 6
- 239000002245 particle Substances 0.000 claims description 19
- 230000006870 function Effects 0.000 claims description 12
- 238000012549 training Methods 0.000 claims description 10
- 238000005516 engineering process Methods 0.000 claims description 9
- 238000009826 distribution Methods 0.000 claims description 7
- 230000001737 promoting effect Effects 0.000 claims description 6
- 238000005070 sampling Methods 0.000 claims description 5
- 230000001808 coupling effect Effects 0.000 claims description 4
- 238000004088 simulation Methods 0.000 claims description 3
- 230000035800 maturation Effects 0.000 claims 1
- 230000000007 visual effect Effects 0.000 abstract description 3
- 230000008901 benefit Effects 0.000 abstract description 2
- 230000009467 reduction Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 5
- 238000005299 abrasion Methods 0.000 description 4
- 238000005520 cutting process Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 230000002159 abnormal effect Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000003908 quality control method Methods 0.000 description 2
- 230000001568 sexual effect Effects 0.000 description 2
- 230000002547 anomalous effect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000003860 storage Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B19/00—Programme-control systems
- G05B19/02—Programme-control systems electric
- G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS], computer integrated manufacturing [CIM]
- G05B19/41875—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS], computer integrated manufacturing [CIM] characterised by quality surveillance of production
-
- G—PHYSICS
- G05—CONTROLLING; REGULATING
- G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
- G05B2219/00—Program-control systems
- G05B2219/30—Nc systems
- G05B2219/32—Operator till task planning
- G05B2219/32368—Quality control
Abstract
The present invention discloses a kind of qualitative forecasting method based on error propagation network and promotion tree algorithm, establishes the manufacturing recourses relational network based on part machining feature and machining element；The multi-process error propagation network that building manufacturing recourses relational network merges with quality sub-network；It determines the feature that outputs and inputs of quality prediction model, constructs based on error propagation network and promoted the quality prediction model of tree algorithm；Hyper parameter is optimized respectively using population and grid-search algorithms；Establish the evaluation index of model accuracy rate and maturity；Data are emulated using the production scene that monte carlo method generates, thus it is speculated that product qualification rate.The present invention realizes the visual modeling of process of producing product, it is stable to devise a kind of predictive ability, parameter is convenient for optimization, efficiency and the high qualitative forecasting method of accuracy rate, solve accurate prediction of the enterprise to product quality, processing quality is accomplished to prevent in advance and control, facilitates the reduction of enterprise-quality loss and the raising of economic benefit.
Description
Technical field
The invention belongs to processing qualities to predict field, and in particular to a kind of based on error propagation network and promotion tree algorithm
Qualitative forecasting method.
Background technique
Combined influence of the manufacturing quality of product by many factors such as people, machine, material, method, ring, surveys, and influence process
It is complicated non-linear process.The mode that enterprise mostly uses the process statuses such as SPC to monitor at present carries out pipe to part processing quality
Control is fed back and produces abnormal conditions present in process by the anomalous discrimination of ideograph, then to abnormal conditions at
Reason, the method for this part processing quality control belong to subsequent control, bring the underproof risk of product quality to enterprise.At present
Qualitative forecasting method based on intelligent algorithm largely uses SVR and BP neural network.The design of BP neural network structure depends on
Personal experience, the amount of training data needed are big.SVR predicts that hyper parameter variation range in need of consideration is big, searching process it is complicated and
It is easily trapped into local optimum.Meanwhile under multi-process complex working condition, the two has that predictive ability is limited, and precision of prediction is not high
The problem of.Enterprise is badly in need of a kind of predictive ability and stablizes, and parameter is convenient for optimization, efficiency and the high qualitative forecasting method of accuracy rate, with
Apply in the actual production under multi-process complex working condition, realize the prediction of product quality and control in advance, reduces economic loss.
Summary of the invention
The purpose of the present invention is to provide a kind of based on error propagation network and promotes the qualitative forecasting method of tree algorithm, with
It solves the above problems.
To achieve the above object, the invention adopts the following technical scheme:
A kind of qualitative forecasting method based on error propagation network and promotion tree algorithm, comprising the following steps:
Step 1, network node is extracted, determines connection relationship, constructs multi-process related with part processing quality characteristic
Error propagation network
Step 2, the error propagation network obtained according to step 1 determines the feature that outputs and inputs of quality prediction model, base
In promoting tree algorithm, the quality prediction model of tree algorithm is constructed based on error propagation network and promoted.
Step 3, the prediction model obtained according to step 2 respectively calculates boosted tree using population and grid-search algorithms
Continuous and discrete hyper parameter carries out optimizing, the quality prediction model after being optimized in method.
Step 4, the evaluation index for establishing model accuracy rate and maturity, to the superiority and inferiority for the model that appraisal procedure 3 obtains,
Guarantee model prediction quality.
Step 5, sample is generated using monte carlo method emulate data, the prediction model mature as training in step 3
Input, product qualification rate is inferred by the predicted value that exports, provides foundation for production adjustment.
Further, in step 1, according to existing between part process, transmitting coupling effect, processing technology is planned and complex web
Network is theoretical, generates manufacturing recourses relational network, specific method is:
101) part machining feature, is defined as network node, the incidence relation (benchmark, Evolvement etc.) between feature is fixed
Justice is the side of network, then part machining feature knot-net can be described as
GF=< F, EF>
In formula, F={ F1,F2,…,FnIndicating machining feature set, n indicates the quantity of manufacturing procedure, EF={ eF1,
eF2,…eFmIndicate machining feature node side collection, m indicate machining feature between existing number of edges.
102) based on, in step 101) to the knot-net of part machining feature, definition machining element is work element
For all factors for influencing machining feature quality in 5M1E theory, each machining element is defined as network node, machining element and
The incidence relation of machining feature is defined as the side of network, and the machining element network of each machining feature can be described as
GDi=< { Fi,Dij},Ei>
In formula, FiFor i-th given of machining feature node, Di={ di1,…,disIt is that the corresponding processing of machining feature is wanted
Plain nodal set, s indicate machining feature FiMachining element nodal point number, Ei={ ei1,…,eisIndicating machining element --- processing is special
Levy the side collection of network node.
103) it, is wanted with the processing of the part machining feature network and each machining feature that are formed in step 101) and step 102)
Based on plain network, the machining element network of part machining feature network and each machining feature is merged, i.e., to GFAnd GDi
It merges, manufacturing recourses relational network can be obtained:
G '={ V ', E ' }=GF∪GD1∪…∪GDn
=< F, EF> ∪ < { F1,D1j},E1> ∪ ... ∪ < { Fn,Dnj},En>
In formula, network node collection V '={ F1∪…∪Fn∪D1j∪…∪Dnj, side collection E '={ e1,…,er′, wherein
R ' is manufacturing recourses relational network number of edges.
Further, it according to the relationship of part machining feature and mass property, constructs related with part processing quality characteristic
Multi-process error propagation network.Machining feature F is based on firstly, establishingiQuality sub-network GQi, it is expressed as
GQi=< { Fi,Qi},EQi>
In formula, Qi={ Qi1,Qi2,…,QilIt is machining feature FiQualitative character collection, l indicate machining feature FiQuality
The quantity of feature.EQi={ eQi1,…,eQil, eQijIt is binary group < Fi,Qij> indicates nonoriented edge, qualitative character point QijSubordinate
In machining feature point Fi。
Finally, manufacturing recourses relational network node identical with quality sub-network is merged, error propagation network, mistake are formed
Journey is described as
G=< V, E >=GF∪GQ1∪…∪GQn=
< { F, D }, EF> ∪ < { F1,Q1},EQ1> ∪ ... ∪ < { Fn,Qn},EQn>
In formula, network node collection V={ F1∪…∪Fn∪D1∪…∪Dn∪Q1∪…∪Qn, network edge collection E={ e1,
e2,…,er, r is error propagation network number of edges.
Further, input, the output feature detailed process of quality prediction model are determined in step 2 are as follows:
201), define machining element be 5M1E theory in it is all survey link, measurable machining element deviation is summarized as
Three classes:
Error of cutter εT: it is characterized with the abrasion loss of cutter, by the abrasion loss for reflecting cutter using time indirect of cutter,
It is expressed as
In formula, the time that t ' expression cutter has used, t representation theory cutter life or time interval of whetting a knife twice.
Machine tool error εM: it is vibrated and is characterized with machine tool chief axis, vibration is divided into 11 grades, marking mode 0 is without friction；
0.1~0.3, light to vibrate, lathe many places are in finishing state；
0.4~0.6, middle vibration, lathe is in roughing, vibrates larger；
0.7~0.9, it regenerates and moves, the serious operating status of machine vibration is bad；
1.0, high vibration.
Jig error εF: jig error is that fixture steps up after part by the inclined of the actual position and ideal position of position location
Difference can be surveyed after part positioning.
Qualitative character is summarized as three classes:
Nominal dimension e1: the basic configuration and spatial position size of part are described, such as length dimension, diameter dimension and height
Size etc..
Geometric tolerance e2: the difference on the description same size shape of part, the difference of the mutual alignment between different sizes,
Form tolerance includes straightness, flatness, circularity and cylindricity etc.；Position of related features includes verticality, the depth of parallelism, gradient, right
Title degree, concentricity and position degree etc..
Surface quality e3: degree of roughness, percent ripple, Surface heat-treatent characteristic of finished surface etc. are described.
202), according to the machining element and mass property concluded in step 201), determine that the output of quality prediction model is special
Sign,
Y=eo
In formula, y is the output feature of prediction model, eoTo need the mass property deviation predicted.
Input feature vector is the machining element and mass property deviation that the mass property deviation is influenced in network,
X={ εi,ei}
In formula, x is the input feature vector set of prediction model, εi={ εTi,εMi,εFiIt is the quality spy for influencing output feature
Sexual deviation set, ei={ ei1,ei2,ei3It is the machining element deviation set for influencing output feature.
Further, quality prediction model of the building based on promoting tree algorithm in step 2, the specific steps are as follows:
21), weak learner, model are expressed as based on CART regression tree:
In formula, θ={ (R1,c1),(R2,c2),…,(RM,cM) indicate regression tree parameter, R={ R1,…,RMIt is input
Space, M are leaf node number, and the input space is divided into M unit, RmFor m-th of unit, cmFor the output of m-th of unit
Value, I (x ∈ Rm) it is indicator function.
22), based on single regression tree of step 21) construction, using preceding to Distribution Algorithm, the boosted tree mould of m step
Type indicates are as follows:
fm(x)=fm-1(x)+T(x；θm)
In formula, fm-1It (x) is "current" model, T (x；θm) it is the m regression tree.
The most optimized parameter of the m regression treeIt indicates are as follows:
In formula, N is training samples number, and r is the residual error that m-1 walks models fitting data.
Further, optimizing is carried out to hyper parameter in tree algorithm is promoted using population and grid-search algorithms in step 3,
Specific step is as follows:
Firstly, the hyper parameter having a major impact to boosted tree estimated performance is summarized as follows:
Learning rate α: successive value, 0 α≤1 <
Weak learner quantity n: discrete integer value, 0 < n
Depth capacity d: discrete integer value, 0 < d
Then, particle swarm optimization algorithm is constructed, particle iteration speed indicates are as follows:
vi+1=w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
In formula, w is inertial factor, c1、c2For Studying factors, rand1、rand2For the random number between two (0,1), vi+1
And viIndicate the speed of particle i+1 and i dimension, pbestiAnd gbestiRefer respectively to the history optimal solution of single particle and entire
The history optimal solution of population.
Particle position, which updates, to be indicated are as follows:
xi+1=xi+vi+1
In formula, xi+1And xiIndicate the position of particle i+1 and i dimension.
Finally, to the hyper parameter λ for taking successive value1Using population optimizing algorithm:
In formula,To optimize discrete hyper parameter, L (y-f (x；λ1)) it is loss function.
To the hyper parameter λ to quantize2Using grid search optimization algorithm:
In formula,To optimize discrete hyper parameter, L (y-f (x；λ2)) it is loss function, wherein λ2∈ N, N are hyper parameter
Valued space calculates loss function using enumeration strategy, the quality prediction model after finally obtaining optimization.
Further, the evaluation index of model accuracy rate and maturity is established in step 4:
Model accuracy rate index SiIt indicates are as follows:
In formula, PiFor the prediction error value of i-th of workpiece quality feature node, RiFor i-th workpiece quality feature node
Actual error.
Models mature degree index M is indicated are as follows:
In formula, N is finished work quantity, and the index expression model that maturity meets setting has already passed through enough samples
Training, can carry out error prediction.
The qualitative character required for general precision is, it is specified that the model that accuracy rate and mature indicator numerical value are more than 85% is expired
Foot training maturity and accuracy requirement can be used for actual mass prediction.Qualitative character higher for required precision, it is specified that
Accuracy rate needs that actual mass prediction can be used for more than 90% or higher with mature indicator numerical value.
Further, the production scene emulation data generated in step 5 using monte carlo method can be expressed as
X '=μ+r
In formula, x '={ x1′,…,xn' it is part input feature vector emulation set, n indicates the input feature vector number of prediction model
Amount；μ is the mean value for the correspondence input feature vector that production process is in stable state down-sampling (machined part) data；R be with
Fluctuation caused by machine factor, also referred to as white Gaussian noise, r obey mean value be 0, variance σ2Normal distribution, σ2For sampling
The variance of the correspondence input feature vector of data；
Input sample by x ' as part quality prediction model obtains corresponding output sample y ', is determined according to qualified product
Standard,
Δmin≤y′≤Δmax
In formula, ΔminFor the lower limit of variation for exporting feature, ΔmaxFor the upper deviation for exporting feature.
Part qualification rate is calculated,
In formula, P indicates qualification rate, and d is qualified product quantity, and D is total simulation times.
According to qualification rate situation, when qualification rate is unsatisfactory for requiring, need to carry out production technology or production status timely
Adjustment, accomplishes to prevent in advance and control to processing quality.
Compared with prior art, the present invention has following technical effect:
A kind of qualitative forecasting method based on error propagation network and promotion tree algorithm provided by the invention, according to part work
There is transmitting coupling effect, processing technology planning and Complex Networks Theory between sequence, by the machining feature of part, machining element and matter
Measure feature is abstracted as network node, in such a way that side connects, multi-working procedure processing course is carried out visual modeling, by being predicted
Mass property determine the network node for influencing mass property, be then based on integrated study thought according to network model, establish
Take CART tree as the boosted tree prediction model of weak learner, realizes that hyper parameter optimizing, building are quasi- by PSO and grid-search algorithms
True rate and mature indicator, the prediction effect of assessment prediction model will emulate number using model Carlow virtual production field data
It is produced according to being extrapolated according to prediction result as the input for having trained mature prediction model in prior art and technology
The qualification rate of product when qualification rate is unsatisfactory for requiring, adjusts production technology or production status in time；
First may be implemented the visual modeling of process of producing product, accomplish to influence processing quality various aspects can survey because
Plain sufficiently comprehensive consideration provides model basis for the prediction of quality of next step；
Second uses boosted tree prediction technique, and input data does not need to pre-process, and algorithm predictive ability is stablized, accuracy rate
Height, hyper parameter optimizing is convenient, the processing quality prediction that can be well adapted under multi-process complex working condition；
Third carries out optimizing to continuous and discrete hyper parameter respectively according to PSO and grid-search algorithms, is conducive to obtain mould
The optimized parameter of type guarantees model prediction quality；
4th utilizes Monte Carlo simulation principle, estimates product qualification rate under prior art condition and production environment, makes a living
It produces adjustment and foundation is provided, realize that processing quality accomplishes to prevent in advance and control, facilitate the reduction and economy of enterprise-quality loss
The raising of benefit.
Detailed description of the invention
Fig. 1 is error propagation network model schematic diagram；
Fig. 2 is the quality prediction model schematic diagram that the present invention designs；
Fig. 3 is single CART regression tree model；
Fig. 4 is the promotion tree-model that the present invention uses；
Fig. 5 is the Model Parameter Optimization flow diagram that the present invention uses；
Specific embodiment
With reference to the accompanying drawing, the present invention is described in more detail.
Fig. 1 to Fig. 5 is please referred to, the processing quality prediction technique in present invention design is proposed based on Digital manufacturing workshop
Prediction of quality new method.With the development of detection technique, the foundation in Contemporary Digital workshop is constantly universal in manufacturing enterprise,
In a large amount of data closely related with production process of the storage insides such as enterprise MES and ERP, the prediction of quality based on intelligent algorithm
Method can the flowing of simulated production process internal error, transmitting complex process, the actual production rule hidden in mining data,
Possess good practical value and development prospect.
Specifically, a kind of qualitative forecasting method based on error propagation network and promotion tree algorithm provided by the invention, packet
Include following steps:
The first step is generated according to there is transmitting coupling effect, processing technology planning and Complex Networks Theory between part process
Manufacturing recourses relational network, specific method are:
101) part machining feature, is defined as network node, incidence relation (benchmark and evolution between part machining feature
Relationship) be defined as the side of network, then part machining feature knot-net can be described as
GF=< F, EF>
In formula, F={ F1,F2,…,FnIndicating machining feature set, n indicates the quantity of manufacturing procedure, EF={ eF1,
eF2,…eFmIndicate machining feature node side collection, m indicate machining feature between existing number of edges.
102) based on, in step 101) to the knot-net of part machining feature, definition machining element is work element
For all factors for influencing machining feature quality in 5M1E theory, each machining element is defined as network node, machining element and
The incidence relation of machining feature is defined as the side of network, and the machining element network node that machining element and machining feature form can be retouched
State for
GDi=< { Fi,Dij},Ei>
In formula, FiFor i-th given of machining feature node, Di={ di1,…,disIt is that the corresponding processing of machining feature is wanted
Plain nodal set, s indicate machining feature FiMachining element nodal point number, Ei={ ei1,…,eisIndicating machining element --- processing is special
Levy the side collection of network node.
103), to be abstracted as base to the knot-net of part machining feature and machining element in step 101) and step 102)
Plinth merges the machining element network of part machining feature network and each machining feature, i.e., to figure GFAnd GDiIt merges,
Manufacturing recourses relational network is obtained, is described as follows
G '={ V ', E ' }=GF∪GD1∪…∪GDn
=< F, EF> ∪ < { F1,D1j},E1> ∪ ... ∪ < { Fn,Dnj},En>
In formula, network node collection V '={ F1∪…∪Fn∪D1j∪…∪Dnj, side collection E '={ e1,…,er′, wherein
R ' is manufacturing recourses relational network number of edges.
Manufacturing recourses relational network node scale are as follows:
In formula, n indicates part machining feature number；M indicates number of the workpiece blank feature as process positioning datum；NiWith
NkRespectively indicate part machining feature collection and machining element set；rki、rkjIndicate machining feature i and j by the same machining element k
It completes the process.
Second step constructs related with part processing quality characteristic according to the relationship of part machining feature and mass property
Multi-process error propagation network is error propagation network model schematic diagram as shown in Figure 1.Machining feature F is based on firstly, establishingi
Quality sub-network GQi, it is expressed as
GQi=< { Fi,Qi},EQi>
In formula, Qi={ Qi1,Qi2,…,QilIt is machining feature FiQualitative character collection, l indicate machining feature FiQuality
The quantity of feature.EQi={ eQi1,…,eQil, eQijIt is binary group < Fi,Qij> indicates the connection of machining feature and mass property
Side, qualitative character point QijIt is subordinated to machining feature point Fi。
Finally, manufacturing recourses relational network node identical with quality sub-network is merged, formed and part processing quality
The related multi-process error propagation network of characteristic, process description are
G=< V, E >=GF∪GQ1∪…∪GQn=
< { F, D }, EF> ∪ < { F1,Q1},EQ1> ∪ ... ∪ < { Fn,Qn},EQn>
In formula, network node collection V={ F1∪…∪Fn∪D1∪…∪Dn∪Q1∪…∪Qn, network edge collection E={ e1,
e2,…,er, r is multi-process error propagation network number of edges.
Third step is illustrated in figure 2 quality prediction model schematic diagram, it is thus necessary to determine that input, the output of quality prediction model
Feature constructs the quality prediction model based on promoting tree algorithm, detailed process are as follows:
301), define machining element be 5M1E theory in it is all survey link, measurable machining element deviation is summarized as
Three classes:
Error of cutter εT: it is characterized with the abrasion loss of cutter, by the abrasion loss for reflecting cutter using time indirect of cutter,
It is expressed as
In formula, the time that t ' expression cutter has used, t representation theory cutter life or time interval of whetting a knife twice.
Machine tool error εM: it is vibrated and is characterized with machine tool chief axis, vibration is divided into 11 grades, marking mode 0 is without friction；
0.1~0.3, light to vibrate, lathe many places are in finishing state；
0.4~0.6, middle vibration, lathe is in roughing, vibrates larger；
0.7~0.9, it regenerates and moves, the serious operating status of machine vibration is bad；
1.0, high vibration.
Jig error εF: jig error is that fixture steps up the inclined of the physical location of position location and theoretical position after part
Difference can be surveyed after part positioning.
Then, processing quality characteristic is summarized as three classes:
Nominal dimension e1: the basic configuration and spatial position size of part are described, such as length dimension, diameter dimension and height
Size etc..
Geometric tolerance e2: the difference on the description same size shape of part, the difference of the mutual alignment between different sizes,
Form tolerance includes straightness, flatness, circularity and cylindricity；Position of related features includes verticality, the depth of parallelism, gradient, symmetrical
Degree, concentricity and position degree etc..
Surface quality e3: degree of roughness, percent ripple, Surface heat-treatent characteristic of finished surface etc. are described.
302), according to the machining element and mass property concluded in step 301), determine that the output of quality prediction model is special
Sign,
Y=eo
In formula, y is the output feature of prediction model, eoTo need the mass property deviation predicted.
Input feature vector is the machining element and mass property that the mass property deviation is influenced in multi-process error propagation network
Deviation,
X={ εi,ei}
In formula, x is the input feature vector set of prediction model, εi={ εTi,εMi,εFiIt is the quality spy for influencing output feature
Sexual deviation set, ei={ ei1,ei2,ei3It is the machining element deviation set for influencing output feature；
303), quality prediction model of the building based on promoting tree algorithm:
The weak learner first based on CART regression tree is illustrated in figure 3 single CART regression tree, and model is expressed as:
In formula, θ={ (R1,c1),(R2,c2),…,(RM,cM) indicate regression tree parameter, R={ R1,…,RMIt is input
Space, M are leaf node number, and the input space is divided into M unit, RmFor m-th of unit, cmFor the output of m-th of unit
Value, I (x ∈ Rm) it is indicator function.
The prediction square error of training data indicates are as follows:
In formula, yiIndicate actual value, T (xi；θ) indicate predicted value.
Optimal cutting variable is found, it is following to solve
In formula,Indicate optimal cutting variable,Indicate optimal cut-off, R1(j, s)=and x | x(j)≤ s } and R2(j,s)
=x | x(j)> s } it is two units that cutting variable j and cut-off s are divided into.
304), as Fig. 4 be promoted tree-model, based on single regression tree of construction, using preceding to Distribution Algorithm, m
The boosted tree model of step is expressed as:
fm(x)=fm-1(x)+T(x；θm)
In formula, fm-1It (x) is "current" model, T (x；θm) it is the m regression tree；
The most optimized parameter of the m regression treeIt indicates are as follows:
In formula, N is training samples number, and r is the residual error that m-1 walks models fitting data.
4th step carries out optimizing, optimizing to hyper parameter in tree algorithm is promoted using population and grid-search algorithms
Journey is as shown in Figure 5, the specific steps are as follows:
Step 401, the hyper parameter having a major impact to boosted tree estimated performance is summarized as follows:
Learning rate α: successive value, 0 α≤1 <；
Weak learner quantity n: discrete integer value, 0 < n；
Depth capacity d: discrete integer value, 0 < d；
Step 402, particle swarm optimization algorithm is constructed, particle iteration speed indicates are as follows:
vi+1=w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
In formula, w is inertial factor, c1And c2For Studying factors, rand1With rand2For the random number between two (0,1),
vi+1And viIndicate the speed of particle i+1 and i dimension, pbestiAnd gbestiRefer respectively to the history optimal solution of single particle and whole
The history optimal solution of a population.
Particle position, which updates, to be indicated are as follows:
xi+1=xi+vi+1
In formula, xi+1And xiIndicate the position of particle i+1 and i dimension.
Step 403, to the hyper parameter λ for taking successive value1Using population optimizing algorithm:
In formula, λ1The hyper parameter that successive value is taken in S401 is represented,To optimize discrete hyper parameter, L (y-f (x；λ1)) be
Loss function.
To the hyper parameter λ to quantize2Using grid search optimization algorithm:
In formula, λ2The hyper parameter quantized in S401 is represented,To optimize discrete hyper parameter, L (y-f (x；λ2)) be
Loss function, wherein λ2∈ N, N are the valued space of hyper parameter, calculate loss function using enumeration strategy；
5th step establishes the evaluation index of model accuracy rate and maturity:
Model accuracy rate index SiIt indicates are as follows:
In formula, PiFor the prediction error value of i-th of workpiece quality feature node, RiFor i-th workpiece quality feature node
Actual error；
Models mature degree index M is indicated are as follows:
In formula, N is finished work quantity, and the index expression model that maturity meets setting has already passed through enough samples
Training, can carry out error prediction.
6th step, the production scene emulation data generated using monte carlo method are expressed as
X '=μ+r
In formula, x '={ x1′,…,xn' it is part input feature vector emulation set, n indicates the input feature vector number of prediction model
Amount；μ is the mean value for the correspondence input feature vector that production process is in stable state down-sampling (machined part) data；R be with
Fluctuation caused by machine factor, also referred to as white Gaussian noise, r obey mean value be 0, variance σ2Normal distribution, σ2For sampling
The variance of the correspondence input feature vector of data；
Input sample by x ' as part quality prediction model obtains corresponding output sample y ', is determined according to qualified product
Standard,
Δmin≤y′≤Δmax
In formula, ΔminFor the lower limit of variation for exporting feature, ΔmaxFor the upper deviation for exporting feature.
Part qualification rate is calculated,
In formula, P indicates qualification rate, and d is qualified product quantity, and D is total simulation times.
According to qualification rate situation, when qualification rate is unsatisfactory for requiring, need to carry out production technology or production status timely
Adjustment, accomplishes to prevent in advance and control to processing quality.
Claims (9)
1. a kind of qualitative forecasting method based on error propagation network and promotion tree algorithm, which comprises the following steps:
Step 1, network node is extracted, determines connection relationship, multi-process error related with part processing quality characteristic is constructed and passes
Pass network；
Step 2, the multi-process error propagation network obtained according to step 1, determine quality prediction model outputs and inputs feature,
Based on tree algorithm is promoted, the quality prediction model of tree algorithm is constructed based on error propagation network and promoted；
Step 3, the quality prediction model obtained according to step 2 respectively calculates boosted tree using population and grid-search algorithms
Continuous and discrete hyper parameter carries out optimizing, the quality prediction model after being optimized in method；
Step 4, the evaluation index for establishing model accuracy rate and maturity, to the superiority and inferiority for the quality model that appraisal procedure 3 obtains,
Guarantee the quality model forecast quality；
Step 5, sample is generated using monte carlo method and emulate data, as the defeated of the prediction model for training maturation in step 3
Enter, product qualification rate is inferred according to the predicted value of prediction model output, provides foundation for production adjustment.
2. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In, in step 1, according between part process exist transmitting coupling effect, processing technology planning and Complex Networks Theory, generate error
Network is transmitted, specific method is:
Step 101, part machining feature and machining element are defined as network node, wherein machining element is defined as 5M1E theory
The middle all factors for influencing machining feature quality, the incidence relation definition between machining feature and between machining element and machining feature
For the side of network, the manufacturing recourses relational network about part machining element and machining feature is obtained；
G'=<{ F, D }, E'>
In formula, side collection E'={ e1,…,er', machining feature node set F={ F1,…,Fn', machining element node set D=
{D1,…,Dm', wherein r' is manufacturing recourses relational network number of edges, and n' is machining feature nodal point number, and m' is machining element node
Number；
Step 102, it establishes and is based on machining feature FiQuality sub-network GQi, expression formula is
GQi=< { Fi,Qi},EQi>
In formula, Qi={ Qi1,Qi2,…,QilIt is machining feature FiQualitative character collection, l indicate machining feature FiQualitative character
Quantity；EQi={ eQi1,…,eQil, eQijIt is binary group < Fi,Qij>, indicate nonoriented edge, qualitative character point QijIt is subordinated to processing
Characteristic point Fi；
Step 103, by the resulting manufacturing recourses relational network of step 101 knot identical with the quality sub-network that step 102 obtains
Point merges, and ultimately forms error propagation network, process description is
G=<V, E>=G' ∪ GQ1∪…∪GQn=<{ F, D }, E'>∪<{ F1,Q1},EQ1>∪…∪<{Fn,Qn},EQn>
In formula, G is error propagation network, V={ F1∪…∪Fn∪D1∪…∪Dm∪Q1∪…∪QlIt is error propagation network
Nodal set, network edge collection E={ e1,e2,…,er, r is error propagation network number of edges, and n, m and l are respectively machining feature, processing
The node quantity of element and mass property.
3. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In determine quality prediction model in step 2 outputs and inputs feature detailed process are as follows:
Step 201, the input value and output valve that model uses is the deviation of machining element and processing quality characteristic, machining elements
For 5M1E theory in it is all survey link, conclude measurable machining element and measurable processing quality characteristic, measurable processing
Element is summarized as three classes, is error of cutter ε respectivelyT, machine tool error εMWith jig error εF；Measurable processing quality characteristic is concluded
It is nominal dimension e respectively for three classes1, geometric tolerance e2With surface quality e3；
Step 202, according to the machining element and processing quality characteristic concluded in step 201, the output of quality prediction model is determined
Feature,
Y=eo
In formula, y is the output feature of prediction model, eoTo need the mass property error amount predicted；
Input feature vector is the error and processing matter that the machining element of the mass property error is influenced in multi-process error propagation network
Flow characteristic deviation,
X={ εi,ei}
In formula, x is the input feature vector set of prediction model, εi={ εTi,εMi,εFiIt is the processing quality characteristic for influencing output feature
Deviation set, ei={ ei1,ei2,ei3It is the machining element deviation set for influencing output feature；
Step 203, based on input feature vector described in step 202 and output feature, the matter based on promoting tree algorithm is constructed
Prediction model is measured, single regression tree quality prediction model is obtained；
Step 204, it based on step 203 single regression tree quality prediction model of gained, using preceding to Distribution Algorithm, is mentioned
Rise tree quality prediction model.
4. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In quality prediction model of the building based on promoting tree algorithm in step 203 is specific as follows:
Firstly, the weak learner based on CART regression tree, single tree model are expressed as:
In formula, θ={ (R1,c1),(R2,c2),…,(RM,cM) indicate regression tree parameter, R={ R1,…,RMIt is that input is empty
Between, M is leaf node number, and the input space is divided into M unit, RmFor m-th of unit, cmFor the output valve of m-th of unit, I
(x∈Rm) it is indicator function；
Then, based on single regression tree, using preceding to Distribution Algorithm, the boosted tree model of m step is expressed as:
fm(x)=fm-1(x)+T(x；θm)
In formula, fm-1It (x) is "current" model, T (x；θm) it is the m regression tree；
The most optimized parameter of the m regression treeIt indicates are as follows:
In formula, N is training samples number, and r is the residual error that m-1 walks models fitting data.
5. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In using population and grid-search algorithms to hyper parameter progress optimizing in promotion tree algorithm in step 3, the specific steps are as follows:
Step 301, the hyper parameter that boosted tree estimated performance has a major impact is concluded, specific as follows:
Learning rate α: successive value, 0 α≤1 <,
Weak learner quantity n: discrete integer value, 0 < n,
Depth capacity d: discrete integer value, 0 < d,
Step 302, particle swarm optimization algorithm is constructed, particle iteration speed indicates are as follows:
vi+1=w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
In formula, w is inertial factor, c1And c2For Studying factors, rand1And rand2For the random number between two (0,1), vi+1And vi
Indicate the speed of particle i+1 and i dimension, pbestiAnd gbestiRefer respectively to single particle history optimal solution and entire particle
The history optimal solution of group；
Particle position, which updates, to be indicated are as follows:
xi+1=xi+vi+1
In formula, xi+1And xiIndicate the position of particle i+1 and i dimension；
Step 303, to the hyper parameter λ for taking successive value1Using population optimizing algorithm:
In formula, To optimize discrete hyper parameter, L (y-f (x；λ1)) it is loss function；
To the hyper parameter λ to quantize2Using grid search optimization algorithm:
In formula,To optimize discrete hyper parameter, L (y-f (x；λ2)) it is loss function, wherein λ2∈ N, N are the value of hyper parameter
Space calculates loss function using enumeration strategy, the quality prediction model after finally obtaining optimization.
6. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In establishing the evaluation index of model accuracy rate and maturity in step 4:
Model accuracy rate index SiIt indicates are as follows:
In formula, PiFor the prediction error value of i-th of workpiece quality feature node, RiFor the reality of i-th of workpiece quality feature node
Error；
Models mature degree index M is indicated are as follows:
In formula, N is finished work quantity, and the index expression model that maturity meets setting has already passed through enough sample instructions
Practice, error prediction can be carried out.
7. the qualitative forecasting method according to claim 6 based on error propagation network and promotion tree algorithm, feature exist
In the qualitative character required for general precision is, it is specified that the model that accuracy rate and mature indicator numerical value are more than 85% meets instruction
Practice maturity and accuracy requirement, can be used for actual mass prediction, qualitative character higher for required precision is, it is specified that accurate
Rate and mature indicator numerical value are not less than 90%, can be used for actual mass prediction.
8. the qualitative forecasting method according to claim 1 based on error propagation network and promotion tree algorithm, feature exist
In the production scene emulation data generated in step 5 using monte carlo method can be expressed as
X '=μ+r
In formula, x'={ x1',…,xn' it is part input feature vector emulation set, n indicates the input feature vector quantity of prediction model；μ
It is that production process is in stable state down-sampling, i.e., the mean value of the correspondence input feature vector of the data of machined part；R is random
Fluctuation caused by factor, also referred to as white Gaussian noise, r obey mean value be 0, variance σ2Normal distribution, σ2For hits
According to correspondence input feature vector variance.
9. the qualitative forecasting method according to claim 8 based on error propagation network and promotion tree algorithm, feature exist
In, using x' as the input sample of part quality prediction model, corresponding output sample y' is obtained, according to qualified product criterion,
Δmin≤y’≤Δmax
In formula, ΔminFor the lower limit of variation for exporting feature, ΔmaxFor the upper deviation for exporting feature；
Part qualification rate is calculated,
In formula, P indicates qualification rate, and d is qualified product quantity, and D is total simulation times；
According to qualification rate situation, when qualification rate is unsatisfactory for requiring, need to adjust production technology or production status in time,
Processing quality is accomplished to prevent in advance and control.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910375901.1A CN110209131B (en) | 2019-05-07 | 2019-05-07 | Quality prediction method based on error transfer network and lifting tree algorithm |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910375901.1A CN110209131B (en) | 2019-05-07 | 2019-05-07 | Quality prediction method based on error transfer network and lifting tree algorithm |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110209131A true CN110209131A (en) | 2019-09-06 |
| CN110209131B CN110209131B (en) | 2020-11-17 |
Family
ID=67785596
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910375901.1A Expired - Fee Related CN110209131B (en) | 2019-05-07 | 2019-05-07 | Quality prediction method based on error transfer network and lifting tree algorithm |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110209131B (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3382938A1 (en) * | 2017-03-31 | 2018-10-03 | Telefonica Digital España, S.L.U. | A computer implemented method, a system and computer programs to forecast the performance of a network |
| CN109214592A (en) * | 2018-10-17 | 2019-01-15 | 北京工商大学 | A kind of Air Quality Forecast method of the deep learning of multi-model fusion |
| CN109214605A (en) * | 2018-11-12 | 2019-01-15 | 国网山东省电力公司电力科学研究院 | Power-system short-term Load Probability prediction technique, apparatus and system |
| CN109523086A (en) * | 2018-11-26 | 2019-03-26 | 浙江蓝卓工业互联网信息技术有限公司 | The qualitative forecasting method and system of chemical products based on random forest |
-
2019
- 2019-05-07 CN CN201910375901.1A patent/CN110209131B/en not_active Expired - Fee Related
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP3382938A1 (en) * | 2017-03-31 | 2018-10-03 | Telefonica Digital España, S.L.U. | A computer implemented method, a system and computer programs to forecast the performance of a network |
| CN109214592A (en) * | 2018-10-17 | 2019-01-15 | 北京工商大学 | A kind of Air Quality Forecast method of the deep learning of multi-model fusion |
| CN109214605A (en) * | 2018-11-12 | 2019-01-15 | 国网山东省电力公司电力科学研究院 | Power-system short-term Load Probability prediction technique, apparatus and system |
| CN109523086A (en) * | 2018-11-26 | 2019-03-26 | 浙江蓝卓工业互联网信息技术有限公司 | The qualitative forecasting method and system of chemical products based on random forest |
Non-Patent Citations (1)
| Title |
|---|
| 江平宇: "基于赋值型误差传递网络的多工序加工质量预测", 《机械工程学报》 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110209131B (en) | 2020-11-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Cus et al. | Approach to optimization of cutting conditions by using artificial neural networks | |
| CN101694652B (en) | Network resource personalized recommended method based on ultrafast neural network | |
| CN106991506A (en) | Intelligent terminal and its stock trend forecasting method based on LSTM | |
| CN108107838A (en) | A kind of numerical control equipment tool wear monitoring method based on cloud knowledge base and machine learning | |
| CN110490415A (en) | A kind of building progress appraisal procedure of multiple no-manned plane vision collaboration | |
| CN104077432B (en) | A kind of adjusting process selection analysis method based on multidimensional correlation function | |
| CN104298818A (en) | Method for predicting and simulating errors of end-milled surface | |
| Ibanez et al. | First benchmark of the unstructured grid adaptation working group | |
| US20230325555A1 (en) | Generative design shape optimization using build material strength model for computer aided design and manufacturing | |
| US20220237344A1 (en) | Generative design shape optimization with damage prevention over loading cycles for computer aided design and manufacturing | |
| CN116305797B (en) | Construction risk assessment system based on BIM and multisource quality information fusion | |
| CN103885867B (en) | Online evaluation method of performance of analog circuit | |
| CN108415884A (en) | A kind of modal parameters real-time tracing method | |
| CN108460191B (en) | Optimization method of machine tool cross beam design scheme based on quality function expansion | |
| CN110472363A (en) | Surrouding rock deformation grade prediction technique and system suitable for Railway Tunnel | |
| Ye | An algorithm for construction project cost forecast based on particle swarm optimization-guided BP neural network | |
| CN116029682A (en) | Construction progress management method and system based on BIM technology | |
| CN115511467A (en) | Project management method and system based on engineering cost | |
| CN115237878A (en) | Process database construction method and medium based on additive manufacturing | |
| CN116703254B (en) | Production information management system for mechanical parts of die | |
| CN110209131A (en) | A kind of qualitative forecasting method based on error propagation network and promotion tree algorithm | |
| Zhang et al. | Key engineering characteristics extraction technology based on QFD | |
| CN113221278A (en) | Light weight method for vehicle-mounted photoelectric platform collimation frame | |
| Zhang et al. | Intelligent STEP-NC-compliant setup planning method | |
| CN106202694A (en) | Combination Kriging model building method based on combination forecasting method |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201117 |