CN112836809A - Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion - Google Patents

Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion Download PDF

Info

Publication number
CN112836809A
CN112836809A CN202110303120.9A CN202110303120A CN112836809A CN 112836809 A CN112836809 A CN 112836809A CN 202110303120 A CN202110303120 A CN 202110303120A CN 112836809 A CN112836809 A CN 112836809A
Authority
CN
China
Prior art keywords
data
neural network
convolutional neural
characteristic
layer
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.)
Withdrawn
Application number
CN202110303120.9A
Other languages
Chinese (zh)
Inventor
唐守伟
张超
唐金鹤
王新
刘继勇
刘海瑞
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jinan Pentium Times Power Technology Co ltd
Original Assignee
Jinan Pentium Times Power Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Jinan Pentium Times Power Technology Co ltd filed Critical Jinan Pentium Times Power Technology Co ltd
Priority to CN202110303120.9A priority Critical patent/CN112836809A/en
Publication of CN112836809A publication Critical patent/CN112836809A/en
Withdrawn legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/25Fusion techniques
    • G06F18/253Fusion techniques of extracted features
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Probability & Statistics with Applications (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

The invention belongs to the technical field of power equipment characteristic analysis, and particularly relates to an equipment characteristic extraction method and a fault prediction method of a convolutional neural network based on differential feature fusion. The extraction method comprises the steps of obtaining operation data of the equipment; carrying out extraction and fusion of differential features and data normalization processing; and constructing a convolutional neural network characteristic extraction model, processing the original characteristic data through the extraction model, and combining a plurality of convolutional layers to obtain dimension reduction data. The predicting step includes labeling the feature matrix; dividing the characteristic matrix; building a classification layer of the network model; setting a loss function to train the network model; and carrying out fault prediction. And a convolutional neural network deep learning algorithm is adopted, data characteristics are extracted deeply, the characteristics of equipment are represented, and the actual requirements are met better.

Description

Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion
Technical Field
The invention belongs to the technical field of power equipment characteristic analysis, and particularly relates to an equipment characteristic extraction method and a fault prediction method of a convolutional neural network based on differential feature fusion.
Background
The equipment characteristic of the generator set represents the running state of the equipment under different running conditions, and is the basis and the premise for effectively diagnosing and predicting the fault of the unit equipment. However, in the actual operation process of the unit, the influence of various factors is received, the operation of the unit equipment has the characteristics of complexity, variability and the like, and the traditional characteristic analysis method of the unit equipment is difficult to comprehensively represent the operation state of the unit equipment, so that the accuracy of fault diagnosis and prediction is influenced, the normal operation of the generator set is directly influenced, and the safe and economic operation of a power plant is seriously influenced. Under the background of current big data, an effective method is applied, and the operation data of the unit equipment is analyzed, so that the equipment characteristic extraction work of the unit equipment is very important.
The traditional unit equipment characteristic analysis modes mainly comprise two types: the first method is as follows: and establishing a mechanism formula model of the operation condition of the equipment according to the mechanism formula and the design curve, wherein the mechanism formula model is used as the characteristic of the equipment changing along with the operation condition and reflects the operation state of the equipment. The second method comprises the following steps: the main method is to carry out a field disturbance test under typical working conditions, obtain test data of the influence of different working condition transformations on the output characteristics of the equipment, and carry out the working condition test, which relates to a series of work such as working condition selection, test design and the like. The basic failure diagnosis and prediction modes are as follows: analyzing the diagnostic rule of the equipment based on the operating characteristics of the equipment, and establishing a diagnostic system based on expert knowledge; or extracting the operating characteristics of the characteristic representation equipment under the fault, and realizing fault diagnosis based on machine learning by using recognition algorithms such as ANN, SVM, fuzzy recognition and the like. The research report related to the invention is not seen yet.
Disclosure of Invention
The invention aims to make up the defects of the existing method and provides a method for extracting the characteristics of equipment. The method is characterized in that a convolutional neural network is improved by a data model method based on fusion difference characteristics, a mode of combining layers of the convolutional neural network is adopted, and the characteristic data is further processed and extracted by combining a pooling layer, so that the characteristics of equipment are extracted and characterized. The method can automatically learn the data rules of the equipment under different working conditions based on mass data without depending on the knowledge and working experience of professionals, represents the operating characteristics of the equipment, provides an effective method for automatic characteristic extraction of the unit equipment and fault diagnosis and prediction, and tests show that the method has important significance for automatically realizing characteristic extraction of the unit equipment.
In order to achieve the purpose, the invention provides the following technical scheme: a device characteristic extraction method of a convolutional neural network based on differential feature fusion comprises the following steps:
acquiring operation data of equipment;
extracting and fusing differential features and carrying out data normalization processing on the operation data to obtain original feature data;
and constructing a convolutional neural network characteristic extraction model, processing the original characteristic data through the convolutional neural network characteristic extraction model, and combining a plurality of convolutional layers to obtain dimension reduction data.
Further, preprocessing the original feature data before processing the original feature data by using a convolutional neural network, including deleting invalid data and missing value processing, specifically including:
detecting the 'straightening data', recording the 'straightening time period for each index of the' straightening data ', and then deleting sample data corresponding to the' straightening time period;
for a certain sample with missing records, if the number of the missing indexes is less than or equal to 2, filling the missing value of the sample, wherein the filling value is the mean value of the indexes corresponding to the missing indexes; and if the number of the missing indexes is more than 2, the sample is considered invalid, and the sample is deleted.
Further, the operation data of the equipment is acquired and included in a system database, indexes which influence the operation of the target equipment are selected, and historical data of the operation of the equipment is read from the system database according to a certain access interval. Wherein the data is normal and fault operation data of the equipment containing a plurality of indexes and is expressed in a matrix form.
Further, the extraction and fusion processing of the difference features of the operation data comprises extracting 1-order difference and 2-order difference features of the operation data, and fusing the 1-order difference and the 2-order difference features with the original feature data; the data normalization processing of the differential characteristics on the operation data comprises adopting a formula according to the specification and the numerical range of each operation data or differential characteristics of the equipment
k=(Max-Min)/(xmax-xmin)
xstandard=Min+k(x-xmin)
Normalizing individual operational data or differential features to [ Min, Max ]]Within a range, where x is operational data or differential characteristics, xstandardNormalized value, x, for the operating data or differential signaturemaxIs the original maximum, x, of the operating data or differential signatureminMax is the maximum value of the designated interval and Min is the minimum value of the designated interval.
Further, the convolutional neural network characteristic extraction model includes a convolutional layer, an activation function and a pooling layer, wherein one convolutional layer has a plurality of convolutional cores, the processing of the raw feature data by the convolutional neural network characteristic extraction model includes a plurality of convolutional cores for performing convolutional operation on the raw feature data, the activation function performs nonlinear transformation on the obtained data, and the pooling layer performs subsampling on input data by a pooling core, wherein the activation function is a linear rectification function ReLU:
Figure BDA0002987099910000031
where x is some input data, yreluAnd x is data after the operation of the ReLU activation function.
Further, constructing the convolutional neural network characteristic extraction model comprises:
constructing an input layer of the convolutional neural network characteristic extraction model, and taking the original characteristic data as input data of the model;
construction of 5 layers of convolutional layers separately denoted C1,C2,C3,C4,C5Wherein each convolution layer adopts the same number of convolution kernels;
merging the 5 convolutional layers C1,C2,C3,C4,C5Recording the merged network as C _ merge;
performing dimension transformation on the upper output data, then constructing a pooling layer, and constructing a pooling layer P by using global maximum pooling operation of globalmaxpool6
Build up of 1 layer of convolutional layer C7
Performing dimension transformation on the upper layer data to construct a pooling layer, and adopting a global maximum pooling operation p8And finally, outputting the dimension reduction data represented by the multi-dimensional feature matrix F _ matrix.
The invention also discloses a method for predicting the fault by using the method for extracting the equipment characteristics of the convolutional neural network based on the differential feature fusion, which comprises the following steps:
labeling the characteristic matrix F _ matrix;
dividing the characteristic matrix into a training sample and a testing sample;
building a classification layer of the network model;
setting a loss function, and training the network model through an optimization algorithm;
analyzing the characteristic matrix by using the trained network model to obtain a probability vector and predicting faults;
further, the classification layer includes: 2 full connection layers and 1 classification layer, wherein the activation function of the full connection layer is a ReLU function, the activation function of the classification layer is Softmax, the input item of the classification layer is the characteristic vector of the training sample, and the output item is the class label corresponding to the training sample.
Further, the loss function is a cross-entropy function, namely:
Figure BDA0002987099910000041
wherein y isiRepresenting the true value of a sample, yi' is the predicted value for that sample.
Further, the probability vector may be represented as:
p=[p1,p2,…,pnum]
wherein num is the number of the types of the samples, and then a label to which the maximum probability of the probability vector belongs is selected as a prediction result of the last sample;
inputting all the test samples into the network model, outputting the predicted values of the test samples, and counting the number n of samples with the predicted values of all the test samples equal to the true valuesrightThen the accuracy of model prediction is:
Figure BDA0002987099910000051
where n is the total number of samples tested.
Compared with the prior art, the invention has the beneficial effects that:
1. the invention provides a characteristic extraction mode of combining multiple convolutional neural network layers, improves the traditional convolutional layer-pooling layer serial mode, can enrich data characteristics and extract the data characteristics more comprehensively, thereby comprehensively representing the equipment characteristics.
2. According to the equipment characteristic extraction method provided by the invention, the data characteristics are further extracted through the convolution layer and the pooling layer by extracting and fusing the characteristics of the equipment data, the characteristics of more dimensions can be fused through the characteristic fusion, and the data characteristics can be enriched through a multi-layer convolution combination mode, so that the equipment characteristics of the unit can be more comprehensively extracted and represented, and the actual requirements can be better met.
3. According to the equipment characteristic extraction method provided by the invention, the difference characteristic of the equipment index is firstly extracted, the numerical value change information of the index can be effectively captured, then the original index characteristic is fused, the convolution operation is further carried out, the defect that the numerical value change of a convolution neural network is not considered is overcome, and the comprehensive characterization of the equipment characteristic is facilitated.
4. The invention applies the characteristics of the unit equipment through the classification model, realizes fault diagnosis, fully excavates data value by applying a modeling method, establishes a high-efficiency practical model to carry out fault diagnosis and prediction on the state of the unit equipment, and is beneficial to ensuring the safe and reliable operation of the unit equipment.
5. The invention can automatically analyze the data according to the equipment operation data, automatically extract the equipment characteristics and has high automation degree.
6. Compared with the traditional shallow machine learning such as a support vector machine and fuzzy recognition, the deep learning algorithm of the convolutional neural network is adopted, the problems of insufficient learning depth and the like can be solved, the data characteristics are extracted deeply, the device characteristics are represented, the deep learning algorithm is suitable for processing large-scale data, and the fault diagnosis and prediction results are more accurate.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
Fig. 1 is a flowchart of an apparatus characteristic extraction method of a convolutional neural network based on differential feature fusion according to an embodiment of the present invention;
fig. 2 is a flowchart of a method for predicting a device failure by using a device characteristic extraction method of a convolutional neural network based on differential feature fusion according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1, a flowchart of an apparatus characteristic extraction method of a convolutional neural network based on differential feature fusion according to an embodiment of the present invention is shown, where the method includes:
s11: and acquiring the operating data of the equipment.
The operation data of the equipment is acquired and included in a system database, indexes influencing the operation of the target equipment are selected, and historical data of the operation of the equipment are read from the system database according to a certain access interval. Wherein the data is normal and fault operation data of the equipment containing a plurality of indexes and is expressed in a matrix form.
Wherein, if M pieces of operation data of N indexes related to the equipment are read from the database, the data size is large, M can be regarded as the number of samples, and N can be regarded as the dimension of the samples. The operational data may be represented in the form of a matrix as follows:
Figure BDA0002987099910000071
s12: and (4) preprocessing data.
The method comprises the following steps of deleting invalid data and processing missing values, and specifically comprises the following steps:
detecting the 'straightening data', recording the 'straightening time period for each index of the' straightening data ', and then deleting sample data corresponding to the' straightening time period. The concept of the "straightened-out data" is as follows: the data value of a certain index is kept still for a long time, and in actual operation production, the data is invalid data or error data.
For a certain sample with missing records, if the number of the missing indexes is less than or equal to 2, filling the missing value of the sample, wherein the filling value is the mean value of the indexes corresponding to the missing indexes; and if the number of the missing indexes is more than 2, the sample is considered invalid, and the sample is deleted.
S13: and carrying out extraction and fusion of differential features and data normalization processing on the operation data to obtain original feature data.
The convolutional neural network structure can automatically extract data characteristics, but change information of indexes is not considered or collected when the characteristics are extracted, and the method adopted by the embodiment of the invention comprises the following steps: before the operation of the convolutional neural network, the characteristic extraction is carried out on part of the indexes.
As a preferred technical solution of the embodiment of the present invention, a specific feature extraction method is: and extracting difference features of certain operation data, including 1-order difference and 2-order difference features of indexes.
And then performing feature fusion, specifically fusing the operating data processed by the S12 and the extracted difference feature data, and sorting the data, for example, deleting the data with the missing value, to obtain fused operating feature data.
Data normalization processing: because the specifications of various indexes or characteristics of the equipment are different, the numerical range may be greatly different, and the data are put together for operation after being subjected to characteristic extraction, which may result in low characteristic precision and failure to well characterize the equipment characteristics, the data of different specifications need to be converted into the data of the same specification, specifically: using a formula
k=(Max-Min)/(xmax-xmin)
xstandard=Min+k(x-xmin)
Normalizing individual operational data or differential features to [ Min, Max ]]Within a range, where x is operational data or differential characteristics, xstandardNormalized value, x, for the operating data or differential signaturemaxIs the original maximum, x, of the operating data or differential signatureminFor the original minimum value of the operation data or the difference characteristic, Max is a designated intervalMin is the minimum value of the specified interval.
And for the processed feature data, normalizing all feature data by adopting the formula to obtain the normalized feature data, wherein the numerical range of each feature is from Min to Max, and then rounding all feature data. And the rounded data is used as input data of the convolutional neural network, namely original characteristic data, and the characteristic data is one-dimensional characteristic data.
S14: and constructing a convolutional neural network characteristic extraction model, processing the original characteristic data through the convolutional neural network characteristic extraction model, and combining a plurality of convolutional layers to obtain dimension reduction data.
The invention improves the convolutional neural network, adopts a mode of merging a plurality of convolutional layers to extract the characteristics of merged data, changes the traditional mode of serially connecting convolutional layers and pooling layers of the convolutional neural network and enriches the characteristics of data.
The convolutional neural network characteristic extraction model comprises convolutional layers, an activation function and pooling layers, wherein one convolutional layer is provided with a plurality of convolutional kernels, the processing of the original characteristic data through the convolutional neural network characteristic extraction model comprises convolution operation of the original characteristic data through a plurality of convolutional kernels, nonlinear transformation of the obtained data through the activation function, and subsampling of input data through the pooling kernels by the pooling layers.
Specifically, in this embodiment of the present invention, the activation function is a linear rectification function ReLU, that is:
Figure BDA0002987099910000081
where x is some input data, yreluThe method is used for data of x after the ReLU activation function operation, so that the problem of gradient diffusion when the deep learning network is deep can be solved.
Specifically, the pooling layer performs sub-sampling on the input data through pooling core, and extracts features while realizing data dimension reduction. The pooling operation is mainly a maximum pooling operation and an average pooling operation. Average pooling averages and fuzzifies the data, and maximum pooling takes the maximum value of the region.
The method steps for predicting the equipment failure are as follows:
s21: and performing labeling processing on the feature matrix F _ matrix.
The method specifically comprises the following steps: and (3) marking a class label 0 for the feature matrix sample corresponding to the normal data sample, marking a label for the feature matrix sample corresponding to the fault data sample according to the fault type mark, for example, marking a class label 1 for the fault x, marking a class label 2 for the fault y, and so on. The fault type is marked as a comprehensive operation rule and is obtained by synthesizing the actual operation records of the equipment.
The tagged feature matrix data can be represented in a matrix form where y is1,y2...ynA label representing each sample.
Figure BDA0002987099910000091
Then label the category y1,y2...ynVectorization, which converts the class label value into a matrix-type representation of a binary (only 0 and 1) vector. Each label is represented by a corresponding row vector of the matrix.
S22: and dividing the characteristic matrix into a training sample and a testing sample.
The method specifically comprises the steps of randomly extracting a part of samples of a certain proportion of feature matrix data as training samples, and dividing the samples by using the rest of samples as test samples.
S23: and building a classification layer of the network model.
The method specifically comprises the following steps: the classification layer model is essentially a deep learning neural network classification model, which comprises the following steps: 2 full connection layers, 1 classification layer, its basic structure is: full connectivity layer-classification layer. Wherein the activation function of the fully-connected layer is a ReLU function, the classification layer activation function is Softmax, and the Softmax function is essentially a function that calculates a conditional probability that a fault belongs to a class under a given signal or characteristic. And the input item of the classification layer is the characteristic vector of the training sample, and the output item is the class label corresponding to the training sample.
S24: and setting a loss function, and training the network model through an optimization algorithm.
When the model is trained, a loss function is defined to evaluate the model, and then the function is optimized to achieve the optimal effect. In a preferred embodiment of the invention, a cross-entropy function is used as the loss function. It is defined as follows:
Figure BDA0002987099910000101
wherein y isiRepresenting the true value of a sample, yi' is the predicted value for that sample.
The basic idea of model training is to obtain the parameters corresponding to the minimized loss function. The method adopts a gradient descent optimization algorithm, continuously iterates, reduces the value of a loss function, further solves the parameters of the model, and then stores the model.
S25: and analyzing the characteristic matrix by using the trained network model to obtain a probability vector and predicting the fault.
The probability vector may be represented as:
p=[p1,p2,…,pnum]
and num is the number of the types of the samples, and then the label to which the maximum probability of the probability vector belongs is selected as the prediction result of the last sample.
And predicting all the test samples, and carrying out statistical calculation on the prediction accuracy. The method specifically comprises the following steps: inputting the characteristics of the test samples into the model, outputting the predicted values of the test samples, and counting the number n of samples with the predicted values equal to the true values of all the test samplesrightThen the accuracy of model prediction is:
Figure BDA0002987099910000111
where n is the total number of samples tested.
Through the steps, the application process of the equipment characteristics, namely equipment fault diagnosis and prediction, is completed.
The following further describes an implementation process of an embodiment of the present invention through detailed explanation of the embodiment of the present invention, with a steam turbine of a certain unit of a certain thermal power plant as an object.
Example (b):
step 1: and acquiring operation data.
The method comprises the steps of obtaining turbine operation data meeting conditions, wherein 56 main relevant measuring point indexes comprise unit load, main steam flow, inlet steam temperature of the A-side turbine, rotor eccentricity, bearing vibration and the like. The data is from 0 point 1/2017 to 24 points 29/12/2017, the counting interval is 30 seconds, the total number of the data is 590660, and the running data is 590660 × 56 matrix data.
Part of the operating data is as follows:
sample number Index 1 Index 2 ... ... Index 55 Index 56
1 102.8 526.3 ... ... 288.2 181.0
2 104.3 525.4 ... ... 287.9 180.8
3 106.3 524.8 ... ... 287.6 180.7
4 108.6 524.3 ... ... 287.5 180.5
5 110.2 524.0 ... ... 287.2 180.4
Step 2: and (4) preprocessing data.
And eliminating invalid data in the running data and processing missing values of the data.
And detecting and processing the operation data, detecting whether the straightening line data and the missing data exist, and preprocessing the sample data according to the step 2 of the equipment characteristic extraction flow. Through detection processing, the running data is found to be good in quality, no straightening line data exists, 10 samples have missing values, the number of the missing indexes of each sample is 1, the missing indexes of the samples are filled, and the size of the sample data after the final processing is 590660 × 56.
And step 3: and extracting, fusing and transforming differential features.
Selecting 4 key indexes such as unit load and main steam flow which influence the operation of the steam turbine to extract features of 1-order difference and 2-order difference, wherein 8 features are obtained in total, original indexes 56 are fused, 64 features are obtained in total, and original feature matrix data with the size of 590660 x 64 are formed. Preferably, when the difference characteristics of the first two sample data are null, the present invention deletes the first two sample data as a whole, and the size of the feature matrix data becomes 590658 × 64.
Further, the characteristic sample data uses a formula (2) to normalize each index to an interval [1,100], and then rounding is carried out on the characteristic number to obtain processed characteristic matrix data.
And 4, step 4: and (5) constructing a convolutional neural network, extracting the characteristics of the equipment and outputting a characteristic matrix.
And 3, constructing the convolutional neural network extraction equipment characteristics by using an improved convolutional neural network method based on the characteristic matrix data obtained in the step 3. And constructing 6 convolutional layers and 2 pooling layers, wherein the convolutional layers are used for further extracting the characteristics of the data, and the pooling layers further abstract and reduce the dimension of the characteristics on the basis of convolutional characteristic extraction, wherein the concrete conditions of each layer and each step are shown in the table.
Figure BDA0002987099910000121
Figure BDA0002987099910000131
Finally, the feature matrix data F _ matrix representing the operating characteristics of the equipment is obtained through the steps, and the size of the feature matrix data F _ matrix is 590658 × 256, wherein the feature dimension of each sample is 256.
The implementation steps of the embodiment of the invention for the power plant steam turbine equipment fault prediction are as follows:
step 1: tagging feature matrix data
Feature matrix data F _ matrix of size 590658 × 256 was obtained and 590658 samples were labeled.
The data of the turbine is analyzed, 505328 pieces of normal data are counted, and 85360 pieces of data are counted under certain faults. Therefore, the feature matrix sample corresponding to the normal data is labeled with class 0, and the fault data is labeled with class 1.
The category labels are then vectorized: tag 0 is converted to vector [0,1] and tag 1 is converted to vector [1,0 ].
Step 2: partitioning training and test samples
70% of the samples of the feature matrix data were randomly drawn as training samples, and the remaining 30% were taken as test samples. The training samples totaled 353730, and the test samples consisted of 151598.
And step 3: building a classification layer of a network model
Building a classification layer of the neural network model: 2 full connection layers and 1 classification layer. The activation function of the front layer 2 is a ReLU function, and the classification layer realizes classification by setting the activation function as a Softmax function. The input of the classification layer model is the feature vector of the training sample, and the output of the model is the class label corresponding to the training sample.
The model order results were: full tie layer F9-full tie layer F10-classification layer F11.
The specific structure and content of each layer are as follows:
content providing method and apparatus Activating a function Number of filters Characteristic size
Full connection layer F9 ReLU function 128 128
Full connection layer F10 ReLU function 256 256
Sorting layer F11 Softmax function —— ——
Wherein the discard rate is set to 0.8 after the full link layer F9 and 0.8 after the full link layer F10. If this setting is considered as a drop layer, the model structure is: full connectivity layer F9-discard layer-full connectivity layer F10-discard layer-sort layer F11.
And 4, step 4: and setting a loss function, and training a model by using an optimization algorithm.
Setting a cross entropy function as a loss function, adopting an Adam optimizer, setting an initial learning rate to be 0.001, training 100 epochs, and training a model. Wherein Adam is a common, efficient gradient descent algorithm. Finally, solving the parameters of the classification model, and then saving the model.
And 5: fault prediction
And (4) according to the stored training model, performing fault prediction on all the test data, and calculating prediction accuracy by using a formula (7). The prediction results of some of the test data are shown in the following table:
Figure BDA0002987099910000141
obtaining model prediction accuracy through calculation: 98.7 percent.
Through the steps, the process of extracting and applying the characteristics of the steam turbine equipment is completed, the accuracy of the prediction of the result surface model is high, and the method is favorable for guiding the actual production operation.
The equipment characteristic extraction method can effectively extract the equipment characteristics and carry out fault diagnosis and prediction, is not influenced by the equipment type, and has popularization.
Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that changes may be made in the embodiments and/or equivalents thereof without departing from the spirit and scope of the invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (10)

1. A method for extracting the device characteristics of a convolutional neural network based on differential feature fusion is characterized by comprising the following steps:
acquiring operation data of equipment;
extracting and fusing differential features and carrying out data normalization processing on the operation data to obtain original feature data;
and constructing a convolutional neural network characteristic extraction model, processing the original characteristic data through the convolutional neural network characteristic extraction model, and combining a plurality of convolutional layers to obtain dimension reduction data.
2. The method for extracting device characteristics of a convolutional neural network based on differential feature fusion as claimed in claim 1, wherein the preprocessing is performed on the raw feature data before the raw feature data is processed by the convolutional neural network, and includes deleting invalid data and missing value processing, specifically including:
detecting the 'straightening data', recording the 'straightening time period for each index of the' straightening data ', and then deleting sample data corresponding to the' straightening time period;
for a certain sample with missing records, if the number of the missing indexes is less than or equal to 2, filling the missing value of the sample, wherein the filling value is the mean value of the indexes corresponding to the missing indexes; and if the number of the missing indexes is more than 2, the sample is considered invalid, and the sample is deleted.
3. The method for extracting device characteristics of a convolutional neural network based on differential feature fusion as claimed in claim 2, wherein the step of obtaining the device operation data comprises selecting an index affecting the operation of the target device from a system database, and reading the device operation history data from the system database according to a certain access interval. Wherein the data is normal and fault operation data of the equipment containing a plurality of indexes and is expressed in a matrix form.
4. The method for extracting device characteristics of a convolutional neural network based on difference feature fusion as claimed in claim 3, wherein the processing of extracting and fusing difference features of the operating data comprises extracting 1-order difference and 2-order difference features of the operating data, and fusing the 1-order difference and 2-order difference features with the original feature data; the data normalization processing of the differential characteristics on the operation data comprises adopting a formula according to the specification and the numerical range of each operation data or differential characteristics of the equipment
k=(Max-Min)/(xmax-xmin)
xstandard=Min+k(x-xmin)
Normalizing individual operational data or differential features to [ Min, Max ]]Within a range, where x is operational data or differential characteristics, xstandardNormalized value, x, for the operating data or differential signaturemaxIs the original maximum, x, of the operating data or differential signatureminMax is the maximum value of the designated interval and Min is the minimum value of the designated interval.
5. The method according to claim 4, wherein the convolutional neural network feature extraction model comprises convolutional layers, an activation function and a pooling layer, one of the convolutional layers has a plurality of convolutional kernels, the processing of the raw feature data by the convolutional neural network feature extraction model comprises a plurality of convolutional kernels performing convolutional operations on the raw feature data, the activation function performs nonlinear transformation on the obtained data, and the pooling layer performs subsampling on input data by the pooling kernel, wherein the activation function is a linear rectification function ReLU:
Figure FDA0002987099900000021
where x is some input data, yreluAnd x is data after the operation of the ReLU activation function.
6. The method for extracting the device characteristics of the convolutional neural network based on the differential feature fusion as claimed in claim 5, wherein constructing the convolutional neural network characteristic extraction model comprises:
constructing an input layer of the convolutional neural network characteristic extraction model, and taking the original characteristic data as input data of the model;
construction of 5 layers of convolutional layers separately denoted C1,C2,C3,C4,C5Wherein each convolution layer adopts the same number of convolution kernels;
merging the 5 convolutional layers C1,C2,C3,C4,C5Recording the merged network as C _ merge;
performing dimension transformation on the upper output data, then constructing a pooling layer, and constructing a pooling layer P by using global maximum pooling operation of globalmaxpool6
Build up of 1 layer of convolutional layer C7
Performing dimension transformation on the upper layer data to construct a pooling layer, and adopting a global maximum pooling operation p8And finally, outputting the dimension reduction data represented by the multi-dimensional feature matrix F _ matrix.
7. A method for predicting device characteristic extraction failure by using the convolutional neural network based on differential feature fusion according to any one of claims 1 to 6, comprising:
labeling the characteristic matrix F _ matrix;
dividing the characteristic matrix into a training sample and a testing sample;
building a classification layer of the network model;
setting a loss function, and training the network model through an optimization algorithm;
analyzing the characteristic matrix by using the trained network model to obtain a probability vector and predicting faults;
8. the method of prediction of equipment failure of claim 7, wherein the classification layer comprises: 2 full connection layers and 1 classification layer, wherein the activation function of the full connection layer is a ReLU function, the activation function of the classification layer is Softmax, the input item of the classification layer is the characteristic vector of the training sample, and the output item is the class label corresponding to the training sample.
9. The method according to claim 8, wherein the loss function is a cross-entropy function, that is, a function of:
Figure FDA0002987099900000031
wherein y isiRepresents the true value, y 'of a sample'iIs the predicted value for that sample.
10. The method for extracting device characteristics of a convolutional neural network based on differential feature fusion as claimed in claim 9, wherein the probability vector can be expressed as:
p=[p1,p2,…,pnum]
wherein num is the number of the types of the samples, and then a label to which the maximum probability of the probability vector belongs is selected as a prediction result of the last sample;
inputting all the test samples into the network model, outputting the predicted values of the test samples, and counting the number n of samples with the predicted values of all the test samples equal to the true valuesrightThen the accuracy of model prediction is:
Figure FDA0002987099900000041
where n is the total number of samples tested.
CN202110303120.9A 2021-03-22 2021-03-22 Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion Withdrawn CN112836809A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110303120.9A CN112836809A (en) 2021-03-22 2021-03-22 Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110303120.9A CN112836809A (en) 2021-03-22 2021-03-22 Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion

Publications (1)

Publication Number Publication Date
CN112836809A true CN112836809A (en) 2021-05-25

Family

ID=75930539

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110303120.9A Withdrawn CN112836809A (en) 2021-03-22 2021-03-22 Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion

Country Status (1)

Country Link
CN (1) CN112836809A (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113688801A (en) * 2021-10-22 2021-11-23 南京智谱科技有限公司 Chemical gas leakage detection method and system based on spectrum video
CN114399066A (en) * 2022-01-15 2022-04-26 中国矿业大学(北京) Mechanical equipment predictability maintenance system and maintenance method based on weak supervision learning
CN115359300A (en) * 2022-08-30 2022-11-18 德玛克(长兴)精密机械有限公司 Intelligent processing and manufacturing system and method for wind driven generator component
WO2023015631A1 (en) * 2021-08-13 2023-02-16 广东技术师范大学 Missing data-based classification model generation method
CN117131348A (en) * 2023-10-27 2023-11-28 深圳中科保泰科技有限公司 Data quality analysis method and system based on differential convolution characteristics

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2023015631A1 (en) * 2021-08-13 2023-02-16 广东技术师范大学 Missing data-based classification model generation method
CN113688801A (en) * 2021-10-22 2021-11-23 南京智谱科技有限公司 Chemical gas leakage detection method and system based on spectrum video
CN114399066A (en) * 2022-01-15 2022-04-26 中国矿业大学(北京) Mechanical equipment predictability maintenance system and maintenance method based on weak supervision learning
CN115359300A (en) * 2022-08-30 2022-11-18 德玛克(长兴)精密机械有限公司 Intelligent processing and manufacturing system and method for wind driven generator component
CN117131348A (en) * 2023-10-27 2023-11-28 深圳中科保泰科技有限公司 Data quality analysis method and system based on differential convolution characteristics
CN117131348B (en) * 2023-10-27 2024-02-09 深圳中科保泰科技有限公司 Data quality analysis method and system based on differential convolution characteristics

Similar Documents

Publication Publication Date Title
CN112836809A (en) Device characteristic extraction method and fault prediction method of convolutional neural network based on differential feature fusion
CN111507376B (en) Single-index anomaly detection method based on fusion of multiple non-supervision methods
CN108985380B (en) Point switch fault identification method based on cluster integration
CN108304567B (en) Method and system for identifying working condition mode and classifying data of high-voltage transformer
CN111882446A (en) Abnormal account detection method based on graph convolution network
CN112949715A (en) SVM (support vector machine) -based rail transit fault diagnosis method
CN111767398A (en) Secondary equipment fault short text data classification method based on convolutional neural network
CN108052504A (en) Mathematics subjective item answers the structure analysis method and system of result
CN112329884B (en) Zero sample identification method and system based on discriminant visual attributes
CN113011763A (en) Bridge damage identification method based on space-time diagram convolution attention
CN116610816A (en) Personnel portrait knowledge graph analysis method and system based on graph convolution neural network
CN113128567A (en) Abnormal electricity consumption behavior identification method based on electricity consumption data
CN117131449A (en) Data management-oriented anomaly identification method and system with propagation learning capability
CN112416732B (en) Hidden Markov model-based data acquisition operation anomaly detection method
CN111737993B (en) Method for extracting equipment health state from fault defect text of power distribution network equipment
CN113378024A (en) Deep learning-based public inspection field-oriented related event identification method
CN117372144A (en) Wind control strategy intelligent method and system applied to small sample scene
CN107122394A (en) Abnormal deviation data examination method and device
CN111626559A (en) Main factor analysis method-based medium-voltage distribution network line loss key characteristic index extraction method and system
CN113393169B (en) Financial industry transaction system performance index analysis method based on big data technology
CN113657726B (en) Personnel risk analysis method based on random forest
CN113705695A (en) Power distribution network fault data identification method based on convolutional neural network
CN114077663A (en) Application log analysis method and device
CN111210147A (en) Sintering process operation performance evaluation method and system based on time sequence feature extraction
CN110569277A (en) Method and system for automatically identifying and classifying configuration data information

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
WW01 Invention patent application withdrawn after publication
WW01 Invention patent application withdrawn after publication

Application publication date: 20210525