CN110209131B - Quality prediction method based on error transfer network and lifting tree algorithm - Google Patents
Quality prediction method based on error transfer network and lifting tree algorithm Download PDFInfo
- Publication number
- CN110209131B CN110209131B CN201910375901.1A CN201910375901A CN110209131B CN 110209131 B CN110209131 B CN 110209131B CN 201910375901 A CN201910375901 A CN 201910375901A CN 110209131 B CN110209131 B CN 110209131B
- Authority
- CN
- China
- Prior art keywords
- quality
- processing
- network
- model
- error
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000000034 method Methods 0.000 title claims abstract description 71
- 238000012546 transfer Methods 0.000 title claims abstract description 34
- 238000012545 processing Methods 0.000 claims abstract description 114
- 238000004519 manufacturing process Methods 0.000 claims abstract description 47
- 239000002245 particle Substances 0.000 claims abstract description 30
- 238000005457 optimization Methods 0.000 claims abstract description 26
- 238000004088 simulation Methods 0.000 claims abstract description 15
- 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
- 238000003754 machining Methods 0.000 claims description 25
- 230000008569 process Effects 0.000 claims description 21
- 230000006870 function Effects 0.000 claims description 12
- 238000012797 qualification Methods 0.000 claims description 9
- 238000009826 distribution Methods 0.000 claims description 7
- 238000012549 training Methods 0.000 claims description 7
- 150000001875 compounds Chemical class 0.000 claims description 5
- 230000001808 coupling effect Effects 0.000 claims description 4
- 230000005540 biological transmission Effects 0.000 abstract description 6
- 230000009286 beneficial effect Effects 0.000 abstract description 2
- 238000013461 design Methods 0.000 abstract description 2
- 230000006872 improvement Effects 0.000 abstract description 2
- 230000009467 reduction Effects 0.000 abstract description 2
- 230000000007 visual effect Effects 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 230000002159 abnormal effect Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000005520 cutting process Methods 0.000 description 2
- 230000002950 deficient Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000010438 heat treatment Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- HDRXZJPWHTXQRI-BHDTVMLSSA-N diltiazem hydrochloride Chemical compound [Cl-].C1=CC(OC)=CC=C1[C@H]1[C@@H](OC(C)=O)C(=O)N(CC[NH+](C)C)C2=CC=CC=C2S1 HDRXZJPWHTXQRI-BHDTVMLSSA-N 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Images
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] or 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] or 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
Landscapes
- Engineering & Computer Science (AREA)
- General Engineering & Computer Science (AREA)
- Manufacturing & Machinery (AREA)
- Quality & Reliability (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- General Factory Administration (AREA)
Abstract
The invention discloses a quality prediction method based on an error transfer network and a lifting tree algorithm, which is used for establishing a manufacturing resource relation network based on part processing characteristics and processing elements; constructing a multi-process error transmission network combining a manufacturing resource relation network and a quality sub-network; determining input and output characteristics of a quality prediction model, and constructing the quality prediction model based on an error transfer network and a lifting tree algorithm; respectively optimizing the hyper-parameters by utilizing a particle swarm algorithm and a grid search algorithm; establishing evaluation indexes of model accuracy and maturity; and (4) inferring the product percent of pass by using the simulation data of the production field generated by the Monte Carlo method. The invention realizes the visual modeling of the production process of the product, designs the quality prediction method with stable prediction capability, convenient parameter optimization and high efficiency and accuracy, solves the problem of accurate prediction of the product quality of enterprises, prevents and controls the processing quality in advance, and is beneficial to the reduction of the quality loss of the enterprises and the improvement of the economic benefit.
Description
Technical Field
The invention belongs to the field of processing quality prediction, and particularly relates to a quality prediction method based on an error transfer network and a lifting tree algorithm.
Background
The production and manufacturing quality of the product is comprehensively influenced by various factors such as human, machine, material, method, ring and measurement, and the influence process is a complex nonlinear process. At present, enterprises mostly adopt SPC and other process state monitoring modes to manage and control the machining quality of parts, abnormal conditions of production existing in the machining process are fed back through abnormal judgment of a pattern diagram, and then the abnormal conditions are processed. At present, most of quality prediction methods based on intelligent algorithms adopt SVR and BP neural networks. The BP neural network structure design depends on personal experience, and the required training data volume is large. The SVR prediction needs to be considered, the change range of the over-parameters is large, the optimization process is complex, and the SVR prediction is easy to fall into local optimization. Meanwhile, under the complex working conditions of multiple processes, both the two have the problems of limited prediction capability and low prediction precision. Enterprises urgently need a quality prediction method with stable prediction capability, convenient parameter optimization and high efficiency and accuracy, so that the method is applied to actual production under the multi-process complex working conditions, the product quality prediction and advanced control are realized, and the economic loss is reduced.
Disclosure of Invention
The present invention provides a quality prediction method based on an error transfer network and a lifting tree algorithm to solve the above problems.
In order to achieve the purpose, the invention adopts the following technical scheme:
a quality prediction method based on an error transfer network and a lifting tree algorithm comprises the following steps:
step 1, extracting network nodes, determining connection relation, and constructing error transmission network of multiple procedures related to quality characteristics of part processing
And 2, determining input and output characteristics of the quality prediction model according to the error transfer network obtained in the step 1, and constructing the quality prediction model based on the error transfer network and the lifting tree algorithm based on the lifting tree algorithm.
And 3, respectively optimizing continuous and discrete hyper-parameters in the lifting tree algorithm by utilizing a particle swarm optimization and a grid search algorithm according to the prediction model obtained in the step 2 to obtain an optimized quality prediction model.
And 4, establishing evaluation indexes of the accuracy rate and the maturity of the model to evaluate the quality of the model obtained in the step 3 and ensure the prediction quality of the model.
And 5, generating sample simulation data by using a Monte Carlo method, using the sample simulation data as input of the mature prediction model trained in the step 3, deducing the product percent of pass from the output predicted value, and providing a basis for production adjustment.
Further, in step 1, a manufacturing resource relation network is generated according to the transmission coupling effect, the processing technology planning and the complex network theory existing among the part processes, and the specific method is as follows:
101) if the feature of the part machining is defined as a network node and the correlation (reference, evolutionary relationship, etc.) between features is defined as an edge of the network, the network of the feature node of the part machining can be described as
GF＝＜F,EF＞
Wherein F ═ F1,F2,…,FnDenotes a processing feature set, n denotes the number of processing steps, EF＝{eF1,eF2,…eFmAnd m represents the number of edges existing between the processing features.
102) Based on the node network of the part processing characteristics in the step 101), defining the processing elements as all the factors influencing the processing characteristic quality in the theory of 5M1E, defining each processing element as a network node, defining the incidence relation between the processing elements and the processing characteristics as the edge of the network, and describing the processing element network of each processing characteristic as the network
GDi＝＜{Fi,Dij},Ei＞
In the formula, FiFor a given ith processing feature node, Di＝{di1,…,disIs a processing element node set corresponding to the processing characteristic, s represents the processing characteristic FiNumber of processed element nodes of (E)i＝{ei1,…,eisAnd represents a processing element, namely an edge set of processing characteristic network nodes.
103) Merging the part processing feature network and the processing element network of each processing feature based on the part processing feature network and the processing element network of each processing feature formed in the steps 101) and 102), namely, merging the part processing feature network and the processing element network of each processing featureFAnd GDiMerging to obtain the manufacturing resource relation network:
G′＝{V′,E′}＝GF∪GD1∪…∪GDn
＝＜F,EF＞∪＜{F1,D1j},E1＞∪…∪＜{Fn,Dnj},En＞
wherein, the network node set V ═ F1∪…∪Fn∪D1j∪…∪DnjE, edge set E ═ E1,…,er′R' is the number of edges of the manufacturing resource relationship network.
Further, a multi-process error transfer network related to the part processing quality characteristics is constructed according to the relation between the part processing characteristics and the quality characteristics. First, a process-based feature F is establishediQuality sub-network G ofQiIs expressed as
GQi＝＜{Fi,Qi},EQi＞
In the formula, Qi＝{Qi1,Qi2,…,QilIs a machining feature FiL represents the machining feature FiThe number of quality features of (a). EQi＝{eQi1,…,eQil}，eQijIs a binary group < Fi,Qij> -ijSubject to machining feature point Fi。
Finally, the nodes with the same manufacturing resource relationship network and quality sub-network are merged to form an error transfer network, and the process is described as
G＝＜V,E＞＝GF∪GQ1∪…∪GQn＝
＜{F,D},EF＞∪＜{F1,Q1},EQ1＞∪…∪＜{Fn,Qn},EQn＞
In the formula, the network node set V ═ F1∪…∪Fn∪D1∪…∪Dn∪Q1∪…∪QnE, network edge set E ═ E1,e2,…,erR is the number of the error transfer network edges.
Further, the specific process of determining the input and output characteristics of the quality prediction model in the step 2 is as follows:
201) defining the processing elements as all measurable links in the theory of 5M1E, and summarizing the measurable deviations of the processing elements into three types:
error of cutting toolT: the wear quantity of the tool is represented by the wear quantity of the tool and indirectly reflected by the service time of the tool, and is expressed as
Where t' represents the time the tool has been used and t represents the theoretical tool life or the double sharpening time interval.
Machine tool errorM: characterizing by using the vibration of a machine tool spindle, dividing the vibration into 11 grades, and marking the vibration in a 0-grade mode without vibration;
0.1-0.3, light vibration, and the machine tool is mostly in a finish machining state;
0.4-0.6, performing medium vibration, performing rough machining on the machine tool, and performing large vibration;
0.7-0.9, heavy vibration, serious machine tool vibration and bad running state;
1.0, vibration vigorously.
Error of the clampF: the clamp error is the deviation of the actual position and the ideal position of the positioned position after the part is clamped by the clamp, and the part can be measured after being positioned.
The quality characteristics are classified into three types:
nominal dimension e1: the basic shape and spatial position dimensions of the part are described, such as length dimension, diameter dimension, height dimension, and the like.
Form and position tolerance e2: describing the difference of the shapes of the parts in the same size and the difference of the mutual positions of the parts in different sizes, wherein the shape tolerance comprises straightness, flatness, roundness, cylindricity and the like; the position tolerance comprises verticality, parallelism, inclination, symmetry, coaxiality, position degree and the like.
Surface quality e3: the roughness, waviness, surface heat treatment characteristics, and the like of the processed surface are described.
202) Determining the output characteristics of the quality prediction model based on the processing elements and the quality characteristics summarized in the step 201),
y＝eo
where y is the output characteristic of the prediction model, eoThe quality characteristic deviation value to be predicted.
The input features are processing elements and quality characteristic deviations in the network that affect the quality characteristic deviations,
x＝{i,ei}
wherein x is the input feature set of the prediction model,i＝{Ti,Mi,Fiis a set of quality characteristic deviations affecting the output characteristic, ei＝{ei1,ei2,ei3And is the set of machining element deviations that affect the output characteristics.
Further, in step 2, a quality prediction model based on a lifting tree algorithm is constructed, and the specific steps are as follows:
21) based on the CART regression tree, the weak learner has the model expression:
wherein θ { (R)1,c1),(R2,c2),…,(RM,cM) Denotes the parameters of the regression tree, R ═ R1,…,RMThe input space is divided into M units, RmIs the m-th unit, cmIs the output value of the m-th cell, I (x ∈ R)m) Is an indicator function.
22) Based on the single regression tree constructed in step 21), adopting a forward distribution algorithm, and expressing the lifting tree model in the mth step as follows:
fm(x)＝fm-1(x)+T(x；θm)
in the formula (f)m-1(x) Is the current model, T (x; θ)m) Is the m regression tree.
Optimization parameters of the mth regression tree
Expressed as:
in the formula, N is the number of training samples, and r is the residual error of the fitting data of the model in the (m-1) step.
Further, in step 3, a particle swarm and a grid search algorithm are used for optimizing the hyper-parameters in the lifting tree algorithm, and the specific steps are as follows:
firstly, the hyper-parameters that have a significant impact on the prediction performance of the lifting tree are summarized as follows:
learning rate α: continuous value, alpha is more than 0 and less than or equal to 1
Number of weak learners n: discrete integer value 0 < n
Maximum depth d: discrete integer value 0 < d
Then, constructing a particle swarm optimization algorithm, wherein the particle iteration speed is expressed as:
vi+1＝w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
wherein w is an inertia factor, c1、c2As a learning factor, rand1、rand2Is a random number between two (0,1), vi+1And viRepresenting the velocity, pbest, of the particle in dimensions i +1 and iiAnd gbestiRespectively, the historical optimal solution of a single particle and the historical optimal solution of the whole particle swarm.
The particle position update is represented as:
xi+1＝xi+vi+1
in the formula, xi+1And xiIndicating the position of the particle in dimensions i +1 and i.
Finally, the hyperparameter lambda of the continuous value is obtained1Adopting a particle swarm optimization algorithm:
in the formula (I), the compound is shown in the specification,
for optimization of the discrete hyperparameters, L (y-f (x; lambda)1) Is a loss function.
For the hyperparametric lambda of taking discrete values2Adopting a grid search optimization algorithm:
in the formula (I), the compound is shown in the specification,
for optimization of the discrete hyperparameters, L (y-f (x; lambda)2) Is a loss function, where2And E, N is a value space of the over-parameter, an enumeration strategy is adopted to calculate a loss function, and finally the optimized quality prediction model is obtained.
Further, establishing evaluation indexes of model accuracy and maturity in step 4:
model accuracy index SiExpressed as:
in the formula, PiIs a predicted error value, R, of the ith workpiece quality feature nodeiThe actual error of the ith workpiece quality characteristic node.
The model maturity index M is expressed as:
in the formula, N is the number of processed workpieces, the maturity meets the set index, the model is represented to be trained by enough samples, and error prediction can be carried out.
For the quality characteristics with general precision requirements, the model with specified accuracy and maturity index numerical value exceeding 85% meets the requirements of training maturity and accuracy, and can be used for actual quality prediction. For quality features with higher precision requirements, the specified accuracy and maturity index values need to exceed 90% or more, so that the method can be used for actual quality prediction.
Further, the simulation data of the production site generated by the Monte Carlo method in step 5 can be expressed as
x′＝μ+r
Wherein x' ═ x1′,…,xn' } is the part input feature simulation set, and n represents the number of input features of the prediction model; μ is the mean of the corresponding input features of the downsampled (machined part) data at steady state during production; r is production fluctuation caused by random factors, also called white Gaussian noise, with the mean value of r obeyed to 0 and the variance to σ2Normal distribution of (a) ("a")2Is the variance of the corresponding input feature of the sampled data;
taking x 'as an input sample of the part quality prediction model to obtain a corresponding output sample y', according to a qualified product judgment standard,
Δmin≤y′≤Δmax
in the formula,. DELTA.minAs a lower limit of deviation of the output characteristic, ΔmaxIs the upper limit of deviation of the output characteristic.
The percent of pass of the parts is calculated,
in the formula, P represents the yield, D represents the number of non-defective products, and D represents the total number of simulation times.
According to the qualification rate, when the qualification rate does not meet the requirement, the production process or the production state needs to be adjusted in time, and the processing quality is prevented and controlled in advance.
Compared with the prior art, the invention has the following technical effects:
the invention provides a quality prediction method based on an error transfer network and a lifting tree algorithm, which abstracts processing characteristics, processing elements and quality characteristics of parts into network nodes according to the existence of transfer coupling effect, processing technology planning and complex network theory among the working procedures of the parts, visually models a multi-procedure processing process in a side connection mode, determines the network nodes influencing the quality characteristics according to a network model by the predicted quality characteristics, then establishes a lifting tree prediction model taking a CART tree as a weak learner based on an integrated learning thought, realizes super-parameter optimization by PSO and a grid search algorithm, constructs accuracy and maturity indexes, evaluates the prediction effect of the prediction model, simulates production field data by using a model Carlo, takes the simulation data as the input of a trained mature prediction model, and calculates the qualification rate of products produced by the existing technology and technology according to the prediction result, when the qualified rate does not meet the requirement, the production process or the production state is adjusted in time;
firstly, the visual modeling of the production process of the product can be realized, all measurable factors influencing the processing quality are fully and comprehensively considered, and a model basis is provided for the next quality prediction;
secondly, a lifting tree prediction method is adopted, input data does not need to be preprocessed, the algorithm prediction capability is stable, the accuracy is high, the optimization of the hyper-parameters is convenient, and the method can be well suitable for the processing quality prediction under the multi-process complex working condition;
the third step is to optimize continuous and discrete hyper-parameters respectively according to PSO and grid search algorithm, which is beneficial to obtaining the optimal parameters of the model and ensuring the prediction quality of the model;
and fourthly, estimating the product percent of pass under the prior art condition and the production environment by utilizing the Monte Carlo simulation principle, providing a basis for production adjustment, realizing advanced prevention and control of processing quality, and contributing to reduction of quality loss of enterprises and improvement of economic benefits.
Drawings
FIG. 1 is a schematic diagram of an error propagation network model;
FIG. 2 is a schematic diagram of a quality prediction model designed by the present invention;
FIG. 3 is a single CART regression tree model;
FIG. 4 is a tree-lifting model employed by the present invention;
FIG. 5 is a schematic flow chart of model parameter optimization employed in the present invention;
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings.
Referring to fig. 1 to 5, the method for predicting the quality of the process according to the present invention is based on a new method for predicting the quality proposed by a digital process plant. With the development of detection technology, the establishment of the current digital workshop is continuously popularized in manufacturing enterprises, a large amount of data closely related to the production process is stored in MES, ERP and the like of the enterprises, the quality prediction method based on the intelligent algorithm can simulate the complex process of error flow and transmission in the production process, and the actual production rule hidden in the data is mined, so that the quality prediction method has good practical value and development prospect.
Specifically, the quality prediction method based on the error transfer network and the lifting tree algorithm provided by the invention comprises the following steps:
the method comprises the following steps of firstly, generating a manufacturing resource relation network according to the transmission coupling effect, the processing technology planning and the complex network theory existing among the part procedures, wherein the specific method comprises the following steps:
101) if the feature of the part machining is defined as a network node and the correlation (reference and evolution) between the feature of the part machining is defined as an edge of the network, the network of the feature node of the part machining can be described as
GF＝＜F,EF＞
Wherein F ═ F1,F2,…,FnDenotes a processing feature set, n denotes the number of processing steps, EF＝{eF1,eF2,…eFmAnd m represents the number of edges existing between the processing features.
102) Based on the node network of the part processing characteristics in the step 101), defining the processing elements as all the factors influencing the processing characteristic quality in the theory of 5M1E, defining each processing element as a network node, defining the incidence relation between the processing elements and the processing characteristics as the edge of the network, and describing the processing element network node formed by the processing elements and the processing characteristics as the network node
GDi＝＜{Fi,Dij},Ei＞
In the formula, FiFor a given ith additionI characteristic node, Di＝{di1,…,disIs a processing element node set corresponding to the processing characteristic, s represents the processing characteristic FiNumber of processed element nodes of (E)i＝{ei1,…,eisAnd represents a processing element, namely an edge set of processing characteristic network nodes.
103) Merging the part processing characteristic network and the processing element network of each processing characteristic based on the node network abstraction of the part processing characteristics and the processing elements in the step 101) and the step 102), namely, merging the part processing characteristic network and the processing element network of each processing characteristic, namely, the graph GFAnd GDiThe merging is performed to obtain a manufacturing resource relationship network, which is described as follows
G′＝{V′,E′}＝GF∪GD1∪…∪GDn
＝＜F,EF＞∪＜{F1,D1j},E1＞∪…∪＜{Fn,Dnj},En＞
Wherein, the network node set V ═ F1∪…∪Fn∪D1j∪…∪DnjE, edge set E ═ E1,…,er′R' is the number of edges of the manufacturing resource relationship network.
The manufacturing resource relation network node size is as follows:
in the formula, n represents the part processing characteristic number; m represents the number of the workpiece blank characteristics as a working procedure positioning reference; n is a radical ofiAnd NkRespectively representing a part processing characteristic set and a processing element set; r iski、rkjIndicating that the machining features i and j are machined by the same machining element k.
And secondly, constructing a multi-process error transfer network related to the part processing quality characteristics according to the relation between the part processing characteristics and the quality characteristics, wherein the error transfer network model schematic diagram is shown in figure 1. First, a process-based feature F is establishediQuality sub-network G ofQiIs expressed as
GQi＝＜{Fi,Qi},EQi＞
In the formula, Qi＝{Qi1,Qi2,…,QilIs a machining feature FiL represents the machining feature FiThe number of quality features of (a). EQi＝{eQi1,…,eQil}，eQijIs a binary group < Fi,Qij>. represents the connecting edge of the processing characteristic and the quality characteristic, the quality characteristic point QijSubject to machining feature point Fi。
Finally, the nodes of the manufacturing resource relation network and the quality sub-network which are the same are merged to form a multi-process error transmission network related to the quality characteristics of the part processing, and the process is described as
G＝＜V,E＞＝GF∪GQ1∪…∪GQn＝
＜{F,D},EF＞∪＜{F1,Q1},EQ1＞∪…∪＜{Fn,Qn},EQn＞
In the formula, the network node set V ═ F1∪…∪Fn∪D1∪…∪Dn∪Q1∪…∪QnE, network edge set E ═ E1,e2,…,erAnd r is the number of edges of the multi-procedure error transfer network.
Thirdly, as shown in fig. 2, a schematic diagram of a quality prediction model is shown, input and output characteristics of the quality prediction model need to be determined, and the quality prediction model based on the lifting tree algorithm is constructed, and the specific process is as follows:
301) defining the processing elements as all measurable links in the theory of 5M1E, and summarizing the measurable deviations of the processing elements into three types:
error of cutting toolT: the wear quantity of the tool is represented by the wear quantity of the tool and indirectly reflected by the service time of the tool, and is expressed as
Where t' represents the time the tool has been used and t represents the theoretical tool life or the double sharpening time interval.
Machine tool errorM: characterizing by using the vibration of a machine tool spindle, dividing the vibration into 11 grades, and marking the vibration in a 0-grade mode without vibration;
0.1-0.3, light vibration, and the machine tool is mostly in a finish machining state;
0.4-0.6, performing medium vibration, performing rough machining on the machine tool, and performing large vibration;
0.7-0.9, heavy vibration, serious machine tool vibration and bad running state;
1.0, vibration vigorously.
Error of the clampF: the clamp error is the deviation of the actual position and the theoretical position of the positioning position after the part is clamped by the clamp, and the part can be measured after being positioned.
Then, the processing quality characteristics are summarized into three types:
nominal dimension e1: the basic shape and spatial position dimensions of the part are described, such as length dimension, diameter dimension, height dimension, and the like.
Form and position tolerance e2: describing the difference of the shapes of the parts in the same size and the difference of the mutual positions of the parts in different sizes, wherein the shape tolerance comprises straightness, flatness, roundness and cylindricity; the position tolerance comprises verticality, parallelism, inclination, symmetry, coaxiality, position degree and the like.
Surface quality e3: the roughness, waviness, surface heat treatment characteristics, and the like of the processed surface are described.
302) Determining the output characteristics of the quality prediction model based on the processing elements and the quality characteristics summarized in the step 301),
y＝eo
where y is the output characteristic of the prediction model, eoThe quality characteristic deviation value to be predicted.
The input features are processing elements and quality characteristic deviations affecting the quality characteristic deviation in the multi-process error transfer network,
x＝{i,ei}
wherein x is the input feature set of the prediction model,i＝{Ti,Mi,Fiis a set of quality characteristic deviations affecting the output characteristic, ei＝{ei1,ei2,ei3The processing element deviation set influencing the output characteristics is adopted;
303) and constructing a quality prediction model based on a lifting tree algorithm:
first, a weak learner is based on a CART regression tree, as shown in fig. 3, a single CART regression tree is represented by:
wherein θ { (R)1,c1),(R2,c2),…,(RM,cM) Denotes the parameters of the regression tree, R ═ R1,…,RMThe input space is divided into M units, RmIs the m-th unit, cmIs the output value of the m-th cell, I (x ∈ R)m) Is an indicator function.
The predicted square error of the training data is expressed as:
in the formula, yiDenotes the actual value, T (x)i(ii) a θ) represents the predicted value.
Find the optimal segmentation variable, solve as follows
In the formula (I), the compound is shown in the specification,
the optimal cut-out variable is represented by,
represents the optimum cut point, R1(j,s)＝{x|x(j)S and R2(j,s)＝{x|x(j)S is the two cells into which the slicing variable j and the slicing point s are divided.
304) As shown in fig. 4, the lifting tree model in the mth step is represented as follows, based on the constructed single regression tree, by using a forward distribution algorithm:
fm(x)＝fm-1(x)+T(x；θm)
in the formula (f)m-1(x) Is the current model, T (x; θ)m) The mth regression tree;
optimization parameters of the mth regression tree
Expressed as:
in the formula, N is the number of training samples, and r is the residual error of the fitting data of the model in the (m-1) step.
Fourthly, optimizing the hyper-parameters in the lifting tree algorithm by utilizing a particle swarm and grid search algorithm, wherein the optimizing process is shown as a figure 5, and the concrete steps are as follows:
step 401, the hyper-parameters that have a significant impact on the prediction performance of the lifting tree are summarized as follows:
learning rate α: continuous value, alpha is more than 0 and less than or equal to 1;
number of weak learners n: discrete integer values, 0 < n;
maximum depth d: discrete integer value, 0 < d;
step 402, constructing a particle swarm optimization algorithm, wherein the particle iteration speed is expressed as:
vi+1＝w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
wherein w is inertiaFactor, c1And c2As a learning factor, rand1And, rand2Is a random number between two (0,1), vi+1And viRepresenting the velocity, pbest, of the particle in dimensions i +1 and iiAnd gbestiRespectively, the historical optimal solution of a single particle and the historical optimal solution of the whole particle swarm.
The particle position update is represented as:
xi+1＝xi+vi+1
in the formula, xi+1And xiIndicating the position of the particle in dimensions i +1 and i.
Step 403, the hyper-parameter λ of the continuous value is obtained1Adopting a particle swarm optimization algorithm:
in the formula, λ1Representing the hyperparameter of successive values in S401,
for optimization of the discrete hyperparameters, L (y-f (x; lambda)1) Is a loss function.
For the hyperparametric lambda of taking discrete values2Adopting a grid search optimization algorithm:
in the formula, λ2Representing the hyper-parameter at S401 taking discrete values,
for optimization of the discrete hyperparameters, L (y-f (x; lambda)2) Is a loss function, where2E, N is the value space of the over-parameters, and an enumeration strategy is adopted to calculate a loss function;
fifthly, establishing evaluation indexes of model accuracy and maturity:
model accuracy index SiExpressed as:
in the formula, PiIs a predicted error value, R, of the ith workpiece quality feature nodeiThe actual error of the ith workpiece quality characteristic node is obtained;
the model maturity index M is expressed as:
in the formula, N is the number of processed workpieces, the maturity meets the set index, the model is represented to be trained by enough samples, and error prediction can be carried out.
Sixthly, the simulation data of the production field generated by the Monte Carlo method is expressed as
x′＝μ+r
Wherein x' ═ x1′,…,xn' } is the part input feature simulation set, and n represents the number of input features of the prediction model; μ is the mean of the corresponding input features of the downsampled (machined part) data at steady state during production; r is production fluctuation caused by random factors, also called white Gaussian noise, with the mean value of r obeyed to 0 and the variance to σ2Normal distribution of (a) ("a")2Is the variance of the corresponding input feature of the sampled data;
taking x 'as an input sample of the part quality prediction model to obtain a corresponding output sample y', according to a qualified product judgment standard,
Δmin≤y′≤Δmax
in the formula,. DELTA.minAs a lower limit of deviation of the output characteristic, ΔmaxIs the upper limit of deviation of the output characteristic.
The percent of pass of the parts is calculated,
in the formula, P represents the yield, D represents the number of non-defective products, and D represents the total number of simulation times.
According to the qualification rate, when the qualification rate does not meet the requirement, the production process or the production state needs to be adjusted in time, and the processing quality is prevented and controlled in advance.
Claims (8)
1. A quality prediction method based on an error transfer network and a lifting tree algorithm is characterized by comprising the following steps:
step 1, extracting network nodes, determining a connection relation, and constructing a multi-process error transfer network related to the machining quality characteristics of parts;
step 2, determining input and output characteristics of a quality prediction model according to the multi-process error transfer network obtained in the step 1, and constructing the quality prediction model based on the error transfer network and a lifting tree algorithm based on the lifting tree algorithm;
step 3, optimizing continuous and discrete hyper-parameters in the lifting tree algorithm by utilizing a particle swarm and a grid search algorithm according to the quality prediction model obtained in the step 2 to obtain an optimized quality prediction model;
step 4, establishing evaluation indexes of model accuracy and maturity, and evaluating the quality of the quality model obtained in the step 3 to ensure the prediction quality of the quality model;
step 5, generating sample simulation data by using a Monte Carlo method, using the sample simulation data as input of the mature prediction model trained in the step 3, deducing the product percent of pass according to the predicted value output by the prediction model, and providing a basis for production adjustment; in step 3, a particle swarm and grid search algorithm is utilized to optimize the hyper-parameters in the lifting tree algorithm, and the specific steps are as follows:
step 301, generalizing the hyper-parameters having important influence on the prediction performance of the lifting tree, specifically as follows:
learning rate α: the continuous value, alpha is more than 0 and less than or equal to 1,
number of weak learners n: discrete integer values, 0 < n,
maximum depth d: discrete integer values, 0 < d,
step 302, constructing a particle swarm optimization algorithm, wherein the particle iteration speed is represented as:
vi+1＝w*vi+c1*rand1*(pbesti-xi)+c2*rand2*(gbesti-xi)
wherein w is an inertia factor, c1And c2As a learning factor, rand1And rand2Is a random number between two (0,1), vi+1And viRepresenting the velocity, pbest, of the particle in dimensions i +1 and iiAnd gbestiRespectively referring to the historical optimal solution of a single particle and the historical optimal solution of the whole particle swarm;
the particle position update is represented as:
xi+1＝xi+vi+1
in the formula, xi+1And xiRepresents the position of the i +1 th and i-th dimensions of the particle;
step 303, the hyper-parameter lambda of the continuous value is obtained1Adopting a particle swarm optimization algorithm:
in the formula (I), the compound is shown in the specification,
for optimization of the discrete hyperparameters, L (y-f (x; lambda)1) Is a loss function;
for the hyperparametric lambda of taking discrete values2Adopting a grid search optimization algorithm:
in the formula (I), the compound is shown in the specification,
for optimization of the discrete hyperparameters, L(y-f(x；λ2) Is a loss function, where2And E, N is a value space of the over-parameter, an enumeration strategy is adopted to calculate a loss function, and finally the optimized quality prediction model is obtained.
2. The method for predicting quality based on the error transfer network and the lifting tree algorithm according to claim 1, wherein in the step 1, the error transfer network is generated according to transfer coupling effect, processing technique planning and complex network theory existing among the part procedures, and the method comprises the following specific steps:
step 101, defining the part processing characteristics and the processing elements as network nodes, wherein the processing elements are defined as all factors influencing the quality of the processing characteristics in the 5M1E theory, and the association relations among the processing characteristics and between the processing elements and the processing characteristics are defined as edges of the network to obtain a manufacturing resource relation network related to the part processing characteristics and the processing characteristics;
G'＝<{F,D},E'>
in the formula, the edge set E ═ E1,…,er'And F, processing a characteristic node set F ═ F1,…,Fn'D, processing element node set D ═ D1,…,Dm'R ' is the number of edges of the manufacturing resource relationship network, n ' is the number of processing characteristic nodes, and m ' is the number of processing element nodes;
step 102, establishing a machining-based feature FiQuality sub-network G ofQiThe expression is
GQi＝<{Fi,Qi},EQi>
In the formula, Qi＝{Qi1,Qi2,…,QilIs a machining feature FiL represents the machining feature FiThe number of quality features of (a); eQi＝{eQi1,…,eQil}，eQijIs a doublet<Fi,Qij>Denotes a nondirectional edge, quality feature point QijSubject to machining feature point Fi；
Step 103, merging the nodes of the manufacturing resource relation network obtained in step 101 and the quality sub-network obtained in step 102 to finally form an error transfer network, wherein the process is described as
G＝<V,E>＝G’∪GQ1∪…∪GQn＝<{F,D},E'>∪<{F1,Q1},EQ1>∪…∪<{Fn,Qn},EQn>
Wherein G is an error transfer network, and V is { F ═ F1∪…∪Fn∪D1∪…∪Dm∪Q1∪…∪QlThe node set of the error transfer network is, and the network edge set E is { E }1,e2,…,erR is the number of edges of the error transfer network, and n, m and l are the number of nodes of the processing characteristic, the processing element and the quality characteristic respectively.
3. The quality prediction method based on the error transfer network and the lifting tree algorithm as claimed in claim 1, wherein the specific process of determining the input and output characteristics of the quality prediction model in the step 2 is as follows:
step 201, the input value and the output value adopted by the model are deviation values of the processing elements and the processing quality characteristics, the processing elements are all measurable links in the theory of 5M1E, the measurable processing elements and the measurable processing quality characteristics are summarized, and the measurable processing elements are summarized into three categories, namely tool errorTError of machine toolMAnd clamping errorF(ii) a The measurable process quality characteristics are classified into three categories, namely the nominal dimension e1Form and position tolerance e2And surface quality e3；
Step 202, determining the output characteristics of the quality prediction model according to the processing elements and the processing quality characteristics summarized in the step 201,
y＝eo
where y is the output characteristic of the prediction model, eoError value of quality characteristic to be predicted;
the input features are errors of processing elements and processing quality characteristic deviations affecting the quality characteristic errors in the multi-process error transfer network,
x＝{i,ei}
wherein x is the input feature set of the prediction model,i＝{Ti,Mi,Fiis a set of process quality characteristic deviations affecting the output characteristics, ei＝{ei1,ei2,ei3The processing element deviation set influencing the output characteristics is adopted;
step 203, constructing a quality prediction model based on a lifting tree algorithm on the basis of the input features and the output features in the step 202 to obtain a single regression tree quality prediction model;
and step 204, obtaining a prediction model for improving the quality of the tree by adopting a forward distribution algorithm on the basis of the single regression tree quality prediction model obtained in the step 203.
4. The method of claim 3, wherein the step 203 is implemented by constructing a quality prediction model based on the lifting tree algorithm, specifically as follows:
first, with a CART regression tree as a basis for a weak learner, a single tree model is represented as:
wherein θ { (R)1,c1),(R2,c2),…,(RM,cM) Denotes the parameters of the regression tree, R ═ R1,…,RMThe input space is divided into M units, RmIs the m-th unit, cmIs the output value of the m-th cell, I (x ∈ R)m) Is an indicator function;
then, based on a single regression tree, adopting a forward distribution algorithm, and expressing the lifting tree model in the mth step as:
fm(x)＝fm-1(x)+T(x；θm)
in the formula (f)m-1(x) Is the current model, T (x; θ)m) The mth regression tree;
optimization parameters of the mth regression tree
Expressed as:
in the formula, N is the number of training samples, and r is the residual error of the fitting data of the model in the (m-1) step.
5. The quality prediction method based on the error transfer network and the lifting tree algorithm as claimed in claim 1, wherein the evaluation indexes of model accuracy and maturity are established in step 4:
model accuracy index SiExpressed as:
in the formula, PiIs a predicted error value, R, of the ith workpiece quality feature nodeiThe actual error of the ith workpiece quality characteristic node is obtained;
the model maturity index M is expressed as:
in the formula, N is the number of processed workpieces, the maturity meets the set index, the model is represented to be trained by enough samples, and error prediction can be carried out.
6. The quality prediction method based on the error transfer network and the lifting tree algorithm as claimed in claim 5, wherein for the quality features with general accuracy requirements, the model with specified accuracy and maturity index value over 85% meets the training maturity and accuracy requirements, and can be used for actual quality prediction, and for the quality features with higher accuracy requirements, the specified accuracy and maturity index value is not less than 90%, so that the model can be used for actual quality prediction.
7. The method of claim 1, wherein the simulation data of the production site generated by the Monte Carlo method in step 5 is expressed as
x’＝μ+r
Wherein x' ═ x1',…,xn' } is the part input feature simulation set, and n represents the number of input features of the prediction model; μ is the mean value of the corresponding input features of the data of the machined part, which is sampled in a stable state during the production process; r is production fluctuation caused by random factors, also called white Gaussian noise, with the mean value of r obeyed to 0 and the variance to σ2Normal distribution of (a) ("a")2Is the variance of the corresponding input feature of the sampled data.
8. The method of claim 7, wherein x 'is used as an input sample of the part quality prediction model to obtain a corresponding output sample y', and according to the qualification criterion,
Δmin≤y’≤Δmax
in the formula,. DELTA.minAs a lower limit of deviation of the output characteristic, ΔmaxIs the upper limit of deviation of the output characteristic;
the percent of pass of the parts is calculated,
in the formula, P represents the qualification rate, D represents the qualified product quantity, and D represents the total simulation times;
according to the qualification rate, when the qualification rate does not meet the requirement, the production process or the production state needs to be adjusted in time, and the processing quality is prevented and controlled in advance.
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 CN110209131A (en) | 2019-09-06 |
| CN110209131B true 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) |
Family Cites Families (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 |
| CN109214592B (en) * | 2018-10-17 | 2022-03-08 | 北京工商大学 | Multi-model-fused deep learning air quality prediction method |
| 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 |
| CN109523086B (en) * | 2018-11-26 | 2021-08-24 | 浙江蓝卓工业互联网信息技术有限公司 | Quality prediction method and system for chemical products based on random forest |
-
2019
- 2019-05-07 CN CN201910375901.1A patent/CN110209131B/en not_active Expired - Fee Related
Also Published As
| Publication number | Publication date |
|---|---|
| CN110209131A (en) | 2019-09-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11761930B2 (en) | Prediction method of part surface roughness and tool wear based on multi-task learning | |
| CN111563301A (en) | Thin-wall part milling parameter optimization method | |
| CN108107838A (en) | A kind of numerical control equipment tool wear monitoring method based on cloud knowledge base and machine learning | |
| Mahdavinejad et al. | Optimization of milling parameters using artificial neural network and artificial immune system | |
| CN104615077A (en) | Efficient energy-saving optimizing method for numerical control milling processing process parameters based on Taguchi method | |
| CN111366123A (en) | Part surface roughness and cutter wear prediction method based on multi-task learning | |
| Li et al. | An effective PSO-LSSVM-based approach for surface roughness prediction in high-speed precision milling | |
| CN107627152B (en) | Numerical control machining chip control method based on BP neural network | |
| CN109885900A (en) | A kind of Cutting parameters optimization method based on gray analysis | |
| CN113408195B (en) | Method for predicting wear state of cutter in process | |
| Ramezani et al. | Surface roughness and cutting force estimation in the CNC turning using artificial neural networks | |
| Topal | The role of stepover ratio in prediction of surface roughness in flat end milling | |
| CN115169453A (en) | Hot continuous rolling width prediction method based on density clustering and depth residual error network | |
| Bhuiyan et al. | Optimization of cutting parameters in turning process | |
| CN110209131B (en) | Quality prediction method based on error transfer network and lifting tree algorithm | |
| Mgherony et al. | Comparison of surface roughness when turning and milling | |
| CN117422373A (en) | Cutter inventory classification comprehensive evaluation method based on combined weighting-ABC | |
| WO2021253689A1 (en) | Multiple regression model-based method and system for predicting price of product processing | |
| CN102930337A (en) | Prediction method for cutting amount in cutting machining | |
| CN109648397B (en) | Broach life prediction method based on correlation between width of cutting edge strip and broaching load | |
| Wang et al. | An accurate tool wear prediction method under different cutting conditions based on network architecture search | |
| Hossain et al. | Development of an artificial neural network algorithm for predicting the surface roughness in end milling of Inconel 718 alloy | |
| Krizbergs et al. | Methods for prediction of the surface roughness 3D parameters according to technological parameters | |
| Li et al. | Tool Wear Monitoring Technology Based on Deep Learning of Machining State Features | |
| Ma et al. | Research on Optimization of NC Milling Machine Process Parameters Based on FBGA Algorithm |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201117 |