CN117556347A - Power equipment fault prediction and health management method based on industrial big data - Google Patents
Power equipment fault prediction and health management method based on industrial big data Download PDFInfo
- Publication number
- CN117556347A CN117556347A CN202311546874.2A CN202311546874A CN117556347A CN 117556347 A CN117556347 A CN 117556347A CN 202311546874 A CN202311546874 A CN 202311546874A CN 117556347 A CN117556347 A CN 117556347A
- Authority
- CN
- China
- Prior art keywords
- data
- feature
- prediction
- value
- fault
- 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
- 230000036541 health Effects 0.000 title claims abstract description 28
- 238000007726 management method Methods 0.000 title claims abstract description 23
- 238000000034 method Methods 0.000 claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 30
- 230000002159 abnormal effect Effects 0.000 claims abstract description 19
- 238000012795 verification Methods 0.000 claims abstract description 13
- 238000012360 testing method Methods 0.000 claims abstract description 11
- 238000012545 processing Methods 0.000 claims abstract description 10
- 238000010606 normalization Methods 0.000 claims abstract description 4
- 230000006870 function Effects 0.000 claims description 36
- 239000011159 matrix material Substances 0.000 claims description 36
- 238000003066 decision tree Methods 0.000 claims description 30
- 239000013598 vector Substances 0.000 claims description 11
- 238000009826 distribution Methods 0.000 claims description 10
- 230000002441 reversible effect Effects 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 230000008901 benefit Effects 0.000 claims description 8
- 238000004364 calculation method Methods 0.000 claims description 8
- 238000000605 extraction Methods 0.000 claims description 8
- 230000007246 mechanism Effects 0.000 claims description 8
- 238000012544 monitoring process Methods 0.000 claims description 7
- 230000009467 reduction Effects 0.000 claims description 7
- 230000006835 compression Effects 0.000 claims description 3
- 238000007906 compression Methods 0.000 claims description 3
- 230000001902 propagating effect Effects 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000004458 analytical method Methods 0.000 description 8
- 238000004519 manufacturing process Methods 0.000 description 6
- 230000002457 bidirectional effect Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 238000010801 machine learning Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 238000009776 industrial production Methods 0.000 description 3
- 238000012843 least square support vector machine Methods 0.000 description 3
- 230000003321 amplification Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000003745 diagnosis Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000002427 irreversible effect Effects 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000012423 maintenance Methods 0.000 description 2
- 238000003199 nucleic acid amplification method Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 102100037651 AP-2 complex subunit sigma Human genes 0.000 description 1
- 101000806914 Homo sapiens AP-2 complex subunit sigma Proteins 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002265 prevention Effects 0.000 description 1
- 230000002035 prolonged effect Effects 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
-
- 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
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/01—Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a power equipment fault prediction and health management method based on industrial big data, which relates to the field of motor control and solves the problems of parameter feature redundancy and unbalanced fault data, and comprises the following steps: s1: constructing a data set of operation parameters containing fault information, S2: carrying out abnormal value and missing value processing and realizing normalization; s3: performing feature selection, and selecting a feature group with stronger correlation with equipment operation index parameters; s4: dividing the data of the feature group into a training set, a testing set and a verification set; s5: an operation parameter prediction model of normal operation of the MSA-BLSTM is established for prediction; s6: performing prediction model verification by using a verification set, taking predicted operation parameters output by the MSA-BLSTM as input, and performing fault judgment by using an isolated forest method model to realize fault prediction of equipment; s7: and carrying out early warning on equipment faults and health state assessment according to the fault prediction result.
Description
Technical Field
The invention relates to the field of electric power, in particular to an electric power equipment fault prediction and health management method based on industrial big data.
Background
In the modern power industry, the safe operation of the power system equipment group has important significance on production, human resources and environment, and along with the development and application of technologies such as sensors, internet of things, artificial intelligence and the like, various equipment groups accumulate industrial acquisition data and simultaneously promote equipment fault prediction and health management to enter an industrial big data era. With the continuous improvement of the intelligent and automatic degree of power system equipment, equipment fault prediction and health management technologies have become one of key technologies for safe and stable operation of a power system. The failure of the power equipment not only can cause shutdown maintenance, but also can bring great influence to production and life, so that the prediction and prevention of the failure of the equipment have become an important subject in the power industry.
Patent CN112462734a discloses a failure prediction analysis method and model of industrial production equipment, comprising the following steps: step S1: dividing equipment levels and collecting operation data; collecting fault rate, fault category and fault reason in each period for general equipment; for key equipment, collecting equipment operation data in real time; step S2: performing equipment failure prediction analysis; the steps are performed for a general device: s2.1, the invention relates to the technical field of industrial production equipment, the fault prediction analysis method and the model of the industrial production equipment collect data through an equipment end, then the data are transmitted to a fault analysis system, the cause of the fault is analyzed through the fault analysis system, fault early warning information is received through a fault early warning unit, and an alarm is sent to corresponding personnel and equipment, so that the fault can be quickly and accurately found, the fault is conveniently maintained by personnel, and the normal operation of the production equipment is prevented from being influenced by the fault. The patent is to collect fault rate, fault category and fault reason in each period of general equipment to match the fault condition of the tested equipment, belongs to a diagnosis method based on an experience model, and judges the system fault type based on priori knowledge and engineering experience. Researchers have validated the feasibility of this approach, but it is inevitable that failure prediction systems based on a priori knowledge have difficulty in building a knowledge base and are less fault tolerant.
Patent CN106096170A is a multivariable fault prediction method of a wind turbine based on data driving, and is realized by adopting the following steps: step 1: collecting state data of the monitored wind turbine generator components; step 2: adopting a five-point moving average method to perform noise reduction treatment on the characteristic quantity; step 3: calculating the correlation degree R between the characteristic quantity and the residual life prediction; step 4: establishing a multivariable least square support vector machine prediction model; step 5: optimizing regularization parameters gamma and kernel parameters sigma 2 of the multivariable least square support vector machine prediction model; step 6: verifying the validity of a multivariable least square support vector machine prediction model; step 7: and predicting the remaining effective life of the wind turbine generator components. In the data-driven fault prediction method, a Least Squares Support Vector Machine (LSSVM) technology in machine learning is applied to proper signal characteristics to extract information about faults to perform fault prediction, but data according to the operation conditions of equipment has time sequence property, and the LSSVM is suitable for simpler regression and classification problems and does not perform well for non-stable time sequence data. The predictive capability of LSSVM has certain limitations in processing non-stationary time series data, such as periodically varying data.
Patent CN108304941a discloses a machine learning based fault prediction method: 1) Acquiring set operation index data of an object to be predicted, and obtaining time sequence data of each set operation index; collecting historical fault data of the object to be predicted; 2) Respectively extracting features of the data acquired in the step 1), and inputting the extracted features into a machine learning system for training to obtain a basic fault prediction model; 3) And collecting real-time data of set operation indexes when the object to be predicted runs, extracting features of the real-time data, inputting the feature extraction data into the basic fault prediction model, and predicting whether the object to be predicted has a fault currently. The patent points out that an intelligent reasoning and deep neural network algorithm is required to be applied, a prediction basic model is constructed based on a feature subset obtained through screening, and super-parameter optimization processing is carried out on the prediction basic model, training is carried out on the basis of normal data and fault data, the comprehensive requirements of a fault training sample are strict, compared with the normal data, sample data which truly generate faults in an actual production environment are few, the model which is difficult to directly learn the faults is difficult to realize, and the engineering practicability of the method is low.
Disclosure of Invention
Aiming at the problems existing in the prior art and overcoming the defects of the prior art, the invention designs a power equipment fault prediction and health management method based on industrial big data aiming at SCADA operation data of a field equipment group of an actual power system and having the problems of parameter feature redundancy and unbalanced fault data, firstly, working condition data under normal conditions are selected to carry out feature selection and simplification of feature parameters by adopting an XGBoost method, then a MSA-BLSTM (multi-head self-attention-two-way long-short-period memory network) prediction equipment operation parameter model is constructed on samples of a time sequence, a multi-head self-attention layer is introduced to further carry out data feature distribution weight aiming at the current problem of feature parameter redundancy under the actual working condition, and on the contrary, the more influenced features of a prediction result are allocated with higher weight, the less weight is allocated, so that the extraction of key features is realized. In order to avoid error amplification, the extracted characteristics according to the weights are used for calculating nonlinearity of the characteristics through two bidirectional BLSTM network layers, so that information is transmitted from the past to the future, and the information combining the history and the future has important reference value for predicting the state at the current moment, so that a more accurate prediction result can be realized.
A power equipment fault prediction and health management method based on industrial big data comprises the following steps:
s1: integrating the equipment operation data and fault log data to construct a data set of operation parameters containing fault information as initialization data;
s2: processing abnormal values and missing values of the initialized data in the step S1, and realizing normalization;
s3: performing feature selection on the normalized data in the step S2, performing dimension reduction on original data indexes, and selecting a feature group with strong correlation with equipment operation index parameters;
s4: the data of the feature group in the step S3 are proportionally divided into a training set, a testing set and a verification set, wherein the training set and the testing set are working condition parameters of normal operation of the equipment, and the verification set comprises operating parameters of normal operation and operating parameters of abnormal operation;
s5: an operation parameter prediction model of normal operation of the MSA-BLSTM is established, and the operation parameters of the equipment are predicted by using a test set;
s6: performing prediction model verification by using a verification set, taking predicted operation parameters output by the MSA-BLSTM as input, performing fault judgment by using an isolated forest method model to realize fault prediction of equipment, and verifying the accuracy of fault judgment;
s7: and carrying out early warning on equipment faults and health state assessment according to the fault prediction result.
Preferably, the characteristic selection in step S3 adopts the XGBoost method.
The specific steps of the step S3 are as follows:
s3.1: extracting n samples from the initialization data to form a training set and establishing a decision tree;
s3.2: repeating the step S3.1, extracting n samples to form a training set, building a second decision tree, and building a plurality of decision trees in the similar way;
s3.3: searching a minimum objective function obj of each decision tree, wherein the value of the minimum objective function obj forms a weak learner in XGBoost;
s3.4: setting f k (x i ) Representing the predicted value of sample i for the kth decision tree,representing the predictive value of the previous k-1 decision trees on the sample, and obtaining the total predictive value of k trees as follows:
s3.5: carrying out calculation in the objective function obj to obtain a formula of the objective function:
wherein obj is (k) Representing a loss function of k decision trees for measuring the difference between model predictive and true values, y i Representing the actual value of the sample i,representation and->Related objective function, Ω (f k ) Representing the model complexity of the kth decision tree;
s3.6: under the precondition that each decision tree is structurally determined, deriving all leaf node values in the decision tree to calculate an objective function minimum obj min The following formula is shown:
wherein G is j Represents the sum of the first derivatives of the samples contained in the current leaf node, j represents the number of the current leaf nodes, H j The method is characterized in that the method comprises the steps of representing the sum of second derivatives of samples contained in current leaf nodes, T represents the number of the leaf nodes, and lambda and gamma respectively represent leaf node hyper-parameters and value hyper-parameters and are used for controlling the complexity of a tree according to actual conditions;
s3.7: determining an optimal splitting point of each feature aiming at each feature variable, splitting left and right new leaf nodes at the optimal splitting point, and obtaining splitting benefit of the feature;
the importance of each feature depends on the difference of the objective functions before and after splitting of each tree, i.e. the post-splitting benefit value, as follows:
wherein G is L And G R Representing the sum of the first derivatives of the samples in the left and right subtrees, H L And H R Representing the sum of the second derivatives of the samples in the left and right subtrees respectively,
the gain represents the reduction of the loss of the objective function caused when a certain feature is used for splitting a sample, the importance of the feature in the model is evaluated by calculating the gain score of each feature, so that feature selection is performed, if gain is large, the feature is indicated to have a larger contribution to the training of the model, and the corresponding feature has a higher importance.
And a multi-head self-attention layer is introduced into the training set to realize the weight distribution of the data characteristics, so that the extraction of key characteristics is further realized.
The method for realizing the weight distribution of the data characteristics by introducing the training set into the multi-head self-attention layer comprises the following steps of:
firstly, inputting according to training set data: with X= [ X ] 1 ,x 2 ,…,x n ]Represents the input vector, n represents the number of input information, and Q is obtained by linear transformation c ,K c ,V c Initial representation of three vectorsThe following formula is shown:
where c represents the current head number, and for each vector X, three coefficients are multiplied respectivelyObtaining Q c ,K c ,V c Three values, Q c ,K c ,V c Respectively represent a query function, a key point matrix and a value matrix, < ->Respectively representing a query weight matrix, a key point weight matrix and a value weight matrix,
wherein Q is c ＝K c ＝V c And introducing self-Attention mechanism weight matrix Attention (Q) c ,K c ,V c ) Is shown as the following formula:
wherein the method comprises the steps ofThe scaling factor is expressed as a constant.
S5.2: multi-head self-attention randomly initializing multiple groups of Q c ,K c ,V c Each set of matrices representing a subspace into which time series vectors are projected simultaneously, helping the model learn different features:
head c ＝Attention(Q c ,K c ,V c ) (c＝1,...,h)
head c representing the c-th self-attention subspace, which is the c-th head in the multi-head attention mechanism, concact () is a connection operation function in the neural network, and each self-attention matrix is spliced together through Concact () into a splicing matrix which is multiplied by a weight matrix W o And then converted into a compression matrix, which is the input of the next layer of BLSTM neural network.
Further, the feature data extracted according to the weight is input into a two-layer bidirectional BLSTM network, two different hidden layers are utilized to realize two-way memory, one is a forward propagation layer used for transmitting information from the past to the future, the other is a backward propagation layer used for transmitting information from the past to the past, and the calculation formula of each network layer of the BLSTM is as follows:
wherein x is t Input at time t, h t-1 A hidden layer state value at the time t-1 is represented;is the output of the forward propagating layer at time t, < >>Is the output of the counter-propagating layer at time t, tanh represents the tangent hyperbolic function,/o>The input gates represent the forward and reverse weight matrix, respectively,>forward and reverse weight matrix representing output gates, respectively, < >>Respectively representing forward and reverse bias matrices, h t Integrate->And->Output value of W o ,b o The weights and bias terms, y, of the output layer, respectively t Is the output value of the BLSTM network at time t.
Preferably, the abnormal condition is pre-warned by monitoring deviation between the operation parameters and the observed values of the prediction equipment obtained by the MSA-BLSTM prediction model.
Further, when observed value v ture_new And a predicted value v pre_new The error between them satisfies the following equation, suggesting a risk of failure:
wherein r is i For the obtained true value of the device data operation parameter, N is the number of data, and μ is the average value of the device operation parameter calculated according to the data.
Further, taking the residual error value v of the equipment operation parameter re ＝ν pre_new -ν ture_new Comparing with the past T points, if the points are all outside the confidence interval, setting the count to be added with 1, and if the count exceeds half of the set time window count, prompting that the fault risk exists, wherein the formula is as follows:
wherein T represents the number of output predicted values, count represents the number of predicted values exceeding the confidence interval, and I represents the logical operator OR.
Compared with the prior art, the invention has the beneficial effects that:
aiming at the problems existing in the prior art and overcoming the defects of the prior art, the invention designs a power equipment fault prediction and health management method based on industrial big data aiming at SCADA operation data of a field equipment group of an actual power system and having the problems of parameter feature redundancy and unbalanced fault data, firstly, working condition data under normal conditions are selected to carry out feature selection and simplification of feature parameters by adopting an XGBoost method, then a MSA-BLSTM (multi-head self-attention-two-way long-short-period memory network) prediction equipment operation parameter model is constructed on samples of a time sequence, a multi-head self-attention layer is introduced to further carry out data feature distribution weight aiming at the current problem of feature parameter redundancy under the actual working condition, and on the contrary, the more influenced features of a prediction result are allocated with higher weight, the less weight is allocated, so that the extraction of key features is realized. In order to avoid error amplification, the extracted characteristics according to the weights are used for calculating nonlinearity of the characteristics through two bidirectional BLSTM network layers, so that information is transmitted from the past to the future, and the information combining the history and the future has important reference value for predicting the state at the current moment, so that a more accurate prediction result can be realized.
1. Aiming at the problem of redundancy of feature vectors in high-dimensional and large-scale actual data, the XGBoost gradient lifting tree algorithm is adopted to perform feature selection and simplify feature parameters, and the XGBoost algorithm can process high-dimensional data and large-scale data sets more efficiently, so that complexity of a model is effectively reduced. The XGBoost algorithm in the prior art is mainly established as a model of fault level, and is different from the feature selection of the invention, the invention abandons the fault prediction thought based on the prior art and only depends on various types of priori knowledge.
2. According to the invention, a time sequence algorithm of the BLSTM is adopted to establish an operation state model of the equipment under normal working conditions aiming at the situation that the time sequence characteristics of actual monitoring data and sample fault data are seriously unbalanced, so that long-term dependency relations in historical and future time sequence data are effectively captured, and the prediction accuracy is improved. The actual nonlinear data also has certain redundant information after feature selection, a self-attention mechanism is introduced to improve the feature extraction capability which is not possessed in the BLSTM network, and because a plurality of time dimensions exist in a time sequence data sample in training, single self-attention weight distribution results are weighted and summed in a multi-head mode, so that a prediction model can pay attention to time sequence information of different positions, and calculation errors of different features are further reduced.
3. Aiming at the practical problem of fewer fault samples in actual data, the invention predicts an operation parameter model of equipment under normal operation by utilizing working condition characteristic data of normal operation and adopting an MSA-BLSTM algorithm based on a time sequence, and carries out fault judgment on residual analysis of a predicted value and an observed value based on the trained equipment normal operation parameter model, thereby integrally forming an MSA-BLSTM fault prediction model.
4. The invention can be used for describing the health state early warning. When an abnormality occurs in the apparatus, the operating parameters of its output deviate from the normal operating interval, and the deviation is irreversible. In order to identify abnormal conditions, the invention performs fault diagnosis by analyzing residual errors and detecting abnormal values of loop ratios, and the MSA-BLSTM prediction model is used for comparing the deviation between the predicted equipment operation parameters and the actual observed values. If the deviation value is large and a trend is formed in a period of time, the abnormal condition of the equipment is indicated, and the health early warning can be made in advance.
Drawings
FIG. 1 is a flow chart of a power equipment fault prediction and health management method based on industrial big data according to the present invention;
FIG. 2 is a diagram of a predictive network architecture of the MSA-BLSTM-independent forest of the present invention;
fig. 3 is a diagram of the internal structure of the BLSTM network of the present invention.
Detailed Description
The invention relates to a power equipment fault prediction and health management method based on industrial big data, which is further described in detail below with reference to the accompanying drawings and the specific implementation method.
The invention is based on SCADA operation data of the equipment group of the actual power production field, firstly adopts the XGBoost method to perform feature selection and simplify feature parameters to further train a prediction model aiming at the problem of parameter feature redundancy, then builds an LSTM equipment normal operation parameter model based on a time sequence aiming at the problem of fewer fault samples of the actual data, and compared with the machine learning algorithm adopted by the patent, the LSTM algorithm has advantages in the aspects of processing nonlinear and non-stable time sequence data, adapting to larger-scale data, processing multidimensional data, performing end-to-end learning, predicting accuracy and the like. And then carrying out abnormal detection on fault data under abnormal working conditions by using the trained normal operation parameter model of the equipment, and judging whether the equipment has faults or not, so that a prediction model is integrally formed to analyze and predict the faults. Finally, fault early warning and state evaluation are carried out on the equipment, and equipment fault problems are solved and maintained early, so that normal operation of the equipment is guaranteed, the service life of the equipment is prolonged, the performance and reliability of the equipment are improved, and the equipment contributes to production and economic benefits of enterprises.
Specifically, as shown in fig. 1, the power equipment fault prediction and health management method based on industrial big data provided by the invention is implemented according to the following steps:
s1: integrating the equipment operation data and the fault log data to construct a data set of operation parameters containing fault information;
s2: the original real data has the real problems of data loss and abrupt change of parameters of various working conditions caused by variable working conditions, the original data is cleaned, abnormal values and missing values are processed, and normalization is realized;
s3: the initialization data input in the S2 is subjected to feature selection by adopting an XGBoost method, the original data index is subjected to dimension reduction, a feature group with stronger correlation with the equipment operation index parameter is selected, and the redundancy of the data features is eliminated, and the method specifically comprises the following steps:
s3.1: extracting n samples from the initialization data to form a training set and establishing a decision tree;
s3.2: repeating the step S3.1, extracting n samples to form a training set, building a second decision tree, and building a plurality of decision trees in the similar way;
s3.3: searching a minimum objective function obj of each decision tree, wherein the value of the minimum objective function obj forms a weak learner in XGBoost;
s3.4: setting f k (x i ) Representing the predicted value of sample i for the kth decision tree,representing the predictive value of the previous k-1 decision trees on the samples, then the total predictive value of the k trees would be:
s3.5: carrying out calculation in the objective function obj to obtain a formula of the objective function:
wherein obj is (k) Representing the loss function of k decision trees, y i Representing the actual value of the sample i,representation and representationRelated objective function, Ω (f k ) Representing the model complexity of the kth decision tree.
S3.6: under the precondition that each decision tree is structurally determined, deriving all leaf node values in the decision tree to calculate an objective function minimum obj min The following formula is shown:
wherein G is j Representing the sum of first derivatives of samples contained in the current leaf node, j represents the number of the current leaf node, hj represents the sum of second derivatives of the samples contained in the current leaf node, T represents the number of the leaf node, and lambda and gamma respectively represent leaf node hyper-parameters and value hyper-parameters for controlling the complexity of the tree according to actual conditions;
s3.7: for each characteristic variable, determining the optimal splitting point of the characteristic through linear scanning, splitting left and right new leaf nodes on the optimal splitting point, and obtaining splitting benefit of the characteristic;
the importance of each feature depends on the difference of the objective functions before and after splitting of each tree, i.e. the post-splitting benefit value, as follows:
wherein G is L And G R Representing the sum of the first derivatives of the samples in the left and right subtrees, H L And H R Representing the sum of the second derivatives of the samples in the left and right subtrees respectively,
the gain represents the reduction of the loss of the objective function caused when a certain feature is used for splitting a sample, the importance of the feature in the model is evaluated by calculating the gain score of each feature, so that feature selection is performed, if gain is large, the feature is indicated to have a larger contribution to the training of the model, and the corresponding feature has a higher importance.
S4: the data selected in the step S3 are proportionally divided into a training set, a testing set and a verification set, wherein the training set and the testing set are working condition parameters of normal operation of equipment, and the verification set comprises normal operation parameters and abnormal operation parameters;
s5, introducing the training data set into a multi-head self-attention layer to realize weight distribution to the data features, and further realizing extraction of key features:
s5.1: firstly, inputting according to training set data: with X= [ X ] 1 ,x 2 ,…,x n ]Representing n input information, obtaining Q by linear transformation c ,K c ,V c The initial representation of the three vectors is shown as follows:
where c represents the current head number, and for each vector X, three coefficients are multiplied respectivelyObtaining Q c ,K c ,V c Three values, Q c ,K c ,V c Respectively represent a query function, a key point matrix and a value matrix, < ->Respectively representing a query weight matrix, a key point weight matrix and a value weight matrix,
where q=k=v, and introduces the Attention mechanism weight matrix Attention (Q c ,K c ,V c ) Is shown as the following formula:
wherein the method comprises the steps ofThe scaling factor is expressed as a constant.
S5.2: the essence of multi-head self-attention is that multiple groups of randomly initialized weight matrixes of Q, K and V, each group of matrixes represents a subspace, and time sequence vectors are projected into the subspaces at the same time, so that models can be helped to learn different characteristics:
head c ＝Attention(Q c ,K c ,V c ) (c＝1,...,h)
head c representing the c-th self-attention subspace, which is the c-th head in the multi-head attention mechanism, concact () is a connection operation function in the neural network, and each self-attention matrix is spliced together through Concact () into a splicing matrix which is multiplied by a weight matrix W o And then converted into a compression matrix, which is the input of the next layer of BLSTM neural network.
S6, inputting the feature data extracted according to the weights into a two-layer bidirectional BLSTM network, and utilizing two different hidden layers to realize two-way memory, wherein one layer is a forward propagation layer used for transmitting information from the past to the future, and the other layer is a backward propagation layer used for transmitting information from the future to the past. The flow chart is shown in fig. 2, and the calculation formula of each network layer of the BLSTM is as follows:
y t ＝W o ·h t +b o
wherein x is t Input at time t, h t-1 A hidden layer state value at the time t-1 is represented;is the output of the forward propagating layer at time t, < >>Is the output of the counter-propagating layer at time t, tanh represents the tangent hyperbolic function,/o>The input gates represent the forward and reverse weight matrix, respectively,>forward and reverse weight matrix representing output gates, respectively, < >>Respectively representing forward and reverse bias matrices, h t Integrate->And->Output value of W o ,b o The weights and bias terms, y, of the output layer, respectively t Is the output value of the BLSTM network at time t.
And obtaining a prediction model of the operation parameters of the power equipment according to the MSA-BLSTM prediction model.
S7, when abnormal conditions occur in the equipment, irreversible conditions of the output operation parameters deviate from the normal working interval, and aiming at the phenomenon, the operation parameters (v) of the prediction equipment obtained through the MSA-BLSTM prediction model are obtained pre_new ) And observed value (v) ture_new ) There will be some deviation value, but the deviation value is larger and the trend in a period of time will indicate that there is an abnormal situation.
S7.1: the residual is analyzed:
firstly, predicting data of a test set and observing an observed value v ture_new And a predicted value v pre_new The error between the two is subjected to statistical confidence test, the probability in the (mu-3 sigma, mu+3 sigma) interval is 99.74 percent according to the 3-sigma principle, and the normal operation parameter interval is judged as shown in the following formula, so that the error data distribution interval is considered to be judged to be abnormal when exceeding the interval;
wherein r is i For acquired device dataThe true value of the operating parameter, N, is the number of data and μ is the mean value of the operating parameter of the device calculated from the data.
S7.2: ring ratio anomaly detection:
aiming at the irreversibility of equipment failure, the phenomenon that the abnormal value continuously or continuously increases in a period of time exists before the equipment fails, the operation parameters in a time window (T) follow a certain trend before the equipment fails, and then we take the residual value v of the operation parameters of the equipment re ＝ν pre_new -ν ture_new Comparing with the past T points, wherein T represents the number of output predicted values, if the output predicted values are all outside the confidence interval, count plus 1 is set, the count represents the number of the predicted values exceeding the confidence interval, and if the count exceeds half of the set time window count, the equipment can be determined to be in fault, and the formula is as follows:
the expression is either a logical operator or.
S8: and (3) performing fault early warning according to S7, forming a prediction network structure diagram of the MSA-BLSTM, performing data processing and prediction related fault analysis report visual display, automatically generating a key monitoring maintenance work order, and performing health management on the power equipment.
The method starts from high-dimensional and large-scale actual monitoring data, firstly, the XGBoost gradient lifting tree algorithm is adopted to conduct feature selection and simplify feature parameters, and the XGBoost algorithm can be used for processing high-dimensional data and large-scale data sets more efficiently, so that complexity of a model is effectively reduced. And then, aiming at the situation that the time sequence characteristics of actual monitoring data and sample fault data are seriously unbalanced, a time sequence algorithm of the BLSTM is adopted to establish an operation state model of the equipment under the normal working condition, so that long-term dependency relations in historical and future time sequence data are effectively captured, and the prediction accuracy is improved. The actual nonlinear data also has certain redundant information after feature selection, a self-attention mechanism is introduced to improve the feature extraction capability which is not possessed in the BLSTM network, and because a plurality of time dimensions exist in a time sequence data sample in training, single self-attention weight distribution results are weighted and summed in a multi-head mode, so that a prediction model can pay attention to time sequence information of different positions, and calculation errors of different features are further reduced. And then taking the output of the normal running state prediction model as input, training the orphan forest model to detect abnormal or fault data, tightly fitting the complex characteristics of the actual monitoring data in the whole algorithm prediction process, and innovating the prediction algorithm. And finally, according to a deep learning algorithm, relevant fault analysis reports are arranged in the steps of data processing and predicting, so that fault early warning and state evaluation are carried out on the equipment.
The present invention provides a power equipment fault prediction and health management method based on industrial big data, and the above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement the same, and not to limit the protection scope of the present invention. All equivalent changes or modifications made in accordance with the spirit of the present invention should be construed to be included in the scope of the present invention.
Claims (9)
1. The power equipment fault prediction and health management method based on the industrial big data is characterized by comprising the following steps of:
s1: integrating the equipment operation data and fault log data to construct a data set of operation parameters containing fault information as initialization data;
s2: processing abnormal values and missing values of the initialized data in the step S1, and realizing normalization;
s3: performing feature selection on the normalized data in the step S2, performing dimension reduction on original data indexes, and selecting a feature group with strong correlation with equipment operation index parameters;
s4: the data of the feature group in the step S3 are proportionally divided into a training set, a testing set and a verification set, wherein the training set and the testing set are working condition parameters of normal operation of the equipment, and the verification set comprises operating parameters of normal operation and operating parameters of abnormal operation;
s5: an operation parameter prediction model of normal operation of the MSA-BLSTM is established, and the operation parameters of the equipment are predicted by using a test set;
s6: performing prediction model verification by using a verification set, taking predicted operation parameters output by the MSA-BLSTM as input, performing fault judgment by using an isolated forest method model to realize fault prediction of equipment, and verifying the accuracy of fault judgment;
s7: and carrying out early warning on equipment faults and health state assessment according to the fault prediction result.
2. The industrial big data-based power equipment fault prediction and health management method according to claim 1, wherein the method comprises the following steps:
and step S3, selecting the characteristic by adopting an XGBoost method.
3. The industrial big data-based power equipment fault prediction and health management method according to claim 2, wherein the method comprises the following steps:
the step S3 comprises the following specific steps:
s3.1: extracting n samples from the initialization data to form a training set and establishing a decision tree;
s3.2: repeating the step S3.1, extracting n samples to form a training set, building a second decision tree, and building a plurality of decision trees in the similar way;
s3.3: searching a minimum objective function obj of each decision tree, wherein the value of the minimum objective function obj forms a weak learner in XGBoost;
s3.4: setting f k (x i ) Representing the predicted value of sample i for the kth decision tree,representing the predictive value of the previous k-1 decision trees on the sample, and obtaining the total predictive value of k trees as follows:
s3.5: carrying out calculation in the objective function obj to obtain a formula of the objective function:
wherein obj is (k) Representing a loss function of k decision trees for measuring the difference between model predictive and true values, y i Representing the actual value of the sample i,representation and->Related objective function, Ω (f k ) Representing the model complexity of the kth decision tree;
s3.6: under the precondition that each decision tree is structurally determined, deriving all leaf node values in the decision tree to calculate an objective function minimum obj min The following formula is shown:
wherein G is j Represents the sum of the first derivatives of the samples contained in the current leaf node, j represents the number of the current leaf nodes, H j The method is characterized in that the method comprises the steps of representing the sum of second derivatives of samples contained in current leaf nodes, T represents the number of the leaf nodes, and lambda and gamma respectively represent leaf node hyper-parameters and value hyper-parameters and are used for controlling the complexity of a tree according to actual conditions;
s3.7: determining an optimal splitting point of each feature aiming at each feature variable, splitting left and right new leaf nodes at the optimal splitting point, and obtaining splitting benefit of the feature;
the importance of each feature depends on the difference of the objective functions before and after splitting of each tree, i.e. the post-splitting benefit value, as follows:
wherein G is L And G R Representing the sum of the first derivatives of the samples in the left and right subtrees, H L And H R Representing the sum of the second derivatives of the samples in the left and right subtrees respectively,
the gain represents the reduction of the loss of the objective function caused when a certain feature is used for splitting a sample, the importance of the feature in the model is evaluated by calculating the gain score of each feature, so that feature selection is performed, if gain is large, the feature is indicated to have a larger contribution to the training of the model, and the corresponding feature has a higher importance.
4. The industrial big data-based power equipment fault prediction and health management method according to claim 1, wherein the method comprises the following steps:
and a multi-head self-attention layer is introduced into the training set to realize the weight distribution of the data characteristics, so that the extraction of key characteristics is further realized.
5. The industrial big data-based power equipment fault prediction and health management method according to claim 4, wherein the method comprises the following steps:
the method for realizing the weight distribution of the data characteristics by introducing the training set into the multi-head self-attention layer comprises the following steps of:
firstly, inputting according to training set data: with X= [ X ] 1 ,x 2 ,…,x n ]Represents the input vector, n represents the number of input information, and Q is obtained by linear transformation c ,K c ,V c The initial representation of the three vectors is shown as follows:
wherein c represents the current number of heads,for each vector X, three coefficients are multiplied respectivelyObtaining Q c ,K c ,V c Three values, Q c ,K c ,V c Respectively represent a query function, a key point matrix and a value matrix, < ->Respectively representing a query weight matrix, a key point weight matrix and a value weight matrix,
wherein Q is c ＝K c ＝V c And introducing self-Attention mechanism weight matrix Attention (Q) c ,K c ,V c ) Is shown as the following formula:
wherein the method comprises the steps ofThe scaling factor is expressed as a constant.
S5.2: multi-head self-attention randomly initializing multiple groups of Q c ,K c ,V c Each set of matrices representing a subspace into which time series vectors are projected simultaneously, helping the model learn different features:
head c ＝Attention(Q c ,K c ,V c )(c＝1,...,h)
head c representing the c-th self-attention subspace, which is the c-th head in the multi-head attention mechanism, concact () is a connected operation function in the neural network, and each self-attention matrix is spliced at the point of the neural network through Concact ()Together into a mosaic matrix, which is multiplied by a weight matrix W o And then converted into a compression matrix, which is the input of the next layer of BLSTM neural network.
6. The industrial big data-based power equipment fault prediction and health management method according to claim 5, wherein the method comprises the following steps:
the feature data extracted according to the weight is input into a two-way BLSTM network of two layers, two different hidden layers are utilized to realize two-way memory, one is a forward propagation layer used for transmitting information from the past to the future, the other is a backward propagation layer used for transmitting information from the past to the past, and the calculation formula of each network layer of the BLSTM is as follows:
y t ＝W o ·h t +b o
wherein x is t Input at time t, h t-1 A hidden layer state value at the time t-1 is represented;is the output of the forward propagating layer at time t, < >>Is the output of the counter-propagating layer at time t, tanh represents the tangent hyperbolic function,/o>The input gates represent the forward and reverse weight matrix, respectively,>forward and reverse weight matrix representing output gates, respectively, < >>Respectively representing forward and reverse bias matrices, h t Integrate->And->Output value of W o ,b o The weights and bias terms, y, of the output layer, respectively t Is the output value of the BLSTM network at time t.
7. The industrial big data-based power equipment fault prediction and health management method according to claim 1, wherein the method comprises the following steps:
and carrying out early warning on abnormal conditions by monitoring deviation between the operation parameters and the observed values of the prediction equipment obtained by the MSA-BLSTM prediction model.
8. The industrial big data-based power equipment fault prediction and health management method according to claim 7, wherein the method comprises the following steps:
when observed value v ture_new And a predicted value v pre_new The error between them satisfies the following equation, suggesting a risk of failure:
wherein r is i For the obtained true value of the device data operation parameter, N is the number of data, and μ is the average value of the device operation parameter calculated according to the data.
9. The industrial big data-based power equipment fault prediction and health management method according to claim 8, wherein the method comprises the following steps:
taking the residual error value v of the equipment operation parameter re ＝ν pre_new -ν ture_new Comparing with the past T points, if the points are all outside the confidence interval, setting the count to be added with 1, and if the count exceeds half of the set time window count, prompting that the fault risk exists, wherein the formula is as follows:
wherein T represents the number of output predicted values, count represents the number of predicted values exceeding the confidence interval, and I represents the logical operator OR.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311546874.2A CN117556347A (en) | 2023-11-20 | 2023-11-20 | Power equipment fault prediction and health management method based on industrial big data |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311546874.2A CN117556347A (en) | 2023-11-20 | 2023-11-20 | Power equipment fault prediction and health management method based on industrial big data |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117556347A true CN117556347A (en) | 2024-02-13 |
Family
ID=89814310
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311546874.2A Pending CN117556347A (en) | 2023-11-20 | 2023-11-20 | Power equipment fault prediction and health management method based on industrial big data |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117556347A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117972533A (en) * | 2024-03-29 | 2024-05-03 | 北京易智时代数字科技有限公司 | Data processing method, device and equipment for industrial equipment |
-
2023
- 2023-11-20 CN CN202311546874.2A patent/CN117556347A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117972533A (en) * | 2024-03-29 | 2024-05-03 | 北京易智时代数字科技有限公司 | Data processing method, device and equipment for industrial equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Smarsly et al. | Machine learning techniques for structural health monitoring | |
| CN111222549B (en) | Unmanned aerial vehicle fault prediction method based on deep neural network | |
| CN116757534B (en) | Intelligent refrigerator reliability analysis method based on neural training network | |
| CN109146246B (en) | Fault detection method based on automatic encoder and Bayesian network | |
| CN111914873A (en) | Two-stage cloud server unsupervised anomaly prediction method | |
| CN113642754B (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
| CN112785091A (en) | Method for performing fault prediction and health management on oil field electric submersible pump | |
| CN108921230A (en) | Method for diagnosing faults based on class mean value core pivot element analysis and BP neural network | |
| CN117556347A (en) | Power equipment fault prediction and health management method based on industrial big data | |
| CN114021932A (en) | Energy efficiency evaluation and diagnosis method, system and medium for wind turbine generator | |
| CN114266278B (en) | Dual-attention network-based equipment residual service life prediction method | |
| CN111079348B (en) | Method and device for detecting slowly-varying signal | |
| CN112257914B (en) | Aviation safety causal prediction method based on random forest | |
| CN115329986A (en) | Wind turbine generator anomaly detection and positioning method based on interpretable graph neural network | |
| CN111191855B (en) | Water quality abnormal event identification and early warning method based on pipe network multi-element water quality time sequence data | |
| CN111122811A (en) | Sewage treatment process fault monitoring method of OICA and RNN fusion model | |
| CN116109039A (en) | Data-driven anomaly detection and early warning system | |
| Smarsly et al. | A decentralized approach towards autonomous fault detection in wireless structural health monitoring systems | |
| CN116595319A (en) | Prediction method and system applied to rail transit motor health state evaluation | |
| CN114611372A (en) | Industrial equipment health prediction method based on Internet of things edge calculation | |
| CN116205265A (en) | Power grid fault diagnosis method and device based on deep neural network | |
| Tang et al. | Prediction of bearing performance degradation with bottleneck feature based on LSTM network | |
| Mao et al. | Comparative study on prediction of fuel cell performance using machine learning approaches | |
| Samhouri et al. | An intelligent machine condition monitoring system using time-based analysis: neuro-fuzzy versus neural network | |
| CN117232809A (en) | Fan main shaft fault pre-diagnosis method based on DEMATEL-ANP-CRITIC combined weighting |
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 |