An engine residual life prediction method of a multi-period rule cascade memory network
By combining an automatically expanded cascaded long short-term memory network with domain knowledge rules, the accuracy bottleneck and volatility issues in engine remaining life prediction are resolved, achieving high-precision and high-efficiency prediction results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-30
- Publication Date
- 2026-03-27
AI Technical Summary
Existing long short-term memory networks suffer from accuracy bottlenecks and unresolved issues of result variability in engine remaining life prediction.
By combining an automatically expanding cascaded long short-term memory network with multiple domain knowledge rules, and through the cascaded structure of multiple sub-modules and rule optimization at different time periods, the prediction accuracy and stability are improved.
It improves the accuracy and stability of engine remaining life prediction, reduces prediction errors, and enhances prediction efficiency.
Smart Images
Figure CN114741958B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of engines and relates to a multi-period rule cascade memory network prediction method suitable for predicting the remaining life of an engine. BACKGROUND
[0002] At present, the prediction of the remaining life of an engine mainly depends on time series data and a long short-term memory network. Although the long short-term memory network has outstanding advantages in learning long time series data due to the special design of the gate structure, the upper limit of the prediction accuracy of the long short-term memory network itself makes it difficult for researchers to break through the prediction accuracy bottleneck of the remaining life of an engine by increasing the number of network layers or the number of training times.
[0003] In addition, although the existing series of variant models related to the long short-term memory network to some extent effectively improve the prediction accuracy of the remaining life of an engine, the results obtained by using these variant models of the long short-term memory network for predicting the remaining life of an engine always have a large fluctuation range and cannot accurately fit the real remaining life of an engine.
[0004] Therefore, a multi-period rule cascade memory network engine remaining life prediction method is proposed. The automatic expansion cascade long short-term memory network used in the method breaks through the prediction accuracy bottleneck caused by the long short-term memory network itself by relying on the structure of multiple sub-module cascades, and reduces the volatility of the engine remaining life prediction results by formulating multiple domain knowledge rules in different periods, effectively solving the difficulty problem that the current mainstream method cannot accurately predict the remaining life of an engine. SUMMARY
[0005] The application proposes a multi-period rule cascade memory network engine remaining life prediction method, which combines the automatic expansion cascade long short-term memory network and multiple domain knowledge rules to accurately predict the remaining life of an engine. The steps in the use process are as follows:
[0006] Step (1): Select appropriate full-life data sets that can cover the whole process of the engine from operation to failure and reflect the change of the remaining life of the engine.
[0007] Step (2): Divide the selected data set into a training set and a test set. The input of the training data is X train , the output of the training data is Y train , the input of the test data is X test , and the output of the test data is Y test .
[0008] Step (3): The automatic expansion cascading long short-term memory network is built in a way of connecting sub-modules of the same structure step by step, each of which is composed of an input layer, a long short-term memory layer, a plurality of fully connected layers and a regression output layer, and the i-th sub-module is denoted as module i, where i is a positive integer, in the order of connecting step by step from front to back;
[0009] Step (4): Through X train and Y train , module 1 is trained, and the trained module 1 is used to predict the output values of X train and X test as the prediction input, respectively. The output results of X train and X test after being predicted by module 1 are Y tpred1 and Y pred1 , respectively, and the two output errors E 1a and E 1b of module 1 are:
[0010] E 1a =Y train -Y tpred1 (1)
[0011] E 1b =Y test -Y pred1 (2)
[0012] Step (5): Module 2 takes X train and X test as the input of the training data and the input of the test data, respectively, and the difference from module 1 is that the output of the training data of module 2 is set as one of the output errors E 1a of module 1, and the output of the test data of module 2 is set as the other output error E 1b of module 1; module 2 is trained through X train and E 1a , and the trained module 2 is used to predict X train and X test ; the output results of X train and X test after being predicted by module 2 are Y tpred2 and Y pred2 , respectively, and the two output errors E 2a and E 2b of module 2 are:
[0013] E 2a =E 1a -Y tpred2 (3)
[0014] E 2b =E 1b -Y pred2 (4)
[0015] Step (6): By continuously taking the output error of the previous module as the training output of the current module, and taking the output error of the current module as the training output of the next module, and so on, when cascaded to module n, the two output errors E na and E nb of module n are calculated respectively as follows:
[0016] E na =E (n-1)a -Y tpredn (5)
[0017] E nb =E (n-1)b -Y predn (6)
[0018] Step (7): Finally, |E na | < ε or |E nb | < ε is taken as the judgment condition for the end of the prediction step of the automatically expanded cascaded long short-term memory network, where ε is a constant for limiting the range of output error; if the judgment condition is met, the module cascading process is automatically terminated, representing that the value of the entire network output error has met the set requirement; if the judgment condition is not met, the next module cascading step is continued and the new output error is calculated until the absolute value of the output error of the last module is less than the set ε;
[0019] The comprehensive formula (1)-(6) is obtained as follows:
[0020]
[0021]
[0022] According to formula (7) and formula (8), the prediction result of the final output of the automatically expanded cascaded long short-term memory network can be obtained by summing each item of the prediction result of each module output, so that the coordinates (X test(j) , Y pred(j) ) of the jth prediction point can be obtained, where and j is a positive integer;
[0023] Step (8): According to the preliminary prediction result of the engine remaining life obtained in step (7), the change characteristics of the engine remaining life in different time periods are analyzed, a plurality of domain knowledge rules suitable for different time period characteristics are formulated, and the following rules are executed in order:
[0024] Rule 1: If Y pred(j)> 149.01, then let Y pred(j) = 150;
[0025] Rule 2: Take 9 predicted values as a group of data, and represent the group of predicted values as Y pred(j) , Y pred(j+1) ,..., Y pred(j+8) in turn; if there is a predicted value equal to 150 among Y pred(j) , Y pred(j+1) ,..., Y pred(j+8) , then let Y pred(j) = 150;
[0026] Rule 3: if Y pred(j) = 150 and Y pred(j+k-1) = 150, then let Y pred(j+1) = 150, Y pred(j+2) = 150,..., Y pred(j+k-2) = 150, where 2 < k < 13 and k is an integer;
[0027] Rule 4: if Y pred(j) < 1.5, then let Y pred(j) = 1;
[0028] Rule 5: round all Y pred(j) to integers by rounding;
[0029] Rule 6: among a group of predicted values, let the coordinates of the 9 predicted points corresponding to the 9 predicted values be (X test(j) , Y pred(j) ), (X test(j+1) , Y pred(j+1) ),..., (X test(j+8) , Y pred(j+8) ); if Y pred(j) < 150, Y pred(j+1) < 150,..., Y pred(j+8) < 150 and Y pred(j) ≥ Y pred(j+1) ≥... ≥ Y pred(j+8) , then the straight line equation L0 fitted by the 9 predicted points can be obtained by using the least square method formula: y = ax + b, where:
[0030] the slope of the fitted straight line
[0031] substitute the point to obtain the intercept of the fitted straight line
[0032]
[0033] the equation of the fitted straight line L0 is: wherein,
[0034] m is a positive integer;
[0035] The X test(j) ,X test(j+1) ,…,X test(j+8) values of the nine predicted points are substituted into the equation of the fitting straight line L0 to calculate the Y test(j) ,Y test(j+1) ,…,Y test(j+8) values of the corrected predicted points, and the new Y test(j) ,Y test(j+1) ,…,Y test(j+8) values are rounded to integers;
[0036] Rule 7: Among a group of predicted values, the jth predicted value is Y pred(j) , the (j+p)th predicted value is Y pred(j+p) , if Y pred(j) -Y pred(j+p) =p, then the (p-1) predicted values between the jth predicted value and the (j+p)th predicted value are arranged in order from Y pred(j) -1 to Y pred(j) -(p-1), i.e. let
[0037] Y pred(j+1) =Y pred(j) -1 (14)
[0038] Y pred(j+2) =Y pred(j) -2 (15)
[0039] Y pred(j+p-1) =Y pred(j) -(p-1) (16)
[0040] wherein p=3, 4, …, 9;
[0041] Step (9): Rules 1 and 4 in step (8) are executed again, and thus the correction of the preliminary predicted result of the engine remaining life is completed, and a high-precision engine remaining life prediction curve is finally obtained.
[0042] The present application has the following advantages and effects relative to the prior art:
[0043] (1) The automatic expansion cascaded long short-term memory network proposed in the application can effectively reduce the output error of the prediction network and improve the prediction accuracy of the engine remaining life by connecting multiple sub-modules in cascade, taking the output error of the previous module as the training output of the current module, and taking the output error of the current module as the training output of the next module.
[0044] (2) The essence of the cascaded structure of the automatic expansion cascaded long short-term memory network is similar to the increase of network width with grouping characteristics, but in terms of calculation time, the automatic expansion cascaded long short-term memory network shortens the time by a multiple of the number of cascades compared to the way of simply increasing the network width, which ensures that the automatic expansion cascaded long short-term memory network improves the prediction accuracy while also ensuring good prediction efficiency.
[0045] (3) Based on the automatic expansion cascaded long short-term memory network, the application formulates a plurality of appropriate domain knowledge rules according to the characteristics of the remaining life prediction results in different time periods, which plays a key optimization role in the prediction results of the engine remaining life and greatly improves the prediction effect of the method. BRIEF DESCRIPTION OF DRAWINGS
[0046] Figure 1 is the automatic expansion cascaded long short-term memory network framework diagram described in the method of the application.
[0047] Figure 2 is the engine remaining life prediction flowchart of the method of the application. DETAILED DESCRIPTION
[0048] The application proposes an engine remaining life prediction method of a multi-period rule cascaded memory network, which is described in detail as follows in combination with the drawings:
[0049] Figure 1 is the automatic expansion cascaded long short-term memory network framework diagram described in the method of the application. By Figure 1As shown, the automatically expanding cascaded long short-term memory network is composed of multiple sub-modules of the same structure connected in stages, each of which includes a long short-term memory layer, multiple fully connected layers, and a regression output layer. The training set data is input in a time series manner and normalized to the range of 0-1, and then trained through the long short-term memory layer, the fully connected layer, and the regression output layer to obtain the trained sub-module. Subsequently, the trained sub-modules at each stage are used to predict the test set data to obtain a series of predicted values. The output error of each sub-module is the difference between the predicted value and the true value. Under the multi-stage cascaded structure of the automatically expanding cascaded long short-term memory network, the output error of the initial sub-module is gradually reduced after multiple predictions by the sub-modules at each stage by continuously taking the output error of the previous stage module as the training output of the current stage module, and taking the output error of the current stage module as the training output of the next stage module. Once the final output error of the entire network is less than a certain set value, it indicates that the predicted value of the data is close to the true value, and the expansion and cascading process of the sub-modules can be automatically ended.
[0050] Figure 2 is a flowchart of the engine remaining life prediction method of the present application. The flowchart first uses the proposed automatically expanding cascaded long short-term memory network to preliminarily predict the remaining life of the engine, and then uses multiple domain knowledge rules to further correct the preliminary prediction result. Pointer variable i in the figure represents the serial number of the sub-module. The training and prediction values input and output each sub-module are time series data. The input of the training data is X train , and the output is Y train ; the input of the test data is X test , and the output is Y test . X train and Y train are used to learn and train module 1, and then the trained module 1 is used to predict the output values of the training input X train and the test input X test respectively. The output results of X train and X test after being predicted by module 1 are Y tpred1 and Y pred1 respectively, from which the two output errors E 1a and E 1b of module 1 can be calculated.
[0051] From module 2, the output error of the previous stage module is taken as the training output of the current stage module, and the output error of the current stage module is taken as the training output of the next stage module, and so on. Module 2 is first trained by X train and E 1a , and then Xtrain and X test Prediction is carried out. train and X test The output result after prediction by module 2 is denoted as Y tpred2 and Y pred2 Similarly, two output errors E 2a and E 2b of module 2 can be calculated.
[0052] Similarly, two output errors E na and E nb of module n can be calculated according to the above idea. Finally, |E na |<ε or |E nb |<ε is used as the judgment condition for ending the prediction step of the automatically expanding cascaded long short-term memory network, where ε is a constant used to limit the range of output error. If the condition is met, the module cascading process is automatically terminated, representing that the value of the output error of the entire network has met the set requirement; if the condition is not met, the next module cascading step is continued and the new output error is calculated until the output error of the last module is less than the set ε. When the cascading is terminated, the final output prediction result of the automatically expanding cascaded long short-term memory network can be obtained by summing each item of the prediction results of each module output, so that the preliminary prediction of the engine remaining life is realized by using the automatically expanding cascaded long short-term memory network.
[0053] After the preliminary prediction of the engine remaining life is completed, seven field knowledge rules are formulated according to the change characteristics of the engine remaining life in different time periods to correct the preliminary prediction result of the engine remaining life, so that the final prediction result of the engine remaining life is closer to its true value, thereby realizing the final prediction of the engine remaining life by the method of the present application.
[0054] The above only describes the preferred embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation according to the content of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.
Claims
1.A method for engine remaining useful life prediction using a multi-period rule cascade memory network, the method comprising: The automatic expansion cascade long short-term memory network and multiple domain knowledge rules are combined to realize accurate prediction of the engine remaining life. Step (1): selecting appropriate full-life data sets capable of covering the whole process from operation to failure of the engine and reflecting the change of the engine remaining life; Step (2): Divide the selected dataset into two parts, training set and test set, the input of training data is X train , the output of training data is Y train ; the input of test data is X test , the output of test data is Y test ; Step (3): building the automatic expansion cascade long short-term memory network in the mode of connecting multiple sub-modules of the same structure in stages, each of which is composed of an input layer, a long short-term memory layer, multiple fully connected layers and a regression output layer, and the i-th sub-module is module i in the order of connecting from front to back, wherein i is a positive integer; Step (4): Through X train and Y train learning training is performed on module 1, and after the trained module 1 is used, training input X train and test input X test are taken as prediction inputs, and respective output values are predicted, X train and X test , and the output results after the prediction of module 1 are Y tpred1 and Y pred1 , and the output errors E 1a and E 1b of module 1 are respectively: E 1a = Y train - Y tpred1 (1) E 1b = Y test - Y pred1 (2) Step (5): Module 2 is trained again with X train and X test as the input of training data and the input of testing data respectively, different from Module 1, Module 2 is to set the output of training data as one of the output errors E 1a of Module 1 and set the output of testing data as another output error E 1b of Module 1; train Module 2 through X train and E 1a , and then use the trained Module 2 to predict X train and X test ; the output results of X train and X test after being predicted by Module 2 are Y tpred2 and Y pred2 respectively, and the two output errors E 2a and E 2b of Module 2 are: E 2a = E 1a - Y tpred2 (3) E 2b = E 1b - Y pred2 (4) Step (6): By continuously taking the output error of the previous module as the training output of the current module, and taking the output error of the current module as the training output of the next module, and so on, when cascaded to module n, the two output errors E na and E nb of module n are calculated as follows: E na = E (n-1)a - Y tpredn (5) E nb = E (n-1)b - Y predn (6) Step (7): final |E na |<ε or |E nb |<ε as the automatic expansion of long short-term memory network prediction step end condition, where ε is a constant for limiting the output error range; if the judgment condition is met, the automatic termination module cascade process is terminated, representing that the value of the entire network output error has met the set requirements; if the judgment condition is not met, the next module cascade step is continued and the new output error is calculated until the output error of the last module is less than the set ε. Comprehensive formula (1)-(6) is obtained: According to the formula (7) and the formula (8), the prediction result of the final output of the automatically expanded cascaded long short-term memory network can be obtained by summing each prediction result of each module output item by item, and thus the coordinates (X test(j) ,Y pred(j) ) of the jth prediction point can be obtained, wherein and j is a positive integer; Step (8): according to the preliminary prediction result of the engine remaining life obtained in step (7), analyzing the change characteristics of the engine remaining life in different time periods, formulating multiple domain knowledge rules suitable for different time period characteristics, and executing the following rules in order: Rule 1: If Y pred(j) > 149.01, then let Y pred(j) = 150; Rule 2: Take 9 predicted values as a group of data, and represent the group of predicted values as Y pred(j) ,Y pred(j+1) ,…,Y pred(j+8) ; if there is a predicted value equal to 150 in Y pred(j) ,Y pred(j+1) ,…,Y pred(j+8) , then let Y pred(j) = 150; Rule 3: If Y pred(j) = 150 and Y pred(j+k-1) = 150, then let Y pred(j+1) = 150, Y pred(j+2) = 150,..., Y pred(j+k-2) = 150, where 2 < k < 13 and k is an integer. Rule 4: If Y pred(j) <1.5, then let Y pred(j) = 1; Rule 5: round all Y's to the nearest integer. pred(j) ; Rule 6: In a set of prediction values, let the coordinates of the prediction points corresponding to the 9 prediction values be (X test(j) ,Y pred(j) ),(X test(j+1) ,Y pred(j+1) ),…,(X test(j+8) ,Y pred(j+8) ) respectively; if Y pred(j) <150,Y pred(j+1) <150,…,Y pred(j+8) <150 and Y pred(j) ≥Y pred(j+1) ≥…≥Y pred(j+8) , then the straight line equation L0 fitted by the 9 prediction points can be obtained by using the least square method formula: y = ax + b, wherein: The slope of the fitted straight line point of substitution obtaining the intercept of the fitted line The equation of the fitted straight line L0 is: Wherein, m is a positive integer; The X-coordinate values X test(j) ,X test(j+1) ,…,X test(j+8) of the nine predicted points are substituted into the equation of the fitting straight line L0 to calculate the Y-coordinate values Y test(j) ,Y test(j+1) ,…,Y test(j+8) of the corrected predicted points test(j) ,Y test(j+1) ,…,Y test(j+8) are rounded to integers; Rule 7: among a set of predicted values, the jth predicted value is Y pred(j) , the (j+p)th predicted value is Y pred(j+p) , if Y pred(j) - Y pred(j+p) = p, then the (p-1) predicted values between the jth predicted value and the (j+p)th predicted value are arranged in order from Y pred(j) -1 to Y pred(j) -(p-1) in decreasing order, i.e. let Y pred(j+1) = Y pred(j) -1 (14) Y pred(j+2) = Y pred(j) -2 (15) Y pred(j+p-1) = Y pred(j) - (p - 1) (16) Wherein p = 3, 4, …, 9; Step (9): executing rule 1 and rule 4 in step (8) again, thus completing the correction of the preliminary prediction result of the engine remaining life, and finally obtaining the high-precision engine remaining life prediction curve.