CN113642779A - ResNet50 network key equipment residual life prediction method based on feature fusion - Google Patents
ResNet50 network key equipment residual life prediction method based on feature fusion Download PDFInfo
- Publication number
- CN113642779A CN113642779A CN202110829782.XA CN202110829782A CN113642779A CN 113642779 A CN113642779 A CN 113642779A CN 202110829782 A CN202110829782 A CN 202110829782A CN 113642779 A CN113642779 A CN 113642779A
- Authority
- CN
- China
- Prior art keywords
- residual
- sample
- residual life
- dimension
- space
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Abstract
The invention discloses a method for predicting the residual service life of ResNet50 network key equipment based on feature fusion, which comprises the steps of firstly, carrying out correlation analysis on data information acquired by each sensor in system equipment, filtering feature variables with lower correlation degree with the service life, and determining high correlation degree feature variables influencing the service life of a system; then, fusing the high-association characteristic variables by using an ISOMAP algorithm to obtain fused characteristic variables and effectively integrate sensor information; and aiming at the fusion characteristic variables and the residual life at the corresponding moment of the system, forming a residual life prediction total sample set, and finally establishing a prediction model and predicting the residual life to obtain a system residual life prediction result. The invention can accurately predict the service life of the equipment in the industrial field and prompt the staff to carry out preventive maintenance on the equipment in time.
Description
Technical Field
The invention belongs to the technical field of monitoring and maintenance of intelligent manufacturing systems, and particularly relates to a method for predicting the residual life of ResNet50 network key equipment based on feature fusion.
Background
Predictive and Health Management (PHM) technology is of critical importance in modern industry and has been widely used in civil aviation, automotive and manufacturing industries. Over time, the performance and health of the equipment will show different degradation trends under the influence of internal factors and various external factors (such as abrasion, external impact, load and operating environment), and the continuous accumulation finally causes the equipment to break down, thereby causing immeasurable economic loss, safety accidents and the like. Residual service life (RUL) prediction is one of the most important components in the modern industrial field Predictive Health Management (PHM) technology, and is defined as the length from the current time of service life to the end of that life. In order to improve the safety and reliability of the system and reduce the maintenance cost, it is necessary to ensure that the equipment is maintained at the optimum time, and parts are replaced periodically to prevent the failure of the equipment due to the accumulation of faults. The data of sensors of all parts of the equipment are monitored, the residual service life of the equipment at a certain moment is accurately predicted, and then the equipment can be preventively maintained at a proper time, so that the equipment downtime is reduced, and the cost is saved.
The prediction of the residual service life of the system is mainly based on a large amount of monitoring data collected by multiple sensors to evaluate the overall degradation state of the system and timely maintain the system. A single sensor signature is not sufficient to accurately determine the health of the system and thus multiple signatures are required. However, too many features cause information redundancy and increase the computational burden. Therefore, the characteristics with low correlation degree with the service life are filtered, and other characteristics of the multiple sensors are fused to obtain better health indexes, improve the accuracy of the prediction method and enable preventive maintenance to be more timely and effective. Aiming at the problem that dimension explosion is caused by excessive features in the residual life prediction research, so that the calculation efficiency is low, the ResNet50 network key equipment residual life prediction method based on feature fusion is provided, and the method not only can enhance the confidence of residual life prediction research, but also plays a critical role in the progress of the subject and the industry.
Disclosure of Invention
The invention aims to provide a ResNet50 network key equipment residual life prediction method based on feature fusion, which can accurately predict the equipment life in the industrial field and prompt the staff to perform preventive maintenance on the equipment in time.
The invention adopts the technical scheme that a ResNet50 network key equipment residual life prediction method based on feature fusion is implemented according to the following steps:
and 4, step 4: establishing a prediction model and predicting the residual life: and (3) training the prediction network by taking the training set of the total residual life prediction sample set obtained in the step (3) as the input of the depth residual error network Resnet50, and predicting the test sample by using the trained network to obtain the residual life prediction result of the system.
The present invention is also characterized in that,
the step 1 is as follows:
step 1.1, numbering data information acquired by a sensor in system equipment, wherein n represents a serial number of the sensor, and x represents a serial number of the sensornFor the monitoring data information of the nth sensor, the monitoring data X is expressed asWherein the content of the first and second substances,for the monitoring data information of the nth sensor at the time i in the system, i is 1,2iIn which yiThe residual life of the system corresponding to the i moment;
Wherein, Cov (x)nY) is monitoring data information xnAnd the covariance between the remaining lifetime Y,andrespectively monitoring data information xnAnd standard deviation of residual life Y;
step 1.3, calculating monitoring data information xnPearson correlation with residual Life YCoefficient of performanceThe method comprises the following specific steps:
and keeping the high-association characteristic variable with the result exceeding more than 0.8.
The step 2 is as follows:
step 2.1, selecting a neighborhood, and constructing a neighborhood graph G: setting a neighbor parameter as k, and constructing a sample set D ═ x by using the high-relevancy feature variable retained in the step 1.31,x2,...,xmM is the number of sample points in a sample set D of the original D-dimensional space, where { x ═ x }1,x2,...,xmArbitrarily choose some two data points xhAnd xjWherein h, j is the number of two data points selected in the sample set D, h, j belongs to 1,2,3hSetting the distance from the adjacent parameter k as Euclidean distance, and setting another data point xjIs xhWhen one of the nearest neighbor parameters k points is found, the nearest neighbor parameters are shown to be adjacent, and the neighborhood graph G has edges, namely the neighborhood graph G is constructed;
step 2.2, calculating a shortest path distance matrix: if two points x on neighborhood graph Gh、xjDefining sample point x when there is an edge joinhAnd xjThe shortest path between is dG(xh,xj)=d(xh,xj) (ii) a Otherwise dG(xh,xj) Infinity, wherein d (x)h,xj) Is a data point xhAnd xjThe Euclidean distance of;
for l ∈ 1,2,3.. m, there are:
dG(xh,xj)=min{dG(xh,xj),dG(xh,xl)+dG(xl,xj)} (3)
wherein x islIs a new sample point in the sample set D, DG(xh,xl) Represents a sample point xhAnd xlShortest path between, dG(xl,xj) Represents a sample point xlAnd xjShortest path between, dG(xh,xj) Represents a sample point xhAnd xjThe shortest path therebetween;
based on this, the shortest path distance matrix can be determined asWherein m is the number of sample points in a sample set D of the D-dimensional original space, h, j and l are the numbers of three data points selected in the sample set D, and h, j and l belong to 1,2 and 3.. m;
step 2.3, constructing a d 'dimension new sample space by reducing the dimension of the d dimension original space, keeping the distance between the d' dimension new sample space and the d dimension original space unchanged, and solving an inner product matrix B of the reduced samples: let B be ZTZ∈Rm×mWherein d is the dimension of the original space, d 'is the dimension of the new sample space, Z is the matrix representation of the d-dimension original space embedded with the d' -dimension new sample space, and R is a real number set;
after the new sample space is constructed, d '< d, and the Euclidean distance of any two samples in the d' dimension new sample space is equal to the distance in the d dimension original space, namely:
||zh-zj||=disthj (4)
wherein z ish,zjRespectively representing sample points xh,xjEuclidean distance, dist, in d' Vietnamese sample spacehjRepresents a sample point xhAnd xjDistance in d-dimensional original space;
equation (4) is expanded and squared across the equality number:
wherein, bhjIs the element of h row and j column in inner product matrix B, BhhAnd bjjBy analogy, disthjRepresents a sample point xhAnd xjDistance in d-dimension original space, disth·、dist·jAnd dist··And so on;
by finding bhj,bhh,bjjThe elements in the inner product matrix B are equal to obtain an inner product matrix B;
step 2.4, distance matrix D for shortest pathGConstructing d-dimension embedding, and finally obtaining a matrix representation Z belonging to the sample and embedding the d' dimension new sample space in the d-dimension original spaced′×m;
And (3) obtaining an eigenvalue matrix and an eigenvector matrix through characteristic decomposition by using the inner product matrix B obtained in the step 2.3, namely:
B=VΛVT (8)
wherein Λ ═ diag (λ)1,λ2,...,λd) Diagonal matrices formed for eigenvalues, λ1≥λ2≥...≥λd,λ1,λ2,...,λdFor decomposing the obtained eigenvalues, V is the corresponding eigenvector matrix.
The formula for embedding the matrix representation Z of the d-dimensional original space into the d' -dimensional new sample space is:the result of calculating Z can be expressed as a matrix:the result at this timeIs the feature variable after fusion;
wherein, there is d*The number of non-zero characteristic values is,diagonal matrices formed for eigenvalues, eigenvaluesV*Is a corresponding feature vector matrix;and (3) the fusion characteristic variable of the m 'th sample point at the i moment in the d' dimension new sample space, wherein m 'is the number of the sample points in the d' dimension new sample space.
The step 3 is as follows:
step 3.1, constructing a sample set: fusing the characteristic variables obtained in the step 2And the remaining life Y of the system is { Y ═ YiReconstructing into a residual life prediction sample set Predicting characteristic variables in the sample set for the reconstructed residual life respectively, and dividing the sample set into a training set and a verification set according to a ratio of 7:3, wherein yiThe residual life of the system corresponding to the i moment;
step 3.2, constructing a sub-training set and a sub-testing set: dividing the training set divided in the step 3.1 into a sub-training set and a sub-testing set according to the proportion of 7: 3;
step 3.3, constructing a total training set and a total testing set: and (3) merging the characteristic variables in the sub-training sets obtained in the step (3.2) into a total training set in a column stacking mode, and simultaneously merging the characteristic variables in the sub-testing sets into a total testing set in a column stacking mode corresponding to the merging sequence of the sub-training sets in the step (3.2), so as to finally form a residual life prediction total sample set consisting of the total training set, the total testing set and the verification set obtained in the step (3.1).
The step 4 is as follows:
step 4.1, constructing a 50-layer depth residual error network ResNet 50: the depth residual error network ResNet50 is composed of 3 parts, namely a convolution pooling part, 4 residual error block parts and a pooling flattening part, wherein the residual error block parts have two basic structures, namely an identical residual error block and a convolution residual error block;
step 4.2, feature extraction: taking the training set of the total sample set of the residual life prediction obtained in the step 3 as the input of a depth residual error network Resnet50, performing feature extraction on the training set by adopting a one-dimensional convolution kernel, and performing maximum pooling on the extracted convolution features to obtain pooled features;
and 4.3, extracting the characteristics of the residual block: taking the pooled features as an input of a residual block, the residual block of the deep residual network Resnet50 is composed of the following four parts: the system comprises a convolution residual block, two identity residual blocks, a convolution residual block, three identity residual blocks, a convolution residual block, five identity residual blocks, a convolution residual block and two identity residual blocks, wherein the four residual blocks are connected in sequence to form a residual network core part;
step 4.4, establishing a prediction model: integrating the output of the residual block through an average pooling layer, and obtaining a prediction model through a flattening layer;
and 4.5, obtaining a residual life prediction result: inputting the test set into a prediction model to obtain a residual life prediction result, obtaining an output prediction result by predicting once, reversely transmitting an error between the prediction result and residual life real data acquired by the system to a depth residual error network Resnet50, adjusting and optimizing each parameter in the network, when the error of the network is less than 0.05, indicating the network convergence, predicting a test sample by using a trained network, finally obtaining a system residual life prediction result, and realizing accurate prediction of the residual life of the system.
The method for predicting the residual service life of the ResNet50 network key equipment based on feature fusion has the advantages that the low-association feature influencing the service life of system equipment is eliminated through a correlation analysis method, the high-association feature is fused by utilizing an ISOMAP algorithm, sensor information is effectively integrated, and the coupling between sensors is eliminated. The ISOMAP algorithm and the depth residual error network Resnet50 are combined, so that the fused sensor characteristic information is used as the input of the depth residual error network Resnet50 to train the prediction network, the hidden layer relation among the characteristics is learned, a residual life prediction model is constructed, and the residual life prediction is realized. The accuracy and the efficiency of the method in the aspect of residual life prediction are verified through experimental simulation.
Drawings
FIG. 1 is a general flow chart of the method for predicting the remaining life of the ResNet50 network key equipment based on feature fusion according to the present invention;
FIG. 2 is a graph showing the results of correlation analysis of characteristic variables of the sensor according to the present invention;
FIG. 3 is a flow chart of feature fusion performed by using ISOMAP algorithm in the present invention;
FIG. 4 is an overall structure diagram of a deep residual error network ResNet50 according to the present invention;
FIG. 5 is a flowchart of the prediction process of the deep residual network ResNet50 according to the present invention;
FIG. 6 is a graph of the predicted results of the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a method for predicting the residual life of ResNet50 network key equipment based on feature fusion, which is implemented by the following steps in particular according to a flow chart shown in figure 1:
with reference to fig. 2 to 5, step 1 is as follows:
step 1.1, numbering data information acquired by a sensor in system equipment, wherein n represents a serial number of the sensor, and x represents a serial number of the sensornFor the monitoring data information of the nth sensor, the monitoring data X is expressed asWherein the content of the first and second substances,for the monitoring data information of the nth sensor at the time i in the system, i is 1,2iIn which yiThe residual life of the system corresponding to the i moment;
Wherein, Cov (x)nY) is monitoring data information xnAnd the covariance between the remaining lifetime Y,andrespectively monitoring data information xnAnd standard deviation of residual life Y;
step 1.3, calculating monitoring data information xnPearson correlation coefficient with remaining lifetime YThe method comprises the following specific steps:
and keeping the high-association characteristic variable with the result exceeding more than 0.8.
the step 2 is as follows:
step 2.1, selecting a neighborhood, and constructing a neighborhood graph G: setting a neighbor parameter as k, and constructing a sample set D ═ x by using the high-relevancy feature variable retained in the step 1.31,x2,...,xmM is the number of sample points in a sample set D of the original D-dimensional space, where { x ═ x }1,x2,...,xmArbitrarily choose some two data points xhAnd xjWherein h, j is the number of two data points selected in the sample set D, h, j belongs to 1,2,3hSetting the distance from the adjacent parameter k as Euclidean distance, and setting another data point xjIs xhWhen one of the nearest neighbor parameters k points is found, the nearest neighbor parameters are shown to be adjacent, and the neighborhood graph G has edges, namely the neighborhood graph G is constructed;
step 2.2, calculating a shortest path distance matrix: if two points x on neighborhood graph Gh、xjDefining sample point x when there is an edge joinhAnd xjThe shortest path between is dG(xh,xj)=d(xh,xj) (ii) a Otherwise dG(xh,xj) Infinity, wherein d (x)h,xj) Is a data point xhAnd xjThe Euclidean distance of;
for l ∈ 1,2,3.. m, there are:
dG(xh,xj)=min{dG(xh,xj),dG(xh,xl)+dG(xl,xj)} (3)
wherein x islIs a new sample point in the sample set D, DG(xh,xl) Represents a sample point xhAnd xlShortest path between, dG(xl,xj) Represents a sample point xlAnd xjShortest path between, dG(xh,xj) Represents a sample point xhAnd xjThe shortest path therebetween;
based on this, the shortest path distance matrix can be determined asWherein m is the number of sample points in a sample set D of the D-dimensional original space, h, j and l are the numbers of three data points selected in the sample set D, and h, j and l belong to 1,2 and 3.. m;
step 2.3, constructing a d 'dimension new sample space by reducing the dimension of the d dimension original space, keeping the distance between the d' dimension new sample space and the d dimension original space unchanged, and solving an inner product matrix B of the reduced samples: let B be ZTZ∈Rm×mWherein d is the dimension of the original space, d 'is the dimension of the new sample space, Z is the matrix representation of the d-dimension original space embedded with the d' -dimension new sample space, and R is a real number set;
after the new sample space is constructed, d '< d, and the Euclidean distance of any two samples in the d' dimension new sample space is equal to the distance in the d dimension original space, namely:
||zh-zj||=disthj (4)
wherein z ish,zjRespectively representing sample points xh,xjEuclidean distance, dist, in d' Vietnamese sample spacehjRepresents a sample point xhAnd xjDistance in d-dimensional original space;
equation (4) is expanded and squared across the equality number:
wherein, bhjIs the element of h row and j column in inner product matrix B, BhhAnd bjjBy analogy, disthjRepresents a sample point xhAnd xjDistance in d-dimension original space, disth·、dist·jAnd dist··And so on;
by finding bhj,bhh,bjjThe elements in the inner product matrix B are equal to obtain an inner product matrix B;
step 2.4, distance matrix D for shortest pathGConstructing d-dimension embedding, and finally obtaining a matrix representation Z belonging to the sample and embedding the d' dimension new sample space in the d-dimension original spaced′×m;
And (3) obtaining an eigenvalue matrix and an eigenvector matrix through characteristic decomposition by using the inner product matrix B obtained in the step 2.3, namely:
B=VΛVT (8)
wherein Λ ═ diag (λ)1,λ2,...,λd) Diagonal matrices formed for eigenvalues, λ1≥λ2≥...≥λd,λ1,λ2,...,λdFor decomposing the obtained eigenvalues, V is the corresponding eigenvector matrix.
The formula for embedding the matrix representation Z of the d-dimensional original space into the d' -dimensional new sample space is:the result of calculating Z can be expressed as a matrix:the result at this timeIs the feature variable after fusion;
wherein, there is d*The number of non-zero characteristic values is,diagonal matrices formed for eigenvalues, eigenvaluesV*Is a corresponding feature vector matrix;and (3) the fusion characteristic variable of the m 'th sample point at the i moment in the d' dimension new sample space, wherein m 'is the number of the sample points in the d' dimension new sample space.
the step 3 is as follows:
step 3.1, constructing a sample set: fusing the characteristic variables obtained in the step 2And the remaining life Y of the system is { Y ═ YiReconstructing into a residual life prediction sample set Predicting characteristic variables in the sample set for the reconstructed residual life respectively, and dividing the sample set into a training set and a verification set according to a ratio of 7:3, wherein yiThe residual life of the system corresponding to the i moment;
step 3.2, constructing a sub-training set and a sub-testing set: dividing the training set divided in the step 3.1 into a sub-training set and a sub-testing set according to the proportion of 7: 3;
step 3.3, constructing a total training set and a total testing set: and (3) merging the characteristic variables in the sub-training sets obtained in the step (3.2) into a total training set in a column stacking mode, and simultaneously merging the characteristic variables in the sub-testing sets into a total testing set in a column stacking mode corresponding to the merging sequence of the sub-training sets in the step (3.2), so as to finally form a residual life prediction total sample set consisting of the total training set, the total testing set and the verification set obtained in the step (3.1).
And 4, step 4: establishing a prediction model and predicting the residual life: and (3) training the prediction network by taking the training set of the total residual life prediction sample set obtained in the step (3) as the input of the deep residual error network Resnet50, predicting the test sample by using the trained network to obtain a system residual life prediction result, and timely prompting a worker to perform preventive maintenance on the whole equipment.
The step 4 is as follows:
step 4.1, constructing a 50-layer depth residual error network ResNet 50: the depth residual error network ResNet50 is composed of 3 parts, namely a convolution pooling part, 4 residual error block parts and a pooling flattening part, wherein the residual error block parts have two basic structures, namely an identical residual error block and a convolution residual error block; the overall structure is shown in fig. 4. The difference of the two residual blocks lies in the existence of a convolution layer at the shortcut connection position, and the effect is to solve the problem of dimension mismatch.
Step 4.2, feature extraction: taking the training set of the total sample set of the residual life prediction obtained in the step 3 as the input of a depth residual error network Resnet50, performing feature extraction on the training set by adopting a one-dimensional convolution kernel, and performing maximum pooling on the extracted convolution features to obtain pooled features;
and 4.3, extracting the characteristics of the residual block: taking the pooled features as an input of a residual block, the residual block of the deep residual network Resnet50 is composed of the following four parts: the system comprises a convolution residual block, two identity residual blocks, a convolution residual block, three identity residual blocks, a convolution residual block, five identity residual blocks, a convolution residual block and two identity residual blocks, wherein the four residual blocks are sequentially connected to form a residual network core part, and as the number of network layers increases, the residual results are close to 0 through the specific shortcut connection of the residual blocks, so that the input loss is reduced;
step 4.4, establishing a prediction model: integrating the output of the residual block through an average pooling layer, and obtaining a prediction model through a flattening layer;
and 4.5, obtaining a residual life prediction result: inputting the test set into the prediction model to obtain a residual life prediction result, wherein a specific life prediction process is shown in fig. 5. And each time of predicting is carried out to obtain an output prediction result, the error between the prediction result and the residual life real data acquired by the system is reversely propagated to a deep residual error network Resnet50, each parameter in the network is optimized, when the error of the network is less than 0.05, the network is converged, the trained network is used for predicting the test sample, and finally the residual life prediction result of the system is obtained, so that the accurate prediction of the residual life of the system is realized.
Examples
In the experiment, a turbine engine system is taken as a research object, sample data of 21 sensors in four fault modes are collected, a data set of one fault mode is taken as an example (20631 sample data of 100 engines), and 17 th engine data (276 life cycle data in total) is randomly selected. After the correlation analysis is fused with the ISOMAP characteristics, the correlation coefficient is kept4 characteristic variables over 0.8. The first 200 life cycle data were used as the total training set, and the last 76 life cycle data were used as the total test set. Based on the above-mentioned data, it is possible to,the method of the invention is adopted to predict the residual life of multiple variables, and in order to describe the experimental result more clearly, the simulation result is visualized, and the result is shown in figure 6. As can be seen from the observation of FIG. 6, the error between the prediction result of the proposed model and the actual data is small, and the effect is good.
Claims (5)
1. A ResNet50 network key equipment residual life prediction method based on feature fusion is characterized by comprising the following steps:
step 1, carrying out correlation analysis on data information acquired by each sensor in system equipment, filtering characteristic variables with low correlation degree with service life, and determining high correlation degree characteristic variables influencing the service life of a system;
step 2, carrying out feature fusion on the high-association feature variables obtained in the step 1: fusing the high-association characteristic variables in the step 1 by using an ISOMAP algorithm to obtain fused characteristic variables and effectively integrate sensor information;
step 3, constructing a data set construction strategy: forming a total sample set of residual life prediction by aiming at the fusion characteristic variables obtained in the step 2 and the residual life at the corresponding moment of the system, and dividing the total sample set into a training set and a testing set;
and 4, step 4: establishing a prediction model and predicting the residual life: and (3) training the prediction network by taking the training set of the total residual life prediction sample set obtained in the step (3) as the input of the depth residual error network Resnet50, and predicting the test sample by using the trained network to obtain the residual life prediction result of the system.
2. The method for predicting the residual life of the ResNet50 network key equipment based on feature fusion as claimed in claim 1, wherein the step 1 is as follows:
step 1.1, numbering data information acquired by a sensor in system equipment, wherein n represents a serial number of the sensor, and x represents a serial number of the sensornFor the monitoring data information of the nth sensor, the monitoring data X is expressed asWherein the content of the first and second substances,for the monitoring data information of the nth sensor at the time i in the system, i is 1,2iIn which yiThe residual life of the system corresponding to the i moment;
Wherein, Cov (x)nY) is monitoring data information xnAnd the covariance between the remaining lifetime Y,andrespectively monitoring data information xnAnd standard deviation of residual life Y;
step 1.3, calculating monitoring data information xnPearson correlation coefficient rho of residual life YxnYThe method comprises the following steps:
3. The method for predicting the residual life of the ResNet50 network key equipment based on feature fusion as claimed in claim 2, wherein the step 2 is as follows:
step 2.1, selecting neighborhoodAnd constructing a neighborhood graph G: setting a neighbor parameter as k, and constructing a sample set D ═ x by using the high-relevancy feature variable retained in the step 1.31,x2,...,xmM is the number of sample points in a sample set D of the original D-dimensional space, where { x ═ x }1,x2,...,xmArbitrarily choose some two data points xhAnd xjWherein h, j is the number of two data points selected in the sample set D, h, j belongs to 1,2,3hSetting the distance from the adjacent parameter k as Euclidean distance, and setting another data point xjIs xhWhen one of the nearest neighbor parameters k points is found, the nearest neighbor parameters are shown to be adjacent, and the neighborhood graph G has edges, namely the neighborhood graph G is constructed;
step 2.2, calculating a shortest path distance matrix: if two points x on neighborhood graph Gh、xjDefining sample point x when there is an edge joinhAnd xjThe shortest path between is dG(xh,xj)=d(xh,xj) (ii) a Otherwise dG(xh,xj) Infinity, wherein d (x)h,xj) Is a data point xhAnd xjThe Euclidean distance of;
for l ∈ 1,2,3.. m, there are:
dG(xh,xj)=min{dG(xh,xj),dG(xh,xl)+dG(xl,xj)} (3)
wherein x islIs a new sample point in the sample set D, DG(xh,xl) Represents a sample point xhAnd xlShortest path between, dG(xl,xj) Represents a sample point xlAnd xjShortest path between, dG(xh,xj) Represents a sample point xhAnd xjThe shortest path therebetween;
based on this, the shortest path distance matrix can be determined asWherein m is the number of sample points in a sample set D of the D-dimensional original space, h, j and l are the numbers of three data points selected in the sample set D, and h, j and l belong to 1,2 and 3.. m;
step 2.3, constructing a d 'dimension new sample space by reducing the dimension of the d dimension original space, keeping the distance between the d' dimension new sample space and the d dimension original space unchanged, and solving an inner product matrix B of the reduced samples: let B be ZTZ∈Rm×mWherein d is the dimension of the original space, d 'is the dimension of the new sample space, Z is the matrix representation of the d-dimension original space embedded with the d' -dimension new sample space, and R is a real number set;
after the new sample space is constructed, d '< d, and the Euclidean distance of any two samples in the d' dimension new sample space is equal to the distance in the d dimension original space, namely:
||zh-zj||=disthj (4)
wherein z ish,zjRespectively representing sample points xh,xjEuclidean distance, dist, in d' Vietnamese sample spacehjRepresents a sample point xhAnd xjDistance in d-dimensional original space;
equation (4) is expanded and squared across the equality number:
wherein, bhjIs the element of h row and j column in inner product matrix B, BhhAnd bjjBy analogy, disthjRepresents a sample point xhAnd xjDistance in d-dimension original space, disth·、dist·jAnd dist··And so on;
by finding bhj,bhh,bjjThe elements in the inner product matrix B are equal to obtain an inner product matrix B;
step 2.4, distance matrix D for shortest pathGConstructing d-dimension embedding, and finally obtaining a matrix representation Z belonging to the sample and embedding the d' dimension new sample space in the d-dimension original spaced′×m;
And (3) obtaining an eigenvalue matrix and an eigenvector matrix through characteristic decomposition by using the inner product matrix B obtained in the step 2.3, namely:
B=VΛVT (8)
wherein Λ ═ diag (λ)1,λ2,...,λd) Diagonal matrices formed for eigenvalues, λ1≥λ2≥...≥λd,λ1,λ2,...,λdFor decomposing the obtained eigenvalues, V is the corresponding eigenvector matrix.
The formula for embedding the matrix representation Z of the d-dimensional original space into the d' -dimensional new sample space is:the result of calculating Z can be expressed as a matrix:the result at this timeIs the feature variable after fusion;
wherein, there is d*A non-zero eigenvalue, Λ*=diag(λ1,λ2,...,λd*) For the formation of characteristic valuesA diagonal matrix of formed eigenvalues λ1≥λ2≥...≥λd*,V*Is a corresponding feature vector matrix;and (3) the fusion characteristic variable of the m 'th sample point at the i moment in the d' dimension new sample space, wherein m 'is the number of the sample points in the d' dimension new sample space.
4. The method for predicting the residual life of the ResNet50 network key equipment based on feature fusion as claimed in claim 3, wherein the step 3 is as follows:
step 3.1, constructing a sample set: fusing the characteristic variables obtained in the step 2And the remaining life Y of the system is { Y ═ YiReconstructing into a residual life prediction sample setyi,Predicting characteristic variables in the sample set for the reconstructed residual life respectively, and dividing the sample set into a training set and a verification set according to a ratio of 7:3, wherein yiThe residual life of the system corresponding to the i moment;
step 3.2, constructing a sub-training set and a sub-testing set: dividing the training set divided in the step 3.1 into a sub-training set and a sub-testing set according to the proportion of 7: 3;
step 3.3, constructing a total training set and a total testing set: and (3) merging the characteristic variables in the sub-training sets obtained in the step (3.2) into a total training set in a column stacking mode, and simultaneously merging the characteristic variables in the sub-testing sets into a total testing set in a column stacking mode corresponding to the merging sequence of the sub-training sets in the step (3.2), so as to finally form a residual life prediction total sample set consisting of the total training set, the total testing set and the verification set obtained in the step (3.1).
5. The method for predicting the residual life of the ResNet50 network key equipment based on feature fusion as claimed in claim 4, wherein the step 4 is as follows:
step 4.1, constructing a 50-layer depth residual error network ResNet 50: the depth residual error network ResNet50 is composed of 3 parts, namely a convolution pooling part, 4 residual error block parts and a pooling flattening part, wherein the residual error block parts have two basic structures, namely an identical residual error block and a convolution residual error block;
step 4.2, feature extraction: taking the training set of the total sample set of the residual life prediction obtained in the step 3 as the input of a depth residual error network Resnet50, performing feature extraction on the training set by adopting a one-dimensional convolution kernel, and performing maximum pooling on the extracted convolution features to obtain pooled features;
and 4.3, extracting the characteristics of the residual block: taking the pooled features as an input of a residual block, the residual block of the deep residual network Resnet50 is composed of the following four parts: the system comprises a convolution residual block, two identity residual blocks, a convolution residual block, three identity residual blocks, a convolution residual block, five identity residual blocks, a convolution residual block and two identity residual blocks, wherein the four residual blocks are connected in sequence to form a residual network core part;
step 4.4, establishing a prediction model: integrating the output of the residual block through an average pooling layer, and obtaining a prediction model through a flattening layer;
and 4.5, obtaining a residual life prediction result: inputting the test set into a prediction model to obtain a residual life prediction result, obtaining an output prediction result by predicting once, reversely transmitting an error between the prediction result and residual life real data acquired by the system to a depth residual error network Resnet50, adjusting and optimizing each parameter in the network, when the error of the network is less than 0.05, indicating the network convergence, predicting a test sample by using a trained network, finally obtaining a system residual life prediction result, and realizing accurate prediction of the residual life of the system.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110829782.XA CN113642779A (en) | 2021-07-22 | 2021-07-22 | ResNet50 network key equipment residual life prediction method based on feature fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110829782.XA CN113642779A (en) | 2021-07-22 | 2021-07-22 | ResNet50 network key equipment residual life prediction method based on feature fusion |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113642779A true CN113642779A (en) | 2021-11-12 |
Family
ID=78418030
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110829782.XA Pending CN113642779A (en) | 2021-07-22 | 2021-07-22 | ResNet50 network key equipment residual life prediction method based on feature fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113642779A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114399440A (en) * | 2022-01-13 | 2022-04-26 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190114511A1 (en) * | 2017-10-16 | 2019-04-18 | Illumina, Inc. | Deep Learning-Based Techniques for Training Deep Convolutional Neural Networks |
CN109726517A (en) * | 2019-01-31 | 2019-05-07 | 西安理工大学 | A kind of equipment method for predicting residual useful life based on multivariable associated data |
CN109855875A (en) * | 2019-01-15 | 2019-06-07 | 沈阳化工大学 | A kind of rolling bearing operational reliability prediction technique |
CN110046409A (en) * | 2019-03-29 | 2019-07-23 | 西安交通大学 | A kind of components of steam turbine health state evaluation method based on ResNet |
CN110096500A (en) * | 2019-05-07 | 2019-08-06 | 上海海洋大学 | A kind of visual analysis method and system towards ocean multidimensional data |
CN110378870A (en) * | 2019-06-06 | 2019-10-25 | 西安交通大学 | A kind of turbine blade erosion degree method of discrimination based on ResNet-GRU network |
US20200210826A1 (en) * | 2018-12-29 | 2020-07-02 | Northeastern University | Intelligent analysis system using magnetic flux leakage data in pipeline inner inspection |
WO2020191800A1 (en) * | 2019-03-27 | 2020-10-01 | 东北大学 | Method for predicting remaining service life of lithium-ion battery employing wde-optimized lstm network |
CN112598144A (en) * | 2020-12-31 | 2021-04-02 | 西安理工大学 | CNN-LSTM burst fault early warning method based on correlation analysis |
-
2021
- 2021-07-22 CN CN202110829782.XA patent/CN113642779A/en active Pending
Patent Citations (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20190114511A1 (en) * | 2017-10-16 | 2019-04-18 | Illumina, Inc. | Deep Learning-Based Techniques for Training Deep Convolutional Neural Networks |
US20200210826A1 (en) * | 2018-12-29 | 2020-07-02 | Northeastern University | Intelligent analysis system using magnetic flux leakage data in pipeline inner inspection |
CN109855875A (en) * | 2019-01-15 | 2019-06-07 | 沈阳化工大学 | A kind of rolling bearing operational reliability prediction technique |
CN109726517A (en) * | 2019-01-31 | 2019-05-07 | 西安理工大学 | A kind of equipment method for predicting residual useful life based on multivariable associated data |
WO2020191800A1 (en) * | 2019-03-27 | 2020-10-01 | 东北大学 | Method for predicting remaining service life of lithium-ion battery employing wde-optimized lstm network |
CN110046409A (en) * | 2019-03-29 | 2019-07-23 | 西安交通大学 | A kind of components of steam turbine health state evaluation method based on ResNet |
CN110096500A (en) * | 2019-05-07 | 2019-08-06 | 上海海洋大学 | A kind of visual analysis method and system towards ocean multidimensional data |
CN110378870A (en) * | 2019-06-06 | 2019-10-25 | 西安交通大学 | A kind of turbine blade erosion degree method of discrimination based on ResNet-GRU network |
CN112598144A (en) * | 2020-12-31 | 2021-04-02 | 西安理工大学 | CNN-LSTM burst fault early warning method based on correlation analysis |
Non-Patent Citations (5)
Title |
---|
彭鸿博;蒋雄伟;: "基于相关向量机的发动机剩余寿命预测", 科学技术与工程 * |
胡伟健;樊杰;杜永兴;李宝山;李灵芳;杨颜博;: "基于深度残差网络的番茄细粒度病症识别", 内蒙古科技大学学报 * |
邓玲;陕振;马连志;: "基于数据筛选的硬盘剩余使用寿命预测方法", 计算机工程与设计 * |
闵宇航: "基于深度残差网络ResNet-50的绿萝状态识别系统", 科学技术创新 * |
魏永合;刘炜;杨艳君;苏君金;: "基于LLE及其改进距离算法的轴承故障诊断模型", 组合机床与自动化加工技术 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114399440A (en) * | 2022-01-13 | 2022-04-26 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
CN114399440B (en) * | 2022-01-13 | 2022-12-13 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111666982B (en) | Electromechanical equipment fault diagnosis method based on deep neural network | |
Byington et al. | Data-driven neural network methodology to remaining life predictions for aircraft actuator components | |
US8036999B2 (en) | Method for analyzing and classifying process data that operates a knowledge base in an open-book mode before defining any clusters | |
CN111259532B (en) | Fault diagnosis method of aeroengine control system sensor based on 3DCNN-JTFA | |
CN109977624A (en) | Photovoltaic plant soft fault monitoring method based on deep neural network | |
WO2006016440A1 (en) | Data processing method, data processing device, diagnosis method, and diagnosis device | |
CN105466693B (en) | The pre- diagnostic method of Fault of Diesel Fuel System based on gray model | |
CN111860775A (en) | Ship fault real-time diagnosis method based on CNN and RNN fusion | |
CN112598144A (en) | CNN-LSTM burst fault early warning method based on correlation analysis | |
CN114297918A (en) | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning | |
CN113642779A (en) | ResNet50 network key equipment residual life prediction method based on feature fusion | |
Huang et al. | An adversarial learning approach for machine prognostic health management | |
KR101505975B1 (en) | Method and system for fault dignosis of engine | |
CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
CN113239610A (en) | Domain self-adaptive rolling bearing fault diagnosis method based on Wasserstein distance | |
Korovin et al. | Real-time diagnostics of oil production equipment using data mining | |
CN116933656A (en) | Ship power system fault diagnosis method based on deep learning | |
Chen et al. | Application of a neural fuzzy system with rule extraction to fault detection and diagnosis | |
Lamoureux et al. | An approach to the health monitoring of the fuel system of a turbofan | |
Raptodimos et al. | An artificial neural network approach for predicting the performance of ship machinery equipment | |
CN115456104A (en) | Bridge inhaul cable abnormity identification and positioning method insensitive to sensor fault | |
CN112613227B (en) | Model for predicting remaining service life of aero-engine based on hybrid machine learning | |
CN114357667A (en) | Engine starting state fault detection method based on RESID recursive structure identification | |
CN108388232B (en) | Method for monitoring operation mode fault in crude oil desalting process | |
Mendonça et al. | Fault detection and isolation of industrial processes using optimized fuzzy models |
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 |