Turnout indication connecting rod abnormity detection method and system
By constructing a fault prediction network model and training a feature extractor using self-supervised learning and prior expert experience, the problem of abnormal detection of turnout connecting rods was solved, enabling early detection of abnormalities in the connecting rods and ensuring safe train operation.
Patent Information
- Application Number
- CN202511511902.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-22
- Publication Date
- 2026-01-23
AI Technical Summary
The switch indicates abnormal faults such as cracks and fractures in the connecting rod caused by long-term strong vibration, which affect the normal train operation plan and are difficult to detect and prevent effectively with existing technology.
A fault prediction network model is constructed, and a feature extractor and classifier are trained using self-supervised learning. Anomaly detection is performed using collected vertical vibration acceleration data. The feature extractor is trained by combining masked samples and prior expert experience to achieve early anomaly detection of the connecting rod.
It can detect minor anomalies in the turnout indicator connecting rod in a timely manner, avoid malfunctions during train operation, ensure train safety, reduce server computation load, and improve the real-time performance of output results.
Smart Images
Figure CN121375901A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of turnout inspection technology, and in particular to a method and system for detecting abnormalities in turnout indicator connecting rods. Background Technology
[0002] Turnouts, which change the position of the switch rails via a switch machine, guide trains to different tracks and are an important component of railway lines. The turnout indicator connecting rod (hereinafter referred to as the "indicator rod") connects the turnout switch rail (spot rail) and the switch machine indicator rod. Its function is to reflect the actual position of the switch rail, indicating the turnout's correct or disconnected position after the switch rail (spot rail) is in close contact with the stock rail (wing rail). When a train passes through the turnout, the wheelset impacts the switch rail and the spot rail, causing the indicator rod to experience strong lateral and vertical vibrations. Prolonged strong vibrations can cause cracks or even breakage of the connecting rod, bushing disengagement, loose nuts, and other abnormal malfunctions, leading to the turnout losing its indication and affecting normal train operation plans. Summary of the Invention
[0003] The purpose of this invention is to detect abnormalities in turnout indicator connecting rods, and to provide a method and system for detecting abnormalities in turnout indicator connecting rods.
[0004] The objective of this invention can be achieved through the following technical solutions: As a first aspect of the present invention, a method for detecting abnormalities in the turnout indicator connecting rod is provided. The method inputs the collected vertical vibration acceleration data of the connecting rod when the train passes through the turnout into the fault prediction network model that has been trained and fine-tuned, and outputs the fault prediction result. The fault prediction network model includes a feature extractor and a classifier composed of a fully connected network. The feature extractor is trained using a self-supervised representation learning method based on a sample set of historical vibration acceleration data of the connecting rod of a normal turnout. The feature extractor and the classifier are trained and fine-tuned based on a sample set of historical vibration acceleration data of the connecting rod of a faulty turnout.
[0005] As a preferred technical solution, the feature extractor training process is as follows: A learning sample set is formed by collecting historical vibration acceleration data of the connecting rod of a normal turnout. A masking operation is performed on the learning samples in the learning sample set to divide the learning samples into multiple sub-samples on an average basis. Based on the set masking rate, multiple segments of data in each sub-sample are randomly selected and set to 0. The masked subsamples are concatenated with the unmasked subsamples in their original order to obtain the masked sample. The masked sample is then input into the feature extractor to obtain the feature vector. The feature vectors obtained by the feature extractor are input into a data reconstructor based on a fully connected neural network to reconstruct sample data, thus obtaining reconstructed samples. Calculate prior expert experience features for the learning samples and obtain prior expert feature vectors; The feature vectors obtained by the feature extractor are input into the prior feature reconstructor based on a fully connected neural network to reconstruct them into prior feature vectors, thus obtaining the reconstructed feature vectors. The feature extractor is trained using the squared error between the reconstructed sample and the original sample, plus the squared error between the reconstructed feature vector and the prior expert feature vector of the original sample, as the loss function.
[0006] As a preferred technical solution, the feature extractor is a one-dimensional deep convolutional network: the first convolutional layer has a kernel size of 64, a stride of 8, and 16 output channels; the second convolutional layer has a kernel size of 3, a stride of 1, and 32 output channels; the third convolutional layer has a kernel size of 3, a stride of 1, and 64 output channels; the fourth, fifth, and sixth convolutional layers have a kernel size of 3, a stride of 1, and 64 output channels; each convolutional layer is followed by a 2×1 pooling layer with a stride of 2; and the feature vector is output through a fully connected layer.
[0007] As a preferred technical solution, the prior expert experience features calculated for the sample include the sample mean, variance, root mean square amplitude, peak value, minimum value, peak index, margin index, kurtosis, mean, root mean square value, standard deviation, maximum value, waveform index, impulse index, and skewness.
[0008] As a preferred technical solution, the feature extractor and classifier are fine-tuned as follows: A training sample set is formed by collecting historical vibration acceleration data of the connecting rod of the faulty turnout. The training sample includes a sequence of vibration acceleration data when a train passes and a fault type label. The vibration acceleration data sequence of the train passing through the training sample is input into the fault prediction network model for prediction. The output of the fault prediction network model is processed by softmax to obtain the predicted fault category. The feature extractor and classifier networks are fine-tuned by using the cross-entropy loss of the predicted fault category and fault type label as the loss function.
[0009] As a second aspect of the present invention, a turnout indicator connecting rod abnormality detection system is provided, the system comprising: a vibration data acquisition device, a client for accessing data, and a server; The vibration data acquisition device is connected to the server and is used to acquire data representing the vibration acceleration of the connecting rod and send the acquired data to the server. The server performs the turnout indicator connecting rod anomaly detection as described above, which is used to process vibration acceleration data. The client connects to the server to allow users to view the test results and the collected vibration data.
[0010] As a preferred technical solution, a fault prediction network model is deployed in the server, and the fault prediction network model includes a feature extractor and a classifier; During the offline model training phase, the fault prediction network model trains the feature extractor using a self-supervised representation learning method based on the historical vibration acceleration data sample set of the normal turnout representative connecting rod; and trains and fine-tunes the feature extractor and classifier based on the historical vibration acceleration data sample set of the faulty turnout representative connecting rod. Online prediction phase: Input the collected vibration acceleration data of the turnout connecting rod into the fault prediction network model, and output the predicted fault type.
[0011] As a preferred technical solution, the specific training process of the feature extractor by the server during the offline model training phase is as follows: A learning sample set is formed by collecting historical vibration acceleration data of the connecting rod of a normal turnout. A masking operation is performed on the learning samples in the learning sample set to divide the learning samples into multiple sub-samples on an average basis. Based on the set masking rate, multiple segments of data in each sub-sample are randomly selected and set to 0. The masked subsamples are concatenated with the unmasked subsamples in their original order to obtain the masked sample. The masked sample is then input into the feature extractor to obtain the feature vector. The feature vectors obtained by the feature extractor are input into a data reconstructor based on a fully connected neural network to reconstruct sample data, thus obtaining reconstructed samples. Calculate prior expert experience features for the learning samples and obtain prior expert feature vectors; The feature vectors obtained by the feature extractor are input into the prior feature reconstructor based on a fully connected neural network to reconstruct them into prior feature vectors, thus obtaining the reconstructed feature vectors. The feature extractor is trained using the squared error between the reconstructed sample and the original sample, plus the squared error between the reconstructed feature vector and the prior expert feature vector of the original sample, as the loss function.
[0012] As a preferred technical solution, the server fine-tunes the feature extractor and classifier during the offline model training phase as follows: A training sample set is formed by collecting historical vibration acceleration data of the connecting rod of the faulty turnout. The training sample includes a sequence of vibration acceleration data when a train passes and a fault type label. The vibration acceleration data sequence of the train passing through the training sample is input into the fault prediction network model for prediction. The output of the fault prediction network model is processed by softmax to obtain the predicted fault category. The feature extractor and classifier networks are fine-tuned by using the cross-entropy loss of the predicted fault category and fault type label as the loss function.
[0013] As a preferred technical solution, the vibration data acquisition device is connected to the server for data exchange. The vibration data acquisition device has a vibration acceleration sensor installed on the connecting rod to collect the vertical acceleration data of the connecting rod when the train passes through the turnout. Offline model training phase: Acceleration data representing each fault type of the connecting rod are collected to form a training sample set; acceleration data representing the fault-free connecting rod are collected to form a learning sample set; Online prediction phase: The collected real-time vibration acceleration data of the turnout indicator connecting rod is transmitted to the server.
[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a method for detecting anomalies in turnout indicator connecting rods. A fault prediction network model is constructed, comprising a feature extractor and a classifier, to predict fault types based on the vertical acceleration data of the connecting rods when a train passes through the turnout. The feature extractor is trained using self-supervised learning. Masking is performed on the training samples to obtain masked samples, and the feature vectors obtained from the masked samples are reconstructed. Simultaneously, prior expert experience is used to reconstruct the feature vectors, resulting in reconstructed feature vectors. The feature extractor is trained using the errors between the reconstructed samples and the original samples, as well as the errors between the reconstructed feature vectors and the original sample feature vectors. This combination of masking and prior expert experience enables the fault prediction network model to detect minor initial anomalies in the turnout indicator connecting rods, preventing train malfunctions from disrupting train schedules. Furthermore, based on the initially trained feature extractor, this application further fine-tunes the feature extractor and classifier using fault data to obtain the final fault prediction network model, enabling training of the classifier using a small number of fault samples. This invention employs offline training of the feature extractor and classifier models, reducing server computation and improving the real-time performance of the output results. Attached Figure Description
[0015] Figure 1 This is a flowchart of the turnout indicator connecting rod abnormality detection method of the present invention.
[0016] Figure 2 This is a flowchart for training a feature extractor for detecting abnormalities in turnout connecting rods. Detailed Implementation
[0017] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0018] Example 1 This invention proposes a method for detecting abnormalities in the connecting rod of a turnout. For example... Figure 1 As shown, a vibration acceleration sensor is installed on the indicator connecting rod to collect vertical acceleration data of the rod when a train passes the turnout. Acceleration data for each fault type of the indicator connecting rod are collected to form a training sample set. Acceleration data for a fault-free indicator connecting rod are collected to form a learning sample set. First, a feature extractor is trained on the learning sample set using a self-supervised representation learning method. Then, the model is trained on the training sample set, using the feature extractor to extract features, and combined with a deep neural network for fault classification, obtaining a fine-tuned feature extractor and classifier as the fault prediction network model. Subsequently, the collected acceleration data of the indicator connecting rod is input into the fault prediction network model, which outputs the predicted fault type. After a fault is detected, turnout maintenance personnel can perform targeted repairs on the indicator connecting rod, replacing components to prevent the indicator connecting rod fault from affecting train operation.
[0019] Collect historical vibration acceleration data of the connecting rod of a normal turnout. This forms the learning sample set. For a certain train passing by The sequence of vibration acceleration data at each moment. The training sample set contains a total of... Group data.
[0020] Collect historical vibration data of the connecting rod of a faulty turnout. This constitutes the training sample set. This is a sequence of vibration acceleration data as a train passes by. The labels are for fault types. The training sample set contains... Group data.
[0021] like Figure 2 As shown, the feature extractor training process is as follows: First, the learning samples Perform a masking operation. The specific implementation is as follows: [The text abruptly ends here, likely due to an incomplete Divide into equal parts Subsamples. The number of segments is set based on the sample length and sensor sampling rate. Randomly selected segments... All sub-segments of data are set to 0, i.e., a masking operation. For the mask rate, choose in practice f =0.5.
[0022] The masked subsamples are concatenated with the unmasked subsamples in their original order to obtain the masked sample. This masked sample is then input into the feature extractor. Obtaining feature vectors . The feature vector Input data reconstructor obtains reconstructed samples .
[0023] The data reconstruction engine uses an n-layer fully connected neural network, with input dimensions and feature vectors. The output dimension is the same as the input learning sample. Without loss of generality, n can be chosen as 5.
[0024] To simultaneously utilize prior expert experience, 15 features, i.e., prior expert feature vectors, are calculated for the samples according to the formulas in Table 1. { Without loss of generality, other prior expert experience features can be calculated.
[0025] Table 1. Prior Feature Samples eigenvectors Input the prior feature reconstructor to obtain the reconstructed feature vector .
[0026] The prior feature reconstructor uses an n-layer fully connected neural network, with input dimension and feature vector. The output dimension is the same as that of the prior expert feature vector. Same. Without loss of generality, n can be 5.
[0027] Feature extractor This is a one-dimensional deep convolutional network. The first convolutional layer has a kernel size of 64, a stride of 8, and 16 output channels; the second convolutional layer has a kernel size of 3, a stride of 1, and 32 output channels; the third convolutional layer has a kernel size of 3, a stride of 1, and 64 output channels; the fourth, fifth, and sixth convolutional layers all have a kernel size of 3, a stride of 1, and 64 output channels. Each convolutional layer is followed by a 2×1 pooling layer with a stride of 2. Finally, a fully connected layer outputs the feature vector.
[0028] Reconstructed Samples Compared with the original sample Squared error plus reconstructed feature vector Compared with the original sample Prior expert feature vectors The squared error is used as the training loss function to train the feature extractor.
[0029] During the training process, 10 samples are randomly selected from the learning sample set to form a batch of sample data. The optimizer is Adam, the number of training rounds is 20, and the parameter learning rate is 0.01.
[0030] The next step is to fine-tune the feature extractor obtained after training.
[0031] The fine-tuning process involves constructing a fault prediction network using a classifier comprised of a feature extractor and a fully connected network. The network output is then processed using softmax to predict the fault category. Cross-entropy loss is used as the loss function.
[0032] Ten samples were selected from the training sample set to form a batch of sample data. The optimizer was Adam, the number of training rounds was 20, and the parameter learning rate was 0.001.
[0033] The feature extractor and classifier networks were obtained through training and fine-tuning.
[0034] For a newly acquired vibration acceleration data sequence The input fault prediction network is used to obtain predicted faults.
[0035] If no fault is predicted, the switch indicates that the connecting rod is currently in normal operating condition. If a fault is predicted, the switch indicates that the connecting rod is currently faulty. Switch maintenance personnel will repair or replace the fault based on the predicted fault type.
[0036] Example 2 As another embodiment of the present invention, this embodiment also provides a turnout indicator connecting rod anomaly detection system for implementing the method described in Embodiment 1 above, including a vibration data acquisition device, a viewing client, and a server. The viewing client is connected to the server and is used to enable users to view detection results, view collected vibration data, modify configurations, etc. The vibration data acquisition device is connected to the server and is used to collect vibration acceleration data of the indicator connecting rod and send the collected data to the server. The server is used to process the vibration acceleration data, detect anomalies in the indicator connecting rod, store data, and query it.
[0037] An acceleration sensor with a 500G range is fixed to the turnout indicator connecting rod using a clamp-type clip. It measures the vibration acceleration of the connecting rod perpendicular to the track bed. The data acquisition device collects sensor data at a frequency of 1000Hz each time a train passes, saves it as a file, and transmits it to a server via Ethernet, mobile network, or wireless LAN. The server has large-capacity storage and is equipped with a database such as MySQL or SQL Server.
[0038] The server executes the turnout indicator connecting rod abnormality detection method as described in Example 1 above. One server can receive and process data from multiple vibration acquisition devices.
[0039] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0040] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A method for detecting abnormalities in a turnout indicator connecting rod, characterized in that, The method inputs the collected vertical vibration acceleration data of the connecting rod when the train passes through the turnout into the fault prediction network model that has been trained and fine-tuned, and outputs the fault prediction result. The fault prediction network model includes a feature extractor and a classifier composed of a fully connected network. Based on the historical vibration acceleration data sample set of the normal turnout representative connecting rod, the feature extractor is trained using a self-supervised representation learning method. Based on the historical vibration acceleration data sample set of the connecting rod of the faulty turnout, the feature extractor and classifier are trained and fine-tuned.
2. The method for detecting abnormalities in a turnout indicator connecting rod according to claim 1, characterized in that, The specific training process for the feature extractor is as follows: A learning sample set is formed by collecting historical vibration acceleration data of the connecting rod of a normal turnout. A masking operation is performed on the learning samples in the learning sample set to divide the learning samples into multiple sub-samples on an average basis. Based on the set masking rate, multiple segments of data in each sub-sample are randomly selected and set to 0. The masked subsamples are concatenated with the unmasked subsamples in their original order to obtain the masked sample. The masked sample is then input into the feature extractor to obtain the feature vector. The feature vectors obtained by the feature extractor are input into a data reconstructor based on a fully connected neural network to reconstruct sample data, thus obtaining reconstructed samples. Calculate prior expert experience features for the learning samples and obtain prior expert feature vectors; The feature vectors obtained by the feature extractor are input into the prior feature reconstructor based on a fully connected neural network to reconstruct them into prior feature vectors, thus obtaining the reconstructed feature vectors. The feature extractor is trained using the squared error between the reconstructed sample and the original sample, plus the squared error between the reconstructed feature vector and the prior expert feature vector of the original sample, as the loss function.
3. The method for detecting abnormalities in a turnout indicator connecting rod according to claim 2, characterized in that, The feature extractor is a one-dimensional deep convolutional network: the first convolutional layer has a kernel size of 64, a stride of 8, and 16 output channels; the second convolutional layer has a kernel size of 3, a stride of 1, and 32 output channels; the third convolutional layer has a kernel size of 3, a stride of 1, and 64 output channels; the fourth, fifth, and sixth convolutional layers have a kernel size of 3, a stride of 1, and 64 output channels; each convolutional layer is followed by a 2×1 pooling layer with a stride of 2; and the feature vector is output through a fully connected layer.
4. The method for detecting abnormalities in a turnout indicator connecting rod according to claim 2, characterized in that, The prior expert experience features calculated for the sample include the sample mean, variance, root mean square amplitude, peak value, minimum value, peak index, margin index, kurtosis, mean, root mean square value, standard deviation, maximum value, waveform index, impulse index, and skewness.
5. The method for detecting abnormalities in a turnout indicator connecting rod according to claim 1, characterized in that, The fine-tuning of the feature extractor and classifier is as follows: A training sample set is formed by collecting historical vibration acceleration data of the connecting rod of the faulty turnout. The training sample includes a sequence of vibration acceleration data when a train passes and a fault type label. The vibration acceleration data sequence of the train passing through the training sample is input into the fault prediction network model for prediction. The output of the fault prediction network model is processed by softmax to obtain the predicted fault category. The feature extractor and classifier networks are fine-tuned by using the cross-entropy loss of the predicted fault category and fault type label as the loss function.
6. A turnout indicator connecting rod abnormality detection system, characterized in that, The system includes: vibration data acquisition equipment, access client, and server; The vibration data acquisition device is connected to the server and is used to acquire data representing the vibration acceleration of the connecting rod and send the acquired data to the server. The server performs abnormal detection of the turnout indicator connecting rod as described in any one of claims 1-5, for the purpose of processing vibration acceleration data; The client connects to the server to allow users to view the test results and the collected vibration data.
7. A turnout indicator connecting rod abnormality detection system according to claim 6, characterized in that, The server is equipped with a fault prediction network model, which includes a feature extractor and a classifier. During the offline model training phase, the fault prediction network model is trained using a self-supervised representation learning method based on a sample set of historical vibration acceleration data of the normal turnout connecting rod. Based on the historical vibration acceleration data sample set of the connecting rod of the faulty turnout, the feature extractor and classifier are trained and fine-tuned. Online prediction phase: Input the collected vibration acceleration data of the turnout connecting rod into the fault prediction network model, and output the predicted fault type.
8. A turnout indicator connecting rod abnormality detection system according to claim 7, characterized in that, The specific training process for the feature extractor during the offline model training phase on the server is as follows: A learning sample set is formed by collecting historical vibration acceleration data of the connecting rod of a normal turnout. A masking operation is performed on the learning samples in the learning sample set to divide the learning samples into multiple sub-samples on an average basis. Based on the set masking rate, multiple segments of data in each sub-sample are randomly selected and set to 0. The masked subsamples are concatenated with the unmasked subsamples in their original order to obtain the masked sample. The masked sample is then input into the feature extractor to obtain the feature vector. The feature vectors obtained by the feature extractor are input into a data reconstructor based on a fully connected neural network to reconstruct sample data, thus obtaining reconstructed samples. Calculate prior expert experience features for the learning samples and obtain prior expert feature vectors; The feature vectors obtained by the feature extractor are input into the prior feature reconstructor based on a fully connected neural network to reconstruct them into prior feature vectors, thus obtaining the reconstructed feature vectors. The feature extractor is trained using the squared error between the reconstructed sample and the original sample, plus the squared error between the reconstructed feature vector and the prior expert feature vector of the original sample, as the loss function.
9. A turnout indicator connecting rod abnormality detection system according to claim 7, characterized in that, The server fine-tunes the feature extractor and classifier during the offline model training phase as follows: A training sample set is formed by collecting historical vibration acceleration data of the connecting rod of the faulty turnout. The training sample includes a sequence of vibration acceleration data when a train passes and a fault type label. The vibration acceleration data sequence of the train passing through the training sample is input into the fault prediction network model for prediction. The output of the fault prediction network model is processed by softmax to obtain the predicted fault category. The feature extractor and classifier networks are fine-tuned by using the cross-entropy loss of the predicted fault category and fault type label as the loss function.
10. A turnout indicator connecting rod abnormality detection system according to claim 6, characterized in that, The vibration data acquisition device is connected to the server. The vibration data acquisition device has a vibration acceleration sensor installed on the connecting rod to collect the vertical acceleration data of the connecting rod when the train passes the turnout. Offline model training phase: Acceleration data representing each fault type of the connecting rod are collected to form a training sample set; acceleration data representing the fault-free connecting rod are collected to form a learning sample set; Online prediction phase: The collected real-time vibration acceleration data of the turnout indicator connecting rod is transmitted to the server.