CN110018670B - Industrial process abnormal working condition prediction method based on dynamic association rule mining - Google Patents
Industrial process abnormal working condition prediction method based on dynamic association rule mining Download PDFInfo
- Publication number
- CN110018670B CN110018670B CN201910244877.8A CN201910244877A CN110018670B CN 110018670 B CN110018670 B CN 110018670B CN 201910244877 A CN201910244877 A CN 201910244877A CN 110018670 B CN110018670 B CN 110018670B
- Authority
- CN
- China
- Prior art keywords
- association rule
- data
- window
- item set
- training
- 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.)
- Active
Links
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
- G05B19/18—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form
- G05B19/406—Numerical control [NC], i.e. automatically operating machines, in particular machine tools, e.g. in a manufacturing environment, so as to execute positioning, movement or co-ordinated operations by means of programme data in numerical form characterised by monitoring or safety
- G05B19/4065—Monitoring tool breakage, life or condition
-
- 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
- G05B2219/37—Measurements
- G05B2219/37616—Use same monitoring tools to monitor tool and workpiece
Landscapes
- Engineering & Computer Science (AREA)
- Human Computer Interaction (AREA)
- Manufacturing & Machinery (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Automation & Control Theory (AREA)
- Testing Or Calibration Of Command Recording Devices (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses an industrial process abnormal working condition prediction method based on dynamic association rule mining, which can be applied to fault prediction and health management of an industrial process. The method utilizes a sliding window mode to mine the dynamic association rule of the operation parameters of the industrial equipment and introduces the dynamic association rule into the prediction of the abnormal working conditions in the industrial process. The invention considers the time characteristic of the association rule, limits the data length by using a sliding window, provides an association rule mining algorithm for mining the dynamic association rule of operating parameters pairwise, and then introduces the association rule mining result into wavelet neural network prediction to continuously update the network by the dynamic association rule so as to obtain a more accurate prediction result. The method has great application value for fault prediction and health management in engineering.
Description
Technical Field
The invention belongs to the technical field of reliability maintenance engineering, and relates to an industrial process abnormal working condition prediction method based on dynamic association rule mining.
Background
With the continuous emergence of complex systems and the increasing demand of real-time monitoring of industrial processes, modern industrial equipment is often equipped with a plurality of sensors to monitor the operation state of the industrial equipment in the operation process. Meanwhile, multiple fault modes may occur in the operation process of the equipment, a certain fault may correspond to a plurality of symptoms, and under the condition, the single sensor information cannot completely reflect the operation state of the equipment, so that fault prediction based on multi-sensor information is generated at the right moment. The failure prediction based on multi-sensor information aims to analyze the operation state of the equipment using comprehensive sensor information, thereby making more reliable equipment diagnosis and prediction. With the continuous development of sensing technology, the use of multiple sensors for condition monitoring, fault diagnosis and prediction of equipment has become a trend.
In the field of fault prediction, the work of combining association rule mining and fault prediction is still fresh at present. Conventional association rule mining algorithms consider that association rules found for an object system are invariant, i.e., valid when analyzing other data sets of the system. The association rule mined in the way does not take the time factor into consideration, so that a static association rule is obtained. In fact, as the data is updated, it may be accompanied by changes in the operating state of the system, and in turn, changes in the characteristics of the data, and thus the association rules. During the operation of the industrial plant, the operating state is not constant. As the data is updated or when the operation state of the equipment changes, the characteristics of the operation parameters of the equipment also change, so that the relevance of the operation parameters may also change. In this case, it is obviously not reasonable to use the association rules mined from the original dataset to guide the subsequent decision.
Disclosure of Invention
Aiming at the current situation of the prior art, the invention aims to solve the problem that dynamic association of sensor data is rarely considered in the existing data-driven prediction technology, provides an equipment abnormal working condition prediction method based on an operation parameter dynamic association rule, and updates a wavelet neural network by using the dynamic association rule and performs abnormal working condition prediction (fault prediction).
The concept of the present invention will now be explained as follows:
the invention introduces dynamic association rule mining aiming at the problem that the association between the operation parameters in the operation process of the industrial equipment is likely to change along with the change of the operation state of the equipment. The invention utilizes a sliding window mode to carry out association rule mining on the data in each window, thereby obtaining the dynamically changed association rule along with the sliding of the window. In order to obtain the complete change condition of the parameter relevance, the method provided by the invention does not set the minimum support degree and confidence degree threshold value during the mining of the association rule of the parameter level, but stores the association rules of all the parameters in pairs and describes the strength of the association rules with the support degree and the confidence degree. And then, introducing the dynamic association rule mining result into the training and updating of the wavelet neural network, and predicting the abnormal working condition through the association parameters obtained by mining the dynamic association rule.
According to the invention concept, the invention provides an industrial process abnormal working condition prediction method based on dynamic association rule mining, which comprises the following specific steps:
step 1: carrying out data preprocessing on a measured value time sequence acquired by a sensor in an industrial process based on a sliding window, and carrying out piecewise linearization representation, clustering and symbolization on the time sequence to generate a transaction set suitable for association rule mining;
step 2: generating a frequent item set by adopting a two-stage frequent item set generation strategy, and mining association rules of every two parameters;
and step 3: performing association rule mining by using an initial data set, and training an initial wavelet neural network model based on an initial association rule mining result;
and 4, step 4: sliding a data window based on data updating, and updating an association rule and a wavelet neural network model for a new data window;
and 5: and predicting abnormal working conditions based on the updated wavelet network model, and continuously updating the association rule and the wavelet neural network model before predicting the occurrence of the abnormal working conditions.
Based on the above scheme, the following implementation manner can be specifically adopted for each step:
preferably, the step 1 is specifically realized by the following sub-steps:
step 1.1: the sensor measurement sequence is recorded asN is the number of sensors, K is the sequence length, K is 1,2, …, K, the length of the sliding window is L, the sliding distance is l each time, and the sliding window is WkThe data contained in the window is
Step 1.2: and performing time series piecewise linearization representation, clustering and symbolization processing on each data window, wherein the specific process is as follows:
for each data window WkPerforming the following processes of 1.2.1-1.2.5:
1.2.1. setting the initial fitting starting point toEnd point ish is 2; fitting error threshold value is omegaE(ii) a Initializing a segmentation point count value 1;
1.1) first calculating end ═ start + h;
1.3) if the fitting error ERR is not greater than the fitting error threshold value omegaEIf h is h +1, skipping to step 1) again;
1.4) if the fitting error ERR is larger than the fitting error threshold value omegaEThen ensureSegment storage point To Pi(ii) a Saving fitted line segmentsRepresenting the fitted line segment by means of a tripletAnd store tokiWhich represents the slope of the line segment,represents the span of the line segment on the time axis, riRepresenting a growth rate of the line segment data; resetting the fitting starting point start to be start + h and resetting h to be 2; updating count + 1;
1.2.3. circularly executing 1.2.2 until end is larger than k + L-1 to obtain a fitted linear time sequenceAnd segmentation pointComposed sequence of segmentation points Pi;
1.2.4. With Wk-1The clustering center of the window is the initial clustering center, and the triple sequence is subjected to K-means clustering algorithmClustering is carried out, and different symbols are distributed to the line segments of different classes to obtain symbolic sequencesWherein in the clustering processTwo line segments s are described by indexes based on Euclidean distanceiAnd sjSimilarity between them:
wherein d isijI.e. representing a line segment siAnd sjSimilarity of (d)ijThe smaller the size, the more similar the change form of the two line segments, ωkAnd ωrRepresenting a weight;
1.2.5. for two operating parameters ViAnd VjRespectively obtaining the symbolized sequencesAndsegment point sequence P combining two operating parametersiAnd PjAnd symbolizing the sequence of the combined segmentation pointsAndperforming segmentation reconstruction to obtain reconstructed symbolic sequenceAndfrom which a transaction set is formed, nij-1 is PiAnd PjThe number of the combined segmentation points.
Preferably, the step 2 is realized by the following sub-steps:
step 2.1: each transaction of the transaction set obtained by step 1 is marked asIn addition willAndthe line segment type symbols included in (1) are respectively marked asAndmiand mjAre respectively asAndthe number of line segment categories contained in (1); setting the minimum support threshold of the frequent item set generation process as min1;
Step 2.2: for an initial window of data W1And (3) performing association rule mining, wherein the specific process is as follows:
2.2.1) scanning the sensor measurement sequence dataset of all sensors to calculate the support of each term, let σ (-) denote the support count of a term or set of terms, initially 0; suppose thatIs denoted by the class symbol tkE represents i or j, t represents c or d;
For each tkJudgment ofIf yes, the method determines that t is truekFor frequent 1-item sets, reserve tkAnd recordCorresponding support degree count σ (t)k);
2.2.2) use the resulting frequent 1-item set tkForm 2-item set and find frequent 2-item set, respectivelypAnd dqRespectively, the symbols of the original line segment class after the step 2.2.1)Andthe item retained in (1);
for each { cp,dqExecuting the operation: for each one exists inC in (2)p,dq}, calculate σ ({ c)p,dq})=σ({cp,dq}) + 1; if it is notNot less than min1Then consider { cp,dqKeep { c) for frequent 2-item setp,dqAnd recording corresponding support degree counts;
2.2.3) use the frequent 2-item set { c) obtained in step 2.2.2p,dqEvery two operation parameters V are calculatediAnd VjThe support in the whole data set is calculated as follows: for every two operating parameters ViAnd VjSet of formed items { Vi,Vj}, calculate σ ({ V)i,Vj})=sum(σ({cp,dq})) two parameters, associated rule supportAnd calculates σ (V)i)=sum(σ(cp));σ(Vj)= sum(σ(dq));
2.2.4) for each group { Vi,VjResults in the following association rules: vj→ViAnd Vi→Vj(ii) a For each association rule, calculate its confidence as
Step 2.3: for each new data window, the following process is adopted for frequent item set generation and association rule generation:
for each k>1 data window WkPerforming steps 2.3.1) to 2.3.3):
2.3.1) transaction to each sliding-out WindowIs marked as { ca1,da2}, judging { ca1,da2Whether the item is a frequent item set or not, if so, updating the sigma ({ c)a1,da2})=σ({ca1,da2})-1;
2.3.2) transactions for each sliding-in windowIs marked as { cb1,db2}, judging { cb1,db2Whether the item set is frequent or not, if yes, update sigma ({ c)b1,db2})=σ({cb1,db2}) +1, if not, in the data window WkInternally calculating the { cb1,db2Support count of σ ({ c)c1,dc2});
2.3.3) for each { c }a1,da2Get it beforeUpdate { c }a1,da2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cb1,db2Get it beforeUpdate { c }b1,db2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cc1,dc2Get it beforeUpdate { c }c1,dc2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set;
2.3.4) update the support and confidence of each association rule using steps 2.2.3) and 2.2.4).
Preferably, the step 3 is implemented by the following sub-steps:
the preset prediction step length is recorded as lpThe association parameter extracted from the association rule mined from the training data set is { V }1,V2,…,VuU is the number of the extracted group of associated parameters; for each parameter ViThe measured value sequence isConstruct the following matrix ItrainFor initial neural network training inputs:
wherein, ItrainEach column in the training input sample is a training input sample; structural training output OtrainIn order to realize the purpose,
training the wavelet neural network by using the training sample constructed by the formula; at network initialization, association rule V is utilizedi→VuCorresponding initial confidence ωiAnd setting an initial weight value between the network input layer and the hidden layer, wherein i is 1,2, … u-1.
Preferably, the step 4 is implemented by the following sub-steps:
step 4.1: the updating method of the association rule comprises the following steps: recording c new data collected by the sensor in the running process of the equipment asc>L, using the most recent L data in the newly collected dataUpdate the association rule through step 2.3, thereby updating association rule Vi→VuThe corresponding confidence is recorded as
Step 4.2: the model updating method comprises the following steps: combining the training data and the newly acquired data to construct a new model training sample, i.e. a data set ofConstruction matrix ItestConstructing O for input of neural network updatetestIn order to be output, the output is,
constructing a new training sample by using the formula, and training the wavelet neural network; at network initialization, the new confidence level is utilizedAnd setting an initial weight value between the network input layer and the hidden layer.
Preferably, the step 5 is implemented by the following sub-steps:
recording a preset abnormal working condition (failure) occurrence threshold value as omegapFor the latest collected data, l is performed by using the model updated in step 4pAnd predicting, namely if the obtained target parameter predicted value exceeds a set threshold relative to the initial normal drift amount, determining that an abnormal working condition occurs.
Preferably, the association rules and models are updated back to step 4 each time a predetermined amount of data is updated before the prediction process is finished.
The method utilizes a sliding window mode to mine the dynamic association rule of the operation parameters of the industrial equipment and introduces the dynamic association rule into the prediction of the abnormal working conditions in the industrial process. The invention considers the time characteristic of the association rule, limits the data length by using a sliding window, provides an association rule mining algorithm for mining the dynamic association rule of operating parameters pairwise, and then introduces the association rule mining result into wavelet neural network prediction to continuously update the network by the dynamic association rule so as to obtain a more accurate prediction result. The method has great application value for fault prediction and health management in engineering.
Drawings
FIG. 1 is a confidence curve of the IDV (13) association rule (v13, v16, v36 → v7) in the embodiment;
FIG. 2 is a confidence curve of the IDV (13) association rule (v35, v36 → v11) in the embodiment;
FIG. 3 shows the predicted result of variable 7 of IDV (13) in the example;
FIG. 4 shows the prediction error of the IDV (13) variable 7 in the example;
FIG. 5 shows the predicted result of the variable 11 of IDV (13) in the example;
FIG. 6 shows the prediction error of IDV (13) variable 11 in an embodiment.
Detailed Description
The embodiments of the present invention will now be further described with reference to the accompanying drawings.
The following embodiment specifically explains the specific operation steps and the effect of the verification method through the tennessee-iseman (TE) process simulation data.
The data set was sampled at 3 minute intervals and recorded the variable measurements taken by each sensor at that sampling interval. Under each operating condition (normal operating state and fault operating state under 21 preset faults), the measurement data of the simulation process will generate two types of data sets, namely a training set and a test set. The acquisition process of the training set is measured values of all 52 variables obtained under the condition that the simulation process runs for 25 hours, wherein, except the training set acquired under the normal running state, the acquisition of the other 21 training set data introduces faults after the simulation process runs for 1 hour, and only the measured data of the following 24 hours are recorded. That is, the training set in the normal operation state has 500 observation samples, and the training sets collected in the remaining 21 fault states are all 480 observation samples. In addition, for 22 test sets, the data is all the variable measurement values collected after the simulation process runs for 48 hours, that is, each test set contains 960 sample data. It should be noted that in the simulation of 21 process faults, the corresponding fault was introduced after the simulation was run for 8 hours. Therefore, for the test set in 21 fault operation states, the first 160 observation samples are normal data, and the last 800 observation samples are fault data. In the TE process simulation model, only IDV (13) is a slowly varying fault, so in this example we use the relevant data of IDV (13) to perform experiments.
The method for predicting the abnormal working condition of the industrial process comprises the following specific implementation processes:
step 1: the method comprises the steps of preprocessing measured value time sequences acquired by sensors in the industrial process based on a sliding window, and performing piecewise linearization representation, clustering and symbolization on the time sequences to generate a transaction set suitable for association rule mining. The step is realized by the following steps:
step 1.1: the sensor measurement sequence is recorded asN is the number of sensors, K is the sequence length, K is 1,2, …, K, the length of the sliding window is L, the sliding distance is l each time, and the sliding window is WkThe data contained in the window isIt should be noted that in the present invention, i and j are numbers indicating sensors as superscripts and are numbers indicating only ordinal numbers as subscripts, regardless of the sensor numbers.
Step 1.2: and performing time series piecewise linearization representation, clustering and symbolization processing on each data window, wherein the specific process is as follows:
for each data window WkPerforming the following processes of 1.2.1-1.2.5:
1.2.1. setting the initial fitting starting point toEnd point ish is 2; fitting error threshold value is omegaE(ii) a Initializing a segmentation point count value 1;
1.1) first calculating end ═ start + h;
1.3) if the fitting error ERR is not greater than the fitting error threshold value omegaEIf h is h +1, skipping to step 1) again;
1.4) if the fitting error ERR is larger than the fitting error threshold value omegaEThen save the segmentation point To Pi(ii) a Saving fitted line segmentsRepresenting the fitted line segment by means of a tripletAnd store tokiWhich represents the slope of the line segment,represents the span of the line segment on the time axis, riRepresenting a growth rate of the line segment data; resetting the fitting starting point start + h and resetting h 2 (resetting the fitting starting point and h); updating count + 1;
1.2.3. circularly executing 1.2.2 until end that end is larger than k + L-1, and obtaining a line-segment time sequence after least square fitting(with multiple line segments fitted therein) and segmentation pointsComposed sequence of segmentation points Pi;
1.2.4. With Wk-1The clustering center of the window is the initial clustering center, and the triple sequence is subjected to K-means clustering algorithmClustering is carried out, and different symbols are distributed to the line segments of different classes to obtain symbolic sequencesWherein, two line segments s are described by indexes based on Euclidean distance in the clustering processiAnd sjSimilarity between them:
wherein d isijI.e. representing a line segment siAnd sjSimilarity of (d)ijThe smaller the size, the more similar the change form of the two line segments, ωkAnd ωrRepresenting a weight;
1.2.5. for two operating parameters ViAnd VjRespectively obtaining the symbolized sequencesAndsegment point sequence P combining two operating parametersiAnd PjAnd symbolizing the sequence of the combined segmentation pointsAndperforming segmentation reconstruction to obtain reconstructed symbolic sequenceAndfrom which a transaction set is formed, nij-1 is PiAnd PjThe number of the combined segmentation points.
Step 2: and generating a frequent item set by adopting a two-stage frequent item set generation strategy, and mining association rules of every two parameters. The step is realized by the following steps:
step 2.1: each transaction of the transaction set obtained by step 1 is marked asIn addition willAndthe line segment type symbols included in (1) are respectively marked asAndmiand mjAre respectively asAndthe number of line segment categories contained in (1); setting the minimum support threshold of the frequent item set generation process as min1(ii) a In this example, the minimum support threshold is set to 0.2;
step 2.2: for an initial window of data W1And (3) performing association rule mining, wherein the specific process is as follows:
2.2.1) scanning the sensor measurement sequence dataset of all sensors to calculate the support of each term, let σ (-) denote the support count of a term or set of terms, initially 0; suppose thatIs denoted by the class symbol tkE represents i or j, t represents c or d (i.e. c)kOr dk);
For each tkJudgment ofIf yes, the method determines that t is truekFor frequent 1-item sets, reserve tkAnd records the corresponding support count σ (t)k);
2.2.2) use the resulting frequent 1-item set tkForm 2-item set and find frequent 2-item set, respectivelypAnd dqRespectively, the symbols of the original line segment class after the step 2.2.1)Andthe item retained in (1);
for each { cp,dqExecuting the operation: for each one exists inC in (2)p,dq}, calculate σ ({ c)p,dq})=σ({cp,dq}) + 1; if it is notNot less than min1Then consider { cp,dqKeep { c) for frequent 2-item setp,dqAnd recording corresponding support degree counts;
2.2.3) use the frequent 2-item set { c) obtained in step 2.2.2p,dqEvery two operation parameters V are calculatediAnd VjThe support in the whole data set is calculated as follows: for every two operating parameters ViAnd VjSet of formed items { Vi,Vj}, calculate σ ({ V)i,Vj})=sum(σ({cp,dq})) two parameters, associated rule supportAnd calculates σ (V)i)=sum(σ(cp));σ(Vj)= sum(σ(dq));
2.2.4) for each group { Vi,VjResults in the following association rules: vj→ViAnd Vi→Vj(ii) a For each association rule, calculate its confidence as
In this example, for IDV (13), the initial association rules are mined through a training set. The support threshold is set to 0.2, and the confidence threshold is set to 0.7, so as to obtain several different sets of association rules, the results of which can be seen in table 1, wherein two sets of association parameters are included, namely (variable 13, variable 16, variable 36 → variable 7) and (variable 35, variable 36 → variable 11).
Step 2.3: for each new data window, the following process is used to perform frequent item set generation and generate association rules. In this example, the length of the sliding window for dynamic association rule mining is set to 100, and the sliding window is set to 1 data point at a time.
For each k>1 data window WkPerforming steps 2.3.1) to 2.3.3):
2.3.1) transaction to each sliding-out WindowIs marked as { ca1,da2}, judging { ca1,da2Whether the item is a frequent item set or not, if so, updating the sigma ({ c)a1,da2})=σ({ca1,da2})-1;
2.3.2) transactions for each sliding-in windowIs marked as { cb1,db2}, judging { cb1,db2Whether the item set is frequent or not, if yes, update sigma ({ c)b1,db2})=σ({cb1,db2}) +1, if not, in the data window WkInternally calculating the { cb1,db2Support count of σ ({ c)c1,dc2});
2.3.3) for each { c }a1,da2Get it beforeUpdate { c }a1,da2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cb1,db2Get it beforeUpdate { c }b1,db2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cc1,dc2Get it beforeUpdate { c }c1,dc2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set;
2.3.4) update the support and confidence of each association rule using steps 2.2.3) and 2.2.4).
With the test set of IDVs (13), starting from the 100 th point, every time 1 sample of data is updated, the first 100 points including this point are used to mine the association rule, and the resulting confidence change situation is shown in fig. 1 and 2. As can be seen, the association rules (confidence) change as the data is updated, but the values for both sets of parameters of IDV (13) are maintained at a higher level.
And step 3: and (5) performing association rule mining by using the initial data set, and training an initial wavelet neural network model based on the initial association rule mining result. The step is realized by the following steps:
the preset prediction step length is recorded as lpIn this case 10. The association parameter extracted from the association rule mined from the training data set is set as V1,V2,…,VuU is the number of the extracted group of associated parameters; for each parameter ViThe measured value sequence isConstruct the following matrix ItrainFor initial neural network training inputs:
wherein, ItrainEach column in the training input sample is a training input sample; structural training output OtrainIn order to realize the purpose,
training the wavelet neural network by using the training sample constructed by the formula; at network initialization, association rule V is utilizedi→VuCorresponding initial confidence ωiAnd setting an initial weight value between the network input layer and the hidden layer, wherein i is 1,2, … u-1.
And 4, step 4: and sliding a data window based on data updating, and updating an association rule and a wavelet neural network model for a new data window. The step is realized by the following steps:
step 4.1: the updating method of the association rule comprises the following steps: recording c new data collected by the sensor in the running process of the equipment asc>L, using the most recent L data in the newly collected dataUpdate the association rule through step 2.3, thereby updating association rule Vi→VuThe corresponding confidence is recorded asNote that c is greater than the preset sliding window length L, that is, the update of the association rule is only started when the device is running and the length of the collected data is greater than L.
Step 4.2: the model updating method comprises the following steps: combining the training data and the newly acquired data to construct a new model training sample, i.e. a data set ofConstruction matrix ItestConstructing O for input of neural network updatetestIn order to be output, the output is,
constructing a new training sample by using the formula, and training the wavelet neural network; at network initialization, the new confidence level is utilizedAnd setting an initial weight value between the network input layer and the hidden layer. In this example, the association rules and models are updated every 10 data updates using the first 300 data of the test set.
And 5: and predicting abnormal working conditions based on the updated wavelet network model, and continuously updating the association rule and the wavelet neural network model before predicting the occurrence of the abnormal working conditions. The step is realized by the following steps:
recording a preset abnormal working condition (failure) occurrence threshold value as omegapIn this example, the threshold value for the occurrence of an abnormal condition (failure) is ωp0.015. For the newly acquired data, l is performed using the model updated in step 4pAnd predicting, namely if the obtained target parameter predicted value exceeds a set threshold relative to the initial normal drift amount, determining that an abnormal working condition occurs. Before the prediction process is finished, returning to step 4 to update the association rule and the model every time U data are updated, wherein the value of U can be adjusted, and U is 10 in this example.
Table 1 association rules
Rule antecedents | Rule clause | Confidence level |
Variable 13 | Variable 7 | 0.7527 |
Variable 16 | Variable 7 | 0.7446 |
Variable 36 | Variable 7 | 0.7017 |
Variable 35 | Variable 11 | 0.7513 |
Variable 36 | Variable 11 | 0.7390 |
TABLE 2 Total prediction error Rate
Introducing dynamic association rules | Introducing association rules | Without introducing association rules | |
Variable 7 | 0.7602 | 1.0247 | 1.7423 |
Variable 11 | 0.8075 | 0.8309 | 1.1884 |
In this example, in order to verify the validity of the method used in this chapter, the prediction result of the dynamic association rule introduced is compared with the case where the association rule is not introduced and the static association rule is introduced, and the comparison between the prediction result and the prediction error is shown in fig. 3 to 6. In fig. 3 and 5, a vertical solid line is an actual failure time under a preset failure threshold, a vertical dotted line is a failure time prediction value of introducing a dynamic association rule, a vertical dotted line is a failure time prediction value of introducing a static association rule, and a vertical dotted line is a failure time prediction value of not introducing an association rule. In addition, to better quantify the prediction error, the present example calculates the total error rate of the two parameter predictions under different conditions, as shown in table 2.
From the prediction results (fig. 3 and 5), compared with the other two cases, the introduction of the dynamic association rule can obtain a more accurate abnormal operating condition prediction result, and especially compared with the case without the introduction of the association rule, the introduction of the dynamic association rule has obvious advantages from the prediction curve. It should be noted that, in fig. 5, the predicted value and the true value have a large deviation when the static association rule is introduced and the association rule is not introduced, which is actually because the failure actual value of the parameter is exactly at a certain extreme point, so that if a slight prediction error exists here, the abnormal operating condition prediction result has a large deviation. In practice, a number of different threshold overruns may be set to monitor the parameters to avoid the risk of such a situation. In addition, from the perspective of prediction error (fig. 4 and 6), compared with the case of introducing a static association rule, introducing a dynamic association rule can reduce the prediction error to a certain extent, and as can be seen from table 2, for the variable 7, the promotion effect of introducing the dynamic association rule is obvious, and for the variable 11, because the change of the association rule is not large, the effect is not obvious; it can be seen from fig. 4 and 6 that the prediction error can be significantly reduced by introducing the dynamic association rule compared to the case of not introducing the association rule, and it can also be more intuitively seen from table 2 that the introduction of the dynamic association rule has a significant advantage compared to the case of not introducing the association rule.
Claims (6)
1. An industrial process abnormal working condition prediction method based on dynamic association rule mining is characterized by comprising the following specific steps:
step 1: carrying out data preprocessing on a measured value time sequence acquired by a sensor in an industrial process based on a sliding window, and carrying out piecewise linearization representation, clustering and symbolization on the time sequence to generate a transaction set suitable for association rule mining;
step 2: generating a frequent item set by adopting a two-stage frequent item set generation strategy, and mining association rules of every two parameters;
and step 3: performing association rule mining by using an initial data set, and training an initial wavelet neural network model based on an initial association rule mining result;
and 4, step 4: sliding a data window based on data updating, and updating an association rule and a wavelet neural network model for a new data window;
and 5: predicting abnormal working conditions based on the updated wavelet network model, and continuously updating the association rule and the wavelet neural network model before predicting the occurrence of the abnormal working conditions;
the step 1 is specifically realized by the following substeps:
step 1.1: the sensor measurement sequence is recorded asN is the number of sensors, K is the sequence length, K is 1,2,., K, the sliding window length is L, the sliding distance is l each time, and the sliding window is denoted as WkThe data contained in the window is
Step 1.2: and performing time series piecewise linearization representation, clustering and symbolization processing on each data window, wherein the specific process is as follows:
for each data window WkPerforming the following processes of 1.2.1-1.2.5:
1.2.1. setting the initial fitting starting point toEnd point isFitting error threshold value is omegaE(ii) a Initializing a segmentation point count value 1;
1.1) first calculating end ═ start + h;
1.3) if the fitting error ERR is not greater than the fitting error threshold value omegaEIf h is h +1, skipping to step 1) again;
1.4) if the fitting error ERR is larger than the fitting error threshold value omegaEThen save the segmentation point To Pi(ii) a Saving fitted line segmentsRepresenting the fitted line segment by means of a tripletAnd store tokiWhich represents the slope of the line segment,represents the span of the line segment on the time axis, riRepresenting a growth rate of the line segment data; resetting the fitting starting point start to be start + h and resetting h to be 2; updating count + 1;
1.2.3. circularly executing 1.2.2 until end is larger than k + L-1 to obtain a fitted linear time sequenceAnd segmentation pointComposed sequence of segmentation points Pi;
1.2.4. With Wk-1The clustering center of the window is the initial clustering center, and the triple sequence is subjected to K-means clustering algorithmClustering is carried out, and different symbols are distributed to the line segments of different classes to obtain symbolic sequencesWherein, two line segments s are described by indexes based on Euclidean distance in the clustering processiAnd sjSimilarity between them:
wherein d isijI.e. representing a line segment siAnd sjSimilarity of (d)ijThe smaller the size, the more similar the change form of the two line segments, ωkAnd ωrRepresenting a weight;
1.2.5. for two operating parameters ViAnd VjRespectively obtaining the symbolized sequencesAndsegment point sequence P combining two operating parametersiAnd PjAnd symbolizing the sequence of the combined segmentation pointsAndperforming segmentation reconstruction to obtain reconstructed symbolic sequenceAndfrom which a transaction set is formed, nij-1 is PiAnd PjThe number of the combined segmentation points.
2. The method for predicting the abnormal working condition of the industrial process based on the dynamic association rule mining as claimed in claim 1, wherein: the step 2 is specifically realized by the following substeps:
step 2.1: each transaction of the transaction set obtained by step 1 is marked asIn addition willAndthe line segment type symbols included in (1) are respectively marked asAndmiand mjAre respectively asAndthe number of line segment categories contained in (1); setting the minimum support threshold of the frequent item set generation process as min1;
Step 2.2: for an initial window of data W1And (3) performing association rule mining, wherein the specific process is as follows:
2.2.1) scanning the sensor measurement sequence dataset of all sensors to calculate the support of each term, let σ (-) denote the support count of a term or set of terms, initially 0; suppose thatIs denoted by the class symbol tkE represents i or j, t represents c or d;
For each tkJudgment ofIf yes, then confirmIs tkFor frequent 1-item sets, reserve tkAnd records the corresponding support count σ (t)k);
2.2.2) use the resulting frequent 1-item set tkForm 2-item set and find frequent 2-item set, respectivelypAnd dqRespectively, the symbols of the original line segment class after the step 2.2.1)Andthe item retained in (1);
for each { cp,dqExecuting the operation: for each one exists inC in (2)p,dq}, calculate σ ({ c)p,dq})=σ({cp,dq}) + 1; if it is notNot less than min1Then consider { cp,dqKeep { c) for frequent 2-item setp,dqAnd recording corresponding support degree counts;
2.2.3) use the frequent 2-item set { c) obtained in step 2.2.2p,dqEvery two operation parameters V are calculatediAnd VjThe support in the whole data set is calculated as follows: for every two operating parameters ViAnd VjSet of formed items { Vi,Vj}, calculate σ ({ V)i,Vj})=sum(σ({cp,dq})) two parameters, associated rule supportAnd calculates σ (V)i)=sum(σ(cp));σ(Vj)=sum(σ(dq));
2.2.4) For each group { Vi,VjResults in the following association rules: vj→ViAnd Vi→Vj(ii) a For each association rule, calculate its confidence as
Step 2.3: for each new data window, the following process is adopted for frequent item set generation and association rule generation:
for each k>1 data window WkPerforming steps 2.3.1) to 2.3.3):
2.3.1) transaction to each sliding-out WindowIs marked as { ca1,da2}, judging { ca1,da2Whether the item is a frequent item set or not, if so, updating the sigma ({ c)a1,da2})=σ({ca1,da2})-1;
2.3.2) transactions for each sliding-in windowIs marked as { cb1,db2}, judging { cb1,db2Whether the item set is frequent or not, if yes, update sigma ({ c)b1,db2})=σ({cb1,db2}) +1, if not, in the data window WkInternally calculating the { cb1,db2Support count of σ ({ c)c1,dc2});
2.3.3) for each { c }a1,da2Get it beforeUpdate { c }a1,da2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cb1,db2Get it beforeUpdate { c }b1,db2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set; for each { cc1,dc2Get it beforeUpdate { c }c1,dc2Counting the support degree of the corresponding item set, and if the support degree is not met, rejecting the corresponding item set;
2.3.4) update the support and confidence of each association rule using steps 2.2.3) and 2.2.4).
3. The method for predicting the abnormal working condition of the industrial process based on the dynamic association rule mining as claimed in claim 2, wherein: the step 3 is specifically realized by the following substeps:
the preset prediction step length is recorded as lpThe association parameter extracted from the association rule mined from the training data set is { V }1,V2,...,VuU is the number of the extracted group of associated parameters; for each parameter ViThe measured value sequence isConstruct the following matrix ItrainFor initial neural network training inputs:
wherein, ItrainEach column in the training input sample is a training input sample; structural training output OtrainIn order to realize the purpose,
training a wavelet neural network by using a training input sample constructed by the formula (1); at network initialization, association rule V is utilizedi→VuCorresponding toInitial confidence ωiAn initial weight between the network input layer and the hidden layer is set, i ═ 1, 2.
4. The method for predicting the abnormal working condition of the industrial process based on the dynamic association rule mining as claimed in claim 3, wherein: the step 4 is specifically realized by the following substeps:
step 4.1: the updating method of the association rule comprises the following steps: recording c new data collected by the sensor in the running process of the equipment asUsing the latest L data in the newly acquired dataUpdate the association rule through step 2.3, thereby updating association rule Vi→VuThe corresponding confidence is recorded as
Step 4.2: the model updating method comprises the following steps: combining the training data and the newly acquired data to construct a new model training sample, i.e. a data set ofConstruction matrix ItestConstructing O for input of neural network updatetestIn order to be output, the output is,
5. The method for predicting the abnormal working condition of the industrial process based on the dynamic association rule mining as claimed in claim 4, wherein the method comprises the following steps: the step 5 is specifically realized by the following substeps:
recording a preset abnormal working condition occurrence threshold value as omegapFor the latest collected data, l is performed by using the model updated in step 4pAnd predicting, namely if the obtained target parameter predicted value exceeds a set threshold relative to the initial normal drift amount, determining that an abnormal working condition occurs.
6. The method for predicting the abnormal working condition of the industrial process based on the dynamic association rule mining as claimed in claim 5, wherein: before the prediction process is finished, the association rules and the model are updated in step 4 every time a predetermined amount of data is updated.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910244877.8A CN110018670B (en) | 2019-03-28 | 2019-03-28 | Industrial process abnormal working condition prediction method based on dynamic association rule mining |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910244877.8A CN110018670B (en) | 2019-03-28 | 2019-03-28 | Industrial process abnormal working condition prediction method based on dynamic association rule mining |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110018670A CN110018670A (en) | 2019-07-16 |
CN110018670B true CN110018670B (en) | 2020-07-10 |
Family
ID=67190170
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910244877.8A Active CN110018670B (en) | 2019-03-28 | 2019-03-28 | Industrial process abnormal working condition prediction method based on dynamic association rule mining |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110018670B (en) |
Families Citing this family (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111060652A (en) * | 2019-11-21 | 2020-04-24 | 西南交通大学 | Method for predicting concentration of dissolved gas in transformer oil based on long-term and short-term memory network |
CN113515554A (en) * | 2020-04-09 | 2021-10-19 | 华晨宝马汽车有限公司 | Anomaly detection method and system for irregularly sampled time series |
CN111858712A (en) * | 2020-07-20 | 2020-10-30 | 上海仪电(集团)有限公司中央研究院 | In-situ water quality inspection data time-space analysis and anomaly detection method and system |
CN112016097B (en) * | 2020-08-28 | 2024-02-27 | 深圳泓越信息科技有限公司 | Method for predicting network security vulnerability time to be utilized |
CN112380274B (en) * | 2020-11-16 | 2023-08-22 | 北京航空航天大学 | Abnormality detection method for control process |
CN112632151B (en) * | 2020-12-25 | 2023-02-10 | 南京邮电大学 | Mobile object infection behavior mining method based on binary filtering |
CN112862347A (en) * | 2021-03-02 | 2021-05-28 | 同济大学 | Equipment abnormity monitoring method and system based on federal learning, storage medium and terminal |
CN112800686A (en) * | 2021-03-29 | 2021-05-14 | 国网江西省电力有限公司电力科学研究院 | Transformer DGA online monitoring data abnormal mode judgment method |
CN112801426B (en) * | 2021-04-06 | 2021-06-22 | 浙江浙能技术研究院有限公司 | Industrial process fault fusion prediction method based on correlation parameter mining |
CN112749370B (en) * | 2021-04-06 | 2021-07-02 | 广东际洲科技股份有限公司 | Fault tracking and positioning method and system based on Internet of things |
CN113297272B (en) * | 2021-05-30 | 2023-01-06 | 福建中锐网络股份有限公司 | Bridge monitoring data association rule mining and health early warning method and system |
CN113591994B (en) * | 2021-08-03 | 2023-06-06 | 北京邮电大学 | Terminal behavior prediction method based on automatic labeling |
CN115689071B (en) * | 2023-01-03 | 2023-05-02 | 南京工大金泓能源科技有限公司 | Equipment fault fusion prediction method and system based on associated parameter mining |
Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101408754A (en) * | 2008-10-30 | 2009-04-15 | 中山大学 | Intelligent house optimizing system based on data excavation |
CN101667197A (en) * | 2009-09-18 | 2010-03-10 | 浙江大学 | Mining method of data stream association rules based on sliding window |
CN201898519U (en) * | 2010-09-01 | 2011-07-13 | 燕山大学 | Equipment maintenance early-warning device with risk control |
CN102818642A (en) * | 2012-07-18 | 2012-12-12 | 辽宁省海洋水产科学研究院 | Disease pre-warning system for stichopus japonicus |
CN103440248A (en) * | 2013-07-22 | 2013-12-11 | 西南交通大学 | Network video event mining framework based on dynamic association rules |
CN103676645A (en) * | 2013-12-11 | 2014-03-26 | 广东电网公司电力科学研究院 | Mining method for association rules in time series data flows |
CN104636479A (en) * | 2015-02-15 | 2015-05-20 | 西安电子科技大学 | Industrial big data driven total completion time prediction method |
CN108873859A (en) * | 2018-05-31 | 2018-11-23 | 浙江工业大学 | Based on the bridge-type grab ship unloader fault prediction model method for improving correlation rule |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7474987B2 (en) * | 2004-02-10 | 2009-01-06 | University Of Maryland | Method and system for optimal data diagnosis |
-
2019
- 2019-03-28 CN CN201910244877.8A patent/CN110018670B/en active Active
Patent Citations (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101408754A (en) * | 2008-10-30 | 2009-04-15 | 中山大学 | Intelligent house optimizing system based on data excavation |
CN101667197A (en) * | 2009-09-18 | 2010-03-10 | 浙江大学 | Mining method of data stream association rules based on sliding window |
CN201898519U (en) * | 2010-09-01 | 2011-07-13 | 燕山大学 | Equipment maintenance early-warning device with risk control |
CN102818642A (en) * | 2012-07-18 | 2012-12-12 | 辽宁省海洋水产科学研究院 | Disease pre-warning system for stichopus japonicus |
CN103440248A (en) * | 2013-07-22 | 2013-12-11 | 西南交通大学 | Network video event mining framework based on dynamic association rules |
CN103676645A (en) * | 2013-12-11 | 2014-03-26 | 广东电网公司电力科学研究院 | Mining method for association rules in time series data flows |
CN104636479A (en) * | 2015-02-15 | 2015-05-20 | 西安电子科技大学 | Industrial big data driven total completion time prediction method |
CN108873859A (en) * | 2018-05-31 | 2018-11-23 | 浙江工业大学 | Based on the bridge-type grab ship unloader fault prediction model method for improving correlation rule |
Non-Patent Citations (5)
Title |
---|
一种面向时空数据的关联规则更新算法;刘伯红等;《计算机与数字工程》;20151020(第10期);第1767-1770、1774页 * |
基于关联规则的故障诊断方法及研究;张建明等;《化工自动化及仪表》;20031030(第05期);第11-14、18页 * |
基于关联规则的火电厂传感器故障检测;鲍文;《中国电机工程学报》;20031231;第170-174页 * |
基于滑动窗口的多元时间序列数据动态关联规则挖掘;熊腾飞;《中国优秀硕士学位论文全文数据库信息科技辑》;20170215;第16-38页 * |
基于神经网络的铝电解混合控制模型研究;汪成亮等;《计算机应用研究》;20100715(第07期);第2536-2539、2559页 * |
Also Published As
Publication number | Publication date |
---|---|
CN110018670A (en) | 2019-07-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110018670B (en) | Industrial process abnormal working condition prediction method based on dynamic association rule mining | |
CN110008253B (en) | Industrial data association rule mining and abnormal working condition prediction method | |
CN110263846B (en) | Fault diagnosis method based on fault data deep mining and learning | |
CN111694879B (en) | Multielement time sequence abnormal mode prediction method and data acquisition monitoring device | |
US9779361B2 (en) | Method for learning exemplars for anomaly detection | |
CN102789545B (en) | Based on the Forecasting Methodology of the turbine engine residual life of degradation model coupling | |
CN110895526A (en) | Method for correcting data abnormity in atmosphere monitoring system | |
Said et al. | Machine learning technique for data-driven fault detection of nonlinear processes | |
CN108985380B (en) | Point switch fault identification method based on cluster integration | |
US11093314B2 (en) | Time-sequential data diagnosis device, additional learning method, and recording medium | |
KR20190072652A (en) | Information processing apparatus and information processing method | |
CN105607631B (en) | The weak fault model control limit method for building up of batch process and weak fault monitoring method | |
CN111931806A (en) | Equipment fault diagnosis method and device for multi-sensor data fusion | |
Lyu et al. | Remaining useful life estimation with multiple local similarities | |
CN114167838B (en) | Multi-scale health assessment and fault prediction method for servo system | |
CN113723010A (en) | Bridge damage early warning method based on LSTM temperature-displacement correlation model | |
CN109298633A (en) | Chemical production process fault monitoring method based on adaptive piecemeal Non-negative Matrix Factorization | |
CN115186762A (en) | Engine abnormity detection method and system based on DTW-KNN algorithm | |
CN114036647A (en) | Power battery safety risk assessment method based on real vehicle data | |
Zhang et al. | Remaining useful life prediction for rolling bearings with a novel entropy-based health indicator and improved particle filter algorithm | |
CN112949735A (en) | Liquid hazardous chemical substance volatile concentration abnormity discovery method based on outlier data mining | |
CN111144018B (en) | Aero-engine complete machine residual performance extraction method based on post-aviation data | |
CN110956155B (en) | Fuzzy clustering identification method for fully mechanized mining face operation procedure based on CO data | |
CN110032585B (en) | Time sequence double-layer symbolization method and device | |
CN115935285A (en) | Multi-element time series anomaly detection method and system based on mask map neural network model |
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 |