CN112001530A - Predictive maintenance method and system for transformer oil chromatography online monitoring device - Google Patents

Abstract

The invention discloses a predictive maintenance method and a predictive maintenance system for an oil chromatogram on-line monitoring device of a transformer, which are characterized in that B-EMD is adopted to perform data fitting and empirical mode analysis on an oil chromatogram to obtain oil chromatogram validity evaluation parameters, a multidimensional DBN-based oil chromatogram data validity analysis network is adopted to evaluate the validity of modal characteristics of oil chromatogram data, and finally a comprehensive evaluation result of the oil chromatogram data validity is obtained by comprehensively weighting different dimensions through a comprehensive weight factor sorting network, so that the evaluation on the operation reliability of the oil chromatogram on-line monitoring device is realized, and the planned maintenance can be carried out in a targeted manner.

Description

Predictive maintenance method and system for transformer oil chromatography online monitoring device

Technical Field

The invention belongs to the field of power equipment fault diagnosis, and particularly relates to a predictive maintenance method for a transformer oil chromatography online monitoring device based on multi-scale deep feature learning.

Background

The reliability of the power transformer, which is the most critical device of the power system, is directly related to the operation safety of the power system. At present, analysis of Dissolved Gas (DGA) (referred to as oil chromatography for short) in transformer oil has been widely applied to monitoring and diagnosing transformer defects and latent faults as an effective means, and a large number of oil chromatography online monitoring devices are also deployed in each provincial power grid. However, the service life of the existing oil chromatography online monitoring device is far shorter than that of the power transformer body, and the oil chromatography online monitoring device is mostly installed outdoors and has a severe operating environment, so that a great number of cases of oil chromatography analysis accuracy reduction and data distortion caused by factors such as dead oil, degasser aging, spectrometer deviation and the like exist, and the reliability and the credibility of oil chromatography monitoring data are greatly reduced. In addition, limiting factors such as complicated instrument operation and high professional requirements exist in field regular calibration work of the oil chromatography device, so that comprehensive regular calibration after field operation of the oil chromatography on-line monitoring device is extremely difficult to popularize practically, and therefore predictive maintenance research on the oil chromatography on-line monitoring device needs to be carried out urgently. Compared with the traditional preventive maintenance, the predictive maintenance is not based on a fixed maintenance period, but utilizes the field online monitoring measurement data and corresponding data evaluation to carry out advanced verification and maintenance on the field high-risk oil chromatography online monitoring equipment, so that the fault rate of the oil chromatography online monitoring equipment is reduced under the condition of limited overhaul resources.

Disclosure of Invention

The purpose of the invention is as follows: in order to fill the gap of the prior art in the predictive maintenance method of the oil chromatography on-line monitoring device, the invention provides the predictive maintenance method of the transformer oil chromatography on-line monitoring device based on multi-scale depth feature learning.

The technical scheme is as follows: a predictive maintenance method for a transformer oil chromatography online monitoring device comprises the following steps:

s100: performing data fitting and empirical mode analysis on oil chromatography data obtained by monitoring of an oil chromatography online monitoring device to obtain a plurality of IMF data;

s200: carrying out layer-by-layer feature extraction on the plurality of IMF data to obtain a feature signal;

s300: and (5) sequencing the weight factors of the characteristic signals in the S200 to obtain an operation reliability result of the oil chromatogram on-line monitoring device.

Further, the S100 specifically includes the following sub-steps:

b spline fitting is carried out on oil chromatography data obtained by monitoring of the oil chromatography on-line monitoring device to obtain a B spline fitting curve;

and performing EMD on the B spline fitting curve to obtain a plurality of IMF data.

Further, the B-spline fitting of the oil chromatogram data to obtain a B-spline fitting curve comprises the following steps:

s110: setting oil chromatogram monitoring sequence u ═ u₀,u₁,u₂,…,u_n+pAnd (6) calculating by adopting a Cox-debor recursive formula to obtain a p-order B spline function B_j,p：

In the formula, B_j,pIs a p-th order B-spline function, B_j,1Is a 1 st order B-spline function, u_jFor the jth oil chromatographic monitoring sequence, u_j+1For the j +1 th oil chromatography monitoring sequence, u_j+pFor the j + p oil chromatographic monitoring sequences, u_j+p-1For the j + p-1 th oil chromatography monitoring sequence, B_j,p-1Is a p-1 th order B spline function, B_j+1,p-1Is a B spline function of p-1 order;

s120: introducing (t), and solving a parameter c to be solved in B spline fitting through least squares of (t)_j：

Wherein y (t) is a raw oil chromatogram data curve, g (t) is a p-th order B-spline fitting curve, c_jFor the parameter to be solved, B_j,pIs a p-order B spline function;

s130: based on p-order B spline function B_j,pParameter c to be solved in B spline fitting_jObtaining a p-order B spline fitting curve g (t):

in the formula, B_j,pIs a P-th order B-splineFunction, c_jThe parameters obtained in S120 are obtained.

Further, the DBN network in S200 comprises a plurality of RBM units, each layer of the RBM units is respectively provided with a visual layer v epsilon {0,1}^DAnd hidden layer h e {0,1}^MWherein D and M are the dimensions of the phase variable.

Further, the S200 includes: respectively inputting the IMF data into corresponding DBN networks to carry out layer-by-layer feature extraction to obtain feature signals; the DBN network comprises a plurality of RBM units, and each layer of the RBM units is formed by a visual layer v epsilon {0,1}^DAnd hidden layer h e {0,1}^MWherein D and M are the dimensions of the phase variable.

The DBN network is established by the following steps:

s210: defining a visual layer v epsilon {0,1} in each layer of RBM units^DAnd hidden layer h e {0,1}^MEnergy function between:

wherein θ is RBM unit parameter θ ═ W, b, a, W_ijRepresenting the ith visual layer node v_iAnd the ith hidden layer node h_jWeight between, a_j,b_iIs the bias coefficient of RBM cell, v_iIs the ith visible layer node, h_jThe node is an ith hidden layer node, D is a visible layer dimension, and M is a hidden layer dimension;

based on the energy function, a single RBM unit is obtained, which is represented as:

wherein p denotes an unnormalized probability function and Z (θ) denotes a distribution function;

s220: constructing a DBN based on a single RBM unit;

s230: pre-training the constructed DBN by adopting an unlabeled IMF sample to obtain a trained DBN;

s240: and fine-tuning the trained DBN through the marked IMF sample to obtain a finally usable DBN.

Further, the pre-training of the constructed DBN network by using the unlabeled IMF sample in S230 to obtain a trained DBN network includes:

s231: training the first layer of RBM units by using an RBM parameter training algorithm to obtain theta ═ W, b, a };

s232: taking a parameter theta in the first-layer RBM unit as the input of the first-layer RBM unit;

s233: training the second layer of RBM units by adopting an RBM parameter training algorithm;

s234: and repeating the steps until the last layer of RBM unit is trained, and obtaining the trained DBN.

Further, the RBM parameter training algorithm is described as:

wherein (·) represents a learning efficiency function;<·>_dataa mathematical expectation function representing the training samples,<·>_k、Δw_ij、Δb_i、Δa_irespectively representing a mathematical expectation function, an updating weight and an updating bias-actuating quantity after k times of Gibbs sampling is carried out on the training sample, v_iIs the ith visible layer node, h_iAnd h_jRespectively an ith hidden layer node and a jth hidden layer node;

and obtaining the RBM unit parameters after training based on the theta ═ W, b, a.

Further, in S240, the trained DBN network is trimmed by marking the sample, so as to obtain a finally usable DBN network, including:

and taking the output of the last layer of RBM unit in the DBN as the input of the trained DBN, and training the trained DBN from back to front through a BP back propagation algorithm to obtain the finally usable DBN.

Further, the S300 specifically includes: inputting the characteristic signals in the S200 into a weight factor sorting network for weight factor sorting to obtain an operation reliability result of the oil chromatography on-line monitoring device;

the weight factor ranking network is represented as:

wherein S is the final result of labeling, Y_nFor confidence output of the DBN network, α_nIs a weight value.

The invention also discloses a predictive maintenance system of the transformer oil chromatography on-line monitoring device, which comprises the following steps:

the data fitting and empirical mode analysis module is used for performing data fitting and empirical mode analysis on the input oil chromatographic data by adopting a B-EMD oil chromatographic data fitting and empirical mode analysis method to obtain a plurality of IMF data;

the DBN network modules are used for carrying out layer-by-layer feature extraction on the IMF data which are correspondingly input to obtain feature signals;

and the weight factor sequencing network module is used for carrying out weight factor sequencing on the characteristic signals output by the DBN network module to obtain an operation reliability result of the oil chromatogram online monitoring device.

Has the advantages that: compared with the prior art, the invention has the following advantages:

1. the invention provides a B-EMD oil chromatographic data fitting and empirical mode analysis method aiming at the chaos of oil chromatographic data, and improves the modal accuracy of the traditional EMD algorithm in the oil chromatographic data decomposition process to perform data fitting and empirical mode analysis on the oil chromatogram;

2. aiming at the problem of effectiveness evaluation of oil chromatographic data, the effectiveness evaluation is carried out on oil chromatographic state monitoring data by establishing an oil chromatographic data reliability analysis network based on a multidimensional DBN;

3. analysis of actual oil chromatogram data shows that the method can effectively judge the effectiveness of the online chromatographic monitoring device, find the fault of the oil chromatogram device in advance, overhaul in time, improve the operation reliability level of the oil chromatogram device and provide support for predictive maintenance of the oil chromatogram monitoring device.

Drawings

FIG. 1 is a B-EMD decomposition diagram of an oil chromatographic signal;

FIG. 2 is a schematic diagram of a DBN network;

FIG. 3 is a schematic view of a RBM structure;

FIG. 4 is a network of oil chromatography signal validity assessment;

FIG. 5 is a flow chart of network training for evaluating validity of oil chromatogram signals;

FIG. 6 is a sample library of oil chromatography monitoring data established based on a sliding window lattice method;

FIG. 7 is a graph of IMF1 network training errors;

FIG. 8 is a graph of IMF2 network training errors;

FIG. 9 is a graph of IMF3 network training errors;

FIG. 10 is a diagram of DBN, RNN, FCN network errors;

FIG. 11 is a graph of accuracy of different levels of network training;

FIG. 12 is a visual comparison of t-SNE of different layers of DBN data.

Detailed Description

The invention is further illustrated below with reference to the figures and examples.

The embodiment provides a predictive maintenance method of a transformer oil chromatography online monitoring device based on multi-scale depth feature learning, which comprises the following steps:

step 1: for reliability evaluation of oil chromatogram data, the embodiment provides a probability generation model based on a Deep Belief Network (DBN), and the DBN estimates conditional probabilities (P (underpservation | Label) and P (Label | obsservation)) of event occurrence by establishing joint probability distribution between measurement data (obsservation) and a Label (Label);

constructing a DBN network as shown in FIG. 2, which sequentially comprises an input layer, a plurality of RBM units and an output layer; referring to FIG. 3, each RBM unit consists of a layer of visual layers v e {0,1}^DAnd a hidden layer h e {0,1}^MWherein D and M are dimensions of phase variables, the energy function of which is defined as:

wherein θ is RBM unit parameter θ ═ W, b, a, W_ijRepresenting the ith visual layer node v_iAnd the ith hidden layer node h_jWeight between, a_j,b_iIs the bias coefficient of RBM cell, v_iIs the ith visible layer node, h_jIs the ith hidden layer node, D is the visible layer dimension, and M is the hidden layer dimension.

Step 2: training the constructed DBN network, wherein the training process comprises the following steps:

step 2.1: B-EMD oil chromatographic data fitting and empirical mode analysis methods are adopted to perform B spline fitting and EMD decomposition on all oil chromatographic data samples to obtain IMF₁、IMF₂、IMF₃Decomposing a sample, referring to fig. 1, in the actual operation process of the oil chromatogram on-line monitoring device, the sampling frequency of the oil chromatogram is mostly 1-3 times per day, considering that each IMF decomposed by the B-EMD can be used as a narrow-band signal, and combining the actual decomposition effect of the B-EMD, error signals caused by the oil chromatogram measuring device are mainly distributed in the IMF₁-IMF₃Performing the following steps;

step 2.2: using unlabelled IMF₁、IMF₂、IMF₃Decomposing a sample to pre-train the DBN network 1, the DBN network 2 and the DBN network 3 respectively to obtain the trained DBN network 1, the trained DBN network 2 and the trained DBN network 3;

step 2.3: after pre-training is complete, marked IMF is used₁-IMF₃Decomposing the sample to finely adjust the trained DBN network 1, DBN network 2 and DBN network 3, wherein the fine adjustment process is described as follows: the output of the last layer RBM in the DBN network is used asAnd (3) carrying out supervised training from back to front through a BP back propagation algorithm to obtain the available DBN (distributed binary network) 1, DBN 2 and DBN 3 for the input of the BP back propagation network.

The DBN network is already applied to transformer fault analysis, insulator fault analysis and fan fault analysis, and compared with other deep learning models, the DBN network has the advantages that before the marked samples are trained, the unsupervised marked samples can be pre-trained through a layer-by-layer unsupervised learning algorithm to obtain network parameter distribution close to an optimal solution, and data are finely adjusted according to the marked samples so as to accelerate the training process of the network.

The training of the DBN network in step 2 specifically includes the following substeps:

visual layer node v_iAnd hidden layer node h_jThe conditional probability distribution in between can be expressed as:

the parameter θ in the RBM unit can be obtained by the maximum likelihood function of p (h; θ) to the edge distribution of θ:

thus, the training algorithm for the RBM parameters can be described as:

wherein (DEG) isLearning an efficiency function;<·>_dataa mathematical expectation function representing the training samples,<·>_k、Δw_ij、Δb_i、Δa_irespectively representing a mathematical expectation function, an updating weight and an updating bias-actuating quantity after k times of Gibbs sampling is carried out on the training sample, v_iIs the ith visible layer node, h_iAnd h_jI, j hidden layer nodes.

The entire DBN pre-training process can be described as:

(1) training a first-layer RBM unit according to equations (9) - (12);

(2) taking the parameters of the hidden layer in the first layer of RBM unit as the input of the second layer of RBM unit;

(3) the second tier RBM units are trained according to equations (9) - (12).

And the whole pre-training process of the DBN network is finished by analogy.

And step 3: and constructing a weight factor sequencing network, taking the output result of the trained DBN network as the input of the weight factor sequencing network, and training the weight factor sequencing network through a BP back propagation algorithm to obtain the usable weight factor sequencing network. The method specifically comprises the following steps:

for the trained IMF₁-IMF₃Network, assuming confidence output of three networks as Y₁、Y₂、Y₃And constructing an importance function.

S＝α₁Y₁+α₂Y₂+α₃Y₃ (14)

For all labeled samples, S is the final result of labeling, Y₁-Y₃is-IMF₃And (4) outputting of the network. Obtaining optimal weight alpha by adopting least square method₁-α₃。

And 4, step 4: performing data fitting and empirical mode analysis on the oil chromatographic data obtained by the oil chromatographic on-line monitoring device by adopting a B-EMD oil chromatographic data fitting and empirical mode analysis method to obtain the first three intrinsic mode functions IMF₁-IMF₃；

And 5: the first three natural mode functionsNumber IMF₁-IMF₃Respectively accessing the DBN network 1, the DBN network 2 and the DBN network 3 to extract features;

step 6: and inputting the output results of the three DBN networks into an available weight factor sorting network for weight factor sorting, and comprehensively obtaining the operation reliability result of the oil chromatogram on-line monitoring device.

The overall structure diagram of the data validity analysis based on the B-EMD and DBN oil chromatography monitoring is shown in FIG. 4, and the training process can be seen in FIG. 5.

Now, the method for fitting the B-EMD oil chromatographic data and performing empirical mode analysis adopted in this embodiment is described as follows: assuming that the oil chromatogram data curve is y (t), the p-order B-spline fitting curve g (t) can be expressed as:

in the formula, c_jFitting the parameters to be solved for B-spline, B_j,pIs a B-spline function of order p.

Suppose the original oil chromatography monitoring sequence u ═ { u ═₀,u₁,u₂,…,u_n+pB spline function of order p_j,pIt can be calculated using the Cox-deBoor recursive formula:

in the formula, B_j,pIs a p-th order B-spline function, B_j,1Is a 1 st order B-spline function, u_jFor the jth oil chromatographic monitoring sequence, u_j+1For the j +1 th oil chromatography monitoring sequence, u_j+pFor the j + p oil chromatographic monitoring sequences, u_j+p-1For the j + p-1 th oil chromatography monitoring sequence, B_j,p-1Is a p-1 th order B spline function, B_j+1,p-1Is a B-spline function of order p-1.

Taking the difference between the B spline fitting curve g (t) and the original oil chromatographic signal y (t) as (t):

(t)＝y(t)-g(t) (4)

therefore, the parameter c to be solved in B-spline fitting_jThe following can be found by least squares of (t):

based on B_j,pAnd c_jObtaining the p-order B-spline fitting curve g (t).

And performing EMD on the p-order B spline fitting curve g (t) to obtain a plurality of IMF data.

Each IMF can be used as a narrow-band signal, and error signals caused by the oil chromatographic measurement device are mainly distributed in the IMF by combining the actual decomposition effect of B-EMD₁-IMF₃In (1).

Example 2:

the gas component detected by the current oil chromatogram on-line monitoring device comprises H₂、CH₄、C₂H₄、C₂H₆、C₂H₂、CO、CO₂7 characteristic gases, however, most of the existing online monitoring devices can only monitor the first 5 gases. In this embodiment, 54 thousands of effective records (removing obvious abnormal data such as data interruption, zero value, infinity, and the like) of 354 oil chromatography online monitoring devices in 2019 of 2007 and osage province are collected, a sample library is established by using a sliding window lattice method, wherein the length of a sliding window lattice is 30 days, the step length is 5 days, and as shown in fig. 6, effective training sample 32774 bars are obtained in total.

The computer used in the embodiment is provided with an 8-core Xeon CPU, a GTX1080TI video card and a 32G memory, relevant oil chromatography online monitoring check record data are searched for 441 pieces in a PMS database, and 4541 pieces of samples are marked according to check results, wherein 3632 pieces of positive samples are marked, and 909 pieces of negative samples are marked. The data of each gas in the sample are respectively disassembled, and the oil chromatography on-line monitoring data validity evaluation network is trained based on the method of example 1, wherein the training error of the signal decomposition network for H2 is shown in FIGS. 7-9.

In order to obtain an optimal oil chromatography monitoring data evaluation model, the present embodiment compares the depth models based on different feature extraction mechanisms and the failure test sample identification accuracy under different model structures and training parameters, and comprehensively analyzes and selects the optimal oil chromatography monitoring data evaluation model.

In this embodiment, feature extraction mechanisms of three networks, namely, a DBN network, an RNN network and an FCN network, are respectively adopted, the same sample is used to test the model, and the recognition accuracy of the evaluation model based on the three types of feature extraction mechanisms on the oil chromatogram data is compared, and the result is shown in fig. 10.

As can be seen from fig. 10, as the training times increase, the recognition accuracy and the training speed of the DBN model are superior to those of the other two models because the DBN network includes an unsupervised pre-training process, and can use unlabeled samples to perform pre-training, and fully explore the potential of unlabeled data, so that the model parameters are close to the optimal solution, and thus can be more quickly degraded to the optimal solution in the training process of labeled samples; and the other two models have no pre-training process due to the limitation of the network structure, the initial parameter distribution is closer to random distribution, the convergence speed is low, the local optimization is easy to fall into, and the optimal solution is difficult to obtain.

Example 3:

on the basis of the embodiment 1, in order to obtain the optimal DBN structure and training parameters, the embodiment compares the evaluation effect of the model on the oil chromatogram monitoring data under different boltzmann machine layer numbers and training cycle numbers, so that the optimal DBN model is constructed according to the actual characteristics of the input sample.

Referring to fig. 11, after the number of network layers of the DBN model reaches 4 and the training period reaches 250, the recognition accuracy of the model tends to converge, and if the number of model layers or the training period is increased, the training and testing time is increased, which affects the calculation efficiency. Therefore, for the practical case of the data of this embodiment, a DBN model of 4 layers is selected, and the number of training cycles is 250.

To verify the validity of the oil chromatography monitoring data of the method of the present invention, the present example was also compared with phase space reconstruction and data driving, the comparison is shown in table 1.

TABLE 1 comparison of effectiveness of different algorithms

Contrast item	Rate of identification accuracy	Is the parameter manually adjusted?	Generalization performance
				The methods as presented herein	95.2％	Whether or not	High strength
Phase space reconstruction	91.5％	Whether or not	Weak (weak)
				Data driving	82.3％	Is that	Weak (weak)

As can be seen from table 1, the model provided by the invention has the highest identification precision on a fault sample, and the phase space reconstruction method is inferior, because the phase space reconstruction technology needs to manually select a time delay parameter, time test statistics and other index parameters, the generalization capability is not strong, and the optimal reconstruction parameters corresponding to the transformers with different operation parameters need to be selected; similarly, the method based on data-driven and multi-fusion criteria also faces the problem of weak generalization, and the method still needs to manually select parameters such as a marker value criterion, a coefficient of variation criterion and the like. The effectiveness evaluation model provided by the invention adopts a feature extraction mechanism of a deep learning DBN model, can automatically mine the oil chromatogram monitoring data features under different faults and normal operation conditions in a mass of samples, avoids the influence of improper manual feature selection, and improves the generalization performance of the model.

The transformer oil dissolved gas data used in this example was decomposed into three modal signals by B-EMD: IMF1, IMF2 and IMF3, in order to analyze the signal characteristics, the present embodiment uses the t-SNE algorithm to map the decomposed three mode signals to a low-dimensional space, as shown in FIG. 12. The embodiment sufficiently reserves abnormal points in the decomposed signals by using the long tail of t distribution in the t-SNE algorithm, thereby being convenient for analyzing the data characteristics of different signals in a low-dimensional space.

Meanwhile, fig. 12 shows that after the DBN layer-by-layer feature extraction, the difference between different feature signal clusters becomes more obvious, so that the DBN model used in this embodiment can effectively extract data features from the original high-dimensional decomposition signal, thereby performing reliability evaluation on the high validity of the transformer oil chromatogram on-line monitoring data.

Claims (10)

1. A predictive maintenance method for a transformer oil chromatography on-line monitoring device is characterized by comprising the following steps: the method comprises the following steps:

s100: performing data fitting and empirical mode analysis on oil chromatography data obtained by monitoring of an oil chromatography online monitoring device to obtain a plurality of IMF data;

s200: carrying out layer-by-layer feature extraction on the plurality of IMF data to obtain a feature signal;

s300: and (5) sequencing the weight factors of the characteristic signals in the S200 to obtain an operation reliability result of the oil chromatogram on-line monitoring device.

2. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 1, wherein: the S100 specifically includes the following substeps:

b spline fitting is carried out on oil chromatography data obtained by monitoring of the oil chromatography on-line monitoring device to obtain a B spline fitting curve;

and performing EMD on the B spline fitting curve to obtain a plurality of IMF data.

3. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 2, characterized in that: the method for performing B-spline fitting on the oil chromatogram data to obtain a B-spline fitting curve comprises the following steps:

s110: setting oil chromatogram monitoring sequence u ═ u₀,u₁,u₂,…,u_n+pAnd (6) calculating by adopting a Cox-debor recursive formula to obtain a p-order B spline function B_j,p：

In the formula, B_j,pIs a p-th order B-spline function, B_j,1Is a 1 st order B-spline function, u_jFor the jth oil chromatographic monitoring sequence, u_j+1For the j +1 th oil chromatography monitoring sequence, u_j+pFor the j + p oil chromatographic monitoring sequences, u_j+p-1For the j + p-1 th oil chromatography monitoring sequence, B_j,p-1Is a p-1 th order B spline function, B_j+1,p-1Is a B spline function of p-1 order;

s120: introducing (t), and solving a parameter c to be solved in B spline fitting through least squares of (t)_j：

Wherein y (t) is a raw oil chromatogram data curve, g (t) is a p-th order B-spline fitting curve, c_jFor the parameter to be solved, B_j,pIs a p-order B spline function;

s130: based on p-order B spline function B_j,pParameter c to be solved in B spline fitting_jObtaining a p-order B spline fitting curve g (t):

in the formula, B_j,pIs a B-spline function of order p, c_jThe parameters obtained in S120 are obtained.

4. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 1, wherein: the S200 includes:

respectively inputting the IMF data into corresponding DBN networks to carry out layer-by-layer feature extraction to obtain feature signals; the DBN network comprises a plurality of RBM units, and each layer of the RBM units is formed by a visual layer v epsilon {0,1}^DAnd hidden layer h e {0,1}^MWherein D and M are the dimensions of the phase variable.

5. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 4, wherein: the DBN network is established by the following steps:

s210: defining a visual layer v epsilon {0,1} in each layer of RBM units^DAnd hidden layer h e {0,1}^MEnergy function between:

wherein θ is RBM unit parameter θ ═ W, b, a, W_ijRepresenting the ith visual layer node v_iAnd the ith hidden layer node h_jWeight between, a_j,b_iIs the bias coefficient of RBM cell, v_iIs the ith visible layer node, h_jThe node is an ith hidden layer node, D is a visible layer dimension, and M is a hidden layer dimension;

based on the energy function, a single RBM unit is obtained, which is represented as:

wherein p denotes an unnormalized probability function and Z (θ) denotes a distribution function;

s220: constructing a DBN based on a single RBM unit;

s230: pre-training the constructed DBN by adopting an unlabeled IMF sample to obtain a trained DBN;

s240: and fine-tuning the trained DBN through the marked IMF sample to obtain a finally usable DBN.

6. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 5, wherein: in S230, pre-training the constructed DBN network by using an unlabeled IMF sample to obtain a trained DBN network, including:

s231: training the first layer of RBM units by using an RBM parameter training algorithm to obtain theta ═ W, b, a };

s232: taking a parameter theta in the first-layer RBM unit as the input of the first-layer RBM unit;

s233: training the second layer of RBM units by adopting an RBM parameter training algorithm;

s234: and repeating the steps until the last layer of RBM unit is trained, and obtaining the trained DBN.

7. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 6, wherein: the RBM parameter training algorithm is described as follows:

wherein (·) represents a learning efficiency function;<·>_dataa mathematical expectation function representing the training samples,<·>_k、Δw_ij、Δb_i、Δa_irespectively representing a mathematical expectation function, an updating weight and an updating bias-actuating quantity after k times of Gibbs sampling is carried out on the training sample, v_iIs the ith visible layer node, h_iAnd h_jRespectively an ith hidden layer node and a jth hidden layer node;

and obtaining the RBM unit parameters after training based on the theta ═ W, b, a.

8. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 5, wherein: in S240, the trained DBN network is fine-tuned by marking the sample, so as to obtain a finally usable DBN network, including:

and taking the output of the last layer of RBM unit in the DBN as the input of the trained DBN, and training the trained DBN from back to front through a BP back propagation algorithm to obtain the finally usable DBN.

9. The predictive maintenance method for the on-line chromatographic monitoring device of transformer oil as claimed in claim 1, wherein: the S300 specifically includes:

inputting the characteristic signals in the S200 into a weight factor sorting network for weight factor sorting to obtain an operation reliability result of the oil chromatography on-line monitoring device;

the weight factor ranking network is represented as:

wherein S is the final result of labeling, Y_nOutput for confidence of DBN network，α_nIs a weight value.

10. The maintenance system of the predictive maintenance method for the on-line chromatographic monitor device of transformer oil according to any one of claims 1 to 9, wherein: the method comprises the following steps:

the data fitting and empirical mode analysis module is used for performing data fitting and empirical mode analysis on the input oil chromatographic data by adopting a B-EMD oil chromatographic data fitting and empirical mode analysis method to obtain a plurality of IMF data;

the DBN network modules are used for carrying out layer-by-layer feature extraction on the IMF data which are correspondingly input to obtain feature signals;

and the weight factor sequencing network module is used for carrying out weight factor sequencing on the characteristic signals output by the DBN network module to obtain an operation reliability result of the oil chromatogram online monitoring device.

Images