Multi-source signal driven turnout switch machine small sample fault diagnosis method and device

Abstract

The invention relates to the technical field of intelligent diagnosis of turnout switch machines, in particular to a multi-source signal driven turnout switch machine small sample fault diagnosis method and device, and the method comprises the steps: collecting the three-phase current data of a turnout switch machine during the conversion operation between a directional direction and a reverse direction; fusion of different weights is carried out on the collected three-phase current data through a self-adaptive fusion strategy; dividing the fused data into a training set and a test set; inputting the fusion data in the training set into a multi-source signal reweighting feature fusion network based on the Meta-ACON, and fusing the extracted features of different levels; generating an original prototype of each fault by using the marked samples in the support set, and expanding the distance between the initial prototypes and enhancing the intra-class aggregation of the initial prototypes by using a class center prototype fine tuning strategy to obtain new prototype distribution; and taking the trained class center fine tuning prototype network as a model algorithm, and carrying out small sample fault diagnosis on the turnout switch machine by using the test set to realize data classification.

Description

Technical Field

[0001] The invention relates to the technical field of intelligent diagnosis of turnout machines, and in particular to a small sample fault diagnosis method and device for turnout machines driven by multi-source signals. Background Art

[0002] Switches, signal machines, and track circuits are the three major pieces of outdoor railway signal equipment, which are related to the safety of railway transportation. Among them, the switch machine is a switch device for turnouts, which is used to indicate the position of turnouts and realize the conversion of turnouts and change the direction of train operation. Due to its complex mechanical structure and circuit system, and its susceptibility to severe outdoor weather, the switch machine has always been a signal device with a high failure rate, which poses a challenge to the railway operation department for inspection and maintenance. With the increase in railway operation density and driving speed, the safety of switch machines has also received great attention. Fault diagnosis technology is an important means to ensure the safe operation of the system and improve the reliability of trains. Therefore, how to achieve high-precision fault diagnosis of switch machines has become the current research focus.

[0003] However, since my country's rail transit still uses the traditional operation and maintenance model to maintain and manage switch machines, the fault status monitoring and fault diagnosis of switch machines have low accuracy. With the development of big data technology and the update and iteration of sensors, the traditional fault diagnosis model can no longer meet the higher-precision diagnosis needs of switch machines. At the same time, in actual scenarios, the fault time of switch machines is far less than the normal time, which leads to only a small number of samples in a certain type of fault mode, and even fewer labeled fault data, which also makes the trained diagnosis model have poor generalization ability and low accuracy.

[0004] Meta-learning follows the scenario paradigm strategy and is an efficient small sample cross-domain learning strategy. As a branch of machine learning, the main idea of ​​meta-learning is to use past experience to quickly adapt to new tasks, and adopt task-based iterative optimization strategies to improve the flexibility and applicability of the model, so that the model can quickly adapt to new tasks in the target field. However, most of the existing meta-learning methods use single-source data for fault diagnosis, without considering the correlation and difference between different types of data. In the process of generating prototypes and adjusting the positions of prototypes, they cannot provide more comprehensive and accurate information about the fault type, resulting in blurred boundaries between various prototypes after multiple iterations, resulting in a high probability of misclassification. Summary of the invention

[0005] The purpose of the present invention is to provide a multi-source signal driven small sample fault diagnosis method and device for turnout switch, aiming to solve the problems of low diagnostic accuracy and poor efficiency caused by the scarcity of fault data and the incomplete characterization of fault characteristics by single-source data in the existing turnout maintenance method. To achieve the above purpose, the present invention adopts the following technical solution.

[0006] According to another aspect of the present invention, the present invention provides a small sample fault diagnosis method for a turnout machine driven by a multi-source signal, comprising the following steps:

[0007] Step 1: Collect the three-phase action current data of the turnout machine when it performs switching operation between directional and reverse directions;

[0008] Step 2: The collected three-phase current data are fused with different weights through an adaptive data fusion strategy;

[0009] Step 3: Divide the fused three-phase current data into a training set and a test set, wherein the training set includes a support set S and a query set Q;

[0010] Step 4: Input the support set S and query set Q contained in the training set into the multi-source signal reweighted feature fusion network based on Meta-ACON, perform feature extraction and feature fusion at different levels, output the fused features in the form of vectors to the high-dimensional vector space to form feature vectors, and generate the original prototype of each category by calculating the average value of the feature vectors of the labeled samples in the support set S;

[0011] Step 5: Fine-tune the center balance strategy to strengthen the repulsion between the original prototypes of each category, so as to expand the distribution distance between the original prototypes of each category, and gradually iterate and optimize the class center fine-tuning prototype network model CFPN to make the boundaries between the original prototypes of each category clear, and obtain a new prototype distribution;

[0012] Step 6: Train and iteratively optimize the class center fine-tuning prototype network model CFPN, and input the query set data into the class center fine-tuning prototype network model CFPN after training and iterative optimization. If the class center fine-tuning prototype network model CFPN can achieve good classification effect on the query set, the class center fine-tuning prototype network model CFPN after training and iterative optimization will be used for small sample fault diagnosis of turnout machines.

[0013] Preferably, the three-phase action current data includes the health status of the switch machine and nine fault states. The switch machine uses a ZYJ7 type electro-hydraulic switch machine, and the collection point is set between the output of the phase failure protector and the first switch starting relay.

[0014] Optionally, in step 2, the three-phase current data is fused with different weights through an adaptive data fusion strategy, including the following steps:

[0015] Step 2.1: Assume the input signal to be {x in(l), i = 1, ..., M; n = 1, ..., N; l = 1, ..., L}, where M represents the number of signal types, N represents the number of samples of each type of signal, and L represents the sequence length. The correlation entropy is used to measure the correlation degree and shared information between two current signals. The improved correlation entropy calculation formula is as follows:

[0016]

[0017] Where log(·) represents the logarithmic function.

[0018] Step 2.2: Then the weight of each signal is adaptively calculated according to the associated entropy, the formula is as follows:

[0019]

[0020] Where exp represents the exponential function.

[0021] Step 2.3: Finally, assign weights to the corresponding signals to obtain the fused signal. The formula is as follows:

[0022]

[0023] Preferably, the process of dividing the fused three-phase current data into a training set and a test set in step 3 comprises the following steps:

[0024] Step 3.0: The fused three-phase data is randomly divided into training set and test set in a ratio of 5:2:

[0025] Step 3.1: Divide the training set into healthy state and 9 fault states, each type contains 50 samples, randomly select 20 samples as the support set S, and the remaining 30 samples are classified as the query set Q;

[0026] Step 3.2: Each time, randomly select 1 or 5 samples in S as labeled samples to generate the original prototype, and then select 5 samples in the support set as unlabeled samples to correct the position of the original prototype.

[0027] Preferably, in step 4, the support set S and the query set Q contained in the training set are input into the multi-source signal reweighted feature fusion network based on Meta-ACON, and features at different levels are extracted and fused, and then the fused features are output to the high-dimensional vector space, and then the original prototype of each category is generated by the support set S, and the steps are as follows:

[0028] Step 4.1: After the data is subjected to one-dimensional convolution and normalization, it is input into three branches for convolution at different levels;

[0029] Step 4.2: Perform maximum pooling on the first branch to obtain local features Fmax , the second branch performs average pooling to obtain the global feature F avg , the third branch performs edge feature extraction F DCMA , where the third branch contains four identical modules, each of which consists of a convolutional neural network (CNN) module and a dual-channel Meta-ACON (DCMN) module. Meta-ACON changes the choice of activation function by switching factors, that is, deciding whether to activate neurons, thereby adaptively controlling whether the activation is linear or nonlinear;

[0030] Step 4.3: The features extracted from the three branches are fused with different weights. The formula is as follows:

[0031] F f =α 1 ·F max +α 2 ·F avg +α 3 ·F DCMA

[0032]

[0033] In the above formula, α i represents the normalized weight of the i-th branch, ∑α i =1. i Represents the initialization weight coefficient of the i-th branch, and the initial value is 1.

[0034] Step 4.4: The fused features are passed through a squeeze-excitation attention module to focus on the key feature information, and the final feature map is flattened into one dimension to generate a more accurate original prototype.

[0035] Preferably, in step 5, the distance between prototypes is fine-tuned by using the "repulsion" between different prototypes. At the same time, the improved center loss is combined with the cross entropy loss to expand the distance between prototypes and enhance the intra-class aggregation, and then the optimization is gradually iterated to make the boundaries between prototypes of each category clear. The steps are as follows:

[0036] Step 5.1: Using the Euclidean distance f d Calculate the labeled sample x i Except for the prototype C k The "repulsion" between other prototypes outside the , and then calculate the distribution probability p between different prototypes through the Softmax function k , the formula is as follows:

[0037]

[0038] In the above formula, Num represents the number of generated prototypes, fθ Indicates mapping the K-dimensional vector to the M-dimensional feature vector C j , which means that except C k Then improve the center loss to make the sample feature distribution of the same category in the space more compact. The formula is as follows:

[0039]

[0040] In the above formula, N L represents the number of labeled samples in the support set, f(x i ) represents x i Eigenvalues ​​in the eigenspace.

[0041] Step 5.2: The improved center loss L c With cross entropy loss L s Combined to get a new loss function L t , which makes prototypes of different categories far away from each other and makes the distribution within the category more compact. The calculation formula is as follows:

[0042] L t =L s +λL c

[0043] In the above formula, λ represents a hyperparameter, which can significantly enhance the discriminative ability of deep features. It is recommended to set λ to 0.1.

[0044] Step 5.3: Repeat steps 5.1-5.2 above and iterate the optimization until the boundaries between the prototypes are clear.

[0045] Preferably, if the Class Center Fine-tuned Prototype Network (CFPN) model can identify the category to which the sample to be detected belongs with an accuracy rate of more than 98%, the network model can be used to realize small sample fault diagnosis of turnout machines.

[0046] The present invention also proposes a small sample fault diagnosis device for a turnout machine driven by a multi-source signal. The fault diagnosis device includes a turnout machine current data acquisition module, a data fusion module, a feature extraction and fusion module, and a fault diagnosis module. The fault diagnosis device fully implements the small sample fault diagnosis steps for a turnout machine driven by a multi-source signal.

[0047] The switch machine current data acquisition module is used to obtain the three-phase induction action current of the switch machine in real time through the Hall current sensor installed in the line between the phase failure protector output and the first switch start relay, and store the current data in the server through the industrial Ethernet;

[0048] The data fusion module uses the improved correlation entropy to calculate the similarity between currents of different phases, and assigns different weights to currents of different phases according to the similarity. Finally, the current data of each phase are fused according to their respective weights to provide input data containing richer information for subsequent model training.

[0049] The feature extraction and fusion module uses a multi-weighted feature fusion network to extract features at different levels from the input fusion data to obtain global features, local features, and edge features. The extracted features are then fused with different weights. The fused features form feature vectors in a high-dimensional vector space to facilitate subsequent generation and adjustment of the original prototype.

[0050] The fault diagnosis module uses the class center fine-tuning prototype network model to perform small sample diagnosis on the turnout machine fault data to be detected to obtain the fault classification result.

[0051] The present invention provides a small sample fault diagnosis method and device for a turnout machine driven by a multi-source signal. The method collects three-phase action current data of a turnout machine when performing a switching operation between a directional direction and a reverse direction in a real scene; then the three-phase current data is fused with different weights through an adaptive data fusion strategy; the fused current data is divided into a support set and a query set, and the support set contains labeled samples and unlabeled samples; the fused current data is input into a multi-source signal reweighted feature fusion network based on Meta-ACON, and features at different levels are extracted and fused; then the fused features are output into a high-dimensional vector space, and then the original prototype of each category is generated through the support set S; then the distance between each type of prototype is expanded through a fine-tuning center balance strategy, and the intra-class aggregation is enhanced, and the boundaries between prototypes of each category are clearly defined through gradual iterative optimization; the class center fine-tuning prototype network model is trained and iteratively optimized, and the optimized class center fine-tuning prototype network model is used to classify the fault categories of the test set data. The present invention adopts the three-phase fused current data of the switch machine in the real scene to enrich the input information of different fault types. The proposed feature fusion network improves the feature extraction ability and generalization performance of the model. The final trained model has high fault diagnosis accuracy and efficiency, and realizes intelligent small sample fault diagnosis of the turnout switch machine.

[0052] In summary, since the present invention adopts the above technical solution, the present invention has the following beneficial technical effects:

[0053] (1) The present invention fuses three-phase current data, then extracts and fuses features at different levels on the fused data, and uses a class center-based fine-tuning prototype network to perform fault diagnosis on the switch machine. The method can be used for small sample intelligent fault diagnosis of multiple types of turnout switch machines, greatly improving the fault diagnosis efficiency and recognition accuracy;

[0054] (2) In this invention, an adaptive data fusion strategy is proposed to allocate weights according to the correlation and difference between different types of data, which effectively solves the problem of incomplete representation of fault information of a single signal. The effectiveness of this strategy is verified by comparing it with a single-source signal.

[0055] (3) In the present invention, a multi-source signal re-weighted feature fusion network is constructed, and the hidden features in the fusion signal are extracted through three branches, which effectively improves the model fault recognition ability;

[0056] (4) In this invention, a fine-tuning center balance strategy is proposed, which can effectively expand the distance between prototypes and enhance the intra-class aggregation, thereby improving the classification accuracy of the model;

[0057] (5) The intelligent fault diagnosis device for turnout machines designed in the present invention has the significant advantages of high diagnostic accuracy, strong generalization performance and high stability, which can meet the actual application needs of turnout machines and also has broad application potential in other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0059] Figure 1 It is a flow chart of a small sample fault diagnosis method of a turnout machine driven by a multi-source signal of the present invention.

[0060] Figure 2 It is a schematic diagram of a three-phase action current curve of a turnout switch machine oriented to a reverse direction under normal conditions provided by the present invention.

[0061] Figure 3 It is a schematic diagram of a three-phase operating current curve when the turnout machine has a relay 12 or 34 coil with poor contact and an operating rod with a gap in the process from directional to reverse provided by the present invention.

[0062] Figure 4 It is a schematic diagram comparing the three-phase current directed to the reverse direction and the fused current curve under normal state provided by the present invention.

[0063] Figure 5 It is a schematic diagram of the working principle of the multi-source signal re-weighted feature fusion network provided by the present invention.

[0064] Figure 6 It is a schematic diagram of the fine-tuning center balance strategy provided by the present invention.

[0065] Figure 7 It is a schematic diagram of a specific execution flow of a small sample fault diagnosis of a turnout machine based on a class-center fine-tuning prototype network provided by the present invention.

[0066] Figure 8 It is a structural block diagram of a small sample fault diagnosis device for a turnout machine based on a class-center fine-tuning prototype network provided by the present invention.

[0067] Fig. 9 It is a structural block diagram of the current data acquisition module provided by the present invention. DETAILED DESCRIPTION

[0068] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0069] See also Figure 1 The present invention provides a small sample fault diagnosis method for a turnout machine driven by a multi-source signal, comprising the following steps:

[0070] S1: Collect the three-phase action current data of the turnout switch when the switch machine performs switching operation between directional and reverse directions; in the present invention, the three-phase action current data includes normal state and nine fault states. The switch machine uses the ZYJ7 type electro-hydraulic switch machine. The collection point is between the output of the phase failure protector and the first turnout starting relay. The three-phase action current of the turnout switch machine refers to the data generated by the switch machine when it switches between the positioning and reverse positions. Since the turnout switch machine is powered by three-phase electricity, the current curves of the A, B, and C phases under different states can be collected. Since the protection system of the switch machine sets the time from the occurrence of a fault to power failure to 13s, a sampling period of the experiment is 13s, and 650 data points are collected for each sample at a sampling frequency of 50Hz; Figure 2 It is a schematic diagram of the three-phase operating current curve of the turnout switch machine under normal conditions provided by the present invention. The A, B, and C phase currents all experience a short-term sudden increase process from 0 to about 5A, and this process is the turnout unlocking stage. Then the three-phase currents all drop by about 2A and last for about 3s. This process indicates that the switch machine is in a switching state. After the conversion is completed, the switch machine enters a locked state. During this process, the B-phase current drops to 0, and the A-phase and C-phase currents drop to about 0.5A. Finally, it enters the slow-release stage of the starting relay 1DQJ, the B-phase current remains at 0, the A-phase and C-phase current values ​​are still 0.5A and last for about 1s, and finally drop to 0; Figure 3It is a schematic diagram of a three-phase operating current curve when the turnout machine relay 12 or 34 coil does not switch levels and the operating rod is notched during the directional to reverse process of the present invention. Figure 3 (a) is the case where the relay 12 or 34 coil does not rotate. It can be seen that the A and B phases have a peak value at about 0.8s, and the current reaches about 0.65A. After that, the current drops to about 0.62A and maintains this value for about 1.5s. Then the current suddenly drops to about 0.1A, and finally slowly drops to 0. As for the C phase, because the relay contacts are in poor contact, the coil does not rotate, and no current is generated. Figure 3 (b) is the case where the operating rod is stuck. It can be seen that the current of phases A, B, and C suddenly increases to about 5.5A at 0.8s, and then the current value quickly drops back to 1.8A and remains unchanged;

[0071] S2: The collected three-phase current data are fused with different weights through an adaptive data fusion strategy;

[0072] S3: dividing the fused three-phase current data into a training set and a test set, and dividing the current data in the training set into a support set S and a query set Q; in the present invention, dividing the fused training set current data into a support set S and a query set Q specifically includes the following sub-steps:

[0073] Step S0301: randomly divide the fused three-phase current data into a training set and a test set at a ratio of 5:2; mark the fused training set current data, and divide the training set into a healthy state and 9 fault states. Taking the turnout switch operating lever oriented to the reverse direction as an example, the healthy state is recorded as C0, the indoor A, B, C phase circuit short circuit or poor contact is recorded as C1, C2, C3 respectively, the 12 or 34 coils of the DQJ do not turn the pole is recorded as C4, the outdoor diode is open circuit is recorded as C5, the foreign body between the outdoor point rail and the basic rail is recorded as C6, the operating rod card gap is recorded as C7, and multiple switch operations and the action rod are broken or not connected are recorded as C8 and C9 respectively; then the current data is divided into 10 types, each type contains 50 samples, among which 20 samples are randomly selected as the support set S, and the remaining 30 samples are classified as the query set Q.

[0074] S0302: Each time, randomly select 1 or 5 samples in the support set as labeled samples for training, and then select 5 samples in the support set as unlabeled samples to correct the position of the original prototype;

[0075] S4: Input the fused current data into the multi-source signal reweighted feature fusion network based on Meta-ACON, that is, input the support set S and query set Q contained in the training set into the multi-source signal reweighted feature fusion network based on Meta-ACON, extract and fuse features at different levels, and then output the fused features into the high-dimensional vector space to form a feature vector, and generate the original prototype of each category by calculating the average value of the feature vector of the labeled sample in the support set S. The overall characteristics of the samples of each category are represented by the prototype. The distance between the new sample and the prototype of each category can be calculated to classify the new sample. The high-dimensional vector space is specifically the feature vector, which prepares for the subsequent prototype generation;

[0076] S5: By fine-tuning the center balance strategy to strengthen the repulsion between the original prototypes of each category, the distribution distance between the original prototypes of each category is expanded, and the intra-category aggregation is enhanced. The class center fine-tuning prototype network model CFPN is gradually iteratively optimized to make the boundaries between the original prototypes of each category clear and obtain a new prototype distribution; CFPN is not in the best state at the beginning, and it needs to be continuously optimized and adjusted to have the best classification performance. Iterative optimization refers to continuously adjusting the hyperparameters in the CFPN model during training and improving the classification ability.

[0077] S6: Train and iteratively optimize the class center fine-tuning prototype network model CFPN, input the query set data into the class center fine-tuning prototype network model CFPN after training and iterative optimization, if the class center fine-tuning prototype network model CFPN can achieve good classification effect on the query set, specifically manifested in good accuracy, stability and generalization ability, train and iteratively optimize the class center fine-tuning prototype network model, at this time the optimized class center fine-tuning prototype network model can classify the fault category of the test set data, then the class center fine-tuning prototype network model CFPN after training and iterative optimization is used for small sample fault diagnosis of turnout machines.

[0078] In step S2, the three-phase current data is fused with different weights through an adaptive data fusion strategy, which specifically includes the following sub-steps:

[0079] S0201: Assume the input signal to be {x in (l), i = 1, ..., M; n = 1, ..., N; l = 1, ..., L}, the correlation entropy is used to measure the correlation degree and shared information between the two current signals. The shared information represents the part of the current signal with similar trends during the entire switch operation process. The improved correlation entropy E i The calculation formula is as follows:

[0080]

[0081] S0202: Then adaptively calculate the weight ω of each signal according to the correlation entropy i (i.e. initialization weight coefficient), the formula is as follows:

[0082]

[0083] S0203: Finally, weights are assigned to corresponding signals to obtain fused signals. The formula is as follows:

[0084]

[0085] In order to verify whether the diagnostic effect of fused data as input data set is better than that of single source data, we input three-phase current data and fused data into ProtoNet and CFPN models for experiments. The experiments were conducted in two action states: directional to reverse and reverse to directional, and 20 samples were trained in 1-shot and 5-shot cases respectively. The results are shown in Tables 1 and 2.

[0086] Table 1 Fault diagnosis accuracy of three-phase current and fusion current in 1-shot case (%)

[0087]

[0088] Table 2 Fault diagnosis accuracy of three-phase current and fusion current in 5-shot case (%)

[0089]

[0090]

[0091] It can be seen from Table 1 above that no matter in which conversion state, when the input data is a fusion signal, the accuracy of ProtoNet and CFPN is improved. In the conversion state from directional to reverse, the accuracy of the A-phase, B-phase and C-phase currents in CFPN are 92.32%, 91.66% and 92.13% respectively, while the accuracy of the fused current can reach 98.78%, which is a maximum increase of 7.12% compared to the single-phase current. Not only that, in the case of reverse to directional, the accuracy of the fused current in CFPN is increased by up to 7.3%. Comparing Table 1 and Table 2, it can be seen that whether in the case of 1-shot or 5-shot, the fused current can effectively improve the accuracy in the PN and CFPN models, and the effect is significant. Figure 4It is a schematic diagram of the comparison between the three-phase current directed to the reverse direction and the fused current curve under normal conditions provided by the present invention. It can be seen that the method of the adaptive data fusion strategy proposed in the application is effective. This is because the data fusion strategy assigns different weights to each type of current based on the correlation entropy, so that the fused current effectively extracts the shared information in the three-phase current and merges the difference parts, maximizing the integration of the effective information in the single-source current, and can provide more effective information for the subsequent model training, thereby improving the diagnostic accuracy of the model.

[0092] In step 4 of the present invention, the support set S and the query set Q contained in the training set are input into a multi-source signal reweighted feature fusion network based on Meta-ACON to perform feature extraction and feature fusion at different levels. In the present invention, the support set S and the query set Q are input into a multi-source signal reweighted feature fusion network that can learn whether to activate neurons (Meta-Activate OrNot, Meta-ACON), and feature extraction and feature fusion at different levels are performed. Then, the fused features are output to a high-dimensional vector space, and finally, the original prototype of each category is generated through the support set S, including the following steps:

[0093] S0401: After the data is subjected to one-dimensional convolution and normalization operations, it is input into three branches for convolution at different levels;

[0094] S0402: Perform maximum pooling on the first branch to obtain local features F max , the second branch performs average pooling to obtain the global feature F avg , the third branch extracts edge features to obtain edge features F DCMA . The third branch contains four identical modules; Figure 5 It is a schematic diagram of the working principle of the multi-source signal reweighted feature fusion network provided by the present invention. Each module is composed of a convolutional neural network CNN module and a dual-channel Meta-ACON (DCMN) module, which is a specific display of the contents of the four identical modules (modules) in the third branch. The dual-channel Meta-ACON module changes the selection of the activation function by switching factors, that is, determines whether to activate neurons, thereby adaptively controlling the activation to be linear or nonlinear;

[0095] S0403: The features extracted from the three branches are fused with different weights. The formula is as follows:

[0096] F f =α 1 ·F max +α 2 ·F avg +α 3 ·FDCMA ;

[0097]

[0098] α i represents the normalized weight of the i-th branch, ∑α i =1; w i represents the initialization weight coefficient of the i-th branch, w j represents the initialization weight coefficient of the j-th branch, and the initial value is 1;

[0099] S0404: After a squeeze-excitation attention module focuses on key feature information, the final feature map is flattened into one dimension to generate a more accurate original prototype.

[0100] In step 5 of the present invention, the class center fine-tuning prototype network CFPN model is gradually iterated and optimized to make the boundaries between the original prototypes of each category clear, and the prototype distribution is obtained, such as Figure 6 The figure shows a schematic diagram of the fine-tuning center balance strategy provided by the present invention, which uses the "repulsion" between different prototypes to fine-tune the distance between prototypes. Then the improved center loss is combined with the cross entropy loss to expand the distance between prototypes A, B, C, and D, while enhancing the intra-class aggregation, and then gradually iterate and optimize to make the boundaries between prototypes of each category clear, which specifically includes the following steps:

[0101] S0501: Using the Euclidean distance function f d Calculate the labeled sample x i (label the sample x i is the support set sample, such as Figure 6 The small circle in the figure is the same as the prototype C k The "repulsion" between other prototypes outside the , and then calculate the distribution probability p between different prototypes through the Softmax function k , the formula is as follows:

[0102]

[0103] In the formula, Num represents the number of generated prototypes, f θ Indicates mapping a K-dimensional vector to an M-dimensional feature vector, C j Indicates the addition of C k The j-th prototype outside, then improve the center loss L c , making the sample feature distribution of the same category in the space more compact, the formula is as follows:

[0104]

[0105] N in the formula Lrepresents the number of labeled samples in the support set, f(x i ) represents x i Eigenvalues ​​in the feature space;

[0106] S0502: N in the above formula L represents the number of labeled samples in the support set, f(x i ) represents x i The eigenvalue in the eigenspace. c With cross entropy loss L s Combined with the above, prototypes of different categories are kept away from each other, while the distribution within the category is more compact. The calculation formula is as follows:

[0107] L t =L s +λL c ;

[0108] In the formula, λ represents a hyperparameter, which can significantly enhance the discriminative ability of deep features. In the present invention, λ is preferably set to 0.1;

[0109] S0503: Repeat the above steps 5.1-5.2, and repeat the iterative optimization until the boundaries between the prototypes are clear.

[0110] In step S6 of the present invention, the support set is used to train and iteratively optimize the class center fine-tuning prototype network model CFPN model, and the optimized class center fine-tuning prototype network model CFPN is used to classify the fault category of the test set data. If the CFPN model can identify the category to which the sample to be detected belongs and the accuracy rate is above 98%, then the network model can be used to realize the small sample fault diagnosis of the turnout switch; the model calculates the Euclidean distance from the query sample to each prototype and classifies it into the category with the closest distance; in the present invention, in order to verify the effectiveness of the turnout switch fault diagnosis method proposed in the present invention, CFPN is compared with ProtoNet, IPN, MWPN, SSMN and ISSML. The data fusion method has been proven to be effective in the previous text, so the present invention is verified by the fusion current data under the two conversion processes of the turnout switch from directional to reverse and reverse to directional. Tables 3 and 4 respectively show the accuracy of each model in the two conversion processes when the training samples are 10, 15 and 20 for 1-shot and 5-shot experiments. N t Indicates the number of training samples for each class.

[0111] Table 3 Accuracy of each model in the case of orientation to reverse direction (%)

[0112]

[0113] Table 4 Accuracy of each model in the reverse to directional case (%)

[0114]

[0115] As can be seen from Tables 3 and 4, CFPN can achieve the highest accuracy in both directional to reverse and reverse to directional situations, which shows the superiority of CFPN. In addition, the standard deviation range of CFPN is relatively small, indicating that it has good stability. This is because the proposed adaptive data fusion strategy can effectively fuse the three-phase currents, and the multi-source signal reweighted feature fusion network can extract features at different levels of the fused signal. The three branches adopt different feature extraction methods and finally fuse according to the weights obtained by training iterations, effectively utilizing the information in the fused signal, making the fault diagnosis model more stable. At the same time, we can see that as the number of training samples increases, the accuracy of each method also increases accordingly. This is because the more training samples, the more information provided to the model training, which makes the generalization performance of the model better and can more accurately identify the category of the samples in the test set. However, the feature extraction capabilities of some methods are limited, and the final accuracy will not reach a very high level. In short, no matter what the conditions are, CFPN can show good fault diagnosis performance and maintain an accuracy of more than 97%. It can be seen that CFPN can complete the small sample fault diagnosis task well.

[0116] Specifically, Figure 7 The figure shows a schematic diagram of the specific execution process of the small sample fault diagnosis of the turnout machine based on the class center fine-tuning prototype network provided by the present invention. First, the three-phase current signals of the turnout machine under different fault types in the real scene are collected. Then, the adaptive data fusion strategy is used to weight the signal to obtain the fused signal, and the fused signal is divided into the support set S and the query set Q. The samples of these two data sets do not intersect. Then, the support set samples are input into the multi-weighted feature fusion network, and then the features of different branches are extracted and weighted fused to obtain the embedded features, thereby generating the original prototype. Then, the class center prototype fine-tuning strategy is used to readjust the distribution between the prototypes in the training task, so that the boundaries between the prototypes are clearer. Finally, the optimized CFPN model is obtained, and the query set Q is used for verification. In the comparative experiment and cross-domain experiment, CFPN has achieved a high accuracy and good classification effect. The three-phase action current data of the turnout machine is collected by the Hall sensor and then exported from the database. The exported data is fused with different weights, and then the fused data is subjected to feature extraction and fusion at different levels, and the CFPN model is continuously trained and optimized. Finally, the data to be tested is input into the CFPN model for fault diagnosis. The maintenance personnel determine the fault type based on the diagnosis results in order to quickly locate the fault site and perform targeted maintenance.

[0117] By conducting comparative experiments using the three-phase fused current of the turnout machine commutation in real scenarios, the effectiveness and superiority of the turnout machine fault diagnosis method based on the class-center fine-tuning prototype network proposed in the present invention are verified, proving that it can perform high-precision fault diagnosis on the turnout machine and meet the needs of actual applications.

[0118] The present invention also proposes a small sample fault diagnosis device for turnout machines driven by multi-source signals, please refer to Figure 8 The small sample fault diagnosis device for the turnout machine includes a current data acquisition module 801, a data fusion module 802, a feature extraction and fusion module 803 and a fault diagnosis module 804;

[0119] The switch machine current data acquisition module 801 uses three Hall sensors to collect real-time switch machine current data. The specific installation position is between the output of the phase failure protector DBQ and the first switch start relay 1DQJ. The collected current data is stored in the server via industrial Ethernet and then exported from the server by the computer;

[0120] The data fusion module 802 calculates the similarity between currents of different phases using the improved correlation entropy, and assigns different weights to currents of different phases according to the similarity. Finally, the current data of each phase are fused according to their respective weights, providing input data containing richer information for subsequent model training.

[0121] The feature extraction and fusion module 803 uses a multi-weighted feature fusion network to extract features at different levels from the input fusion data to obtain global features, local features and edge features, and then fuses the extracted features with different weights. The fused features form feature vectors in a high-dimensional vector space to facilitate subsequent generation and adjustment of the original prototype.

[0122] Fault diagnosis module 804 uses the CFPN model that has been trained and optimized to perform small sample intelligent fault diagnosis on the fault data of the turnout machine to be detected, and obtains the fault classification result. The specific process is to input the current data to be detected into the trained CFPN model, and then output the corresponding prediction label, and determine the fault type of the input current data according to the number of the prediction label. The healthy state is recorded as C0, the three-phase circuit faults of A, B, and C are recorded as C1, C2, and C3 respectively, the 12 or 34 coils of DQJ are not polarized and recorded as C4, the outdoor diode is open and recorded as C5, the foreign matter between the outdoor point rail and the basic rail is recorded as C6, the operating rod is notched and recorded as C7, and the multiple switch operations and the action rod are broken or not connected and recorded as C8 and C9 respectively.

[0123] Fig. 9It is a structural block diagram of the current data acquisition module provided by the present invention. The current data acquisition module includes: a phase failure protector (DBQ) 901, a protection relay (BHJ) 902, an acquisition card 903, a turnout start relay 904, and a ZYJ7 type electro-hydraulic switch machine 905. The acquisition card 903 includes three Hall sensors (9031, 9032, 9033), a signal terminal 9034, a GND terminal 9035, and a +24V terminal 9036. The ZYJ7 type electro-hydraulic switch machine 905 includes a circuit controller 9051, a display rod 9052, and an action rod 9053. The connection method between each module is as follows: Fig. 9 . The three-phase current is transmitted to the phase failure protector (DBQ) 901 by the three resistors R1, R2, and R3 respectively. The acquisition card inputs the received signal to the turnout relay 904, and finally outputs it to the circuit controller 8051 in the switch machine, driving the indicating rod 9052 and the action rod 9053 to realize the reversing operation. Among them, the phase failure protector (901) and the protection relay (902) prevent the circuit from being burned by excessive current. In the Hall sensor, AN and AP respectively represent the negative terminal and the positive terminal of the first device, and BN and BP respectively represent the negative terminal and the positive terminal of the second device. There is only one device in this circuit. Finally, the three-phase current data collected is transmitted to the server via the industrial Ethernet RJ45 (9037) and stored. The obtained three-phase current data is exported from the server by a computer, and the three-phase current data is fused and used as input data to iteratively optimize the CFPN model. Then, the optimized CFPN model is used to perform fault diagnosis on the data to be detected, and the fault classification result is obtained, which guides the maintenance personnel to quickly locate the fault location for maintenance and prevention. The present invention fuses different weights according to the correlation and difference between the three-phase currents, and then effectively extracts and fuses the data features, effectively improving the classification ability and classification accuracy of the CFPN fault diagnosis model, and realizing intelligent fault diagnosis of turnout machines.

[0124] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A small sample fault diagnosis method for a turnout machine driven by multi-source signals, characterized in that: The following steps are involved: Step 1: Collect the three-phase action current data of the turnout machine when it performs switching operation between directional and reverse directions; Step 2: The collected three-phase current data are fused with different weights through an adaptive data fusion strategy; Step 3: Divide the fused three-phase current data into a training set and a test set, wherein the training set includes a support set S and a query set Q; Step 4: Input the support set S and query set Q contained in the training set into the multi-source signal reweighted feature fusion network based on Meta-ACON, perform feature extraction and feature fusion at different levels, output the fused features in the form of vectors to the high-dimensional vector space to form feature vectors, and generate the original prototype of each category by calculating the average value of the feature vectors of the labeled samples in the support set S; Step 5: Strengthen the repulsion between the original prototypes by fine-tuning the center balance strategy, so as to expand the distribution distance between the original prototypes of each category, and gradually iterate and optimize the class center fine-tuning prototype network model CFPN to make the boundaries between the original prototypes of each category clear, and obtain a new prototype distribution; Step 6: Train and iteratively optimize the class center fine-tuning prototype network model CFPN, and input the query set data into the class center fine-tuning prototype network model CFPN after training and iterative optimization. If the class center fine-tuning prototype network model CFPN can achieve good classification effect on the query set, the class center fine-tuning prototype network model CFPN after training and iterative optimization will be used for small sample fault diagnosis of turnout machines.

2. A multi-source signal driven turnout machine small sample fault diagnosis method as claimed in claim 1, characterized in that: The three-phase action current data include the health status of the switch machine and nine fault states. The switch machine uses a ZYJ7 type electro-hydraulic switch machine, and the collection point is set between the output of the phase failure protector and the first switch starting relay.

3. A small sample fault diagnosis method for a turnout machine driven by a multi-source signal as claimed in claim 1, characterized in that: In step 2, fusing the collected three-phase current data with different weights through an adaptive data fusion strategy includes the following steps: Step 2.1: Assume the input three-phase current data signal to be {x in (l), i = 1, ..., M; n = 1, ..., N; l = 1, ..., L}, where M represents the number of signal types, N represents the number of samples of each type of signal, and L represents the sequence length. The correlation entropy is used to measure the correlation degree and shared information between two current signals. The improved correlation entropy E i The calculation formula is as follows: Where log(·) represents the logarithmic function, H(·) represents the distance function; Step 2.2: Adaptively calculate the weight of each current signal according to the correlation entropy, the formula is as follows: Where exp represents the exponential function; Step 2.3: Assign weights to the corresponding current signals to obtain fused signals. The formula is as follows:

4. A multi-source signal driven turnout machine small sample fault diagnosis method as claimed in claim 1, characterized in that: In step 3, the fused three-phase data is divided into a training set and a test set, including the following steps: Step 3.0: The fused three-phase current data is randomly divided into a training set and a test set in a ratio of 5:2; Step 3.1: Divide the training set into healthy state and 9 fault states, each type contains 50 samples, randomly select 20 samples as the support set S, and the remaining 30 samples are classified as the query set Q; Step 3.2: Each time, randomly select 1 or 5 samples in the support set S as labeled samples to generate the original prototype, and then select 5 samples in the support set S as unlabeled samples to correct the position of the original prototype.

5. A small sample fault diagnosis method for a turnout machine driven by multi-source signals as claimed in claim 1, characterized in that: In step 4, the support set S and the query set Q contained in the training set are input into the multi-source signal reweighted feature fusion network based on Meta-ACON to perform feature extraction and feature fusion at different levels, and feature extraction and feature fusion at different levels are performed, including the following steps: Step 4.1: First, perform one-dimensional convolution and normalization operations on the training set, and then input the operated data into the three branches for convolution at different levels; Step 4.2: Perform maximum pooling on the first branch to obtain local features F max , the second branch performs average pooling to obtain the global feature F avg , the third branch extracts edge features to obtain edge features F DCMA , where the third branch contains four identical modules, each of which consists of a convolutional neural network CNN module and a dual-channel Meta-ACON module; Step 4.3: The features extracted from the three branches are fused with different weights. The formula is as follows: F f =α1·F max +α2·F avg +α3·F DCMA ; Among them, α i represents the normalized weight of the i-th branch, ∑α i =1; w i represents the initialization weight coefficient of the i-th branch, w j represents the initialization weight coefficient of the j-th branch; In step 4.4, the fused features are passed through a squeeze-excitation attention module, and the final feature map is flattened into a one-dimensional vector to generate a more accurate original prototype.

6. A small sample fault diagnosis method for a turnout machine driven by multi-source signals as claimed in claim 1, characterized in that: In step 5, the class center fine-tuning prototype network CFPN model is gradually iterated and optimized so that the boundaries between the original prototypes of each category are clear, and the prototype distribution is obtained, including the following steps: Step 5.1: Using the Euclidean distance function f d Calculate the labeled sample x i Except for the prototype C k The "repulsion" between other prototypes outside the , and then calculate the distribution probability p between different prototypes through the Softmax function k , the formula is as follows: In the formula, Num represents the number of generated prototypes, f θ Indicates mapping a K-dimensional vector to an M-dimensional feature vector, C j Indicates the addition of C k The j-th prototype outside, then improve the center loss L c , making the sample feature distribution of the same category in the space more compact, the formula is as follows: N in the formula L represents the number of labeled samples in the support set, f(x i ) represents x i Eigenvalues ​​in the feature space; Step 5.2: The improved center loss L c With cross entropy loss L s Combined to get a new loss function L t , which makes prototypes of different categories far away from each other and makes the distribution within the category more compact. The calculation formula is as follows: THE t =L s +λL c ; In the formula, λ represents a hyperparameter; Step 5.3: Repeat steps 5.1-5.3 above and iterate the optimization until the boundaries between the prototypes are clear.

7. The multi-source signal driven turnout machine small sample fault diagnosis method according to claim 1, characterized in that: In step 6, if the class center fine-tuning prototype network model CFPN can identify the category to which the sample to be detected belongs and the accuracy rate is above 98%, then the network model can be used to realize small sample fault diagnosis of turnout machines.

8. A small sample fault diagnosis device for turnout machines driven by multi-source signals, characterized in that: The fault diagnosis device comprises a switch machine current data acquisition module, a data fusion module, a feature extraction and fusion module and a fault diagnosis module, and the fault diagnosis device completely implements the small sample fault diagnosis steps of the turnout machine driven by multi-source signals as described in claims 1 to 7; The switch machine current data acquisition module is used to obtain the three-phase induction action current of the switch machine in real time through the Hall current sensor installed in the line between the phase failure protector output and the first switch start relay, and store the current data in the server through the industrial Ethernet; The data fusion module uses the improved correlation entropy to calculate the similarity between currents of different phases, and assigns different weights to currents of different phases according to the similarity. Finally, the current data of each phase are fused according to their respective weights to provide input data containing richer information for subsequent model training. The feature extraction and fusion module uses a multi-weighted feature fusion network to extract features at different levels from the input fusion data to obtain global features, local features, and edge features. The extracted features are then fused with different weights. The fused features form feature vectors in a high-dimensional vector space to facilitate subsequent generation and adjustment of the original prototype. The fault diagnosis module uses the class center fine-tuning prototype network model CFPN to perform small sample diagnosis on the turnout machine fault data to be detected to obtain the fault classification result.

Images