CN113125095A - Universal circuit breaker contact system residual mechanical life prediction method based on deep learning - Google Patents
Universal circuit breaker contact system residual mechanical life prediction method based on deep learning Download PDFInfo
- Publication number
- CN113125095A CN113125095A CN202110416581.7A CN202110416581A CN113125095A CN 113125095 A CN113125095 A CN 113125095A CN 202110416581 A CN202110416581 A CN 202110416581A CN 113125095 A CN113125095 A CN 113125095A
- Authority
- CN
- China
- Prior art keywords
- formula
- value
- circuit breaker
- time
- effective
- 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
Links
- 238000000034 method Methods 0.000 title claims abstract description 135
- 238000013135 deep learning Methods 0.000 title claims abstract description 20
- 239000013598 vector Substances 0.000 claims description 40
- 230000006870 function Effects 0.000 claims description 28
- 230000008569 process Effects 0.000 claims description 27
- 238000005070 sampling Methods 0.000 claims description 27
- 238000000605 extraction Methods 0.000 claims description 25
- 239000011159 matrix material Substances 0.000 claims description 25
- 238000012360 testing method Methods 0.000 claims description 22
- 230000009471 action Effects 0.000 claims description 18
- 238000000354 decomposition reaction Methods 0.000 claims description 16
- 230000004913 activation Effects 0.000 claims description 15
- 238000004364 calculation method Methods 0.000 claims description 15
- 238000011176 pooling Methods 0.000 claims description 12
- 239000007787 solid Substances 0.000 claims description 12
- 239000012634 fragment Substances 0.000 claims description 9
- 230000007246 mechanism Effects 0.000 claims description 9
- 210000002569 neuron Anatomy 0.000 claims description 9
- 238000004146 energy storage Methods 0.000 claims description 6
- 238000001914 filtration Methods 0.000 claims description 6
- 238000009432 framing Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 210000004027 cell Anatomy 0.000 claims description 5
- 230000037433 frameshift Effects 0.000 claims description 4
- 238000009499 grossing Methods 0.000 claims description 4
- 230000006399 behavior Effects 0.000 claims description 3
- 230000000875 corresponding effect Effects 0.000 claims 2
- 108010069898 fibrinogen fragment X Proteins 0.000 claims 2
- 230000002596 correlated effect Effects 0.000 claims 1
- 238000012549 training Methods 0.000 abstract description 23
- 238000006731 degradation reaction Methods 0.000 abstract description 18
- 230000015556 catabolic process Effects 0.000 abstract description 14
- 238000011160 research Methods 0.000 abstract description 12
- 238000013528 artificial neural network Methods 0.000 abstract description 4
- 230000015654 memory Effects 0.000 abstract description 3
- 238000010586 diagram Methods 0.000 description 19
- 238000013527 convolutional neural network Methods 0.000 description 12
- 238000012423 maintenance Methods 0.000 description 8
- 238000012795 verification Methods 0.000 description 8
- 238000004458 analytical method Methods 0.000 description 5
- 230000008859 change Effects 0.000 description 5
- 238000009826 distribution Methods 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000007787 long-term memory Effects 0.000 description 4
- 238000010200 validation analysis Methods 0.000 description 4
- 230000002159 abnormal effect Effects 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000036541 health Effects 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 230000000306 recurrent effect Effects 0.000 description 3
- 230000005653 Brownian motion process Effects 0.000 description 2
- 230000002238 attenuated effect Effects 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- WHXSMMKQMYFTQS-UHFFFAOYSA-N Lithium Chemical compound [Li] WHXSMMKQMYFTQS-UHFFFAOYSA-N 0.000 description 1
- 238000005299 abrasion Methods 0.000 description 1
- 239000000654 additive Substances 0.000 description 1
- 230000000996 additive effect Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000000872 buffer Substances 0.000 description 1
- 238000005520 cutting process Methods 0.000 description 1
- 230000003628 erosive effect Effects 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000009545 invasion Effects 0.000 description 1
- 229910052744 lithium Inorganic materials 0.000 description 1
- 230000033001 locomotion Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 238000012216 screening Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M7/00—Vibration-testing of structures; Shock-testing of structures
- G01M7/02—Vibration-testing by means of a shake table
- G01M7/022—Vibration control arrangements, e.g. for generating random vibrations
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M7/00—Vibration-testing of structures; Shock-testing of structures
- G01M7/02—Vibration-testing by means of a shake table
- G01M7/025—Measuring arrangements
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R15/00—Details of measuring arrangements of the types provided for in groups G01R17/00 - G01R29/00, G01R33/00 - G01R33/26 or G01R35/00
- G01R15/14—Adaptations providing voltage or current isolation, e.g. for high-voltage or high-current networks
- G01R15/20—Adaptations providing voltage or current isolation, e.g. for high-voltage or high-current networks using galvano-magnetic devices, e.g. Hall-effect devices, i.e. measuring a magnetic field via the interaction between a current and a magnetic field, e.g. magneto resistive or Hall effect devices
- G01R15/202—Adaptations providing voltage or current isolation, e.g. for high-voltage or high-current networks using galvano-magnetic devices, e.g. Hall-effect devices, i.e. measuring a magnetic field via the interaction between a current and a magnetic field, e.g. magneto resistive or Hall effect devices using Hall-effect devices
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R19/00—Arrangements for measuring currents or voltages or for indicating presence or sign thereof
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/327—Testing of circuit interrupters, switches or circuit-breakers
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention relates to a universal circuit breaker contact system residual mechanical life prediction method based on deep learning. The method adopts a deep learning method to carry out life prediction research, and firstly, a concept of predicting effective segments of vibration signals by mechanical life is provided; secondly, introducing a VMD algorithm and automatically calibrating the interval of the double thresholds based on short-time energy; constructing a multi-channel convolution self-encoder network (MCCAE) again, training in an unsupervised learning mode, and extracting deep degradation time sequence characteristics of effective segments; and finally, constructing a long-time and short-time memory neural network (LSTM), taking the time sequence characteristics as input, adopting a supervised learning training mode, completing the steps of prediction and the like, and effectively completing the prediction of the residual mechanical life of the universal circuit breaker contact system.
Description
Technical Field
The technical scheme of the invention relates to the technical field of residual service life prediction of a contact system of a circuit breaker, in particular to a universal circuit breaker contact system residual mechanical service life prediction method based on deep learning.
Background
The low-voltage distribution system is used as the tail end of the whole power system, is the ring closest to users, and is inseparable from the life safety of people and the stable operation of the society. The universal circuit breaker is used as key power equipment of a low-voltage power distribution system, is used as scheduling control equipment of the power system on one hand, and is used for putting in or cutting off specific lines according to the operation requirements of a power grid, and plays a role in protecting the power system on the other hand. Due to the complexity of a mechanical structure and the influence of other uncertain factors, the performance and the health state of the universal circuit breaker are inevitably degraded in the operation process, and finally the system is failed. As a critical component of a circuit breaker, the contact system is more prone to failure. The contact system is an actuating mechanism in the switching-on and switching-off process of the circuit breaker, when the circuit breaker is abnormal, the switching-on and switching-off action is slow if the circuit breaker is abnormal, and the circuit breaker can not be broken timely if the circuit breaker is abnormal, so that a major safety accident is caused. Furthermore, if the residual service life of the contact system can be accurately predicted in time according to monitoring information in the performance degradation process of the contact system, the reliability of the circuit breaker can be greatly improved, the occurrence of faults is avoided, meanwhile, the downtime can be effectively reduced, the maintenance period is shortened, the maintenance steps are simplified, and therefore the maintenance cost is reduced. Therefore, the method has important significance for researching the residual service life of the contact system of the universal circuit breaker.
As a core of Prediction and Health Management (PHM) technology, Remaining life (RUL) prediction has been widely applied and achieved with great success in the fields of aeroengines, bearings, lithium batteries, etc., but in the field of circuit breakers, there is only a clear concern about mechanical life prediction of universal circuit breakers, and most of them are life research about high-voltage and vacuum circuit breakers, etc. For example, Payman et al (Dehghanian P, Guan Y, Kezunovic M.real-time life-cycle assessment of high-voltage circuit breakers for main using online condition monitoring data [ J ]. IEEE Transactions on Industrial Applications,2019,55(2):1135-1146.) propose a method for estimating the life cycle of a high-voltage circuit breaker in real time by using online condition monitoring data, which is helpful for determining the probability that each high-voltage circuit breaker is in different degradation stages within the life thereof, and further is helpful for more exactly finding the optimal usage scheme of the circuit breaker; vacuum circuit breaker reliability analysis [ J ] based on Weibull distribution, high-voltage electrical appliances 2020,56(01):30-35.) of Muyan et al (Muyan, Wangyongxing, Zhouyi, et al) establishes a residual life prediction model of the vacuum circuit breaker based on Weibull distribution, and reasonably predicts the average mechanical life of the vacuum circuit breaker. At present, the service life prediction research of the universal circuit breaker is only carried out on an operation accessory and a brake separating action mechanism, and the service life prediction of a contact system is not carried out. In addition, the method is a statistical data driving method, and the method usually needs to select an artificial degradation index method according to knowledge and experience, so that the prediction precision is not high, and the generalization capability is lacked.
Disclosure of Invention
The invention aims to provide a universal circuit breaker contact system residual mechanical life prediction method based on deep learning, aiming at the characteristics of the prior art that a statistical data driving method has defects and the universal circuit breaker is degraded. According to the method, through research on a contact system, a deep learning method is adopted for life prediction research, and firstly, an effective segment concept of mechanical life prediction vibration signals is provided; secondly, introducing a VMD algorithm and automatically calibrating the interval of the double thresholds based on short-time energy; constructing a multi-channel convolution self-encoder network (MCCAE) again, training in an unsupervised learning mode, and extracting deep degradation time sequence characteristics of effective segments; and finally, constructing a long-time and short-time memory neural network (LSTM), taking the time sequence characteristics as input, adopting a supervised learning training mode, completing the steps of prediction and the like, and effectively completing the prediction of the residual mechanical life of the universal circuit breaker contact system.
The technical scheme of the invention is as follows:
a method for predicting the residual mechanical life of a conventional breaker contact system based on deep learning comprises the following steps:
the method comprises the following specific scheme:
firstly, a service life test system of a contact system of the universal circuit breaker is used for measuring a closing vibration signal X (t) ═ x in the action process of the universal circuit breaker1,…,xi,…,xN) Where N is the number of sampling points, xi-xNN sampling points;
secondly, performing VMD decomposition on the closing original vibration signal X (t) to extract effective modal components; the method comprises the following specific steps:
2-1, initializing the number K of pre-decomposition modal components to be 2, and performing VMD decomposition on the brake original vibration signal X (t) to obtain K inherent modal components (IMFs);
2-2 respectively obtaining the cross-correlation coefficients rho of K IMFs and the original vibration signal X (t)kCalculating the difference c between adjacent mode and original signal cross-correlation coefficient according to formula (1) and formula (2)k-1Obtaining K cross correlation coefficients rhokAnd K-1 cross correlation coefficient differences ck-1;
ck-1=ρk-ρk-1 (2)
In the formula (1), N is the number of sampling points, xn、Is a switching-on vibration signal X (t) sampling point and its mean value, ukn、The Kth IMF component element and the mean value thereof; in the formula (2), ρk,ρk-1The cross-correlation coefficients of the kth and kth-1 IMF components with X (t), respectively;
2-3 presetting a minimum threshold value c ═ Δ, Δ ranging from 0 to 0.05, if c is obtainediIf the value is larger than a preset threshold value c, K is equal to K +1, and the operation is repeated for 2-1 and 2-2 in a circulating mode; if obtained ciIf the difference is smaller than a preset threshold value c, the two modes are considered to have high similarity, namely, an over-decomposition phenomenon occurs, and if the circulation is stopped, K is equal to K-1;
2-4, after the K value is determined, selecting the IMF with the maximum correlation coefficient under the current K value as an effective modal component u' (t), and regarding the modal signal as the most relevant to the original signal as the main characteristic information for representing the original signal;
thirdly, combining the effective modal component u' (t) to complete automatic calibration of the effective segment; the method comprises the following steps:
3-1 framing the active mode component u' (t) according to equation (3),
zi(n)=ω(n)*u′((i-1)*inc+n) (3)
in the formula, ω (n) is a window function; z is a radical ofi(n) is a frame number, n is 1,2, …, L.i is 1,2, …, fnL is the frame length; inc is the frame shift length; f. ofnThe total frame number after framing; wherein omega (n) is selected from a Hainin window according to formula (4),
3-2 calculating z of the i frame effective modal componenti(n) short-time energy formula according to formula (5):
in the formula, zi(n) is the value of a frame of data, E (i) is the short-time energy of a frame of data;
3-3, completing effective segment calibration by a double-threshold method:
obtaining a short-time energy value E (i) of the effective modal component through 3-2, and further drawing a short-time energy waveform; then setting two threshold values Q according to the wave crest and the wave trough of the drawn short-time energy waveform1And Q2,Q1<Q2Wherein Q is1The short-time energy value corresponding to the wave trough is set to be 0-0.1; q2Setting the short-time energy value corresponding to the second vibration peak value within the range of 0.1-0.5; when the energy value E (i) is higher than Q2The signal segment interval of the contact impact is determined, and the energy value of the frame data is higher or lower than Q1To determine the specific end point E of the effective segment1,E2,E1,E2The original vibration signal in between is the valid segment X' (t);
fourthly, taking the effective vibration segment X' (t) of the mechanical life of the contact system as input, and establishing a single-channel convolution self-encoder (CAE) feature extraction model; the method comprises the following specific steps:
4-1 constructing an Encoder (Encoder) part of a Convolutional Auto Encoder (CAE);
4-1-1 is provided with an effective segment X' (t) action X of ithi′={x1,x2,…,xt,…,xn}(xtCharacteristic value indicating time t) by Xi' for input, build a convolutional layer, according to equation (6):
H1=σf(W*Xi′+b1) (6)
in the formula, Xi′∈R1×nIs the effective fragment sequence of the ith row, and n is the length of the effective fragment sequence; denotes a one-dimensional convolution operation; w and b1Respectively representing a one-dimensional convolution kernel and an offset in the encoding process; sigmafAdopting a Leakyrelu activation function for the encoder activation function;
4-1-2 convolution according to formula (6) to obtain a signature sequence H1Constructing a pooling layer, wherein the expression is as follows:
Zi=p(H1) (7)
p represents a pooling operation, the invention adoptsMacropooling (max-pooling); ziFeature vectors of low dimensionality, ZiDimension much smaller than Xi′;
4-2 constructing a Decoder (Decoder) part of a Convolutional Autocoder (CAE);
4-2-1 obtains a low-dimensional feature vector Z by the formula (7)iConstructing an upper sampling layer, wherein the expression is as follows:
H2=S(Zi) (8)
wherein S represents upsampling, H2Features obtained for up-sampling;
4-2-2 up-sampling operation with the formula (8) to obtain a feature vector H2Constructing an deconvolution layer, wherein the expression is as follows:
X″i=σe(U*H2+b2) (9)
in the formula, U and b2Respectively representing a one-dimensional convolution kernel and an offset in the decoding process, representing a one-dimensional convolution operation, X ″)iIs input Xi' reconstruction data, σeIs the activation function in the decoder, which is also the Leakyrelu function;
fifthly, further combining a plurality of parallel single-channel convolution self-encoder networks according to the low-dimensional deep representative feature vector Z obtained in the fourth step to construct a multi-channel convolution self-encoder feature extraction Model (MCCAE) and complete the extraction of time sequence features;
the 5-1MCCAE characteristic extraction steps are as follows:
firstly, the effective segment X at the 1 st time is divided1Effective segment X from time' to tt' simultaneously inputting into the improved convolutional self-encoder network; signal X at time 1i' Via CAE channel 1, extract a segment of feature Z1(ii) a Signal S at time 22Through the channel 2, a section of the feature Z is extracted2(ii) a And so on until the feature Z at the t-th moment is extractedt(ii) a Finally, the extracted features Z1~ZtSplicing the time sequence characteristics into a time sequence characteristic F with a time dimension according to the time sequence;
5-2 with 5-1, based on the time sequence feature extraction principle, the MCCAE unit specifically comprises the following calculation processes: the length of the valid segment X' (t) obtained in the fourth step is n, that is, X ═ X1,x2,…,xt,…,xnDetermining the number a of parallel channels, and calculating X in the first time1′,X2′,X3′,…,XaTaking out to form an a multiplied by n matrix, inputting the matrix to an MCCAE unit for convolution operation and feature extraction; the size of the extracted feature matrix F is a multiplied by m; inputting the feature matrix into LSTM unit for calculation, and outputting a predicted value pi(ii) a Taking X for the second time2′~Xa+1' input MCCAE unit to calculate, and so on, and take X each timei′~Xi+a' a line vectors are calculated and predicted, and one line is pushed down after calculation is completed until X is takenn-a′~XnWhen the time sequence characteristic matrix F is obtained, the calculation is finished, and n-a +1 time sequence characteristic matrices F are obtained;
5-3 obtaining a time sequence feature matrix F according to 5-2, wherein F is a low-dimensional feature vector Z of a adjacent momentsiCombinations of (1) thus for ZiF is normalized by normalization;
let i' th behavior Z of Zi={z1,z2,…,zt,…zmIn which Zi∈R1×m,m<n,ztRepresenting the feature value extracted by CAE at the t-th time), and Z is calculated according to the equation (10)iCarrying out maximum and minimum normalization to obtain normalized feature vectorObtaining the normalized time sequence characteristic matrix
In the formula, ZiIs the ith row feature vector, Zmin、ZmaxRespectively the minimum value and the maximum value of the feature values in the whole feature vectors,is normalized feature vector expressed as
Sixthly, constructing an LSTM prediction model;
6-1 constructing an LSTM cell by the following steps:
6-1-1 normalized time series characteristic matrix obtained by 5-3As an input, a forgetting gate layer is constructed, and the expression is as follows:
in the formula ht-1In order to output the signals at the last moment,for input at the present moment, Wf、bfRespectively an input layer weight and a bias vector; σ is sigmoid activation function, adopt, ftOutputting for the forgetting gate layer;
6-1-2 forgets gate output according to the formula (11), and simultaneously constructs an input gate layer and a tanh layer according to the formulas (12), (13) and (14), and calculates the information needing to be updated:
in the formula: σ is sigmoid function, WC、bCThe weights and bias vectors of the state update layer are separately,new candidate value calculated for tanh layer, CtFor updated candidate value, itInputting for an input gate layer;
6-1-3 constructing an output gate layer according to equation (15) and equation (16) by the updated candidate value obtained in equation (14):
ht=ot*tanh(Ct) (16)
wherein σ is sigmoid function, Wo,boWeight and offset vector, o, of the output gate layer, respectivelytFor output of the output gate layer, htIs the final determined output;
6-2 constructing an LSTM prediction model and obtaining a residual life prediction value
An LSTM unit is constructed through 6-1, and an LSTM hidden layer neuron is obtained; further, constructing three layers of LSTM networks, wherein neurons of a hidden layer are 8, 16 and 8 respectively, and finally constructing a full-connection layer according to a formula (17), wherein the number of the neurons is 1, and the activation function of each layer is tanh; simultaneously, a dropout mechanism is added behind each layer, the parameter is set to be 0.1, the prediction result is subjected to smooth filtering according to the formula (18), and the obtained final prediction life is p'i;
pi=W*ht+b (17)
In the formula (17), W and b are the weight and bias of the full connection layer, respectivelyPut in, htFor the output obtained from the last layer of LSTM, piOutputting the predicted value of the residual service life for the full connection layer;
in formula (18), p'iAnd N is the sliding window length, and the final residual life predicted value is obtained after smoothing filtering.
The service life test system of the universal circuit breaker contact system comprises an industrial personal computer, a PCL-720+ board card, an acquisition card, a Hall current sensor, a Hall voltage sensor, a vibration sensor, a solid relay group and a switching power supply; the industrial personal computer is respectively connected with the board card and the acquisition card, the board card is connected with the solid state relay group, the acquisition card is respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor, and the switching power supply is also respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor; the solid relay group is connected with an energy storage loop, a closing loop and an opening loop of the universal circuit breaker, the Hall current sensor is connected with an opening and closing accessory of the universal circuit breaker, and the Hall voltage sensor and the vibration sensor are respectively connected with a contact system of the universal circuit breaker.
The invention has the substantive characteristics that:
when the service life of the conventional universal circuit breaker is researched, a statistical data driving method is mainly adopted for modeling only on an operation accessory and a brake separating action mechanism, and the method needs to artificially select degradation indexes according to knowledge and experience, so that not only is time and energy consumed, but also the problem of inaccurate index selection is caused; secondly, as the degradation process of the universal circuit breaker is a nonlinear process, the accuracy of the statistical data driving method for fitting the service life degradation curve is not high; finally, statistical data-driven methods lack generalization capability and are not practical. Therefore, the method provided by the invention takes the defects of the statistical data driving method and the advantages of the deep learning method into consideration, and establishes a contact system service life prediction model based on the multi-channel convolution self-encoder long-term memory network (MCCAE-LSTM), so that the method provided by the invention has certain innovativeness.
Compared with the prior art, the method for predicting the residual life provided by the invention has the following remarkable improvements:
(1) the method selects the vibration signal in the action process of the universal circuit breaker to solve the problem of residual service life prediction of the contact system. The vibration signal has the advantage of non-invasive measurement, and the on-line residual mechanical life prediction of the contact system can be completed under the condition of not damaging the internal mechanical structure of the circuit breaker. In addition, the vibration signal has abundant mechanical characteristics, can effectively reflect degradation information of mechanical parts of the circuit breaker, and is favorable for completing prediction of residual mechanical life.
(2) The invention provides a concept of effective vibration signal segments by considering the particularity of a prediction object and aiming at the characteristic that a vibration signal generated by a universal circuit breaker belongs to a multi-source vibration signal. The idea is specifically to strip the vibrator event of the contact system from the overall vibration signal to achieve more accurate life prediction and improve the accuracy and effectiveness of the prediction.
(3) The invention provides a specific method for realizing automatic calibration of effective fragments. Firstly, processing an original vibration signal by a VMD method based on a cross-correlation coefficient to complete the separation of effective modal components; secondly, a double-threshold method based on short-time energy is introduced, effective modal components are combined, automatic calibration of effective segments is achieved, as can be seen from fig. 11, the original vibration signals are used as the double-threshold method to input the effective segment interval which cannot be identified, and otherwise, the effective segment interval is accurately identified through the effective modal components.
(4) The invention considers that the characteristics extracted by the single-channel CAE network lack time dimension information and cannot achieve higher precision during LSTM prediction, thereby constructing an MCCAE-LSTM life prediction model. Extracting time sequence characteristics through an MCCAE unit at the front end of the model; and at the rear end of the model, an LSTM network is constructed at the same time, and the time sequence characteristics are used as input to complete the service life prediction of the contact system.
(6) The subject currently completes data acquisition of two breaker test articles in the whole life cycle, and in order to verify the effectiveness of the method, a cross-training verification method is adopted to complete model prediction, namely, a test article 1 is used for training, a test article 2 is used for prediction, and vice versa. In order to verify the effectiveness of the method compared with the traditional method, the method is compared with CNN, LSTM and CAE-LSTM similar networks respectively, for example, FIG. 16 is a prediction result of four models of a sample 1, and FIG. 17 is a prediction result of a model of a sample 2. It can also be seen from fig. 16 and 17 that the method of the present invention can more accurately fit the life degradation curve.
(7) In order to reflect the effectiveness of the method more objectively, the method carries out error analysis and coincidence degree evaluation respectively in the stage of evaluating the prediction result. In error analysis, by introducing an average error (e)mean) And maximum error (e)max) The predicted life value and the real life value are analyzed, and taking the sample 2 as an example, the MCCAE-LSTM average error of the method provided by the invention is respectively reduced by 51.35%, 42.63% and 29.8% relative to CNN, LSTM and CAE-LSTM, and the maximum error is respectively reduced by 39.13%, 52.17% and 26.62% relative to CNN. In the evaluation of the degree of agreement, by introducing a statistical goodness-of-fit (R)2) Evaluating the model prediction result, taking the prediction result of the sample 2 as an example, R of MCCAE-LSTM2Compared with CNN, LSTM and CAE-LSTM, the service life degradation curve is improved by 16.5%, 10.1% and 4.6%, which shows that the method provided by the invention can fit the service life degradation curve more accurately, thereby improving the prediction accuracy. Through error analysis and coincidence degree evaluation, the MCCAE-LSTM prediction result can reduce prediction errors and improve the coincidence degree of prediction curves, and the effectiveness of the method provided by the invention is verified visually.
(8) The method solves the key problems of the conventional circuit breaker contact system prediction and health management technology, provides technical guidance for the circuit breaker maintenance according to the situation, promotes the change of the circuit breaker maintenance technology from the timing inspection to the maintenance technology according to the situation, improves the reliability of the circuit breaker, and greatly reduces the daily maintenance cost due to the safety and the utilization rate.
Drawings
FIG. 1 is a diagram of the general scheme of the remaining mechanical life of a conventional circuit breaker contact system in an embodiment;
FIG. 2 is a schematic diagram of a circuit breaker mechanical life testing system according to an embodiment;
FIG. 3 is a diagram of a CAE basic structure in an embodiment;
FIG. 4 is a diagram of a method for extracting timing characteristics according to an embodiment;
FIG. 5 is a diagram of an MCCAE unit in accordance with an embodiment;
FIG. 6 is a diagram of CAE network basic parameters in an embodiment;
FIG. 7 is the basic structure diagram of the LSTM network in the embodiment
Fig. 8 is a diagram of VMD decomposition when K is 6 and K is 5 in example verification;
FIG. 9 is a graph comparing an original closing vibration signal with an effective modal component in an example verification; wherein, fig. 9a is an original closing vibration signal diagram; FIG. 9b is a graph of effective modal components;
FIG. 10 is a graph of the effective components of the closing vibration signal and its short-term energy in case verification; wherein, fig. 10a is a diagram of effective components of a closing vibration signal; FIG. 10b is a short time energy diagram thereof;
fig. 11 is a diagram of end point identification of effective segments by a short-time energy-based double-threshold method in example verification, wherein, in fig. 11(a), when an input signal is an original vibration signal, end point identification is performed on the effective segments by the short-time energy-based double-threshold method; fig. 11(b) illustrates the end point identification of the effective segment based on the short-time energy dual-threshold method when the input signal is the effective modal component;
FIG. 12 is a graph comparing a reconstructed signal with an original signal in an example verification; wherein, fig. 12a is a reconstructed signal diagram; FIG. 12b is a graph of raw signals;
FIG. 13 is a graph of the comparison of the original signal, features, at time 250 in an example validation; wherein, fig. 13a is the 250 th input original signal; FIG. 13b is a graph of timing characteristics extracted after the 250 th input;
FIG. 14 is a 1750 th original signal, feature comparison graph in example verification; wherein, fig. 14a is a diagram of the original signal inputted at the 1750 th time; FIG. 14b is a timing diagram extracted after the 1750 th input;
FIG. 15 is a 3500 th primary signal, feature comparison plot in an example validation; wherein, fig. 15a is a diagram of the original signal inputted at the 3500 th time; FIG. 15b is a graph of timing characteristics extracted after the 3500 th input;
FIG. 16 is a graph of the predicted results of four models of sample 1 in example validation; wherein, fig. 16(a) is a CNN prediction result graph; FIG. 16(b) is a graph showing the predicted result of LSTM; FIG. 16(c) is a diagram showing the CAE-LSTM prediction result; FIG. 16(d) is a graph showing MCCAE-LSTM prediction results;
FIG. 17 is a graph of the predicted results of four models of sample 2 in example validation; wherein, 17(a) is a CNN prediction result graph; FIG. 17(b) is a graph showing the predicted result of LSTM; FIG. 17(c) is a diagram showing the CAE-LSTM prediction result; FIG. 17(d) is a graph showing the prediction results of MCCAE-LSTM.
Detailed Description
In the prior art, when the service life prediction research is carried out on the switching-on and switching-off accessory, the adopted research signal is a switching-on and switching-off accessory current signal, and the adopted method is a wiener process in a statistical data driving method.
When the service life prediction research is carried out on the brake-separating action mechanism, the adopted research signal is a brake-separating vibration signal, and the adopted method is a VMD method for carrying out original signal denoising, a short-time energy-based double-threshold method for extracting characteristic parameters and wiener process modeling for carrying out service life prediction.
The universal circuit breaker mainly has three processes of energy storage, closing and opening in the operation process, and the mechanical actions generated in different processes are greatly different, so that the generated related signals are also greatly different. The opening and closing vibration signals are mainly generated by the action of the opening accessories and the impact between the opening buffers, and the closing vibration signals are mainly generated by the action of the closing accessories and the impact between the contact systems, so that the opening and closing vibration signals are different, and the generation mechanisms of the opening and closing vibration signals are different. In addition, the closing vibration signal is more closely related to the mechanical characteristics of the contact system.
The invention selects the closing vibration signal generated by the mechanical characteristics of the contact system for research, and introduces the VMD method to select effective components, which is different from denoising. The initial purpose of the method is that the vibration signals are multi-source signals, and the frequencies of the vibration signals generated by different vibration sources are different, so the invention considers that the VMD method is introduced to separate the natural modal components, namely, to separate the modal signals generated by different vibration sources, and the mode related to the impact of the contact system is found according to the cross correlation coefficient method, and is taken as the effective modal component.
The embodiment shown in fig. 1 shows that the general flow of the method for predicting the residual life of the contact system of the universal circuit breaker based on deep learning provided by the invention is as follows: collecting vibration signals, contact state signals and accessory current signals in the action process of the universal circuit breaker in a full life cycle → taking the vibration signals as input, introducing a VMD method based on cross-correlation coefficients to extract effective modal components → combining the effective modal components, completing the calibration of effective segments according to a double-threshold method based on short-time energy → constructing an MCCAE unit, taking the effective segments as input, introducing an unsupervised learning training mode, extracting deep degradation time sequence characteristics of the signals → constructing an LSTM unit at the rear end, adopting a supervised learning training mode, and taking the time sequence characteristics as input to complete the life prediction of a contact system.
The structural schematic diagram of the universal circuit breaker contact system life test system for executing the universal circuit breaker contact system residual life prediction method based on deep learning is shown in fig. 2, the mechanical life of the circuit breaker is mechanical wear resistance, and is represented by the number of unloaded operation cycles of normal operation of the circuit breaker before maintenance or replacement of mechanical parts is needed. The invention takes the DW15-1600 circuit breaker as a test sample and constructs a mechanical life experiment system of the circuit breaker. The system mainly comprises a motion control part, a signal detection part, a data acquisition part and the like,
the service life test system of the universal circuit breaker contact system comprises an industrial personal computer, a PCL-720+ board card, an acquisition card, a Hall current sensor, a Hall voltage sensor, a vibration sensor, a solid relay group and a switching power supply; the industrial personal computer is respectively connected with the board card and the acquisition card, the board card is connected with the solid state relay group, the acquisition card is respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor, and the switching power supply is also respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor; the solid state relay group is connected with an energy storage, closing and opening loop of the universal circuit breaker, the Hall current sensor is connected with a switching-on and closing accessory of the universal circuit breaker, and the Hall voltage sensor and the vibration sensor are respectively connected with a contact system of the universal circuit breaker;
the board card is a PCL-720+ board card; the acquisition card is a high-speed acquisition card USB-7648A; the vibration sensor is an LCO159 vibration sensor; the solid state relay group is an ACSSR solid state relay group;
in the control part, an industrial personal computer controls a PCL-720+ board card to send out energy storage, closing and opening instructions through an ISA bus, and each instruction directly operates a solid state relay set to control the energy storage, closing, opening and undervoltage processes of a circuit breaker; in the detection part, a high-precision wide-range LCO159 vibration sensor is adopted to measure vibration signals in the action process, the sensor is arranged on a beam, the vibration signals in the action process can be fully collected, and the current of a coil of a switching-on and switching-off accessory and the state voltage of a contact are measured through a Hall current sensor and a voltage sensor; in the data acquisition part, the measured vibration signal, current signal and voltage signal are acquired by a high-speed acquisition card USB-7648A at a sampling frequency of 20 kHz.
The acquired vibration signals are processed by MATLAB software to obtain effective vibration segments; and then the residual service life of the contact system is predicted by Pycharm software processing.
The invention introduces a double-threshold method based on short-time energy to determine the effective segment interval, which is different from the method for extracting the characteristic parameters of the circuit breaker, and the invention respectively takes effective modal components and original signals as input, so that the effectiveness of the method provided by the invention can be clearly seen in a figure.
The present invention is further described with reference to the following drawings and examples, but the scope of the claims of the present invention is not limited thereto.
The invention provides a universal circuit breaker contact system residual life prediction method (a prediction method for short) based on deep learning, which comprises the following steps:
the first step is as follows: a mechanical life experiment system is constructed by taking the universal circuit breaker as an object, and a vibration signal X (t) ═ x in the action process of the universal circuit breaker in a whole life cycle is collected1,…,xi,…,x3000) (where N is 3000, i.e. the number of sampling points is 3000, x)i-x30003000 sampling points), contact status signal and on-off brake attachmentAnd (4) carrying out current signal and carrying out the life prediction research of the contact system according to the vibration signal. (the sampling points refer to the sampling amplitudes of the vibration signals at different moments, namely the vibration signals at different times)
Aiming at the service life problem of the contact system, the characteristics that the vibration signal has non-invasion and online measurement are considered, and the mechanical characteristic change of the contact system is closely related to the vibration signal generated in the action process, so the service life prediction research of the contact system is carried out by detecting the vibration signal.
Secondly, performing VMD decomposition on the closing original vibration signal X (t) to extract effective modal components; the method comprises the following specific steps:
2-1, initializing the number K of pre-decomposition modal components to be 2, and performing VMD decomposition on the brake original vibration signal X (t) to obtain K inherent modal components (IMFs);
2-2 respectively obtaining the cross-correlation coefficients rho of K IMFs and the original vibration signal X (t)kCalculating formula as formula (1), and calculating difference c between adjacent mode and original signal cross-correlation coefficient according to formula (2)k-1Obtaining K cross correlation coefficients rhokAnd K-1 cross correlation coefficient differences ck-1;
ck-1=ρk-ρk-1 (2)
In the formula (1), N is the number of sampling points, xn、Is a switching-on vibration signal X (t) sampling point and its mean value, ukn、The Kth IMF component element and the mean value thereof; in the formula (2), ρk,ρk-1The cross-correlation coefficients of the kth and kth-1 IMF components with X (t), respectively;
2-3 presetting a poleThe small threshold c is Δ, Δ is preset to 0.02. If obtained ciIf the value is larger than a preset threshold value c, K is equal to K +1, and the operation is repeated for 2-1 and 2-2 in a circulating mode; if obtained ciIf the difference is smaller than a preset threshold value c, the two modes are considered to have high similarity, namely, an over-decomposition phenomenon occurs, and if the circulation is stopped, K is equal to K-1;
2-4, after the K value is determined, selecting the IMF with the maximum correlation coefficient under the current K value as an effective modal component u' (t), and regarding the modal signal as the most relevant to the original signal as the main characteristic information for representing the original signal;
the specific algorithm steps of the VMD decomposition are calculated and analyzed according to the method steps disclosed by Dragomirksky K and the like (Dragomirksky D. spatial mode decomposition [ J ]. IEEE Transactions on Signal Processing,2014,62(3): 531-544);
thirdly, combining the effective modal component u' (t) to complete automatic calibration of the effective segment; the method comprises the following steps:
3-1 framing the active mode component u' (t), as in equation (3),
zi(n)=ω(n)*u′((i-1)*inc+n) (3)
in the formula, ω (n) is a window function; z is a radical ofi(n) is a frame number, n is 1,2, …, L.i is 1,2, …, fnL is the frame length; inc is the frame shift length; f. ofnThe total frame number after framing; wherein omega (n) is selected from a Hainin window as shown in formula (4),
3-2 calculating z of the i frame effective modal componenti(n) the short-time energy formula is as follows:
in the formula, zi(n) is the value of a frame of data, E (i) is the short-time energy of a frame of data;
3-3, completing effective segment calibration by a double-threshold method:
obtaining a short-time energy value E (i) of the effective modal component through 3-2, and further drawing a short-time energy waveform; setting two thresholds Q according to the wave crest and the wave trough of the drawn short-time energy waveform1And Q2,Q1<Q2Wherein Q is1The short-time energy value corresponding to the wave trough is set to be 0.01; q2The short-time energy value corresponding to the second vibration peak value is set to be 0.13; when the energy value E (i) is higher than Q2The signal segment interval of the contact impact is determined, and the energy value of the frame data is higher or lower than Q1To determine the specific end point E of the effective segment1,E2,E1,E2The original vibration signal in between is an effective segment X' (t), and the length of the effective segment after the implementation is 400;
and fourthly, combining the convolutional neural network with the encoder by the convolutional self-encoder, and the convolutional self-encoder is mainly different from a common encoder in that convolutional and pooling operations are introduced into an encoder, and deconvolution and anti-pooling operations are introduced into a decoder. The basic structure is shown in figure 3.
Establishing a single-channel convolution self-encoder (CAE) feature extraction model by taking a contact system mechanical life effective vibration segment X' (t) as input; the method comprises the following specific steps:
4-1 constructing an Encoder (Encoder) part of a Convolutional Auto Encoder (CAE);
4-1-1 is provided with an effective segment X' (t) action X of ithi′={x1,x2,…,xt,…,x400}(xtCharacteristic value indicating time t) by Xi' As an input, a convolutional layer is constructed, as in formula (6):
H1=σf(W*Xi′+b1) (6)
in the formula, Xi′∈R1×nIs the effective fragment sequence of the ith row, and n is the length of the effective fragment sequence; denotes a one-dimensional convolution operation; w and b1Respectively representing a one-dimensional convolution kernel and an offset in the encoding process; sigmafAdopting a Leakyrelu activation function for the encoder activation function;
4-1-2 convolution according to formula (6) to obtain a signature sequence H1Constructing a pooling layer, wherein the expression is as follows:
Zi=p(H1) (7)
p represents pooling operation, and the invention adopts max-pooling; ziFeature vectors of low dimensionality, ZiDimension much smaller than Xi' the extraction characteristic length after the concrete implementation is 50;
4-2 constructing a Decoder (Decoder) part of a Convolutional Autocoder (CAE);
4-2-1 obtains a low-dimensional feature vector Z by the formula (7)iConstructing an upper sampling layer, wherein the expression is as follows:
H2=S(Zi) (8)
wherein S represents upsampling, H2Features obtained for up-sampling;
4-2-2 up-sampling operation with the formula (8) to obtain a feature vector H2Constructing an deconvolution layer, wherein the expression is as follows:
X″i=σe(U*H2+b2) (9)
in the formula, U and b2Respectively representing a one-dimensional convolution kernel and an offset in the decoding process, representing a one-dimensional convolution operation, X ″)iIs input Xi' reconstruction data, σeIs the activation function in the decoder, which is also the Leakyrelu function;
4-3 training a feature extraction model of the convolutional self-encoder, wherein an optimizer is Adadelta, and a loss function is continuously adjusted through a back propagation algorithm to optimize parameters of a one-dimensional convolutional self-encoder (CAE), and is expressed as a formula (10):
in the formula, LrIs Mean Square Error (MSE), X ″iTo reconstruct the valid fragment sequence feature values, Xi' effective segment original input data.
Fifthly, further combining a plurality of parallel single-channel convolution self-encoder networks according to the low-dimensional deep representative feature vector Z obtained in the fourth step to construct a multi-channel convolution self-encoder feature extraction Model (MCCAE) and complete the extraction of time sequence features;
5-1MCCAE time sequence feature extraction is shown in figure 4, and the specific steps are as follows:
firstly, the effective segment X at the 1 st time is divided1Effective segment X from time' to tt' simultaneously inputting into the improved convolutional self-encoder network; signal X at time 11' Via CAE channel 1, extract a segment of feature Z1(ii) a Signal S at time 22Through the channel 2, a section of the feature Z is extracted2(ii) a And so on until the feature Z at the t-th moment is extractedt(ii) a Finally, the extracted features Z1~ZtSplicing the time sequence characteristics into a time sequence characteristic F with a time dimension according to the time sequence;
5-2 is based on the 5-1 time sequence feature extraction principle, and the MCCAE unit specifically calculates the process as follows. The number of parallel channels of the MCCAE unit is different, and here, 5 channels are taken as an example, and the structure is shown in fig. 5. The length of the valid segment X' (t) obtained in the fourth step is 400, that is, X ═ X1,x2,…xt,…x400}. For the first calculation, X is added1′,X2′,X3′,X4′,X5' take out, form a 5 x 400 matrix, input to an MCCAE unit for convolution operation and feature extraction. The extracted feature matrix F has a size of 5 × 50. Inputting the feature matrix into LSTM unit for calculation, and outputting a predicted value pi. Taking X for the second time2′~X6' input MCCAE unit to calculate, and so on, and take X each timei′~Xi+4' five rows of vectors goLine calculation and prediction, wherein, after calculation is completed, one line is pushed down until X is takenn-4′~XnWhen the time sequence characteristic matrixes are obtained, the calculation is finished, and n-5+1 time sequence characteristic matrixes F are obtained.
5-3MCCAE parameter settings. As shown in fig. 6, the parameters of the CAE network of a certain channel in the multiple channels are set as follows, and the other channels of the MCCAE network are consistent with the setting types of the parameters of the channel. In fig. 6, COV represents a convolutional layer, MP represents a maximum pooling layer, UP represents an upsampling layer, and the parameter setting principle is the number of convolutional kernels × the size of the convolutional kernels × the convolution step length, and a feature fusion means is introduced to fuse 16 × 50 multidimensional features into a 1 × 50 feature sequence.
5-4 obtaining a time sequence feature matrix F according to 5-2, wherein F is a low-dimensional feature vector Z of 5 adjacent momentsiCombinations of (1) thus for ZiF is normalized by normalization;
let i' th behavior Z of Zi={z1,z2,…,zt,…z50Z according to formula (11)iCarrying out maximum and minimum normalization to obtain normalized feature vectorObtaining the normalized time sequence characteristic matrix
In the formula, ZiIs the ith row feature vector, Zmin、ZmaxRespectively the minimum value and the maximum value of the feature values in the whole feature vectors,is normalized feature vector expressed as
Sixthly, constructing an LSTM prediction model;
the LSTM recurrent neural network is also formed by recursively connecting a plurality of recurrent units, and has a structure shown in fig. 7, in which the recurrent unit of the LSTM includes three gate structures, namely a forgetting gate, an input gate, and an output gate. The forgetting gate is used for screening and leaving important information in the long-term memory state and is a structure for updating the long-term memory state; the input gate may determine which information to use to update the long-term memory state, a structure used to calculate the state of the candidate cell; the output gate is the structure used to calculate the short-term cell memory state, which is the summary of the previous stage work and determines the output information. The RUL prediction problem of the contact system of the circuit breaker belongs to the time sequence problem, vibration signals at adjacent moments are related, and the LSTM is used as a rear-end RUL prediction model in the invention because the three gate control structures in the LSTM can learn the related information of the vibration signals generated in the degradation process of the contact system of the circuit breaker.
6-1 constructing an LSTM cell by the following steps:
6-1-1 normalized time series characteristic matrix obtained by 5-3As an input, a forgetting gate layer is constructed, and the expression is as follows:
in the formula ht-1In order to output the signals at the last moment,for input at the present moment, Wf、bfRespectively an input layer weight and a bias vector; σ is sigmoid activation function, adopt, ftOutputting for the forgetting gate layer;
6-1-2 forgets gate output according to the formula (12), and simultaneously constructs an input gate layer and a tanh layer according to the formulas (13), (14) and (15), and calculates the information needing to be updated:
in the formula: σ is sigmoid function, WC、bCThe weights and bias vectors of the state update layer are separately,new candidate value calculated for tanh layer, CtFor updated candidate value, itInputting for an input gate layer;
6-1-3 constructing an output gate layer according to equation (16) and equation (17) by the updated candidate value obtained in equation (15):
ht=ot*tanh(Ct) (17)
wherein σ is sigmoid function, Wo,boWeight and offset vector, o, of the output gate layer, respectivelytFor output of the output gate layer, htIs the final determined output;
6-2, constructing an LSTM prediction model;
an LSTM unit is constructed through 6-1, and an LSTM hidden layer neuron is obtained; further, constructing three layers of LSTM networks, wherein neurons of a hidden layer are 8, 16 and 8 respectively, and finally constructing a full-connection layer according to a formula (18), wherein the number of the neurons is 1, and the activation function of each layer is tanh; and simultaneously, adding a dropout mechanism behind each layer, setting parameters to be 0.1, and performing smooth filtering on a prediction result.
pi=W*ht+b (18)
Wherein W and b are weight and offset of full connection layer, htFor the output obtained from the last layer of LSTM, piOutputting the predicted value of the residual service life for the full connection layer;
6-3 training an LSTM prediction model;
the low-dimensional deep layer representative features extracted by CAE and normalized by the maximum and minimum values are obtained by the step 5-3And obtaining a normalized time sequence characteristic matrixIs easy to know Zi *Is a 1 × 50 feature vector, Fi *A 5 × 50 timing feature matrix; a supervised learning training mode is adopted, and the specific training steps are as follows:
6-3-1 sets a normalized remaining life label y per lineiAs equation (19), then each timing feature matrix Fi *Y of the last row insideiA lifetime label as the timing characteristic;
wherein i represents the number of rows of the line, and n represents the total number of rows of the sample;
6-3-2 obtaining the predicted Life value p from the equation (18)iEquation (19) yields the true lifetime value yiCalculating a predicted lifetime value p according to equation (20)iAnd the real life value yiThe optimizer adopts Adam to optimize the error between the predicted value and the true value in a back propagation gradient descent mode;
in the formula, n represents a sampleTotal number of rows, yiIs the true residual life value, piPredicting a residual life value, wherein MSE is mean square error;
and 6-3-3 inputting a training sample to the model, continuously optimizing the error between the predicted value and the true value according to 6-3-2, and if the errors of the five continuous rounds do not change obviously, determining that the LSTM model is converged. The model is brought into a test sample after being converged to obtain a predicted value p of the residual life of the contact systemi. Smoothing the prediction result according to the formula (21), wherein N is 11, and the obtained final predicted life is p'i
In formula (II) p'iAnd N is the sliding window length, and the final residual life predicted value is obtained after smoothing filtering.
Example 1
The contact system mounted on the conventional circuit breaker model DW15-1600 was used as a test object in this example. The contact system is mainly composed of a main contact and an arc contact as a key actuating mechanism of the opening and closing of the circuit breaker, the closing sequence of the contact system is from arc to arc, the effective mechanical service life vibration signal segment of the contact system is obtained in the specific implementation mode, and the service life prediction model is constructed by taking the effective mechanical service life vibration signal segment as input. This section verifies the effectiveness of the above theory.
The method for predicting the residual life of the contact system of the universal circuit breaker based on deep learning is adopted to predict the residual life of the contact system of the universal circuit breaker, and the method comprises the following specific steps:
first, data are collected. The method comprises the steps of collecting vibration signals, contact state signals and closing accessory current signals in a closing process by using a universal circuit breaker contact system life test system, setting the test operation frequency to be 20 times/h according to relevant standards of a circuit breaker, setting the sampling frequency of a data acquisition card to be 20kHz, and setting the sampling time to be 150ms, so that 3000 sampling points are obtained.
And secondly, sorting data. At present, according to the experimental system, signal acquisition and test article of two universal circuit breakers in the whole life cycle are completed4440 pieces of valid data are collected for 1, and 4450 pieces of valid data are collected for 2. Taking sample 1 as an example, the sample data set is sorted into 4440 rows, where the row number is equal to the number of times of collection, and the column number is equal to the length of data collected once, and the length of each row is 3000. Assigning a lifetime label y to each row of dataiAnd the residual service life of the circuit breaker corresponding to the row of data is shown. By the method, the neural network can be conveniently solved by using a gradient descent algorithm during optimization, and the prediction precision of the residual service life can be improved. The expression of the label is shown as (1)
(1) Where i represents the number of rows in the row and n represents the total number of rows.
During training, training the model by adopting an alternating training method, and when the data of the test sample 1 is used for training, the data of the test sample 2 is used for testing; when the data of the sample 2 is used for training, the data of the sample 1 is used for prediction, so that the validity of the MCCAE-LSTM model is verified
And thirdly, extracting the VMD effective component. According to the second step of the embodiment, the invention presets the range of K to be 3-8 and the threshold c to be 0.02, and by decomposing the closing vibration signal, the cross-correlation coefficient between each mode and the original signal is obtained as shown in table 1:
TABLE 1 Cross-correlation coefficient Table
The absolute value of the difference between the cross-correlation coefficients of adjacent modes is further found, as shown in table 2:
TABLE 2 mutual coefficient difference table
As can be seen from tables 1 and 2, when K is 6, the cross-correlation coefficients of IMF2 and IMF3 with the original signal are both 0.29, and the difference between the two is 0, which indicates that the two are highly similar to the original signal, and the mode aliasing phenomenon has occurred, so K is selected to be 5.
As shown in fig. 8, when K is 6 and K is 5, the original closing vibration signal is decomposed. At K6, IMF2 is seen to have high similarity with IMF3 signal, modal aliasing has occurred, and at K5, the modal signals are greatly different and have prominent self-characteristics, which further verifies the accuracy of the second step theory of the specific embodiment.
Further, when K is 5, the cross correlation coefficient between the IMF3 component and the original signal is the largest, and according to the second step of the embodiment, the cross correlation coefficient is directly selected as the effective component, as shown in fig. 9, which is a comparison graph of the IMF3 component and the original signal, it can be found by observation that the IMF3 component highlights the energy peak value of the circuit breaker contact system at the time of impact, which indicates that the IMF3 component is the main mode at the time of contact impact and is the effective mode component at the time of contact impact of the original signal.
And fourthly, completing effective segment calibration based on a short-time energy double-threshold method. According to the third step of the method in the specific implementation manner, in order to better reflect the non-stationary characteristic of the vibration signal, the window length is selected to be 10, the frame shift is selected to be 5, and as shown in fig. 10, the window length is a graph of the effective component of the closing vibration signal and the short-time energy thereof, and it can be known from the graph that the mutation point is effectively enhanced after the effective component is processed by the short-time energy method, which is more beneficial to the calibration of the effective segment by the double-threshold method.
As shown in fig. 11, effective segments are respectively calibrated for the original vibration signal and the effective component signal by a short-time energy-based dual-threshold method, and it can be known from fig. 11 that an effective segment interval cannot be effectively detected for the original vibration signal, otherwise, the effective interval is calibrated by the effective component. The starting threshold of the double thresholds can accurately identify the descending moment of the contact state signal, namely the moment when the moving arc contact is just contacted with the static arc contact; and as the energy value of the vibration signal is attenuated, at the termination threshold, the energy value is already attenuated to a minimum value point, and the signal area in the visible interval represents the energy reflected by the contact impact, that is, the signal area E in embodiment 3-31,E2The signal in between, i.e. the effective segment。
And fifthly, extracting the time sequence characteristics of the MCCAE unit. In order to ensure the integrity of the features, the invention directly uses the effective segment of the original vibration signal as the input of the MCCAE to perform feature extraction, as shown in fig. 12, which is a reconstructed signal and an original signal, it can be seen that the original signal is substantially restored, and the extracted low-dimensional features have certain representativeness. Taking the extracted characteristics of the sample 1 as an example, the change of the mechanical property of the contact system of the circuit breaker in different periods is explained.
Considering that the break-in process exists in the early period and the signal fluctuation is large, the effective segment from the 250 th round for five times is selected as the input of the MCCAE. Fig. 13 shows the valid segment of the original signal and the signal characteristics. As can be seen from fig. 13, the signal is relatively stable in the early stage, and the peak value of the vibration signal is relatively concentrated, which indicates that the contact system of the circuit breaker is not degraded at this time, and is at the initial stage of degradation of mechanical properties.
The middle signal has mechanical wear, and the mechanical wear process has a plateau period and a period with obvious degradation. To highlight the trend of the features, the 1750 th group of signals was selected as the input of the MCCAE. As shown in fig. 14, which is a 1750 th time effective segment of the original signal, signal characteristics, and as can be seen from fig. 14, the signal has changed greatly in the middle period, the peak value of the vibration signal is reduced to a certain extent in the earlier period, the contact system is degraded, and more serious mechanical wear is generated.
The contact is slightly deformed to a certain degree due to erosion of mechanical abrasion to contact materials in the later period, and the contact rebounds due to the change of the shape of the contact. Comprehensively, the 3500 th signal is selected as the input of the MCCAE, and fig. 15 shows the effective segment and signal characteristics of the original signal. As can be seen from fig. 15, the vibration signal becomes dense, indicating that the inter-contact bounce phenomenon is severe and the mechanical structure of the contact system has degraded to an end.
The section carries out time sequence feature extraction on the effective segments, and as can be seen from fig. 13, 14 and 15, according to the effective segments and feature extraction comparison thereof, the defined effective segments and the extracted time sequence features can reflect the mechanical states of the contact system in different periods, which also verifies the effectiveness of the method provided by the invention. In addition, the features extracted by the MCCAE have not only a certain time sequence in the macro but also contain richer degradation information in the micro.
And sixthly, predicting a result and evaluating the performance. The MCCAE-LSTM model is mainly composed of an improved CAE network and an LSTM network, and because the CAE network mainly performs feature extraction through CNN, the model is compared with a single CNN network and a single LSTM network in order to verify the effectiveness of the model, and meanwhile, in order to verify the advantages of the MCCAE unit relative to single-channel CAE, the model is compared with the CAE-LSTM network, and it needs to be explained that the compared method is more commonly used in the field of life prediction. The input and output parameters are compared for four models, as shown in table 3.
TABLE 3 comparison of input and output parameters of four models
And (4) adopting an alternate training method for verification, wherein when the test product 1 is a training set, the test product 2 is a testing set, and vice versa. And (4) respectively training and predicting the four models. The results of prediction of sample 1 are shown in fig. 16, and the results of prediction of sample 2 are shown in fig. 17. In the figure, the abscissa represents the number of times, the ordinate represents the ratio of the remaining life to the total life, the predicted remaining life is represented by a dotted line, and the actual life is represented by a solid line.
As can be seen from fig. 16 and 17, the MCCAE-LSTM prediction method has improved accuracy compared with other three prediction models widely and mature in the life prediction field, and the prediction curve is closer to the actual life curve. In order to evaluate the prediction errors of the four models more intuitively, average errors and maximum errors are introduced to respectively calculate the four models, and the formulas of the average errors and the maximum errors are as shown in formula (2):
pito predict the RUL value, yiIs the true RUL value.
TABLE 4 mean error and maximum error for the four models
As can be seen from table 4: by combining error analysis, the method of the invention is superior to the other three methods, taking sample 2 as an example, the average error is respectively reduced by 51.35%, 42.63% and 29.8% relative to CNN, LSTM and CAE-LSTM, and the maximum error is respectively reduced by 39.13%, 52.17% and 26.62% compared with CNN. This is because MCCAE-LSTM introduces a feature with temporal order, thereby reducing prediction error.
In order to further evaluate the degree of coincidence between the residual life prediction result and the actual life, a goodness of fit R in statistics is introduced2To further evaluate the predicted result, the calculation formula (3) is as follows:
wherein SStotIs a sum of squares, SSregTo return the sum of squares, SSresAs a sum of squared residuals, goodness of fit R2The value range of (1) is 0-1, and the value size can reflect the curve fit degree. The goodness of fit of the four model predictions is shown in table 5:
TABLE 5 goodness of fit of predicted results to true results for the four models
As can be seen from Table 5, the goodness of fit of the proposed model is greatly improved compared with the original model. Taking sample 2 as an example, R of MCCAE-LSTM2Compared with CNN, LSTM and CAE-LSTM, the content of the additive is respectively increased by 16.5%, 10.1% and 4.6%. R of predicted results of two samples2The average improvement is 10%. Further, the method can accurately fit the life degradation process and accurately complete the residual life prediction.
As can be seen from table 6, although the proposed model has a certain increase in training and prediction time compared to other methods due to the increase in complexity, it still can meet the requirements of practical engineering.
TABLE 6 training and prediction times for the four models
The steps are realized by adopting LabVIEW, MATLAB and Pycharm software.
The LabVIEW, MATLAB and Pycharm software used in the above examples is well known to those skilled in the art.
The percentages in the above examples are all numerical percentages.
Compared with the prior art, the method for predicting the residual service life of the universal circuit breaker contact system based on deep learning provided by the invention is provided. Firstly, considering the particularity of a prediction object and the fact that a vibration signal generated by a circuit breaker belongs to a multi-source vibration signal, a concept of an effective vibration fragment of a mechanical life is provided; secondly, because the vibration signal belongs to a non-stationary signal and is difficult to carry out data preprocessing by a single algorithm, the invention introduces a VMD method based on cross-correlation coefficients to obtain effective modal components; thirdly, the effective component is used as input, and the effective segment calibration is automatically completed by combining a short-time energy-based double-threshold method; and finally, constructing a hybrid prediction model, wherein the model is divided into two parts, an MCCAE characteristic extraction model is constructed at the front end by considering time sequence information among signals, and an LSTM network is introduced at the rear end to complete the service life prediction of the contact system. Experimental results show that the method has high prediction precision and is suitable for predicting the residual life of the contact system of the universal circuit breaker.
Nothing in this specification is said to apply to the prior art.
Claims (6)
1. A method for predicting the residual mechanical life of a conventional breaker contact system based on deep learning is characterized by comprising the following steps:
firstly, a service life test system of a contact system of the universal circuit breaker is used for measuring a closing vibration signal X (t) ═ x in the action process of the universal circuit breaker1,…,xi,…,xN) Where N is the number of sampling points, xi-xNN sampling points;
secondly, performing VMD decomposition on the closing original vibration signal X (t) to extract effective modal components;
thirdly, combining the effective modal component u' (t) to complete automatic calibration of the effective segment; the method comprises the following steps:
3-1 framing the active mode component u' (t) according to equation (3),
zi(n)=ω(n)*u′((i-1)*inc+n) (3)
in the formula, ω (n) is a window function; z is a radical ofi(n) is a frame number, n is 1,2, …, L.i is 1,2, …, fnL is the frame length; inc is the frame shift length; f. ofnThe total frame number after framing; wherein omega (n) is selected from a Hainin window according to formula (4),
3-2 calculating z of the i frame effective modal componenti(n) short-time energy formula according to formula (5):
in the formula, zi(n) is the value of a frame of data, E (i) is the short-time energy of a frame of data;
3-3, completing effective segment calibration by a double-threshold method;
fourthly, taking the effective vibration segment X' (t) of the mechanical life of the contact system as input, and establishing a single-channel convolution self-encoder (CAE) feature extraction model; the method comprises the following specific steps:
4-1 constructing an Encoder (Encoder) part of a Convolutional Auto Encoder (CAE);
4-1-1 is provided with an effective segment X' (t) action X of ithi′={x1,x2,…,xt,…,xn}(xtCharacteristic value indicating time t) by Xi' for input, build a convolutional layer, according to equation (6):
H1=σf(W*X′i+b1) (6)
in formula (II), X'i∈R1×nIs the effective fragment sequence of the ith row, and n is the length of the effective fragment sequence; denotes a one-dimensional convolution operation; w and b1Respectively representing a one-dimensional convolution kernel and an offset in the encoding process; sigmafAdopting a Leakyrelu activation function for the encoder activation function;
4-1-2 convolution according to formula (6) to obtain a signature sequence H1Constructing a pooling layer, wherein the expression is as follows:
Zi=p(H1) (7)
p represents pooling operation, and the invention adopts max-pooling; ziFeature vectors of low dimensionality, ZiDimension much smaller than Xi′;
4-2 constructing a Decoder (Decoder) part of a Convolutional Autocoder (CAE);
4-2-1 obtains a low-dimensional feature vector Z by the formula (7)iConstructing an upper sampling layer, wherein the expression is as follows:
H2=S(Zi) (8)
wherein S represents upsampling, H2Features obtained for up-sampling;
4-2-2 up-sampling operation with the formula (8) to obtain a feature vector H2Constructing an deconvolution layer, wherein the expression is as follows:
X″i=σe(U*H2+b2) (9)
in the formula, U and b2Respectively representing a one-dimensional convolution kernel and an offset in the decoding process, representing a one-dimensional convolution operation, X ″)iIs input Xi' reconstruction data, σeIs an activation function in a decoderThe number is the Leakyrelu function;
fifthly, further combining a plurality of parallel single-channel convolution self-encoder networks according to the low-dimensional deep representative feature vector Z obtained in the fourth step to construct a multi-channel convolution self-encoder feature extraction Model (MCCAE) and complete the extraction of time sequence features;
sixthly, constructing an LSTM prediction model;
6-1 constructing an LSTM cell by the following steps:
6-1-1 normalized time series characteristic matrix obtained by 5-3As an input, a forgetting gate layer is constructed, and the expression is as follows:
in the formula ht-1In order to output the signals at the last moment,for input at the present moment, Wf、bfRespectively an input layer weight and a bias vector; σ is sigmoid activation function, adopt, ftOutputting for the forgetting gate layer;
6-1-2 forgets gate output according to the formula (11), and simultaneously constructs an input gate layer and a tanh layer according to the formulas (12), (13) and (14), and calculates the information needing to be updated:
in the formula: σ is sigmoid function, WC、bCThe weights and bias vectors of the state update layer are separately,new candidate value calculated for tanh layer, CtFor updated candidate value, itInputting for an input gate layer;
6-1-3 constructing an output gate layer according to equation (15) and equation (16) by the updated candidate value obtained in equation (14):
ht=ot*tanh(Ct) (16)
wherein σ is sigmoid function, Wo,boWeight and offset vector, o, of the output gate layer, respectivelytFor output of the output gate layer, htIs the final determined output;
6-2, constructing an LSTM prediction model and obtaining a predicted value of the residual life.
2. The method for predicting the residual mechanical life of the contact system of the universal circuit breaker based on deep learning as claimed in claim 1, characterized in that VMD decomposition is carried out on a brake original vibration signal X (t) to extract an effective modal component; the method comprises the following specific steps:
2-1, initializing the number K of pre-decomposition modal components to be 2, and performing VMD decomposition on the brake original vibration signal X (t) to obtain K inherent modal components (IMFs);
2-2 respectively obtaining the cross-correlation coefficients rho of K IMFs and the original vibration signal X (t)kCalculating the difference c between adjacent mode and original signal cross-correlation coefficient according to formula (1) and formula (2)k-1Obtaining K cross correlation coefficients rhokAnd K-1 cross correlation coefficient differences ck-1;
ck-1=ρk-ρk-1 (2)
In the formula (1), N is the number of sampling points, xn、Is a switching-on vibration signal X (t) sampling point and its mean value, ukn、The Kth IMF component element and the mean value thereof; in the formula (2), ρk,ρk-1The cross-correlation coefficients of the kth and kth-1 IMF components with X (t), respectively;
2-3 presetting a minimum threshold value c ═ Δ, Δ ranging from 0 to 0.05, if c is obtainediIf the value is larger than a preset threshold value c, K is equal to K +1, and the operation is repeated for 2-1 and 2-2 in a circulating mode; if obtained ciIf the difference is smaller than a preset threshold value c, the two modes are considered to have high similarity, namely, an over-decomposition phenomenon occurs, and if the circulation is stopped, K is equal to K-1;
2-4, after the K value is determined, selecting the IMF with the maximum correlation coefficient under the current K value as an effective modal component u' (t),
the modal signal can be considered to be most correlated with the original signal as the main characteristic information for characterizing the original signal.
3. The method for predicting the residual mechanical life of the conventional circuit breaker contact system based on deep learning as claimed in claim 1, wherein in the step 3-3, the effective segment calibration is completed by a double-threshold method, which comprises the following steps:
obtaining a short-time energy value E (i) of the effective modal component through 3-2, and further drawing a short-time energy waveform; then setting two threshold values Q according to the wave crest and the wave trough of the drawn short-time energy waveform1And Q2,Q1<Q2Wherein Q is1The short-time energy value corresponding to the wave trough is set to be 0-0.1; q2Setting the short-time energy value corresponding to the second vibration peak value within the range of 0.1-0.5; when the energy value E (i) is higher than Q2The signal segment interval of the contact impact is determined, and the energy value of the frame data is higher or lower than Q1To determine the specific end point E of the effective segment1,E2,E1,E2The original vibration signal in between is the valid segment X' (t).
4. The method for predicting the remaining mechanical life of the conventional circuit breaker contact system based on deep learning as claimed in claim 1, wherein the fifth step comprises:
the 5-1MCCAE characteristic extraction steps are as follows:
firstly, the effective fragment X 'at the 1 st moment'1Valid fragment X 'to time t'tSimultaneously inputting the data into an improved convolutional self-encoder network; signal X 'at time 1'1Extracting a section of characteristic Z through a CAE channel 11(ii) a Signal S at time 22Through the channel 2, a section of the feature Z is extracted2(ii) a And so on until the feature Z at the t-th moment is extractedt(ii) a Finally, the extracted features Z1~ZtSplicing the time sequence characteristics into a time sequence characteristic F with a time dimension according to the time sequence;
5-2 based on the 5-1 time sequence feature extraction principle, the MCCAE unit specifically comprises the following calculation processes: the length of the valid segment X' (t) obtained in the fourth step is n, that is, X ═ X1,x2,…,xt,…,xnDetermining the number a of parallel channels, and calculating X in the first time1′,X2′,X3′,…,XaTaking out to form an a multiplied by n matrix, inputting the matrix to an MCCAE unit for convolution operation and feature extraction; the size of the extracted feature matrix F is a multiplied by m; inputting the feature matrix into LSTM unit for calculation, and outputting a predicted value pi(ii) a Taking X for the second time2′~Xa+1' input MCCAE unit for calculation, and so on, take each timeXi′~Xi+a' a line vectors are calculated and predicted, and one line is pushed down after calculation is completed until X is takenn-a′~XnWhen the time sequence characteristic matrix F is obtained, the calculation is finished, and n-a +1 time sequence characteristic matrices F are obtained;
5-3 obtaining a time sequence feature matrix F according to 5-2, wherein F is a low-dimensional feature vector Z of a adjacent momentsiCombinations of (1) thus for ZiF is normalized by normalization;
let i' th behavior Z of Zi={z1,z2,…,zt,…zmIn which Zi∈R1×m,m<n,ztRepresenting the feature value extracted by CAE at the t-th time), and Z is calculated according to equation (12)iCarrying out maximum and minimum normalization to obtain normalized feature vectorObtaining the normalized time sequence characteristic matrix
5. The method for predicting the remaining mechanical life of the conventional circuit breaker contact system based on deep learning as claimed in claim 1, wherein the step 6-2 of constructing an LSTM prediction model and obtaining the predicted value of the remaining life comprises the following steps:
an LSTM unit is constructed through 6-1, and an LSTM hidden layer neuron is obtained; further, constructing three layers of LSTM networks, wherein neurons of a hidden layer are 8, 16 and 8 respectively, and finally constructing a full-connection layer according to a formula (17), wherein the number of the neurons is 1, and the activation function of each layer is tanh; simultaneously, a dropout mechanism is added behind each layer, the parameter is set to be 0.1, the prediction result is subjected to smooth filtering according to the formula (18), and the obtained final prediction life is p'i;
pi=W*ht+b (17)
In the formula (17), W and b are weight and offset of the full connection layer, htFor the output obtained from the last layer of LSTM, piOutputting the predicted value of the residual service life for the full connection layer;
in formula (18), p'iAnd N is the sliding window length, and the final residual life predicted value is obtained after smoothing filtering.
6. The method for predicting the residual mechanical life of the universal circuit breaker contact system based on deep learning as claimed in claim 1, wherein the universal circuit breaker contact system life test system comprises an industrial personal computer, a PCL-720+ board card, a collection card, a Hall current sensor, a Hall voltage sensor, a vibration sensor, a solid relay set and a switching power supply; the industrial personal computer is respectively connected with the board card and the acquisition card, the board card is connected with the solid state relay group, the acquisition card is respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor, and the switching power supply is also respectively connected with the Hall current sensor, the Hall voltage sensor and the vibration sensor; the solid relay group is connected with an energy storage loop, a closing loop and an opening loop of the universal circuit breaker, the Hall current sensor is connected with an opening and closing accessory of the universal circuit breaker, and the Hall voltage sensor and the vibration sensor are respectively connected with a contact system of the universal circuit breaker.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110416581.7A CN113125095B (en) | 2021-04-19 | 2021-04-19 | Universal circuit breaker contact system residual mechanical life prediction method based on deep learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110416581.7A CN113125095B (en) | 2021-04-19 | 2021-04-19 | Universal circuit breaker contact system residual mechanical life prediction method based on deep learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113125095A true CN113125095A (en) | 2021-07-16 |
CN113125095B CN113125095B (en) | 2022-03-22 |
Family
ID=76777569
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110416581.7A Expired - Fee Related CN113125095B (en) | 2021-04-19 | 2021-04-19 | Universal circuit breaker contact system residual mechanical life prediction method based on deep learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113125095B (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113779820A (en) * | 2021-07-26 | 2021-12-10 | 国网辽宁省电力有限公司电力科学研究院 | On-load tap-changer fault identification method based on neural network response surface model |
CN114021477A (en) * | 2021-11-18 | 2022-02-08 | 江苏科技大学 | Relay service life prediction method based on wavelet packet decomposition and 1D-CNN |
CN114184833A (en) * | 2021-10-27 | 2022-03-15 | 中国科学院微电子研究所 | Spin Hall device, Hall voltage acquisition method and maximum pooling method |
CN114646891A (en) * | 2022-03-10 | 2022-06-21 | 电子科技大学 | Residual life prediction method combining LSTM network and wiener process |
CN115902615A (en) * | 2023-01-09 | 2023-04-04 | 佰聆数据股份有限公司 | Method and device for analyzing defects of power circuit breaker |
CN116796890A (en) * | 2023-05-29 | 2023-09-22 | 南方电网能源发展研究院有限责任公司 | Carbon emission right cost prediction method, device, equipment, medium and product |
CN117330816A (en) * | 2023-12-01 | 2024-01-02 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20180065498A (en) * | 2016-12-08 | 2018-06-18 | 한국항공대학교산학협력단 | Method for deep learning and method for generating next prediction image using the same |
CN109387779A (en) * | 2018-09-28 | 2019-02-26 | 河北工业大学 | A kind of omnipotent breaker operation attachment method for predicting residual useful life based on statistical data driving |
CN109883699A (en) * | 2018-12-20 | 2019-06-14 | 上海理工大学 | A kind of rolling bearing method for predicting residual useful life based on long memory network in short-term |
CN110021165A (en) * | 2019-03-18 | 2019-07-16 | 浙江工业大学 | A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model |
CN112285546A (en) * | 2020-10-19 | 2021-01-29 | 广东电网有限责任公司佛山供电局 | Contact life evaluation system and method based on multi-source data and neural network |
CN112347898A (en) * | 2020-11-03 | 2021-02-09 | 重庆大学 | Rolling bearing health index construction method based on DCAE neural network |
-
2021
- 2021-04-19 CN CN202110416581.7A patent/CN113125095B/en not_active Expired - Fee Related
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
KR20180065498A (en) * | 2016-12-08 | 2018-06-18 | 한국항공대학교산학협력단 | Method for deep learning and method for generating next prediction image using the same |
CN109387779A (en) * | 2018-09-28 | 2019-02-26 | 河北工业大学 | A kind of omnipotent breaker operation attachment method for predicting residual useful life based on statistical data driving |
CN109883699A (en) * | 2018-12-20 | 2019-06-14 | 上海理工大学 | A kind of rolling bearing method for predicting residual useful life based on long memory network in short-term |
CN110021165A (en) * | 2019-03-18 | 2019-07-16 | 浙江工业大学 | A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model |
CN112285546A (en) * | 2020-10-19 | 2021-01-29 | 广东电网有限责任公司佛山供电局 | Contact life evaluation system and method based on multi-source data and neural network |
CN112347898A (en) * | 2020-11-03 | 2021-02-09 | 重庆大学 | Rolling bearing health index construction method based on DCAE neural network |
Non-Patent Citations (3)
Title |
---|
SHUGUANG SUN等: "Quantitative Evaluation of Electrical Life of AC Contactor Based on Initial Characteristic Parameters", 《IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT》 * |
孙曙光等: "基于动作过程振动检测的低压断路器机械寿命预测", 《仪器仪表学报》 * |
陈自强: "基于LSTM网络的设备健康状况评估与剩余寿命预测方法的研究", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113779820A (en) * | 2021-07-26 | 2021-12-10 | 国网辽宁省电力有限公司电力科学研究院 | On-load tap-changer fault identification method based on neural network response surface model |
CN114184833A (en) * | 2021-10-27 | 2022-03-15 | 中国科学院微电子研究所 | Spin Hall device, Hall voltage acquisition method and maximum pooling method |
CN114021477A (en) * | 2021-11-18 | 2022-02-08 | 江苏科技大学 | Relay service life prediction method based on wavelet packet decomposition and 1D-CNN |
CN114646891A (en) * | 2022-03-10 | 2022-06-21 | 电子科技大学 | Residual life prediction method combining LSTM network and wiener process |
CN114646891B (en) * | 2022-03-10 | 2023-05-30 | 电子科技大学 | Residual life prediction method combining LSTM network and wiener process |
CN115902615A (en) * | 2023-01-09 | 2023-04-04 | 佰聆数据股份有限公司 | Method and device for analyzing defects of power circuit breaker |
CN116796890A (en) * | 2023-05-29 | 2023-09-22 | 南方电网能源发展研究院有限责任公司 | Carbon emission right cost prediction method, device, equipment, medium and product |
CN117330816A (en) * | 2023-12-01 | 2024-01-02 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
CN117330816B (en) * | 2023-12-01 | 2024-01-26 | 南京中旭电子科技有限公司 | Monitoring data optimization method for Hall current sensor |
Also Published As
Publication number | Publication date |
---|---|
CN113125095B (en) | 2022-03-22 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113125095B (en) | Universal circuit breaker contact system residual mechanical life prediction method based on deep learning | |
CN112116198A (en) | Data-driven process industrial state perception network key node screening method | |
CN116881685A (en) | Bearing life prediction method for improving TCN (train control network) under mixed attention mechanism | |
CN116610998A (en) | Switch cabinet fault diagnosis method and system based on multi-mode data fusion | |
CN110794254B (en) | Power distribution network fault prediction method and system based on reinforcement learning | |
CN114266301A (en) | Intelligent power equipment fault prediction method based on graph convolution neural network | |
CN113780401A (en) | Composite insulator fault detection method and system based on principal component analysis method | |
CN113988210A (en) | Method and device for restoring distorted data of structure monitoring sensor network and storage medium | |
CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
CN117471252A (en) | DC fault arc diagnosis method based on ICEEMDAN-KPCA feature extraction and SOA-KELM | |
CN117272102A (en) | Transformer fault diagnosis method based on double-attention mechanism | |
Sun et al. | Remaining useful life prediction for circuit breaker based on opening-related vibration signal and SA-CNN-GRU | |
CN117407770A (en) | High-voltage switch cabinet fault mode classification and prediction method based on neural network | |
ul Hassan et al. | Online static security assessment for cascading failure using stacked De-noising Auto-encoder | |
CN116756575A (en) | Non-invasive load decomposition method based on BGAIN-DD network | |
CN114519305A (en) | Method for predicting residual life of circuit breaker based on stage attention mechanism network model | |
CN116307773A (en) | Reliability estimation method for secondary equipment of transformer substation | |
CN115712064A (en) | Excitation system fault diagnosis method based on LSTM-CNN hybrid neural network | |
CN113989070A (en) | Short-term energy consumption prediction method and system for power utilization end of smart power grid | |
CN112883639A (en) | GIS equipment service life prediction device and method based on machine learning | |
Zheng et al. | Research on predicting remaining useful life of equipment based on health index | |
CN111044808A (en) | Power utilization information acquisition system operation and maintenance quality reliability assessment system and method | |
Xie et al. | Evaluation algorithm of high voltage switchgear based on bayesian fusion | |
Du et al. | Fault diagnosis method of automation equipment in independent and controllable substation based on deep reinforcement learning | |
CN117829822B (en) | Power transformer fault early warning method and system |
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 | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20220322 |