CN117150340A - Method and device for diagnosing faults of small samples of switch machine - Google Patents
Method and device for diagnosing faults of small samples of switch machine Download PDFInfo
- Publication number
- CN117150340A CN117150340A CN202311092199.0A CN202311092199A CN117150340A CN 117150340 A CN117150340 A CN 117150340A CN 202311092199 A CN202311092199 A CN 202311092199A CN 117150340 A CN117150340 A CN 117150340A
- Authority
- CN
- China
- Prior art keywords
- prototype
- data
- switch machine
- vibration data
- support set
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 53
- 238000003745 diagnosis Methods 0.000 claims abstract description 49
- 239000013598 vector Substances 0.000 claims abstract description 32
- 230000008569 process Effects 0.000 claims abstract description 26
- 238000006243 chemical reaction Methods 0.000 claims abstract description 12
- 238000012549 training Methods 0.000 claims abstract description 7
- 238000003062 neural network model Methods 0.000 claims abstract 4
- 230000001133 acceleration Effects 0.000 claims description 19
- 238000000605 extraction Methods 0.000 claims description 17
- 230000014509 gene expression Effects 0.000 claims description 14
- 230000009471 action Effects 0.000 claims description 11
- 238000005259 measurement Methods 0.000 claims description 11
- 230000009467 reduction Effects 0.000 claims description 7
- 238000004891 communication Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000012937 correction Methods 0.000 claims description 4
- 230000004927 fusion Effects 0.000 claims description 3
- 230000007246 mechanism Effects 0.000 claims description 3
- 238000011176 pooling Methods 0.000 claims description 3
- 230000007704 transition Effects 0.000 claims description 2
- 238000002405 diagnostic procedure Methods 0.000 claims 1
- 238000010586 diagram Methods 0.000 description 20
- 238000013528 artificial neural network Methods 0.000 description 9
- 230000000903 blocking effect Effects 0.000 description 4
- 238000012423 maintenance Methods 0.000 description 4
- 239000000284 extract Substances 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 238000004321 preservation Methods 0.000 description 2
- 230000011664 signaling Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004422 calculation algorithm Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0895—Weakly supervised learning, e.g. semi-supervised or self-supervised learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computational Linguistics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention discloses a method and a device for diagnosing faults of a small sample of a turnout point machine, wherein the diagnosis method comprises the steps of collecting vibration data of the turnout point machine in the conversion process; dividing the acquired vibration data into a support set S and a query set Q, inputting the support set S and the query set Q into a double-scale neural network model to extract deep features and shallow features respectively, calculating the average value of labeled sample feature vectors in each state in the support set S as a prototype by using a semi-supervised weighted prototype updating strategy, training and optimizing the semi-supervised weighted prototype network model by using the support set data, inputting the query set data into the trained semi-supervised weighted prototype network model, and identifying different categories on the query set. The invention uses the scarce data of the faults in the real scene, improves the fault diagnosis and identification precision and the generalization performance of the model, can realize the fault diagnosis of the small sample of the switch machine, and provides theoretical basis for the fault diagnosis of the cross-equipment and the small sample.
Description
Technical Field
The invention belongs to the technical field of intelligent diagnosis of a switch machine, and particularly relates to a method and a device for diagnosing faults of a small sample of the switch machine.
Background
The signal system is a complex control system that includes a number of mechanical devices and software. These devices and software work cooperatively to enable the safe and efficient operation of the train. A minor fault may cause a major safety incident due to the complexity and specificity of the signaling system. Some studies have shown that switch machine failures account for about 40% of the total failure of the signaling system. The switch machine is an important signal base device for controlling the running direction of the train by switching the direction of the switch rail. However, the switch machine has some faults under the severe working environment, and the safe running of the train is seriously threatened. Therefore, in order to ensure safe operation of the train, it is important to study the fault diagnosis technology of the switch machine.
At present, the operation maintenance management of the turnout point switch of urban rail transit in China adopts a traditional operation and maintenance mode to realize maintenance support. However, the conventional operation and maintenance mode cannot provide more accurate state information of the switch machine and more accurate fault recognition due to the lag of the monitoring means and the defect of the diagnostic model. In addition, since a large amount of fault data is difficult to obtain in a practical scene, the requirement of the existing point machine fault diagnosis method for enough marking data is difficult to meet, so that the diagnosis precision is low and the generalization performance is poor. The meta learning can be used as an effective means to solve the problem that the deep learning model has poor generalization performance due to less fault data of the switch machine, but the limitation that the prototype changes greatly in different iterations still exists. Therefore, in order to improve driving efficiency and safety in the actual scene of rare faults of the switch machine, a switch fault diagnosis method is urgently needed to realize real-time state monitoring and high-efficiency and accurate fault identification.
Disclosure of Invention
The invention aims at: the method and the device for diagnosing the faults of the small sample of the switch machine are provided, and are used for solving the problems of low fault recognition precision and poor model generalization performance caused by the lag of a monitoring means and the scarcity of fault data in the prior art. To achieve the above object, the present invention employs the following methods and apparatuses:
the invention provides a method for diagnosing faults of small samples of a switch machine, which comprises the following steps:
s01: the method comprises the steps that vibration data in the conversion process are collected through a turnout point machine collecting module, the vibration data are vibration data signals in the vertical direction, a sensor is arranged at the tail end of a point machine action rod, and the types of the vibration data signals in the vertical direction comprise normal state data and fault state data of the point machine;
s02: dividing collected vibration data of the switch machine into a support set S and a query set Q, wherein the support set S comprises a label sample and a label-free sample;
s03: inputting data of a support set and a query set of the switch machine into a double-scale neural network (DSNN) model to extract deep and shallow features respectively, and outputting the extracted deep and shallow features into a feature measurement space in a vector form;
s04: in the feature measurement space, a semi-supervised weighting prototype updating strategy is used, the average value of the labeled sample feature vectors in each state in the support set S is calculated to serve as a prototype, the square Euclidean distance between the unlabeled sample feature vectors in each state and the prototype is calculated and weighted, so that the position of the prototype is finely adjusted, the square Euclidean distances among different prototypes are calculated and weighted to enlarge the distance among the prototypes, and the stability and the distinguishing capability of a semi-supervised prototype network are improved;
s05: the support set data is used for training and optimizing a semi-supervised weighted prototype network (SSWPN) model, query set data is input into the trained semi-supervised weighted prototype network (SSWPN) model, and if the semi-supervised weighted prototype network (SSWPN) model can identify different categories on the query set, the method is particularly good in accuracy, stability and generalization, the method can be used for achieving small sample fault diagnosis of the switch machine. The optimized SSWPN is used as a pattern recognition algorithm for diagnosing the faults of the small samples of the switch machine, and the small sample data of the switch machine to be diagnosed is diagnosed to obtain the fault classification result of the switch machine; the invention uses the scarce data of the faults in the real scene, improves the fault diagnosis and identification precision and the generalization performance of the model, can realize the fault diagnosis of the small sample of the switch machine, and provides theoretical basis for the fault diagnosis of the cross-equipment and the small sample.
In the above scheme, in the step S02, the vibration data of the switch machine is divided into a support set S and a query set Q, which specifically includes the following sub-steps:
s0201: marking vibration data, wherein the vibration data comprises 5 types, namely the vibration data comprises 1 turnout point machine normal state data and 4 fault state data, taking the reverse movement process of the turnout point machine positioning as an example, the label of the normal state is 0, the label of the action bar breaking fault state is 1, the label of the blocking fault state is 2, the label of the blocking notch fault state is 3, and the label of the multiple switching state is 4; each type contains 50 samples; taking 20 samples in each type as a support set, and taking the rest 30 samples as a query set;
s0202: taking 1 sample or 5 samples each time in the support set for training; furthermore, 5 samples at a time were taken in the support set as unlabeled sample correction prototype centers;
in a further preferred embodiment of the foregoing solution, in step S03, the extracting deep and shallow features in the switch machine specifically includes the following sub-steps:
s0301: and the vibration data in the support set and the query set are subjected to convolution operation through a convolution layer with the convolution kernel size of 1 multiplied by 1 to obtain a feature map, the feature map is input into the convolution layer with the convolution kernel size of 1 multiplied by 1 to expand the channel number to 2 times, and the features after the channel number expansion are divided into two groups of features of the same channel on average.
S0302: the method comprises the steps of inputting a first group of features into a deep relation of capturing features in a convolution layer with a convolution kernel size of 3 multiplied by 1, inputting a second group of features into a shallow relation of capturing features in a convolution layer with a convolution kernel size of 1 multiplied by 1, and respectively learning the relevance inside the two groups of features by using an effective channel attention mechanism.
S0303: and (3) carrying out channel fusion on the two groups of features and carrying out dimension reduction on the two groups of features through a convolution layer with the convolution kernel size of 1 multiplied by 1, so that the expression capability and the extraction capability of the features in different scales are enhanced.
S0304: and adding the feature map and the features subjected to S0303 dimension reduction to improve the feature preservation capacity of the DSNN model, obtaining feature vectors with discrimination type attribute information through a maximum pooling layer and a full connection layer, and outputting the feature vectors into a high-dimension space, so that a prototype network can effectively measure the feature vectors.
In a further preferred embodiment of the foregoing aspect, in step S04, the prototype tuning and updating process specifically includes the following sub-steps:
s0401: the invention utilizes unlabeled data to fine tune the prototype position of the support set, where the unlabeled data is also considered part of the support set, and the support set S is redefined as:
wherein U represents a union symbol, specifically means that samples of a left set and a right set of symbols are combined, L represents a labeled dataset, U represents an unlabeled dataset, and x n,i Represents the ith sample, K, in the nth class L Is to support the number of samples concentrated with tag data, K U The method comprises the steps that the number of samples of the unlabeled data in the support set is N, the total number of types is represented by N, the semi-supervised weighted prototype generates N-type prototypes by using the labeled data set L in the support set S, and then the position of the prototypes is finely adjusted by using the unlabeled data set U.
S0402: using the average value of each class in the labeled dataset L as the initial prototype P of the corresponding class n As shown in expression (2), h (·) represents the feature extraction process by the DSNN model.
S0403: by calculating the squared Euclidean distance between the feature vector of the unlabeled exemplar to the original prototype, i.e. the square of the difference between the two vectors, and calculating the weight of each distance, then converting the weighted distance between the two into a probabilityThe calculation process is shown in expressions (3) to (5).
d i ＝E(h(x n,i ),P n ) (3)
Wherein x is n,i E U is the unlabeled exemplar set, E (-) represents the squared Euclidean distance, ω i Representing the distance weighting coefficients of the i-th unlabeled exemplar and the n-th prototype,representing the weighted distance transition probabilities assigned to each class for unlabeled exemplars.
S0404: calculating the feature vector and weighted distance conversion probability of unlabeled exemplarsIs a product of (1) to get a prototype guided by unlabeled samples +.>Recalculating->With the original prototype P n The distance between the two prototypes is increased so as to increase the distinguishing capability between the prototypes of different types;
s0405: introducing Gaussian distribution exp (-t) 2 2) calculating a prototype guided by unlabeled exemplarsWith the original prototype P n Squared Euclidean distance between->Instead of t in Gaussian distribution 2 Obtaining prototype weight c corresponding to the distance between each prototype n As shown in expressions (6) to (7).
Wherein c n Representing the weight of the prototype of the n-th class,representing updated prototypes->Representing a collection of prototypes containing N classes of updates,
s0406: by combiningAssign to the set of initial prototypes P, i.e. +.>Assignment to P n As shown in expression (8), the updating process is then continuously performed, i.e., steps S0403 to S0405, to effectively correct the prototype position, thereby improving the discrimination capability for the failure samples.
In the above scheme, in the step S05, if the semi-supervised weighted prototype network shows a good fault diagnosis accuracy on the query set, it indicates that the constructed small sample fault diagnosis model has obvious advantages in the actual scene of scarce switch machine data, and can meet the actual application requirements.
According to another aspect of the present invention, the present invention also provides a railroad switch machine small sample fault diagnosis apparatus, the apparatus includes a vibration data acquisition module 701, a feature extraction module 702, and a small sample fault diagnosis module 703;
the vibration data acquisition module 701 is provided with a triaxial vibration acceleration sensor (HWT 605) on the tail end of the action bar of the ZYJ7 electrohydraulic switch machine for acquiring vibration data in real time, and stores three orthogonal vibration data into computer equipment through serial communication, preferably adopting an RS-232 communication mode;
the feature extraction module 702 is configured to extract deep and shallow features, in which a dual-scale neural network (DSNN) model is embedded, and vibration data in a support set and a query set automatically extract feature vectors after passing through the feature extraction module, so that measurement is conveniently performed in a feature measurement space;
the small sample fault diagnosis module 703 is configured to perform fault diagnosis on the switch machine data to be tested by using the trained semi-supervised weighting (SSWPN) model, so as to obtain a fault diagnosis result.
The invention has the following beneficial technical effects:
(1) The invention extracts deep and shallow layer characteristics of the vibration signals of the switch machine on different scales, can be used for fault diagnosis of small samples of various switch machine equipment through a semi-supervision weighting prototype network, and greatly improves fault diagnosis precision and generalization performance of a model;
(2) In the invention, a double-scale neural network is provided, and the expression capability and the extraction capability of different scale features of the switch machine are enhanced;
(3) In the invention, a semi-supervised weighted prototype updating strategy is provided, so that the stability and distinguishing capability of the label-free data optimization prototype are improved;
(4) In the invention, a semi-supervised weighted prototype SSWPN model is established, and the recognition accuracy and generalization performance of the model under a limited fault sample are improved;
(5) The small sample fault diagnosis device for the turnout switch machine has the remarkable advantages of high diagnosis speed, high diagnosis precision, good generalization performance and less required training samples, can meet the actual application requirements of the turnout switch machine, and has wide application potential in other fields.
Drawings
Fig. 1 is a flow chart of a method for diagnosing faults of a small sample of a switch machine according to the invention.
Fig. 2 is a schematic diagram of vertical vibration data of the switch machine without faults.
Fig. 3 is a schematic diagram of vertical vibration data of a broken action bar of the railroad switch machine.
Fig. 4 is a schematic diagram of vertical vibration data of the switch machine with jamming fault.
Fig. 5 is a schematic diagram of vertical vibration data of a railroad switch machine with a snap fault.
Fig. 6 is a schematic diagram of vertical vibration data of the switch machine with multiple operating faults.
Fig. 7 is a schematic diagram of a double-scale neural network of the switch machine provided by the invention.
Fig. 8 is a schematic diagram of a semi-supervised weighted prototype update strategy provided by the present invention.
Fig. 9 is a schematic diagram of a small sample fault diagnosis process of a switch machine based on a semi-supervised weighted prototype network.
Fig. 10 is a block diagram of a small sample fault diagnosis device of a switch machine based on a semi-supervised weighted prototype network.
Fig. 11 is a block diagram of a vibration data acquisition module according to the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be clearly and completely described below by way of example only with reference to the accompanying drawings of the present invention. However, it should be noted that the examples in the specification are preferred examples, and not all examples. Furthermore, many of the details set forth in the specification are provided solely for the purpose of providing a thorough understanding of the key issues of the present invention, and the present invention may be practiced without these specific details.
The invention provides a method for diagnosing faults of a small sample of a switch machine, and a flow diagram is shown in fig. 1, and the method comprises the following steps:
s01: the method comprises the steps that vibration data in the conversion process are collected through a turnout point machine collecting module, wherein the vibration data comprise normal state data and fault state data of the turnout point machine, namely the type of collecting vertical vibration signals comprises normal state data and fault state data of the turnout point machine;
s02: dividing collected vibration data of the switch machine into a support set S and a query set Q, wherein the support set S comprises a label sample and a label-free sample;
s03: inputting data of a support set and a query set of the switch machine into a double-scale neural network (DSNN) model to extract deep and shallow features respectively, and outputting the extracted deep and shallow features into a feature measurement space in a vector form to form feature vectors;
s04: in the feature measurement space, calculating the average value of the feature vectors of the labeled samples in each state in the support set S as a prototype by using a semi-supervised weighting updating strategy, calculating the square Euclidean distance between the feature vectors of the unlabeled samples in each state and the prototype, weighting the square Euclidean distance, fine tuning the position of the prototype, calculating the square Euclidean distance between different prototypes, and weighting the square Euclidean distance to enlarge the distance between the prototypes;
s05: and training and optimizing the SSWPN model by using support set data, and then inputting query set data into the trained SSWPN model, wherein if the SSWPN model can identify different categories on the query set, the SSWPN model is particularly good in accuracy, stability and generalization, the SSWPN model can be used for realizing the fault diagnosis of small samples of the switch machine.
The vibration data of the switch machine in the conversion process refers to vibration acceleration data acquired by a triaxial vibration acceleration sensor in the conversion process from positioning to inversion or from inversion to positioning. The vibration data collected by the triaxial vibration acceleration sensor comprises vibration acceleration data in three orthogonal directions. Since the vibration acceleration data in the vertical direction can well reflect the state information of the movement process of the switch rail, the invention preferably adopts the vibration data in the vertical direction for diagnosing the faults of small samples of the switch machine. In particular, the data is recorded when the switch machine switching process is started, and the data is stored after the switch machine is completely switched.
Fig. 2 is a schematic diagram of vertical vibration data of the switch machine without faults. At about 1s, the vibration data has wave peaks along with the starting of the switch machine; at about 4.5s, the switch rail is switched to the other side, and a small peak appears in the vibration signal; during the closing process of the switch machine, the vibration data again show small wave peaks, then tend to stabilize and end the collection as the three-phase motor stops.
Fig. 3 is a schematic diagram of vertical vibration data of a broken action bar of the railroad switch machine. The vibration data appear wave peaks along with the starting of the switch machine and then tend to be stable; when the switch machine is locked, the vibration data appears a peak again, then the vibration data tends to be stable and the acquisition is ended along with the stop of the three-phase motor; the acquisition time of the vibration data was about 6s.
Fig. 4 is a schematic diagram of vertical vibration data of the switch machine with jamming fault. The vibration data has wave peaks along with the starting of the switch machine, and then has a plurality of wave peaks; when the foreign matter is blocked in the switch rail conversion process and cannot be converted to the other side of the track in about 4.5s, the amplitude of vibration data becomes large, and then the vibration data tends to be stable and is collected as the three-phase motor stops; the acquisition time of the vibration data was about 14s.
Fig. 5 is a schematic diagram of vertical vibration data of a railroad switch machine with a snap fault. The vibration data appear wave peaks along with the starting of the switch machine and then tend to be stable; because of the fault of the notch, the switch is stopped when the switch rail is not completely switched to the other side of the track, the amplitude of the peak wave appearing in vibration data is smaller, and then the vibration data tends to be stable and the collection is ended along with the stop of the three-phase motor; the acquisition time of the vibration data was about 9s.
Fig. 6 is a schematic diagram of vertical vibration data of the switch machine with multiple operating faults. Because the switch machine is started for a plurality of times, the vibration data correspondingly represents a plurality of wave peaks, the vibration data tend to be stable after the locking is finished, and the acquisition is finished along with the stop of the three-phase motor; the acquisition time of the vibration data was about 13s.
For the step S02, taking the ZYJ7 electrohydraulic switch machine as an example, the vibration data of the switch machine is divided into a support set S and a query set Q, specifically:
marking vibration data, taking the reverse movement process of the switch machine positioning as an example, wherein the label of a normal state is 0, the label of a breaking fault state of an action bar is 1, the label of a blocking fault state is 2, the label of a blocking notch fault state is 3, and the label of a multiple switching state is 4; in particular, the vibration data of the switch machine conversion process is collected to comprise 5 types, each containing 50 samples; preferably, the present invention takes 20 samples in each type as a support set, the remaining 30 samples as a query set. Furthermore, the present invention takes 5 samples at a time from the support set as unlabeled exemplars to correct and update the position of the prototype.
Fig. 7 is a schematic diagram of a dual-scale neural network of a switch machine, which is used for extracting deep and shallow features in vibration data of the switch machine, and specifically includes:
the invention adopts a convolution layer with the convolution kernel size of 1 multiplied by 1 to carry out convolution dimension-increasing operation on the vibration data in the support set and the query set to obtain a feature map, then the feature map is input into the convolution layer with the convolution kernel size of 1 multiplied by 1 to expand the channel number to 2 times, and the features after the channel number expansion are divided into two groups of features with the same channels on average, and the specific steps are as follows:
in this example, the features of the deep layer relationship are captured by adopting a convolution layer with a convolution kernel size of 3×1 for the features of the first group, the features of the shallow layer relationship are captured by adopting a convolution layer with a convolution kernel size of 1×1 for the features of the second group, the internal relevance of the two groups of features is learned by using an effective channel attention mechanism respectively,
and (3) carrying out channel fusion on the two groups of features and carrying out dimension reduction on the two groups of features through a convolution layer with the convolution kernel size of 1 multiplied by 1, so that the expression capability and the extraction capability of the features in different scales are enhanced.
And adding the feature map obtained by the convolution operation in the step with the feature after dimension reduction, namely adding the feature before dimension increase and the feature after dimension reduction, improving the feature preservation capacity of a dual-scale neural network (DSNN) model, obtaining a feature vector with discrimination type attribute information through a maximum pooling layer and a full connection layer, and outputting the feature vector into a high-dimension space, so that a prototype network can effectively measure the feature vector.
The invention also provides a semi-supervised weighted prototype updating strategy, and referring to fig. 8, the fine tuning and updating process of the prototype specifically comprises:
the invention utilizes unlabeled data to fine tune the prototype position of the support set, where the unlabeled data is also considered part of the support set, and the support set S is redefined as:
it should be noted that U representsUnion symbols, in particular, merging samples of left and right sets of symbols, L representing a labeled dataset, U representing an unlabeled dataset, x n,i Represents the ith sample, K, in the nth class L Is to support the number of samples concentrated with tag data, K U The method comprises the steps that the number of samples of the unlabeled data in the support set is N, the total number of types is represented by N, the semi-supervised weighted prototype generates an N-type prototype by using the labeled data set L in the support set S, and then the prototype is subjected to fine tuning by using the unlabeled data set U. In this example, preferably, 5 samples are selected from each class as the unlabeled dataset U.
Using the average value of each class in the labeled dataset L as the initial prototype P of the corresponding class n As shown in expression (2), h (·) represents the feature extraction process by a dual-scale neural network (DSNN) model.
By calculating the squared Euclidean distance between the feature vector of the unlabeled exemplar to the original prototype, and calculating the weight of each distance, and then converting the weighted distance between them into probabilityThe calculation process is shown in expressions (3) to (5).
d i ＝E(h(x n,i ),P n ) (3)
X is the number n,i E U is the unlabeled exemplar set, E (-) represents the squared Euclidean distance, ω i Representing the distance between the ith unlabeled exemplar and the nth class prototypeFrom the weighting coefficients of the weights,representing the probability that unlabeled exemplars are assigned to each class.
Calculating the feature vector and weighted distance conversion probability of unlabeled exemplarsIs a product of (1) to get a prototype guided by unlabeled samples +.>Recalculating->With the original prototype P n The distance between them to increase the discrimination between different types of prototypes.
Introducing Gaussian distribution exp (-t) 2 2) calculating a prototype guided by unlabeled exemplarsWith the original prototype P n Squared Euclidean distance between->Instead of t in Gaussian distribution 2 Obtaining prototype weight c corresponding to the distance between each prototype n As shown in expressions (6) to (7).
Wherein c n Representing the weight of the prototype of the n-th class,the prototype of the update is represented by the model,/>representing a set of prototypes containing N classes of updates;
by assembling prototypesAssign to the set of initial prototypes P, i.e. +.>Assignment to P n And as shown in the expression (8), the updating process is continuously executed, so that the effective correction of the prototype position is realized, and the distinguishing capability of fault samples is improved.
For the step S05, if the semi-supervised weighted prototype network shows good fault diagnosis accuracy on the query set, it is indicated that the constructed small sample fault diagnosis model has obvious advantages in the actual scene of scarcity of the switch machine data, and can meet the actual application requirements.
Fig. 9 is a schematic diagram of a small sample fault diagnosis flow of a switch machine based on a semi-supervised weighted prototype network, which is provided by the invention, by collecting new vibration data of the switch machine, realizing intelligent diagnosis of the small sample by using a trained SSWPN model, performing fault location according to a fault result, sending out fault alarm information, and guiding professionals to maintain quickly.
The invention provides a structural block diagram of a small sample fault diagnosis device of a switch machine, which is shown in fig. 10, and comprises a switch machine vibration data acquisition module 701, a feature extraction module 702 and a small sample fault diagnosis module 703;
the vibration data acquisition module 701 is provided with a HWT605 vibration acceleration sensor at the tail end of the action bar of the ZYJ7 electrohydraulic switch machine to acquire vibration data in real time, and stores three orthogonal vibration acceleration data into computer equipment through serial communication, preferably adopting an RS-232 communication mode;
the feature extraction module 702 is embedded with a dual-scale neural network (DSNN) model, and is configured to automatically extract feature vectors of vibration data in a support set and a query set, so as to facilitate measurement in a measurement space;
the small sample fault diagnosis module 703 is configured to perform fault diagnosis on switch machine data to be tested by using the trained semi-supervised weighted prototype network (SSWPN) model, so as to obtain a fault diagnosis result.
Fig. 11 is a block diagram of a vibration data acquisition module according to the present invention. The vibration data acquisition module comprises: the ZYJ7 electrohydraulic switch machine 801, a signal collector 802, a vibration acceleration sensor 803 and a computer 804, wherein the ZYJ7 electrohydraulic switch machine 801 comprises a circuit controller 8011, an action bar 8012 and a presentation bar 8013, and the collector 802 comprises a signal converter 8021 and a USB collecting interface 8022. The circuit controller 8011 is an electrical component, directly connected to the power supply of the switch machine, and the actuating rod 8012 and the indicating rod 8013 are mechanical components, driven by the motor of the switch machine. The circuit controller 8011 is used to control the rotation of the motor in the switch machine to drive the translational movement of the actuator rod 8012 and the indicator rod 8013, which in turn drives the switching of the switch rails. The collector 802 is used for collecting vibration data in the switching process of the switch machine, the signal converter 8021 and the USB collecting interface 8022 are connected with each other, the signal converter 8021 is used for converting sensor signals into vibration data, and the USB collecting interface 8022 is used for outputting the vibration data. In this case, the vibration acceleration sensor 803 is preferably a triaxial vibration acceleration sensor of HWT605 type, and is further connected to the signal converter 8021 by an RS232 serial communication manner, and the connection manner between the vibration acceleration sensor 803 and the signal converter 8021 is a four-wire system for converting data, which can be specifically described as: the 5V positive power line, the GND negative power line, the TXD transmit signal line, and the RXD receive signal line, wherein the 5V positive power line and the GND negative power line are used for supplying power to the vibration acceleration sensor 803, the TXD transmit signal line is used for transmitting a transmitted electrical signal, and the RXD receive signal line is used for transmitting a received electrical signal. The collection principle is that when the switch starts to move, the computer 804 sends an electric signal to the vibration acceleration sensor 803 through the TXD transmitting signal line, the vibration acceleration sensor 803 starts to collect weak current generated by vibration in the switch rail conversion process, the weak current is amplified through the RXD receiving signal line and the signal collector, then the signal converter 8021 is utilized to convert the current signal into vibration data, and the vibration data is stored in the computer 804 in a USB collection interface 8022 reading data form. And the obtained vertical vibration data are used as a data source for fault diagnosis of a small sample of the follow-up switch machine.
The foregoing is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that modifications, equivalents, variations, etc. made by those skilled in the art without departing from the principles of the present invention are also considered as being within the scope of the present invention.
Claims (7)
1. A method for diagnosing faults of small samples of a switch machine is characterized in that: the diagnostic method comprises the steps of:
s01: collecting vibration data of the switch machine in the switching process, wherein the vibration data comprises normal state data and fault state data of the switch machine;
s02: dividing the acquired vibration data into a support set S and a query set Q, wherein the support set S comprises a label sample and a label-free sample;
s03: inputting a support set S and a query set Q into a dual-scale neural network model to extract deep features and shallow features respectively, and outputting the extracted deep features and shallow features into a feature measurement space in the form of vectors to form feature vectors;
s04: in the feature measurement space, calculating the square Euclidean distance between the label-free sample feature vector and the prototype in each state by using the average value of the label sample feature vector in each state in the semi-supervised weighting prototype updating strategy calculation support set S as a prototype, and then calculating the square Euclidean distance between different prototypes and weighting;
s05: and training and optimizing the semi-supervised weighted prototype network model by using the support set data, inputting the query set data into the trained semi-supervised weighted prototype network model, and if the semi-supervised weighted prototype network model identifies different categories on the query set, realizing the fault diagnosis of the small sample of the switch machine.
2. The method for diagnosing a small sample fault of a switch machine according to claim 1, wherein: in the step S02, vibration data of the switch machine is divided into a support set S and a query set Q, and specifically includes the following sub-steps:
s0201: marking vibration data, wherein the vibration data comprises 5 types, each type comprises 50 samples, 20 samples are taken as supporting sets in each type, and the rest 30 samples are taken as query sets;
s0202: taking 1 sample or 5 samples each time in the support set for training; and taking 5 samples at a time in the support set as unlabeled exemplar correction prototype positions.
3. The method for diagnosing a small sample fault of a switch machine according to claim 1, wherein: in the step S03, the extraction of deep and shallow features in the switch machine specifically includes the following sub-steps:
s0301: the vibration data in the support set and the query set are subjected to convolution operation through a convolution layer with the convolution kernel size of 1 multiplied by 1 to obtain a feature image, the feature image is input into the convolution layer with the convolution kernel size of 1 multiplied by 1 to expand the channel number to 2 times, and the feature image with the expanded channel number is divided into two groups on average;
s0302: inputting the first group of features into a deep relation of capturing features in a convolution layer with the convolution kernel size of 3 multiplied by 1, inputting the second group of features into a shallow relation of capturing features in a convolution layer with the convolution kernel size of 1 multiplied by 1, and respectively learning the relevance of the interiors of the two groups of features by using an effective channel attention mechanism;
s0303: channel fusion is carried out on the two groups of features, and dimension reduction is carried out on the two groups of features through a convolution layer with the convolution kernel size of 1 multiplied by 1, so that the expression capacity and the extraction capacity of the features in different scales are enhanced;
s0304: and adding the feature map obtained by the convolution operation in the S0301 with the feature obtained by the dimension reduction in the step S0303, and obtaining the feature vector with the discrimination type attribute information through the maximum pooling layer and the full connection layer.
4. The method for diagnosing a small sample fault of a switch machine according to claim 1, wherein: in said step S04, the process of fine tuning and updating the prototype specifically comprises the following sub-steps:
s0401: the prototype position of the support set is fine-tuned with the unlabeled data, which is then also considered as part of the support set, and the support set S is redefined as:
wherein U represents a union symbol, specifically means that samples of a left set and a right set of symbols are combined, L represents a labeled dataset, U represents an unlabeled dataset, and x n,i Represents the ith sample, K, in the nth class L Representing the number of samples supporting the concentrated tag data, K U The method comprises the steps that the number of samples of unlabeled data in a support set is represented, N represents the total number of types, a semi-supervised weighted prototype network model generates N types of prototypes by using a labeled data set L in a support set S, and then the position of the prototypes is finely adjusted by using an unlabeled data set U;
s0402: using the average value of each class of feature vectors in the labeled dataset L as the initial prototype set P of the corresponding class n The following expression is satisfied;
wherein, h (-) represents the characteristic extraction process through a double-scale neural network model;
s0403: by calculating the squared Euclidean distance between the feature vector of the unlabeled exemplar to the original prototype, i.e. the square of the difference between the two vectors, and calculating the weight of each distance, then converting the weighted distance between the two into a probabilityThe conversion calculation process is as follows:
d i ＝E(h(x n,i ),P n )， (3)；
wherein x is n,i E U is the unlabeled exemplar set, E (-) represents the squared Euclidean distance, ω i Representing the distance weighting coefficients of the i-th unlabeled exemplar and the n-th prototype,a weighted distance transition probability representing the assignment of unlabeled exemplars to each class;
s0404: calculating the feature vector and weighted distance conversion probability of unlabeled exemplarsIs a product of (1) to get a prototype guided by unlabeled samples +.>Recalculating the unlabeled sample guide +.>With the initial prototype set P n A distance therebetween;
s0405: calculation of label-free sample guided prototypes with gaussian distributionWith the initial prototype set P n Squared Euclidean distance between->Obtaining prototype weight c corresponding to the distance between prototypes n And satisfies the following expression:
wherein c n Representing the weight of the prototype of the n-th class,representing updated prototypes->Representing a set of prototypes containing N classes of updates;
s0406: by assembling prototypesAssigning values to the set of initial prototypes P, and then continuously executing the updating process from step S0403 to step S0405 to realize effective correction of prototype positions, where the set of initial prototypes P satisfies:
5. the method for diagnosing a small sample fault of a switch machine according to claim 1, wherein: in the step S05, if the semi-supervised weighted prototype network shows good fault diagnosis accuracy on the query set, it is indicated that the constructed small sample fault diagnosis model has obvious advantages in the actual scene of the switch machine data scarcity, and the actual application requirements are satisfied.
6. A railroad switch machine small sample fault diagnosis device comprising: a switch machine vibration data acquisition module, a characteristic extraction module and a small sample fault diagnosis module, wherein the small sample fault diagnosis device is operated to realize the steps of the switch machine small sample fault diagnosis method according to any one of claims 1 to 5;
the vibration data acquisition module is used for installing a triaxial vibration acceleration sensor on the tail end of the action bar of the ZYJ7 electrohydraulic switch machine for acquiring vibration data in real time and storing three orthogonal vibration data into computer equipment through serial communication;
the feature extraction module is used for extracting deep and shallow features, a double-scale neural network model is embedded in the feature extraction module, and feature vectors are automatically extracted after vibration data in a support set and a query set pass through the feature extraction module, so that different types of feature vectors can be measured in a feature measurement space conveniently;
and the small sample fault diagnosis module is used for carrying out fault diagnosis on the data of the switch machine to be detected by using the trained semi-supervised weighted prototype network model to obtain a fault diagnosis result.
7. The railroad switch machine small sample fault diagnosis device according to claim 6, wherein: the vibration data acquisition module comprises: the system comprises a turnout point machine, a signal collector, a vibration acceleration sensor, a computer, an action bar and a signal converter, wherein the turnout point machine is used for effectively collecting and storing vibration signals in the switching process of the turnout point machine; the vibration acceleration sensor is arranged at the tail end of the action bar of the turnout point machine, when the turnout point machine starts to move, the vibration acceleration sensor starts to collect current generated by vibration, the current is amplified by the signal collector, and then the current is converted into vibration data by the signal converter and is stored in the computer.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311092199.0A CN117150340A (en) | 2023-08-28 | 2023-08-28 | Method and device for diagnosing faults of small samples of switch machine |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311092199.0A CN117150340A (en) | 2023-08-28 | 2023-08-28 | Method and device for diagnosing faults of small samples of switch machine |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117150340A true CN117150340A (en) | 2023-12-01 |
Family
ID=88907328
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311092199.0A Pending CN117150340A (en) | 2023-08-28 | 2023-08-28 | Method and device for diagnosing faults of small samples of switch machine |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117150340A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117951632A (en) * | 2024-03-27 | 2024-04-30 | 合肥工业大学 | PU contrast learning anomaly detection method and system based on multi-mode prototype network |
-
2023
- 2023-08-28 CN CN202311092199.0A patent/CN117150340A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117951632A (en) * | 2024-03-27 | 2024-04-30 | 合肥工业大学 | PU contrast learning anomaly detection method and system based on multi-mode prototype network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Fault diagnosis of bearings based on multi-sensor information fusion and 2D convolutional neural network | |
| CN111046583B (en) | Point machine fault diagnosis method based on DTW algorithm and ResNet network | |
| CN117150340A (en) | Method and device for diagnosing faults of small samples of switch machine | |
| Li et al. | Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data | |
| CN113505664B (en) | Fault diagnosis method for planetary gear box of wind turbine generator | |
| CN113030723A (en) | Alternating current asynchronous motor state monitoring system | |
| CN112113755A (en) | Mechanical fault intelligent diagnosis method based on deep convolution-kurtosis neural network | |
| CN111597948A (en) | Fault detection and classification method for bearing vibration signals | |
| CN112614130A (en) | Unmanned aerial vehicle power transmission line insulator fault detection method based on 5G transmission and YOLOv3 | |
| CN1908987A (en) | Method for diagnosing accidents of waveshape recognition based substation locomotive traction circuit | |
| CN115905832A (en) | Transformer fault diagnosis method based on vibration signal clustering identification | |
| CN114638318A (en) | Power grid fault diagnosis method and system based on SE-inclusion network | |
| CN117786507A (en) | Rolling bearing unknown fault detection method based on global and local feature coupling guidance | |
| CN112434979A (en) | Health assessment method for turnout system | |
| CN115374829A (en) | Deep learning-based bearing fault diagnosis method and system | |
| CN114233581A (en) | Intelligent patrol alarm system for fan engine room | |
| CN114065651A (en) | Fault time prediction method for rotary equipment | |
| CN113674329A (en) | Vehicle driving behavior detection method and system | |
| CN111476383A (en) | Pump station unit state maintenance dynamic decision method | |
| CN113723592A (en) | Fault diagnosis method based on wind power gear box monitoring system | |
| CN113420387B (en) | Migration diagnosis method and system for rolling bearing of compacting machine | |
| CN118170125B (en) | Mobile device for diagnosing new energy automobile faults based on embedded AI and diagnosis method | |
| Bian et al. | Degradation detection method for railway point machines | |
| CN118155024B (en) | Automatic generation method and system for large model image sample | |
| CN114487819B (en) | Demagnetizing fault diagnosis system and method for double-stator permanent magnet linear motor |
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 |