The content of the invention
For above-mentioned technical deficiency, the present invention proposes a kind of method for diagnosing faults towards semiconductor manufacturing equipment, its mesh
Be:The redundancy of time of rete algorithms and structural similarity thought are dissolved into fuzzy reasoning, are used to overcome and are pushed away fuzzy
The relatively low problem of matching efficiency caused by the explosive combination of Li Shiyin knowledge, and then improve fuzzy reasoning efficiency;Build fuzzy
The self study correction mechanism of knowledge base, using in system day-to-day operation accumulate fault diagnosis sample, carry out rule intensity from
Whole fuzzy rule base is improved in study amendment, amendment, fuzzy reasoning is more had engineering practicability.
The technical scheme is that:A kind of method for diagnosing faults towards semiconductor manufacturing equipment, comprises the following steps:
1) Fuzzy processing, the fuzzy fact of generation are carried out for semiconductor equipment real-time monitoring parameter;
2) fuzzy rule base is set up;
3) the fuzzy fact is matched with the rule in fuzzy rule base using rete algorithms, is obtained fuzzy reasoning knot
Really;
4) fuzzy reasoning result is carried out into de-fuzzy and draws fault diagnosis result;
5) sample set is built according to fault diagnosis result and actual feedback result, rule intensity self-study is carried out based on sample set
Practise amendment.
Described carries out Fuzzy processing for semiconductor manufacturing equipment real-time monitoring parameter, is subordinate to using using Gauss
Degree function method, is realized by below equation:
Above formula is Gauss membership functions of the independent variable x for the monitoring parameter fuzzy collection A of semiconductor manufacturing equipment, and μ is
The Mean Parameters of monitoring of equipment parameter, σ is the variance parameter of monitoring of equipment parameter.
The fuzzy fact with the rule in fuzzy rule base match comprising the following steps by described use rete algorithms:
(1) rete fuzzy patterns are built:[PF] represents fuzzy rule former piece element, and P is parameter name, and F is fuzzy quantifier,
Then P and F are test domain, and P and F is connected with each other, and just constitute rete fuzzy patterns;
(2) rete connection networks are built:Connection net using in fuzzy rule with " and " relation rete fuzzy patterns as
One set, realizes the structure to fuzzy rule former piece, and each connection net sets a reteflag mark to record the connection
Whether the match is successful for net, if the match is successful, reteflag is true;Conversely, being then false;
(3) rete net mates:The fuzzy fact is matched with the test domain of rete fuzzy patterns, and by the match is successful
Fuzzy true storage in corresponding α registers, then the update status further according to α registers carry out reteflag more
Newly, it is the conclusion of the fuzzy rule of true using reteflag as fuzzy reasoning result.
Described is matched the fuzzy fact with the test domain of rete fuzzy patterns, and by the fuzzy fact that the match is successful
In storing corresponding α registers, then the update status further according to α registers carry out the renewal of reteflag includes following step
Suddenly:
It is the Ingress node of rete fuzzy pattern networks with parameter name test domain, proceeds by rete fuzzy pattern networks
Matching:
If the match is successful for parameter name test domain, find following fuzzy quantifier along matched chain and test domain node,
Proceed matching;
If all, the match is successful, by the fuzzy true storage of correspondence to α registers;
After the completion of rete fuzzy pattern net mate processes, just start to connect net mate:
After the α registers of certain fuzzy pattern are updated, the α registers of all patterns of traversal correspondence connection network,
And whether the match is successful to judge all patterns;If all the match is successful for the corresponding all patterns of the connection network, will
Reteflag is updated to true, using the conclusion of correspondence fuzzy rule as fuzzy reasoning result;If the connection network is corresponding
All the match is successful for all patterns, then reteflag is updated into false.
It is described rule intensity self study amendment is carried out based on sample set to comprise the following steps:
A. the error amount of the actual feedback result in sample set and result is calculated;
Whether B errors in judgement value exceeds the upper limit;
If without departing from the upper limit, amendment terminates;
Otherwise continue to determine whether beyond maximum iteration, if it was exceeded, amendment terminates;If without departing from entering
Row back-propagation, corrects the rule intensity of fuzzy rule, then recalculates the error amount of sample set;
C. return to step B, untill the error amount of sample set is reduced to higher limit or reaches maximum iteration.
The error amount is obtained by below equation:
In above formula, E is error amount, yIt is theoreticalIt is fault diagnosis result;yIt is actualIt is actual feedback result.
Described to carry out back-propagation, the rule intensity for correcting fuzzy rule is realized by following formula:
For the conclusion y of fuzzy reasoning resultj, the rule intensity w of corresponding certain the regular i of the conclusionijRegulation formula is as follows:
wij(n+1)=wij(n)+ηΔwijI=1,2 ..., M (3)
Wherein, Δ wijIt is the variable quantity of each rule intensity amendment, η is Learning Step, and n is current iteration number of times, and i is rule
Then sequence number, M is conclusion yjCorresponding fuzzy rules.
The variation delta w of each rule intensity amendmentijObtained by below equation:
Wherein, E is error amount, yIt is theoreticalIt is fault diagnosis result;yIt is actualIt is actual feedback, f (yj) it is conclusion yjCorresponding mould
Paste collection central value, yjIt is the corresponding decision confidence of conclusion, aijIt is conclusion yjThe former piece confidence level of corresponding regular i, wijIt is knot
By yjThe rule intensity of corresponding regular i.
The invention has the advantages that and advantage:
1. the present invention is obscured using Gauss membership function method when parameter fuzzy information is built to parameter
Change, with good antijamming capability, and obfuscation result is closer to the characteristics of cognition of the mankind.
2. fuzzy reasoning of the present invention based on rete algorithms is sufficiently used redundancy of time and structural similarity, saves
The memory space of fuzzy rule base, and improve Reasoning Efficiency so that fuzzy reasoning method can be suitably used for will to real-time
Seek engineering field higher.
3. the present invention corrects the rule intensity of fuzzy rule using BP algorithm, fully can diagnose sample based on fault routine
This, propagates to realize the amendment to rule intensity, and then make fuzzy rule base more conform to daily fortune using reverse error gradient
Market condition, improves the engineering practicability of expert system.
Specific embodiment
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is described in further detail.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and without
It is of the invention in limiting.
The present invention relates to a kind of method for diagnosing faults towards semiconductor manufacturing equipment, its job step is as follows:Step one,
The structure of parameter fuzzy information;The foundation of step 2, fuzzy rule base;Step 3, the fuzzy reasoning based on rete algorithms;Step
Rapid four, de-fuzzy and fault diagnosis result is drawn;Step 5, rule intensity self study amendment.With existing fuzzy reasoning side
Method is compared, and the present invention makes expert system be provided with preliminary self-learning capability by application rule intensity self study correction algorithm,
By the Reasoning Efficiency that expert system is improve using rete algorithms.Designed according to this invention expert system not only efficiency high,
Accuracy is high, and can be corrected based on routine use situation and improve fuzzy rule base, for engineer applied provide it is a kind of more
Plus practical and reliable method for diagnosing faults.
Present invention can apply to field of semiconductor manufacture, the failure towards semiconductor manufacturing equipment proposed by the present invention is examined
Disconnected method carries out software realization, then carries out software by carrier of the PC at semiconductor technology scene, server or industrial computer
Install, software is made inferences by input of the monitoring of equipment such as the light intensity of litho machine, luminous power parameter, realized by online mode
For the fault diagnosis of litho machine.
As shown in figure 1, technical scheme, specific design step is as follows:
Step one:The structure of parameter fuzzy information, specific method is:
The structure that litho machine monitors parameter fuzzy information is carried out first.Fuzzy introduction uses Gauss membership function
Method, it not only has good antijamming capability, and obfuscation result is closer to the characteristics of cognition of the mankind, its degree of membership
The expression formula of function is:
Above formula is Gauss membership functions of the independent variable x for the monitoring parameter fuzzy collection A of litho machine, and μ is supervised for litho machine
The Mean Parameters of parameter are surveyed, σ is the variance parameter that litho machine monitors parameter.
Step 2:The foundation of fuzzy rule base, specific method is:
The fuzzy rule of use represents method, and core concept is sought to traditional production rule " IF conditions THEN knots
By ", including condition obfuscation, conclusion obfuscation and Rules control, Fuzzy Production Rule can be by being expressed as below:
Rule:IFA1(f1,t1)andA2(f2,t2)and...andAn(fn,tn)THEN B(tB)CF
A1,A2,...,AnIt is knowledge former piece part, B is conclusion, they are litho machine monitoring parameter and fuzzy quantifier
Combination, we are represented with " PF ", and P is parameter name, for example light intensity, luminous power etc., and F is fuzzy quantifier, such as higher, relatively low
Deng f1,f2,...,fnBe the membership function of regular former piece, the parameter of membership function select be it is corresponding with its former piece,
t1,t2,...,tnThe former piece confidence level drawn after being calculated for membership function, tBIt is the confidence level of conclusion, CF is rule intensity, n
It is former piece number.
Step 3:Based on the fuzzy reasoning of rete algorithms, specific method is:
1) fuzzy reasoning based on rete algorithms
The litho machine monitoring parameter value that will be collected first carries out obfuscation and obtains the fuzzy fact, obscures true by ID, ginseng
Several titles, fuzzy quantifier and true confidence level are constituted, wherein true confidence level is to be input to degree of membership by by monitoring parameter value
Function is calculated.
Then the fuzzy fact is matched with the rule in fuzzy rule base using rete algorithms, is obtained fuzzy reasoning knot
Really, step is as follows:
(1) rete fuzzy patterns are built
Rete fuzzy patterns are connected in sequence by one or more test domains of fuzzy rule former piece element, each
Rete fuzzy patterns are owned by a α register, and the fact is successfully obscured with the pattern match for storing.
Fuzzy rule former piece element is represented with [PF], P is parameter name, and F is that fuzzy quantifier, then P and F are test
Domain, P and F is connected with each other, and just constitutes rete fuzzy patterns.
For certain rete fuzzy pattern, with α1As the register of the rete fuzzy patterns, when certain fuzzy true success
Have matched the P of the pattern1And F1During two test domains, then by the fuzzy true ID storages to α1In, while by the fuzzy mould
Formula is subordinate to angle value μ1It is entered as the fact that this is fuzzy true confidence level.
In order to avoid the structural redundancy of rete fuzzy patterns, and Reasoning Efficiency is improved, when rete fuzzy patterns are built,
Same test domain between each pattern can be shared.
(2) rete connection networks are built
After completing the structure of rete fuzzy patterns, then the foundation that rete connects network, rete connection networks pair are proceeded by
Should be in whole fuzzy rule base, it is made up of connection net, and each connection net corresponds to a fuzzy rule, therefore builds
The core of rete connection networks is the structure for connecting net.
Connection net using in fuzzy rule with " and " relation rete fuzzy patterns as one set, and then realize it is right
The structure of fuzzy rule former piece, each connection net can set a reteflag mark to record whether the connection net matches into
Work(, if the match is successful, reteflag is true;Conversely, being then false.
(3) rete net mates
Rete net mates are that the fuzzy fact is matched with the test domain of rete fuzzy patterns, and by the match is successful
The fuzzy fact is stored in corresponding α registers, and then the update status further according to α registers carry out the renewal of reteflag.
It is the Ingress node of rete fuzzy pattern networks with parameter name test domain, proceeds by rete fuzzy pattern networks
Matching.
If the match is successful for parameter name test domain, find following fuzzy quantifier along matched chain and test domain node,
Proceed matching;
If all, the match is successful, and correspondence is obscured into true ID stores in α registers.Due to that can be total between pattern
Identical test domain is enjoyed, therefore the matching efficiency of rete fuzzy patterns will be greatly improved.
After the completion of rete fuzzy pattern net mate processes, just start to connect net mate, mainly the α according to change is posted
Storage corresponds to reteflag to update, and embodies redundancy of time, and then improve Reasoning Efficiency.Its main flow is:When certain
After the α registers of individual fuzzy pattern are updated, the α registers of all patterns of traversal correspondence connection network, if certain pattern
α registers in have fuzzy fact ID, then pattern match success, otherwise pattern match failure is determined whether correspondingly
Whether the match is successful for all patterns of connection network;If all the match is successful for the corresponding all patterns of the connection network, will
Reteflag is updated to true, using the conclusion of correspondence fuzzy rule as fuzzy reasoning result;If the connection network is corresponding
All the match is successful for all patterns, then reteflag is updated into false.
In rete net mate processes, using Mamdani methods come the confidence level of composition rule former piece, for conclusion yjCorrespondence
Certain regular i, if it has n former piece subitem, between former piece subitem all with " and " connect, conclusion yjThe former piece of corresponding regular i
Confidence level composite formula is as follows:
WhereinFor each former piece subitem is subordinate to angle value in regular i, if the corresponding rule sums of conclusion j
It is M, conclusion yjThe rule intensity of corresponding regular i is wij, then the confidence level of conclusion j be:
Step 4:De-fuzzy simultaneously draws fault diagnosis result, and specific method is:
The system carries out de-fuzzy and processes and draw fault diagnosis result using gravity model appoach, and it takes full advantage of fuzzy pushing away
All information in reason result, the clear value for obtaining has good robustness.Computing formula is as follows:
Wherein N is the number of fuzzy reasoning result, f (yj) it is conclusion yjCorresponding fuzzy set central value, fuzzy set central value
Set according to practical situations.
Step 5:Rule intensity self study correction algorithm, specific method is:
The reverse error gradient that rule intensity self study amendment has incorporated BP (Back Propagation) algorithm propagates think of
Think, according to actual feedback result and the difference situation of fault diagnosis result, error to de-fuzzy and is drawn into reasoning knot successively
Fruit link, the fuzzy reasoning link based on rete algorithms are transmitted, and then realize the amendment to rule intensity, constantly instead
Multiple said process, until error convergence is to minimum value.
Rule intensity self study amendment flow is comprised the steps of:
A. the error amount of the actual feedback result in sample set and fault diagnosis result is calculated;
Whether B errors in judgement value exceeds the upper limit;
If without departing from the upper limit, amendment terminates;
Otherwise continue to determine whether beyond maximum iteration, if it was exceeded, amendment terminates;If without departing from entering
Row back-propagation, corrects the rule intensity of fuzzy rule, then recalculates the error total value of sample set;
C. return to step B, untill the error amount of sample set is reduced to higher limit or reaches maximum iteration.
The rule intensity self study correction algorithm of design uses batch processing mode, is with the correction algorithm of single sample below
Example illustrates its principle, and the amendment of batch processing sample is exactly adding up for single sample amendment, principle is similar.
The error function E of single sample conclusion is:
In above formula, yIt is theoreticalIt is fault diagnosis result;yIt is actualIt is actual feedback result.
For the conclusion y of fuzzy reasoning resultj, corresponding certain the regular i of the conclusion rule intensity regulation formula it is as follows:
wij(n+1)=wij(n)+ηΔwijI=1,2 ..., M (6)
In above formula, Δ wijIt is the variable quantity of each rule intensity amendment, η is Learning Step, and n is current iteration number of times, and i is
Number of regulation, M is conclusion yjCorresponding fuzzy rules.
Due to the system towards varying number level end value when, be modified using identical Learning Step, therefore be
The efficient learning efficiency of guarantee and preferably convergence effect, by the reasoning results value a common standard can be mapped to before amendment
In interval, therefore only need to set Learning Step η according to standard interval.
The quasi- interval of bidding is { std_down, std_up }, and wherein std_down is standard interval limit, and std_up is standard
The interval upper limit;If it is { obj_down, obj_up } that the codomain of current goal is interval, wherein obj_down is standard interval limit,
Obj_up is the standard interval upper limit, then mapping equation is as follows:
In above formula, obj_value is to be mapped to the numerical value before standard interval, after std_value is to be mapped to standard interval
Numerical value.
According to Feedback error algorithm, reverse error gradient is successively as shown in formula (8), formula (9).In formula (9), according to
Center-of-gravity defuzzifier formula, as N=1, by yjWith f (yj) merged, it is mainly in view of at following 2 points:
1) in view of f (yj) it is individual and yjRelated discrete function, it is impossible to derivation;
2)f(yj) the order of magnitude it is different, f (yj) inherently zooming in or out to a kind of constant amount of error.
Work as N>It is according to the feature of Center-of-gravity defuzzifier, each conclusion value is credible with conclusion with the difference of center-of-gravity value when 1
The result that degree summation is divided by approximately is transmitted as error gradient, so both take into account each conclusion value relative to center-of-gravity value
Bias factor, discrete function f (y are avoided againj) derivation problem, and confidence level is incorporated into consideration.
E is error amount, yIt is theoreticalIt is fault diagnosis result;yIt is actualIt is actual feedback, f (yj) it is conclusion yjCorresponding fuzzy set
Central value, yjIt is the corresponding decision confidence of conclusion, aijIt is conclusion yjThe former piece confidence level of corresponding regular i, wijIt is conclusion yj
The rule intensity of corresponding regular i.
As can be seen that the partial derivative in formula (8) is certainly existed.In formula (9), if denominatorIt is formula if zero
(9) partial derivative is non-existent, but in general, denominatorIt is y if zeroIt is theoreticalAlso it is just zero, illustrates that the conclusion is basic
It is not activated, it will not be also fed back, therefore the partial derivative of formula (9) is also what is existed.Due to f (yj) mark can be mapped to
Without being zero in quasi- interval, so the partial derivative in formula (10) is also certainly existed.
The workflow of the fault diagnosis expert system towards semiconductor manufacturing equipment that Fig. 1 is provided for the present invention, specifically
Method mainly includes following four step:
1) structure of parameter fuzzy information
By taking litho machine as an example, its monitoring parameter is light intensity and luminous power, is represented with A and B respectively, and fault diagnosis result is electricity
Road probability of damage, is represented with Y;
Set up the obfuscation information of parameter, the obfuscation information of light intensity and luminous power respectively as shown in table 1 to table 2, mainly
Setting including Mean Parameters and variance parameter in Gauss membership function, the obfuscation information such as table 3 of circuit probability of damage
It is shown, the corresponding central value of each fuzzy quantifier is mainly set;
The obfuscation information of the A of table 1
|
It is low |
In |
It is high |
Mean Parameters |
100 |
150 |
200 |
Variance parameter |
20.8 |
20.8 |
20.8 |
The obfuscation information of the B of table 2
|
It is low |
In |
It is high |
Mean Parameters |
100 |
150 |
200 |
Variance parameter |
20.8 |
20.8 |
20.8 |
The obfuscation information of the Y of table 3
It is extremely low |
It is relatively low |
It is moderate |
It is higher |
It is high |
0.1 |
0.3 |
0.5 |
0.7 |
0.9 |
2) structure of fuzzy rule base
For litho machine, following three typical fuzzy rules, the former piece of fuzzy rule support " bracket ", " and " and
" or " three kinds of logical symbols, when the storage of fuzzy rule is carried out, can first solve bracket, then according to before " or " left and right two ends
Fuzzy rule is split as two rules by part;
ID1:Low high rule intensity=0.9 of THEN Y=of IF A=and B=high
ID2:Moderate rule intensity=0.8 of THEN Y=in or B=in IF A=
ID3:The low and B=of IF A=extremely low rule intensity=0.9 of THEN Y=high
3) fuzzy reasoning
Data based on initial input carry out fuzzy reasoning, then de-fuzzy and draw fault diagnosis result, initial defeated
Enter as shown in table 4 with fault diagnosis result;
The initial input of table 4 and fault diagnosis result
Initial input A |
Initial input B |
The reasoning results Y |
200 |
32 |
0.814 |
160 |
40 |
0.708 |
140 |
43 |
0.546 |
As can be seen from Table 4, initial input and fault diagnosis result conform generally to the rule described by fuzzy rule, i.e. A
Ratio with B is higher, then Y is higher.
4) rule intensity self study amendment
When fault diagnosis reasoning result and actual feedback in table 4 are not inconsistent, actual feedback, such as table 5 can be submitted to
It is shown;
The contrast of the fault diagnosis result of table 5 and actual feedback
Final the reasoning results |
Actual feedback |
0.814 |
0.7 |
0.708 |
0.58 |
0.546 |
0.47 |
Then the difference according to final the reasoning results and actual feedback is come modification rule intensity, wherein, the rule of ID1 is strong
It is regular strong due to not being activated that degree has been adapted to 0.9996, ID3 by 0.9 rule intensity for being adapted to 0.15, ID2 by 0.9
Degree keeps constant, a fuzzy reasoning is then carried out again, as a result as shown in table 6;
The contrast of the revised fault diagnosis result of table 6 and desired output
Revised final the reasoning results |
Actual feedback |
0.7 |
0.7 |
0.579 |
0.58 |
0.508 |
0.47 |
As can be seen from Table 6, revised fault diagnosis result more closing to reality value of feedback, demonstrates rule intensity certainly
Study correction algorithm has preferable effect.