CN113434970A - Health index curve extraction and service life prediction method for mechanical equipment - Google Patents

Health index curve extraction and service life prediction method for mechanical equipment Download PDF

Info

Publication number
CN113434970A
CN113434970A CN202110608119.7A CN202110608119A CN113434970A CN 113434970 A CN113434970 A CN 113434970A CN 202110608119 A CN202110608119 A CN 202110608119A CN 113434970 A CN113434970 A CN 113434970A
Authority
CN
China
Prior art keywords
data
health index
decoder
mechanical equipment
health
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.)
Granted
Application number
CN202110608119.7A
Other languages
Chinese (zh)
Other versions
CN113434970B (en
Inventor
李红辉
段宇航
张宁
白岩慧
张春
刘峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Jiaotong University
Original Assignee
Beijing Jiaotong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Jiaotong University filed Critical Beijing Jiaotong University
Priority to CN202110608119.7A priority Critical patent/CN113434970B/en
Publication of CN113434970A publication Critical patent/CN113434970A/en
Application granted granted Critical
Publication of CN113434970B publication Critical patent/CN113434970B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/17—Mechanical parametric or variational design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/04—Ageing analysis or optimisation against ageing

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Geometry (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Computer Hardware Design (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Complex Calculations (AREA)

Abstract

The invention provides a method for extracting a health index curve and predicting the service life of mechanical equipment. The method comprises the following steps: acquiring mechanical sensing degradation data from a healthy state to a fault state of a training example of mechanical equipment to obtain a data block sequence; extracting a health index curve reflecting the mechanical health state of the training example from the data block sequence by using a Transformer encoder-decoder; establishing a mapping relation between a health index curve of the mechanical equipment and the sensing reading by using a linear regression model; and inputting the health index curve of the test example of the mechanical equipment to be evaluated into the mapping relation to obtain the health index curve of the test example of the mechanical equipment to be evaluated, and performing similarity calculation on the health index curve of the test example of the mechanical equipment to be evaluated and the health index curve library to obtain the estimation result of the residual life of the mechanical equipment to be evaluated. According to the method, the health index curve reflecting the mechanical health state is extracted from the degradation time sequence, so that the residual service life of the mechanical equipment is effectively predicted.

Description

Health index curve extraction and service life prediction method for mechanical equipment
Technical Field
The invention relates to the technical field of health management of mechanical equipment, in particular to a method for extracting a health index curve and predicting the service life of the mechanical equipment.
Background
The purpose of PHM (fault prediction and Health Management) is to improve the safety and stability of complex systems and reduce maintenance costs. RUL (Remaining Useful Life) prediction, one of the important tasks of the PHM, can evaluate the health of a degraded system, dynamically maintaining mechanical equipment to prevent performance degradation and sudden failure. The mechanical life prediction generally comprises four technical processes, namely data acquisition, Health index (Health indicators) construction, Health stage (Health stage) division and residual life prediction.
In 2016, Malhotra et al proposed an unsupervised sequence-to-sequence similarity matching residual life prediction model. Wherein, an LSTM (Long Short-Term Memory network) based encoder converts multidimensional sensing time sequence data into a low-dimensional embedded vector representation. The change in the embedded vector from the initial to the end of the run is then computed by the computing system to construct a health indicator curve. Gugulothu et al replaced the LSTM cells with the original RNN structure and achieved the extraction of the health index curves on the engine dataset. The method is characterized in that a two-way LSTM is used for improvement on the basis of Gugulothu research by the Yunlian et al, the feature extraction capability of a one-way LSTM encoder is improved, the health state information of a low-dimensional embedded vector is enriched, and a more accurate health index curve is extracted.
In a conventional health index construction method based on an RNN (Recurrent Neural Network) encoder-decoder, prediction performance is limited by time feature extraction capabilities of different RNN structures. In general, the LSTM structure is superior to the GRU (please provide full english and chinese) structure and the RNN structure. The predicted performance of the bi-directional RNN structure is superior to that of the uni-directional RNN structure. In addition, each time step of the multidimensional sensing time sequence data has the same contribution to the structure of the health index, and the influence of the key time step information on the structure of the health index is not highlighted.
Disclosure of Invention
The embodiment of the invention provides a method for extracting a health index curve and predicting the service life of mechanical equipment, so as to effectively predict the residual service life of the mechanical equipment.
In order to achieve the purpose, the invention adopts the following technical scheme.
A health index curve extraction and life prediction method for mechanical equipment comprises the following steps:
collecting mechanical sensing degradation data from a healthy state to a fault state of a training example of mechanical equipment, carrying out data preprocessing to obtain time sequence data, and dividing the time sequence data by using a sliding time window to obtain a data block sequence;
constructing a Transformer decoder-encoder comprising a multi-head attention layer, a residual connecting layer and a feedforward neural network layer, and extracting a health index curve reflecting the mechanical health state of a training example from the data block sequence by using the Transformer decoder-encoder;
constructing a health index curve library by the health index curves of all the training examples, and establishing a mapping relation between the health index curve of the mechanical equipment and the sensing reading of a transform encoder-decoder by using a linear regression model;
and inputting the sensing reading of the test example of the mechanical equipment to be evaluated into the mapping relation to obtain a health index curve of the test example of the mechanical equipment to be evaluated, and performing similarity calculation on the health index curve of the test example of the mechanical equipment to be evaluated and the health index curve library to obtain an estimation result of the residual life of the mechanical equipment to be evaluated.
Preferably, the acquiring mechanical sensing degradation data of mechanical equipment from a healthy state to a fault state and performing data preprocessing to obtain time series data, and dividing the time series data by using a sliding time window to obtain a data block sequence includes:
collecting time series data X of mechanical health state from healthy operation to fault of training example of mechanical equipment1,x2,...,xtAnd after the time series data X is subjected to noise reduction and smooth data preprocessing operation, dividing the time series data X by using a sliding time window w with a fixed length, moving the time window with the fixed length by one time step each time to obtain a time window data block sequence, and obtaining a data block sequence omega with the same time window length as { omega ═ omega }1,Ω2,...,ΩtThe data shape of each data block sample is [ N, T, F ]]Where N represents the number of sample time window sequences, T represents the length of the time window, and F represents the dimension of the monitoring sensor.
Preferably, the method for constructing a transform decoder-encoder comprising a multi-head attention layer, a residual connecting layer and a feedforward neural network layer comprises the following steps:
constructing a Transformer decoder-encoder, which consists of a multi-head attention layer, a residual error connection layer and a feedforward neural network layer;
the multi-head attention layer highlights key time step information in the time sequence data by using the received time sequence data and adopting a scaling-dot product attention calculation mode;
the residual error connecting layer splices the shallow features obtained by the shallow neural network and the deep features obtained by the deep neural network through cross connection to supplement information in the middle of the network layer;
and the feedforward neural network performs dimension transformation on the characteristics of the network middle layer according to the vector calculation requirement. If the number of the transform coding units is not 1, the feedforward neural network needs to perform dimension transformation of depth features according to the input requirement of the next layer of coding units.
Preferably, the extracting, by using the transform encoder-decoder structure, a health indicator curve reflecting a machine health status of a training instance from the data block sequence includes:
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtThe input is input to a coder of a transform coder, which is set to { Ω } in a sequence of data blocks1,Ω2,...,ΩtAdding position coding information into each data block in the sequence to obtain a data sequence XposData sequence XposThe depth feature is extracted by an encoder input into a transform encoder, firstly, a data sequence X is extracted through a multi-head attention layerposThe key time step information in the time step information is linked and supplemented with the characteristic information between network layers through residual errors, finally the characteristic output of coding units is obtained through a batch normalization layer, if the number of the coding units exceeds one, the output of the previous coding unit is used as the input of the next coding unit, and the last coding unit outputs a depth characteristic vector reflecting the current time step health state information;
inputting the depth eigenvector into a decoder of a transform coder, wherein the decoder multiplies the depth eigenvector by a unit lower triangular matrix to be used as input data X of a multi-head attention layer, the decoder outputs a reconstructed time sequence Y after the input data X is subjected to a multi-head attention calculation process, the reconstructed time sequence Y and the input data X are subjected to error loss calculation, the network parameter value of the transform coder-decoder is adjusted by using a back propagation algorithm according to the error loss calculation result, and the transform coder-decoder with the adjusted network parameter value is used as a trained transform coder-decoder;
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtAnd inputting the trained encoder in a transform encoder-decoder to obtain a depth feature vector sequence reflecting the time step health state information of the mechanical equipment, and carrying out normalization calculation on the depth feature vector sequence to obtain a health index curve in a 0-1 data range of the training example of the mechanical equipment.
Preferably, the extracting, by using the transform encoder-decoder structure, a health indicator curve of a machine health status of a training instance from the data block sequence includes:
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtThe input is input to a coder of a transform coder, which is set to { Ω } in a sequence of data blocks1,Ω2,...,ΩtAdding position coding information into each data block in the data block, wherein the formula is as follows:
Figure BDA0003094854200000041
pos in the formula represents the bit number of the data in the time sequence, i represents the ith sensing channel, and F represents the dimension number of the monitoring sensor;
obtaining a data sequence X after adding the position information in the data blockposThe encoder pairs the data sequence XposPerforming a Multi-headed attention calculation, data series XposIs transformed into [ N, T, M ] through linear layer]Wherein M is the number of hidden units of the linear layer, for the data sequence XposIs divided into three dimensions, and is disassembled and formed into [ N, H, T, H _ dim [ ]]The multi-head input vector (v) includes a Query vector (Q), a Key vector (K), and value (v), and the calculation formula is as follows:
Figure BDA0003094854200000051
wherein, KTRepresents the vector transposition with the shape [ N, H, H _ dim, T]
Inputting the multi-head input vector as a shallow feature into a residual connecting layer, wherein the calculation formula of the residual connecting layer is as follows:
Redisual connection＝Fshallow network+FDeep network
FShallow networkRepresenting the shallow feature representation extracted by the shallow network model, FDeep networkRepresenting the depth feature representation extracted by the deep network model, and combining the depth feature representation and the depth feature representation to be input into a deep network for feature extraction operation;
then, obtaining the output of the coding unit through residual connection and normalization layer again, inputting the output to the next unit, and outputting the depth characteristic vector reflecting the current time step health status information by the last coding unit;
inputting the depth eigenvector into a decoder of a transform coder, wherein the decoder multiplies the depth eigenvector by a unit lower triangular matrix to be used as input data X of a multi-head attention layer, the decoder outputs a reconstruction time sequence Y after the input data X is subjected to a multi-head attention calculation process, and the reconstruction time sequence Y and the input data X are subjected to one-norm loss function calculation:
ei＝Xi-Yi,i∈{1,2,...,T}
Figure BDA0003094854200000052
adjusting the network parameters of a Transformer encoder-decoder by using a back propagation algorithm according to the calculation result of the first normal form loss function, and taking the Transformer encoder-decoder with the adjusted network parameter values as a trained Transformer encoder-decoder;
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtEach data block omega intInput to a trained Transformer encoder-decoder that outputs an embedded vector sequence reflecting machine health status, Z ═ Z { (Z })1,z2,...,ztAnd f, setting an embedded vector Z obtained at an initial time stepnorm＝{z1,z2,z3The method can represent the complete health state of the mechanical system, and calculate the change value of the embedded vector sequence of the running time step and the embedded vector of the health state, and the calculation formula is as follows:
Figure BDA0003094854200000061
Figure BDA0003094854200000062
where N represents the number of embedded vectors in a fully healthy state, dtIndicating the degree of departure of the system from health at time t, dmax，dminMaximum and minimum values, h, respectively, representing the deviation of the system from the healthy statetA health index curve representing a 0-1 data range of a normalized training example of a machine.
Preferably, the step of constructing a health index curve library by using the health index curves of all the training examples, and establishing a mapping relationship between the health index curve of the mechanical equipment and the sensing reading of the transducer encoder-decoder by using a linear regression model includes:
acquiring a plurality of time series data of machine health states from healthy operation to failure of a plurality of training examples of mechanical equipment, acquiring a health index curve corresponding to each time series data by using a Transformer encoder-decoder, and forming the health index curves of all the training examples into a health index curve library;
the mapping between the health indicator curve and the sensory readings of the transform encoder-decoder is expressed using a linear regression model, which has the following functional form:
htr＝θ0+θxtr
wherein h istrIndicates the value of the health index of the training unit, xtrA piece of multi-dimensional sensory data, theta, representing a training unit0A sensing independent variable coefficient and a bias coefficient value respectively representing a linear regression function;
and training the linear regression model by using a least square method, and obtaining a mapping relation between a health index curve and the sensing reading of a Transformer encoder-decoder after training.
Preferably, the inputting the sensing reading of the test case of the mechanical device to be evaluated into the mapping relationship to obtain a health index curve of the test case of the mechanical device to be evaluated, and performing similarity calculation on the health index curve of the test case of the mechanical device to be evaluated and the health index curve library to obtain an estimation result of the remaining life of the mechanical device to be evaluated includes:
inputting the sensing reading of the Transformer encoder-decoder of the test example of the mechanical equipment to be evaluated into the mapping relation between the health index curve and the sensing reading of the Transformer encoder-decoder to obtain the health index curve of the test example of the mechanical equipment to be evaluated;
calculating the similarity between the health index curve of the test example of the mechanical equipment to be evaluated and the health index curve of each training example in the health index curve library by using a similarity measurement formula, wherein the similarity calculation formula is as follows:
Figure BDA0003094854200000071
Figure BDA0003094854200000072
wherein HI' represents a health index curve of a test case of a mechanical device to be evaluated, HI(j)Representing a health index curve of a jth training example in the health index curve library, wherein tau represents time shift of a test example of the mechanical equipment to be evaluated, lambda is a relaxation factor used for adjusting the size of a similarity metric value, and d (·) represents an Euclidean distance between the two curves;
forming similarity measure vector Sim ═ Sim of similarities between the health index curves of all training examples in the health index curve library and the health index curves of the test examples1,Sim2,...,Simn]；
The remaining life estimate for the mechanical device under evaluation at time shift τ for each test example was calculated using the following equation:
RUL(j,τ)＝Tj-T'-τ
wherein, TjThe T' respectively represents the running time of the jth training example and the running time of the testing example in the health index curve library;
forming residual life estimation values of the mechanical equipment to be evaluated, which are calculated according to the health index curves of all the training examples in the health index curve library, into a residual life estimation vector RUL ═ RUL1,RUL2,...,RULn]；
According to the similarity metric vector Sim ═ Sim1,Sim2,...,Simn]And said residual lifetime estimation vector RUL ═ RUL1,RUL2,...,RULn]And predicting the residual service life of the test example of the mechanical equipment to be evaluated by using a weighted summation mode, wherein the formula is as follows:
Figure BDA0003094854200000081
wherein RUL ^ represents the estimated value of the residual service life of the mechanical equipment,
Figure BDA0003094854200000082
representing the maximum degree of similarity, β controls the number of training instances that participate in the final life estimation.
According to the technical scheme provided by the embodiment of the invention, the method extracts the scheme of the health index sequence for reflecting the mechanical health state from the complete degradation time sequence, and further effectively predicts the residual service life of the mechanical equipment. The method is suitable for any collected mechanical sensing degradation data from a healthy state to a fault state completely.
Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the description of the embodiments are briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings based on these drawings without creative efforts.
FIG. 1 is a block diagram of a transform encoder-decoder according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of an implementation of a method for extracting a health index curve and predicting a life of a mechanical device according to an embodiment of the present invention;
FIG. 3 is a flowchart illustrating a method for extracting a health index curve and predicting a life of a mechanical device according to an embodiment of the present invention;
fig. 4 is a schematic diagram of a sliding window data processing method according to an embodiment of the present invention;
FIG. 5 is a block diagram of a multi-headed attention layer in a transform encoder-decoder according to an embodiment of the present invention;
FIG. 6 is a diagram illustrating a result of a health indicator curve according to an embodiment of the present invention;
fig. 7 is a method for estimating remaining life of similarity matching according to an embodiment of the present invention.
Detailed Description
Reference will now be made in detail to embodiments of the present invention, examples of which are illustrated in the accompanying drawings, wherein like reference numerals refer to the same or similar elements or elements having the same or similar function throughout. The embodiments described below with reference to the accompanying drawings are illustrative only for the purpose of explaining the present invention, and are not to be construed as limiting the present invention.
As used herein, the singular forms "a", "an", "the" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and/or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof. It will be understood that when an element is referred to as being "connected" or "coupled" to another element, it can be directly connected or coupled to the other element or intervening elements may also be present. Further, "connected" or "coupled" as used herein may include wirelessly connected or coupled. As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.
It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
For the convenience of understanding the embodiments of the present invention, the following description will be further explained by taking several specific embodiments as examples in conjunction with the drawings, and the embodiments are not to be construed as limiting the embodiments of the present invention.
The embodiment of the invention utilizes a Transformer encoder-decoder structure in an unsupervised mode to extract a scheme of a health index curve for reflecting the health state of a machine from a complete degradation time sequence. The method is suitable for any collected mechanical sensing degradation data from a healthy state to a fault state completely.
The decoder unit of the transform of the embodiment of the present invention has one more layer of mask multi-headed attention layer and residual concatenated layer than the encoder unit. The input of the transform encoder-decoder module is multidimensional sensing time series data, and the data shape can be represented as X ═ N, T, F. Where N represents the number of data samples, T represents the length of the time window, and F represents the number of dimensions of the sensor data. Since the Transformer model does not have a convolution calculation module and a loop calculation module, and cannot directly use time series information in a time series, position coding information needs to be added before the Transformer model is input. In the invention, corresponding position coding information is added on different time steps of different characteristic channels by using sine functions with different frequencies, so that the problem that a transform structure cannot directly utilize time sequence information is solved.
Data X after position information additionposAnd inputting the depth features into an encoder of a transform encoder for depth feature extraction. Firstly, through a multi-head attention layer, extracting key time step information in a time sequence. Then, the characteristic information between the network layers is supplemented through residual linking. And finally, obtaining the characteristic output of the coding unit through a batch normalization layer. And if the number of the coding units exceeds one, outputting the coding unit of the previous coding unit as the input of the next coding unit. The depth feature output of the last coding unit is the required depth feature vector, i.e. the embedding vector, which can reflect the current time step health status information.
Unlike the input to the coding unit, the input to the transform decoder structure is first multiplied by a unit lower triangular matrix. The method aims to shield the information of the future time step by only utilizing the information of the historical time step and the current time step when the depth characteristic of the current time step is calculated. In the second multi-head attention calculation process of the decoder, the embedded vector output by the encoder participates in the calculation as an input to the attention layer. And finally, performing error loss calculation on the reconstructed time sequence Y output by the decoder and the input X, and adjusting the network parameter values of the transform encoder-decoder by using a back propagation algorithm.
Once training of the transform encoder-decoder network is completed, time series data input to the transform encoder may obtain an embedded vector sequence that reflects the machine health. We consider that the embedded vector acquired early in the operation of the machine reflects a completely healthy machine condition. As machine operating time increases, the embedded vector may gradually deviate from the healthy state embedded vector. The state deviation of the embedded vector is calculated by using the Euclidean distance, and a health index curve in a 0-1 data range can be obtained through normalization calculation.
The structure of the transform encoder-decoder module according to the embodiment of the present invention is shown in fig. 1, and the transform encoder-decoder module is composed of a multi-head attention layer, a residual connection layer and a feedforward neural network layer.
The multi-head attention layer is used for highlighting key time step information in the time sequence data by using the received time sequence data in a scaling-dot product attention calculation mode. The time series after multi-head attention calculation can give more weight information to the key time step, and the role of the time series in embedding the vector is highlighted.
The residual connecting layer is used for splicing the shallow features obtained by the shallow neural network with the deep features obtained by the deep neural network through cross connection. The method aims to relieve the phenomena of gradient loss of a Transformer structure and network learning capability degradation by supplementing information in the middle of a network layer.
The feed-forward neural network has the function of carrying out dimension transformation on the characteristics of the network middle layer according to vector calculation requirements. If the number of the transform coding units is not 1, the feedforward neural network needs to perform dimension transformation of depth features according to the input requirement of the next layer of coding units.
The invention is suitable for completely collected multidimensional sensing sequence data from a mechanical health state to a fault state, and aims to provide a method for constructing a mechanical health index with strong generalization, the implementation principle of the method for extracting a health index curve and predicting the service life of mechanical equipment provided by the embodiment of the invention is shown in figure 2, the specific processing flow is shown in figure 3, and the method comprises the following processing steps:
step S310: the method comprises the steps of collecting mechanical sensing degradation data from a healthy state to a fault state of a training example of mechanical equipment, carrying out data preprocessing to obtain time series data, and dividing the time series data by using a sliding time window to obtain a data block sequence.
Acquiring time series data X of mechanical health state from healthy operation to failure of training example of complete mechanical equipment1,x2,...,xt}. After the time series data X is subjected to data preprocessing operations such as noise reduction and smoothing, the time series data X is divided by using a sliding time window w with a fixed length to obtain a data block sequence omega with the same time window length { omega ═ omega }1,Ω2,...,Ωt}. The sliding time window processing process provided by the embodiment of the invention is shown in fig. 4, and the time window with the fixed length is moved each time from the initial stage to the end stage of the operation of the mechanical systemAnd moving a time step to obtain a time window data block sequence. At this point, data pre-processing is complete and the data shape of each data block sample is [ N, T, F ]]Where N represents the number of sample time window sequences, T represents the length of the time window, and F represents the dimension of the monitoring sensor.
Step S320: constructing a Transformer decoder-encoder comprising a multi-head attention layer, a residual connecting layer and a feedforward neural network layer, and extracting a health index curve reflecting the mechanical health state of a training example from the data block sequence in an unsupervised mode by using the Transformer decoder-encoder.
After the data preprocessing is finished, firstly adding position coding information in the data blocks in the data block sequence in a transform encoder part, wherein the formula is as follows:
Figure BDA0003094854200000131
because the structure of the Transformer is simple, no convolution calculation operator or loop calculation operator exists, and the time sequence order information in the time sequence data cannot be directly utilized. Therefore, we add position-coded information at different time steps for different sensing channels using sinusoidal functions of different frequencies. Pos in the formula represents the bit order of the data in the time series, i represents the ith sensing channel, and F represents the number of dimensions of the monitoring sensor.
Obtaining X after adding position information in data blockposThe data shape is still [ N, T, F]. Subsequently, XposInputting the data into a Transformer coding unit for multi-head attention calculation.
FIG. 5 is a block diagram of a multi-headed attention layer in a transform encoder-decoder according to an embodiment of the present invention, first XposIs transformed into [ N, T, M ] through linear layer]Wherein M is the number of hidden units of the linear layer. Then, for XposIs divided into three dimensions, and is disassembled and formed into [ N, H, T, H _ dim [ ]]Wherein N represents the number of samples, H represents the number of heads of attention, T represents the length of the time window, and H _ dim represents d _ modeThe number of units obtained by dividing l by H (H _ dim ═ d _ model// H). In FIG. 5, the Query vector (Q), Key vector (K) and value (V) vectors are equal to Xpos. Decomposing into H head input vector and changing into Query head vector (Q)head) Key head vector (K)head) And Value (V)head) Head vector, three vectors are in the shape of [ N, H, T, H _ dim]. On the left side of fig. 5 is a detailed zoom-dot-multiply-attention-calculation layer structure. The calculation formula is as follows:
Figure BDA0003094854200000132
wherein, KTRepresents the vector transposition with the shape [ N, H, H _ dim, T]. Different from the standard point-by-point attention calculation, the attention calculation result value is subjected to H _ dim scaling to prevent the dot product result from becoming large, so that the gradient in the network training process becomes too small, and the phenomenon of local optimization rather than global optimization is caused.
The residual concatenation is then used to supplement the information and pass it through a bulk normalization layer. Then, we convert to the required feature dimension through the feedforward layer. The calculation formula of the residual connection layer is as follows:
Redisual connection＝Fshallow network+FDeep network
FShallow networkRepresenting the shallow feature representation extracted by the shallow network model, FDeep networkRepresenting the extracted depth feature representation of the deep-level network model. The two are combined and input into a deep network for feature extraction operation.
The output of this coding unit is then taken again through the residual concatenation and normalization layer and input to the next unit. The output of the last coding unit is a compressed representation of the entire input time sequence, i.e. the embedded vector.
Unlike the input to the encoding unit, the decoding unit is at input XposIt is necessary to multiply the corresponding position by a unit lower triangular matrix to mask future information after the current computation time step. The calculation is schematically as follows:
Figure BDA0003094854200000141
thus, when calculating the t time step information, only the information of the current time step and the historical time step can be focused.
The output reconstruction time series Y of the decoding unit and the input X are subjected to a normal form loss function (objective function) calculation:
ei＝Xi-Yi,i∈{1,2,...,T}
Figure BDA0003094854200000142
the parameters of the transform encoder-decoder network structure are then trained using a back propagation algorithm. In the model training process, three measures can be taken to avoid overfitting:
(1) randomly selecting 20% of samples in the training set as a validation set. During the model training process, if the loss of the verification set does not drop obviously for 20 consecutive rounds, the network training is exited.
(2) And setting the regularization parameter value of L2 in a training optimizer to be 0.01, and limiting the value of the model parameter so that the trained model is relatively stable.
(3) In the gradient back propagation process, the gradient is cut, so that the model training process is smoother, and the obtained model is more stable.
Once the Transformer network structure training is complete. Time window sequence Ω ═ Ω1,Ω2,...,ΩtEach data block omega intAn embedded vector sequence reflecting the mechanical health state can be obtained by inputting the vector sequence into a Transformer encoder, and Z is { Z ═ Z }1,z2,...,zt}。
In the operation process of the mechanical system, the mechanical system is generally considered to be in a complete health state in an initial operation state, and as the operation time of the machine increases, the mechanical system gradually moves to a stable state and finally enters a fault state. Therefore, in this bookSeveral embedded vectors Z obtained at an initial time step considered in the present inventionnorm＝{z1,z2,z3Can characterize the state of the mechanical system as being completely healthy. Therefore, the health status evaluation of the current running time step can be obtained by calculating the variation value of the embedded vector sequence of the running time step and the embedded vector of the health status. The calculation formula is as follows:
Figure BDA0003094854200000151
Figure BDA0003094854200000152
where N represents the number of embedded vectors in a fully healthy state. dtIndicating the extent to which the system deviates from a healthy state at time t. dmax，dminRespectively representing the maximum and minimum values of the deviation of the system from the healthy state. h istA health indicator curve representing a normalized training example. So far, a health index curve of the mechanical equipment is obtained in an unsupervised training mode. Fig. 6 is a structural result example of a health index of a mechanical device according to an embodiment of the present invention.
Step S330: and forming a health index curve library by the health index curves of all the training examples, and establishing a mapping relation between the health index curves of the mechanical equipment and the sensing readings of the transducer encoder-decoder by using a linear regression model.
And sequentially calculating the similarity and the estimated value of the residual life of each training set example and the online test set example by using a similarity matching technology. Finally, the residual life estimation value of the test example is given in a weighted summation mode.
We take the NASA public data set (CMAPSS engine data subset FD001) as an example, and specifically describe the remaining life prediction calculation process of the test example. In the training set of the FD001 data subset, there are 100 complete airplane engine multidimensional sensing time series data instances running to failure. Each training data instance is processed by the time window of fig. 5 to obtain a time window sequence, and is input into the proposed Transformer model for unsupervised training. Each training instance may obtain a Health indicator curve (HIs). The health index curves of the training examples form an offline health index curve library (HIs-lib). The linear regression model is then used to express the mapping between the health indicator curve and the sensory readings of the transform encoder-decoder. The functional form of the linear regression model is as follows:
htr＝θ0+θxtr
wherein h istrRepresenting a health index value, x, of the training unittrA piece of multi-dimensional sensory data representing a training unit. Theta, theta0The values of the sense independent variable and the bias coefficient respectively represent linear regression functions. And training the linear regression model by using a least square method, and obtaining a mapping relation between a health index curve and the sensing reading of a Transformer encoder-decoder after training. Model training is performed by using a least square method, and an objective function is as follows: h istr-(θ0+θxtr). When the loss of the objective function reaches the minimum, the required linear regression model can be obtained.
Subsequently, the sensing data of the test case can be directly input into the mapping relation to obtain the corresponding health index curve.
Step S340: and inputting the sensing reading of the mechanical equipment to be evaluated into the mapping relation to obtain a health index curve of the mechanical equipment to be evaluated, and performing similarity calculation on the health index curve of the mechanical equipment to be evaluated and the health index curve library to obtain an estimation result of the residual life of the mechanical equipment to be evaluated.
In the FD001 test set, there are 100 instances of the multidimensional sense time series (run-to-fault truncated) that truncate run cycles. For the residual life prediction of a single test case, firstly, the sensing reading of the mechanical equipment to be evaluated is input into a trained linear regression model to obtain a truncated health index curve. Second, the similarity of the test health indicator curve to each of the training health indicator curves in HIs-lib is calculated using a similarity metric formula. The similarity calculation formula is as follows:
Figure BDA0003094854200000171
Figure BDA0003094854200000172
wherein HI', HI(j)Respectively representing the health index curve of the test example and the health index curve of the jth training example in HIs-lib. τ represents the time shift of the test case. λ is a relaxation factor for adjusting the magnitude of the similarity metric value. d (-) represents the Euclidean distance of the two curves.
Forming similarity measure vector Sim ═ Sim of similarities between the health index curves of all training examples in the health index curve library and the health index curves of the test examples1,Sim2,...,Simn]；
Then, the remaining life estimation value at each time shift τ is calculated using the following equation.
RUL(j,τ)＝Tj-T'-τ
Wherein, TjAnd T' represents the running time length of the jth training example in HIs-lib and the running time length of the test example respectively. To intuitively explain the remaining life calculation process, the calculation process of matching the similarity of the health index curve of the jth training example and the test example with the remaining service life is shown in fig. 7.
Forming residual life estimation values of the mechanical equipment to be evaluated, which are calculated according to the health index curves of all the training examples in the health index curve library, into a residual life estimation vector RUL ═ RUL1,RUL2,...,RULn]；
Each training instance is calculated to obtain a similarity metric vector Sim ═ Sim1,Sim2,...,Simn]And residual lifetime estimation vector RUL ═ RUL1,RUL2,...,RULn]I.e. each training instance will be tested onlineThe test example calculates similarity and remaining life estimates.
Finally, according to the similarity metric vector Sim ═ Sim1,Sim2,...,Simn]And said residual lifetime estimation vector RUL ═ RUL1,RUL2,...,RULn]And predicting the residual service life of the test case of the mechanical equipment to be evaluated in a weighted summation mode. The formula is as follows:
Figure BDA0003094854200000181
wherein RUL ^ represents the estimated value of the residual service life of the mechanical equipment,
Figure BDA0003094854200000182
representing the maximum degree of similarity, β controls the number of training instances that participate in the final life estimation.
In summary, the method of the embodiment of the present invention extracts a scheme of a health index sequence for reflecting the health status of the machine from the complete degradation time sequence, and further effectively predicts the remaining service life of the machine. The method is suitable for any collected mechanical sensing degradation data from a healthy state to a fault state completely.
The method has strong generalization capability and can be suitable for mining the residual life of complete degraded mechanical data. For example: aircraft engines, cutter wear, etc.
The invention can provide more accurate residual life prediction time for mechanical equipment maintenance personnel to make dynamic maintenance plans. Compared with the traditional regular timing operation and maintenance, the method is more flexible in time, unnecessary maintenance is reduced, and cost is saved.
Those of ordinary skill in the art will understand that: the figures are merely schematic representations of one embodiment, and the blocks or flow diagrams in the figures are not necessarily required to practice the present invention.
From the above description of the embodiments, it is clear to those skilled in the art that the present invention can be implemented by software plus necessary general hardware platform. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which may be stored in a storage medium, such as ROM/RAM, magnetic disk, optical disk, etc., and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method according to the embodiments or some parts of the embodiments.
The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for apparatus or system embodiments, since they are substantially similar to method embodiments, they are described in relative terms, as long as they are described in partial descriptions of method embodiments. The above-described embodiments of the apparatus and system are merely illustrative, and the units described as separate parts may or may not be physically separate, and the parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of the present embodiment. One of ordinary skill in the art can understand and implement it without inventive effort.
The above description is only for the preferred embodiment of the present invention, but the scope of the present invention is not limited thereto, and any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (7)

1. A health index curve extraction and service life prediction method for mechanical equipment is characterized by comprising the following steps:
collecting mechanical sensing degradation data from a healthy state to a fault state of a training example of mechanical equipment, carrying out data preprocessing to obtain time sequence data, and dividing the time sequence data by using a sliding time window to obtain a data block sequence;
constructing a Transformer decoder-encoder comprising a multi-head attention layer, a residual connecting layer and a feedforward neural network layer, and extracting a health index curve reflecting the mechanical health state of a training example from the data block sequence by using the Transformer decoder-encoder;
constructing a health index curve library by the health index curves of all the training examples, and establishing a mapping relation between the health index curve of the mechanical equipment and the sensing reading of a transform encoder-decoder by using a linear regression model;
and inputting the sensing reading of the test example of the mechanical equipment to be evaluated into the mapping relation to obtain a health index curve of the test example of the mechanical equipment to be evaluated, and performing similarity calculation on the health index curve of the test example of the mechanical equipment to be evaluated and the health index curve library to obtain an estimation result of the residual life of the mechanical equipment to be evaluated.
2. The method of claim 1, wherein the collecting mechanical equipment mechanical sensing degradation data from healthy to faulty states and performing data preprocessing to obtain time series data, and dividing the time series data by using a sliding time window to obtain a data block sequence comprises:
collecting time series data X of mechanical health state from healthy operation to fault of training example of mechanical equipment1,x2,...,xtAnd after the time series data X is subjected to noise reduction and smooth data preprocessing operation, dividing the time series data X by using a sliding time window w with a fixed length, moving the time window with the fixed length by one time step each time to obtain a time window data block sequence, and obtaining a data block sequence omega with the same time window length as { omega ═ omega }1,Ω2,...,ΩtThe data shape of each data block sample is [ N, T, F ]]Where N represents the number of sample time window sequences,t represents the length of the time window and F represents the dimension of the monitoring sensor.
3. The method of claim 1, wherein constructing a transform decoder-encoder comprising a multi-head attention layer, a residual connection layer, and a feedforward neural network layer comprises:
constructing a Transformer decoder-encoder, which consists of a multi-head attention layer, a residual error connection layer and a feedforward neural network layer;
the multi-head attention layer highlights key time step information in the time sequence data by using the received time sequence data and adopting a scaling-dot product attention calculation mode;
the residual error connecting layer splices the shallow features obtained by the shallow neural network and the deep features obtained by the deep neural network through cross connection to supplement information in the middle of the network layer;
and the feedforward neural network performs dimension transformation on the characteristics of the network middle layer according to the vector calculation requirement. If the number of the transform coding units is not 1, the feedforward neural network needs to perform dimension transformation of depth features according to the input requirement of the next layer of coding units.
4. The method according to claim 3, wherein said extracting a health indicator curve reflecting a machine health status of a training instance from said data block sequence by using said transform encoder-decoder structure comprises:
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtThe input is input to a coder of a transform coder, which is set to { Ω } in a sequence of data blocks1,Ω2,...,ΩtAdding position coding information into each data block in the sequence to obtain a data sequence XposData sequence XposThe depth feature is extracted by an encoder input into a transform encoder, firstly, a data sequence X is extracted through a multi-head attention layerposThe key time step information in the process, the characteristic information between network layers is supplemented through residual error linkage, and finally the process is batchedThe quantity normalization layer obtains the feature output of the coding units, if the number of the coding units exceeds one, the output of the previous coding unit is used as the input of the next coding unit, and the last coding unit outputs a depth feature vector reflecting the current time step health state information;
inputting the depth eigenvector into a decoder of a transform coder, wherein the decoder multiplies the depth eigenvector by a unit lower triangular matrix to be used as input data X of a multi-head attention layer, the decoder outputs a reconstructed time sequence Y after the input data X is subjected to a multi-head attention calculation process, the reconstructed time sequence Y and the input data X are subjected to error loss calculation, the network parameter value of the transform coder-decoder is adjusted by using a back propagation algorithm according to the error loss calculation result, and the transform coder-decoder with the adjusted network parameter value is used as a trained transform coder-decoder;
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtAnd inputting the trained encoder in a transform encoder-decoder to obtain a depth feature vector sequence reflecting the time step health state information of the mechanical equipment, and carrying out normalization calculation on the depth feature vector sequence to obtain a health index curve in a 0-1 data range of the training example of the mechanical equipment.
5. The method of claim 4, wherein said extracting a health indicator curve of the mechanical health status of the training instance from the data block sequence using the transform encoder-decoder structure comprises:
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtThe input is input to a coder of a transform coder, which is set to { Ω } in a sequence of data blocks1,Ω2,...,ΩtAdding position coding information into each data block in the data block, wherein the formula is as follows:
Figure FDA0003094854190000031
pos in the formula represents the bit number of the data in the time sequence, i represents the ith sensing channel, and F represents the dimension number of the monitoring sensor;
obtaining a data sequence X after adding the position information in the data blockposThe encoder pairs the data sequence XposPerforming a Multi-headed attention calculation, data series XposIs transformed into [ N, T, M ] through linear layer]Wherein M is the number of hidden units of the linear layer, for the data sequence XposIs divided into three dimensions, and is disassembled and formed into [ N, H, T, H _ dim [ ]]The multi-head input vector (v) includes a Query vector (Q), a Key vector (K), and value (v), and the calculation formula is as follows:
Figure FDA0003094854190000041
wherein, KTRepresents the vector transposition with the shape [ N, H, H _ dim, T]
Inputting the multi-head input vector as a shallow feature into a residual connecting layer, wherein the calculation formula of the residual connecting layer is as follows:
Redisual connection＝Fshallow network+FDeep network
FShallow networkRepresenting the shallow feature representation extracted by the shallow network model, FDeep networkRepresenting the depth feature representation extracted by the deep network model, and combining the depth feature representation and the depth feature representation to be input into a deep network for feature extraction operation;
then, obtaining the output of the coding unit through residual connection and normalization layer again, inputting the output to the next unit, and outputting the depth characteristic vector reflecting the current time step health status information by the last coding unit;
inputting the depth eigenvector into a decoder of a transform coder, wherein the decoder multiplies the depth eigenvector by a unit lower triangular matrix to be used as input data X of a multi-head attention layer, the decoder outputs a reconstruction time sequence Y after the input data X is subjected to a multi-head attention calculation process, and the reconstruction time sequence Y and the input data X are subjected to one-norm loss function calculation:
ei＝Xi-Yi,i∈{1,2,...,T}
Figure FDA0003094854190000042
adjusting the network parameters of a Transformer encoder-decoder by using a back propagation algorithm according to the calculation result of the first normal form loss function, and taking the Transformer encoder-decoder with the adjusted network parameter values as a trained Transformer encoder-decoder;
setting the data block sequence omega to { omega ═ omega1,Ω2,...,ΩtEach data block omega intInput to a trained Transformer encoder-decoder that outputs an embedded vector sequence reflecting machine health status, Z ═ Z { (Z })1,z2,...,ztAnd f, setting an embedded vector Z obtained at an initial time stepnorm＝{z1,z2,z3The method can represent the complete health state of the mechanical system, and calculate the change value of the embedded vector sequence of the running time step and the embedded vector of the health state, and the calculation formula is as follows:
Figure FDA0003094854190000051
Figure FDA0003094854190000052
where N represents the number of embedded vectors in a fully healthy state, dtIndicating the degree of departure of the system from health at time t, dmax，dminMaximum and minimum values, h, respectively, representing the deviation of the system from the healthy statetHealth index profile representing a 0-1 data range of a normalized training instance of a machineA wire.
6. The method of any one of claims 1 to 5, wherein the step of constructing the health index curves of all the training examples into a health index curve library and the step of using a linear regression model to establish a mapping relationship between the health index curves of the mechanical equipment and the sensing readings of the transducer encoder-decoder comprises:
acquiring a plurality of time series data of machine health states from healthy operation to failure of a plurality of training examples of mechanical equipment, acquiring a health index curve corresponding to each time series data by using a Transformer encoder-decoder, and forming the health index curves of all the training examples into a health index curve library;
the mapping between the health indicator curve and the sensory readings of the transform encoder-decoder is expressed using a linear regression model, which has the following functional form:
Figure FDA0003094854190000053
wherein h istrIndicates the value of the health index of the training unit, xtrA piece of multi-dimensional sensory data, theta, representing a training unit0A sensing independent variable coefficient and a bias coefficient value respectively representing a linear regression function;
and training the linear regression model by using a least square method, and obtaining a mapping relation between a health index curve and the sensing reading of a Transformer encoder-decoder after training.
7. The method according to claim 6, wherein the inputting the sensing readings of the test case of the mechanical equipment to be evaluated into the mapping relationship to obtain a health index curve of the test case of the mechanical equipment to be evaluated, and performing similarity calculation on the health index curve of the test case of the mechanical equipment to be evaluated and the health index curve library to obtain the estimation result of the remaining life of the mechanical equipment to be evaluated comprises:
inputting the sensing reading of the Transformer encoder-decoder of the test example of the mechanical equipment to be evaluated into the mapping relation between the health index curve and the sensing reading of the Transformer encoder-decoder to obtain the health index curve of the test example of the mechanical equipment to be evaluated;
calculating the similarity between the health index curve of the test example of the mechanical equipment to be evaluated and the health index curve of each training example in the health index curve library by using a similarity measurement formula, wherein the similarity calculation formula is as follows:
Figure FDA0003094854190000061
Figure FDA0003094854190000062
wherein HI' represents a health index curve of a test case of a mechanical device to be evaluated, HI(j)Representing a health index curve of a jth training example in the health index curve library, wherein tau represents time shift of a test example of the mechanical equipment to be evaluated, lambda is a relaxation factor used for adjusting the size of a similarity metric value, and d (·) represents an Euclidean distance between the two curves;
forming similarity measure vector Sim ═ Sim of similarities between the health index curves of all training examples in the health index curve library and the health index curves of the test examples1,Sim2,...,Simn]；
The remaining life estimate for the mechanical device under evaluation at time shift τ for each test example was calculated using the following equation:
RUL(j,τ)＝Tj-T'-τ
wherein, TjThe T' respectively represents the running time of the jth training example and the running time of the testing example in the health index curve library;
all trainings in the library of curves according to the health indexThe residual life estimation value of the mechanical equipment to be evaluated calculated by the health index curve of the exercise example forms a residual life estimation vector RUL ═ RUL1，RUL2，...，RULn]；
According to the similarity metric vector Sim ═ Sim1，Sim2，...，Simn]And said residual lifetime estimation vector RUL ═ RUL1，RUL2，...，RULn]And predicting the residual service life of the test example of the mechanical equipment to be evaluated by using a weighted summation mode, wherein the formula is as follows:
Figure FDA0003094854190000071
s.t.Sim(j，τ)≥β*(maxj，τSim(j，τ))
wherein RUL ^ represents the estimated value of the residual service life of the mechanical equipment,
Figure FDA0003094854190000072
representing the maximum degree of similarity, β controls the number of training instances that participate in the final life estimation.
CN202110608119.7A 2021-06-01 2021-06-01 Health index curve extraction and service life prediction method for mechanical equipment Active CN113434970B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110608119.7A CN113434970B (en) 2021-06-01 2021-06-01 Health index curve extraction and service life prediction method for mechanical equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110608119.7A CN113434970B (en) 2021-06-01 2021-06-01 Health index curve extraction and service life prediction method for mechanical equipment

Publications (2)

Publication Number Publication Date
CN113434970A true CN113434970A (en) 2021-09-24
CN113434970B CN113434970B (en) 2023-01-17

Family

ID=77803425

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110608119.7A Active CN113434970B (en) 2021-06-01 2021-06-01 Health index curve extraction and service life prediction method for mechanical equipment

Country Status (1)

Country Link
CN (1) CN113434970B (en)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114035098A (en) * 2021-12-14 2022-02-11 北京航空航天大学 Lithium battery health state prediction method integrating future working condition information and historical state information
CN114186475A (en) * 2021-10-28 2022-03-15 南京工业大学 Pivoting support service life prediction method based on Attention-MGRU
CN114298443A (en) * 2022-03-03 2022-04-08 北京航空航天大学杭州创新研究院 Industrial equipment predictive maintenance method and device based on health state index and electronic equipment
CN114580982A (en) * 2022-05-07 2022-06-03 昆仑智汇数据科技(北京)有限公司 Method, device and equipment for evaluating data quality of industrial equipment
CN114936517A (en) * 2022-04-28 2022-08-23 上海波士内智能科技有限公司 Metal welding signal characteristic curve characteristic modeling method based on deep learning
CN115392360A (en) * 2022-08-11 2022-11-25 哈尔滨工业大学 Transformer-based large bridge temperature-response related pattern recognition and health diagnosis method
CN115713044A (en) * 2023-01-09 2023-02-24 佰聆数据股份有限公司 Method and device for analyzing residual service life of electromechanical equipment under multi-working-condition switching

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107101828A (en) * 2017-05-27 2017-08-29 辽宁大学 The bearing residual life Forecasting Methodology that a kind of encoder health degree is evaluated
CN109141881A (en) * 2018-07-06 2019-01-04 东南大学 A kind of rotating machinery health evaluating method of depth autoencoder network
US20190042675A1 (en) * 2017-08-07 2019-02-07 The Boeing Company System and method for remaining useful life determination
CN112364567A (en) * 2020-11-18 2021-02-12 浙江大学 Residual life prediction method based on consistency check of similarity of degraded tracks
CN112418277A (en) * 2020-11-03 2021-02-26 西安电子科技大学 Method, system, medium, and apparatus for predicting remaining life of rotating machine component

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107101828A (en) * 2017-05-27 2017-08-29 辽宁大学 The bearing residual life Forecasting Methodology that a kind of encoder health degree is evaluated
US20190042675A1 (en) * 2017-08-07 2019-02-07 The Boeing Company System and method for remaining useful life determination
CN109141881A (en) * 2018-07-06 2019-01-04 东南大学 A kind of rotating machinery health evaluating method of depth autoencoder network
CN112418277A (en) * 2020-11-03 2021-02-26 西安电子科技大学 Method, system, medium, and apparatus for predicting remaining life of rotating machine component
CN112364567A (en) * 2020-11-18 2021-02-12 浙江大学 Residual life prediction method based on consistency check of similarity of degraded tracks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
李劲松等: "基于退化曲线相似性的剩余使用寿命估计方法", 《应用科技》 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114186475A (en) * 2021-10-28 2022-03-15 南京工业大学 Pivoting support service life prediction method based on Attention-MGRU
CN114186475B (en) * 2021-10-28 2023-05-12 南京工业大学 Slewing bearing service life prediction method based on Attention-MGRU
CN114035098A (en) * 2021-12-14 2022-02-11 北京航空航天大学 Lithium battery health state prediction method integrating future working condition information and historical state information
CN114298443A (en) * 2022-03-03 2022-04-08 北京航空航天大学杭州创新研究院 Industrial equipment predictive maintenance method and device based on health state index and electronic equipment
WO2023165006A1 (en) * 2022-03-03 2023-09-07 北京航空航天大学杭州创新研究院 Predictive maintenance method and apparatus for industrial equipment based on health status index, and electronic device
CN114936517A (en) * 2022-04-28 2022-08-23 上海波士内智能科技有限公司 Metal welding signal characteristic curve characteristic modeling method based on deep learning
CN114580982A (en) * 2022-05-07 2022-06-03 昆仑智汇数据科技(北京)有限公司 Method, device and equipment for evaluating data quality of industrial equipment
CN114580982B (en) * 2022-05-07 2022-08-05 昆仑智汇数据科技(北京)有限公司 Method, device and equipment for evaluating data quality of industrial equipment
CN115392360A (en) * 2022-08-11 2022-11-25 哈尔滨工业大学 Transformer-based large bridge temperature-response related pattern recognition and health diagnosis method
CN115392360B (en) * 2022-08-11 2023-04-07 哈尔滨工业大学 Transformer-based large bridge temperature-response related pattern recognition and health diagnosis method
CN115713044A (en) * 2023-01-09 2023-02-24 佰聆数据股份有限公司 Method and device for analyzing residual service life of electromechanical equipment under multi-working-condition switching
CN115713044B (en) * 2023-01-09 2023-04-25 佰聆数据股份有限公司 Method and device for analyzing residual life of electromechanical equipment under multi-condition switching

Also Published As

Publication number Publication date
CN113434970B (en) 2023-01-17

Similar Documents

Publication Publication Date Title
CN113434970B (en) Health index curve extraction and service life prediction method for mechanical equipment
US11334407B2 (en) Abnormality detection system, abnormality detection method, abnormality detection program, and method for generating learned model
CN109948117A (en) A kind of satellite method for detecting abnormality fighting network self-encoding encoder
Che et al. Hybrid multimodal fusion with deep learning for rolling bearing fault diagnosis
CN112036042B (en) Power equipment abnormality detection method and system based on variational modal decomposition
Xia et al. Multiscale similarity ensemble framework for remaining useful life prediction
CN115688035A (en) Time sequence power data anomaly detection method based on self-supervision learning
Deng et al. LSTMED: An uneven dynamic process monitoring method based on LSTM and Autoencoder neural network
CN115618732B (en) Nuclear reactor digital twin key parameter autonomous optimization data inversion method
CN112147432A (en) BiLSTM module based on attention mechanism, transformer state diagnosis method and system
CN115169430A (en) Cloud network end resource multidimensional time sequence anomaly detection method based on multi-scale decoding
CN114297918A (en) Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning
Son et al. Deep learning-based anomaly detection to classify inaccurate data and damaged condition of a cable-stayed bridge
CN116679211A (en) Lithium battery health state prediction method
CN117076171A (en) Abnormality detection and positioning method and device for multi-element time sequence data
CN117763494A (en) Bearing diagnosis method and system based on multi-mode and multi-scale fusion network
Xie et al. Adaptive symbolic transfer entropy and its applications in modeling for complex industrial systems
Huang et al. A hybrid bayesian deep learning model for remaining useful life prognostics and uncertainty quantification
Wu et al. Multiscale reduction clustering of vibration signals for unsupervised diagnosis of machine faults
Luo et al. A novel method for remaining useful life prediction of roller bearings involving the discrepancy and similarity of degradation trajectories
Yang et al. Dual-frequency enhanced attention network for aircraft engine remaining useful life prediction
Guo et al. MHT: A multiscale hourglass-transformer for remaining useful life prediction of aircraft engine
CN114792026A (en) Method and system for predicting residual life of aircraft engine equipment
CN117688496B (en) Abnormality diagnosis method, system and equipment for satellite telemetry multidimensional time sequence data
Gao et al. Construction method of turbine engine health indicator based on deep learning

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
GR01 Patent grant
GR01 Patent grant