Catenary service life prediction method based on data enhancement and spatial feature extraction

Abstract

The invention belongs to the technical field of high-speed rail data processing, and provides a catenary service life prediction method based on data enhancement and spatial feature extraction, which comprises the following steps: collecting a catenary dynamic and static data set and a catenary defect record data table to obtain a to-be-processed data set; extracting an optimal feature data set; extracting a to-be-enhanced data set from the optimal feature data set, and constructing a generative adversarial network to form an expanded data set; constructing a graph structure, and generating a plurality of adjacent matrixes to construct a multi-scale graph convolutional layer; and based on a multi-scale image convolution layer, predicting the extended data set to obtain a residual service life prediction result. By adopting the method, the prediction accuracy of the remaining service life of the overhead line system can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of high-speed rail data processing, and in particular to a method for predicting the life of a contact network based on data enhancement and spatial feature extraction. Background Art

[0002] As a vital component of modern transportation, the safe operation and maintenance of high-speed railways are crucial for protecting passengers' lives and property, as well as socioeconomic development. During high-speed operation, high-speed trains are powered by a pantograph-catenary system. The catenary system is crucial for providing stable and reliable power to high-speed trains, and its performance stability and reliability are directly linked to the normal operation of the trains.

[0003] To ensure the normal operation of trains, fault prediction and health management of the catenary system are crucial. Remaining service life prediction is a key technology in this area. Accurately predicting the remaining service life of the catenary system can reduce the number of ineffective repairs, enabling efficient maintenance while also avoiding serious problems such as train downtime, electrical fires, and economic losses caused by catenary failures. Existing methods for predicting the remaining service life of the catenary system fall into two main categories: physical model-based methods and data-driven model-based methods. Physical model-based methods predict the remaining service life by establishing a physical degradation model of the equipment. These methods are suitable for equipment systems with clear failure mechanisms and sufficient domain knowledge. However, the high complexity of high-speed rail catenary systems makes comprehensive quantification of physical models difficult. Data-driven model-based methods use collected condition detection data to construct a remaining service life prediction model using mathematical statistics and artificial neural networks. Therefore, data-driven model-based methods are widely used in the field of remaining service life prediction.

[0004] However, the prediction accuracy of data-driven models depends on the quantity and quality of data and the quality of feature extraction. Since the maintenance of high-speed rail contact networks is relatively strict and frequent, the amount of fault data is relatively small. Existing models are difficult to fully capture the global structural information of objects such as contact networks with complex structures and multi-scale features. When extracting features, they tend to ignore the relationship between samples, resulting in insufficient accuracy in the prediction results. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting the life of a contact network based on data enhancement and spatial feature extraction. By performing data enhancement and spatial feature extraction, the present invention can obtain more accurate results for predicting the remaining service life of the contact network. This invention aims to address the technical problem of insufficient accuracy in existing methods for predicting the remaining service life of the contact network.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: The method for predicting the life of the contact network based on data enhancement and spatial feature extraction includes the following steps: Collecting a dynamic and static data group of the contact network and a data table of defect records of the contact network, and obtaining a data set to be processed based on the dynamic and static data group of the contact network and the data table of defect records of the contact network; Preprocessing the data set to be processed to form a basic data set, and extracting the optimal feature data set from the basic data set using a filtering method; Extracting a dataset to be enhanced from the optimal feature dataset, constructing a generative adversarial network based on penalty gradient and Wasserstein distance, and performing data enhancement on the dataset to be enhanced through the generative adversarial network to form an expanded dataset; Based on the expanded data set, a graph structure is constructed, and several adjacency matrices are generated to construct a multi-scale graph convolution layer; Based on the multi-scale graph convolution layer, the expanded data set is predicted to obtain a remaining useful life prediction result.

[0007] Furthermore, the contact network defect record data table includes a plurality of defect data, and the step of obtaining a data set to be processed based on the contact network dynamic and static data group and the contact network defect record data table includes: Establishing the contact network pillar corresponding to the defect data and establishing the kilometer mark corresponding to the contact network pillar; According to the kilometer markers, a contact network parameter data set is extracted from the contact network dynamic and static data group. The data types of the contact network parameter data set include speed, guide height, pull-out value, hard point, network pressure, spark time, contact force and height difference within the span. Several kilometer markers and the contact network parameter data sets constitute a data set to be processed.

[0008] Furthermore, the step of preprocessing the data set to be processed to form a basic data set includes: Drawing a plurality of scatter plots based on the data set to be processed to identify a plurality of outliers in the data set to be processed, and removing the outliers to form a first data set; Filling in a number of missing values ​​in the first data set using linear interpolation to form a second data set; The second data set is normalized to form a basic data set.

[0009] Furthermore, the step of extracting the optimal feature data set from the basic data set by using a filtering method includes: setting a RUL value according to a maintenance time interval of the overhead line equipment, establishing a plurality of relevant features based on a data type of the basic data set, where the data type of the basic data set is consistent with the data type of the overhead line parameter data set, extracting a plurality of feature values ​​corresponding to the relevant features from the basic data set to calculate a Pearson correlation coefficient between the relevant features and the RUL value; The absolute values ​​of the Pearson correlation coefficients are sorted by size to select a number of optimal features from the number of related features, and a number of feature values ​​corresponding to the number of optimal features are combined into an optimal feature data set.

[0010] Furthermore, the steps of constructing a generative adversarial network based on penalty gradient and Wasserstein distance include: Based on the GAN model, Wasserstein distance is used as the loss function in the GAN model to form a basic adversarial network; A Lipschitz constraint is performed on the basic adversarial network based on the penalty gradient to construct a generative adversarial network.

[0011] Furthermore, the generative adversarial network includes a generator and a discriminator, and the loss function of the generator is:

[0012] in, represents the loss function of the generator, represents random noise, express The sampling sample, express The expectation of the joint distribution of Represents the discriminator output result, Represents the output result of the generator; The loss function of the discriminator is:

[0013] in, represents the loss function of the discriminator, represents the real contact network sample data, express of the sampling sample.

[0014] Furthermore, the step of constructing a graph structure based on the expanded data set and generating a plurality of adjacency matrices to construct a multi-scale graph convolution layer includes: Taking the catenary support as a node and the connection between two nodes as an edge, constructing a graph structure consisting of a plurality of nodes and a plurality of edges based on the expanded data set; Based on the plurality of edges, a Euclidean distance formula is used to perform calculation and analysis to obtain a plurality of adjacency matrices; Based on the plurality of adjacency matrices and the plurality of nodes, a multi-scale graph convolution layer is constructed.

[0015] Furthermore, the step of setting a multi-scale graph convolution layer based on the plurality of adjacency matrices and the plurality of nodes includes: Based on the plurality of adjacency matrices, obtaining a plurality of neighbor nodes corresponding to each of the nodes; Constructing a plurality of subgraph convolutional layers based on the plurality of nodes and the plurality of neighboring nodes corresponding to the plurality of nodes; Based on the number of nodes and the number of neighbor nodes corresponding to the nodes, the scale of the subgraph convolution layer is set, so as to group the subgraph convolution layers into a multi-scale graph convolution layer according to the scales.

[0016] Furthermore, the step of predicting the expanded data set based on the multi-scale graph convolution layer to obtain a remaining useful life prediction result includes: Based on the attention mechanism, weights are assigned to all the nodes in the subgraph convolutional layer to form a weight matrix, and dynamic attention weights are calculated between all the nodes in the subgraph convolutional layer and the neighboring nodes corresponding to the nodes; Calculating aggregate features of a plurality of the subgraph convolutional layers based on the plurality of weight matrices, the plurality of dynamic attention weights, and the plurality of adjacency matrices; A weighted sum is performed on several of the aggregated features to obtain several final node outputs, and a fully connected layer mapping is performed on the several final node outputs to obtain RUL prediction values ​​to obtain the remaining useful life prediction result.

[0017] Compared with the prior art, the beneficial effects of the present invention are: through preprocessing and the filtering method, the collected data is processed, outliers in the data are eliminated, and irrelevant features are reduced to improve the quality of the data and the quality of feature extraction; by constructing the generative adversarial network, the problem of insufficient contact network defect data is effectively solved, and the generative adversarial network constructed by penalizing gradients and Wasserstein distance overcomes the problems of pattern collapse and gradient disappearance that are prone to occur in traditional adversarial networks, thereby ensuring the quality of the generated data; the spatial characteristics of the contact network are analyzed by Euclidean distance, and dynamic attention weights are used when extracting features. The prediction results are obtained through the multi-scale graph convolution layer, which effectively improves the accuracy of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1This is a flowchart of a method for predicting the life of a contact network based on data enhancement and spatial feature extraction in a first embodiment of the present invention; Figure 2 This is a structural block diagram of a contact network life prediction system based on data enhancement and spatial feature extraction in a second embodiment of the present invention; The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. DETAILED DESCRIPTION

[0019] To facilitate understanding of the present invention, the present invention will be described more fully below with reference to the accompanying drawings. The drawings illustrate several embodiments of the present invention. However, the present invention may be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and comprehensive understanding of the present invention.

[0020] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be an intermediate element. When an element is referred to as being "connected to" another element, it may be directly connected to the other element or there may be an intermediate element. The terms "vertical," "horizontal," "left," "right," and similar expressions used herein are for illustrative purposes only.

[0021] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0022] See also Figure 1 The first embodiment of the present invention provides a method for predicting the life of a contact network based on data enhancement and spatial feature extraction, comprising the following steps: Step S10: collecting a catenary dynamic and static data group and a catenary defect record data table, and obtaining a data set to be processed based on the catenary dynamic and static data group and the catenary defect record data table; Preferably, the contact network dynamic and static data groups can be collected by IC dynamic inspection vehicles, with a collection period of half a month, and the contact network defect record data table is obtained through fault repair reports.

[0023] The step S10 includes: S110: Determine the contact network pillar corresponding to the defect data, and determine the kilometer mark corresponding to the contact network pillar; Preferably, the high-speed railway contact network is composed of a plurality of contact network pillars, and the contact network pillars are connected by lines.

[0024] S120: According to the kilometer mark, a contact network parameter data set is extracted from the contact network dynamic and static data group. The data types of the contact network parameter data set include speed, lead height, pull-out value, hard point, network pressure, spark time, contact force and height difference within the span. Several kilometer marks and the contact network parameter data sets constitute a data set to be processed.

[0025] Understandably, the maintenance work of high-speed rail contact network is more rigorous and frequent, so the defect data is relatively small.

[0026] Step S20: pre-processing the data set to be processed to form a basic data set, and extracting the optimal feature data set from the basic data set using a filtering method; It can be understood that preprocessing the data set to be processed, removing outliers in the data, and extracting features that are highly correlated with the remaining service life of the contact network can effectively improve the quality of the data and the quality of feature extraction.

[0027] The step S20 includes: S210: Drawing a plurality of scatter plots based on the data set to be processed to identify a plurality of outliers in the data set to be processed, and removing the outliers to form a first data set; Preferably, a scatter plot is established according to each type of data. For example, a scatter plot corresponding to the guide height is established, the vertical axis is the value of the guide height, and the horizontal axis is the serial number of the data. Several guide height values ​​are arranged in sequence in the graph to form the scatter plot corresponding to the guide height, and the abnormal values ​​are eliminated according to the data point graph in the scatter plot.

[0028] S220: Filling in a number of missing values ​​in the first data set using a linear interpolation method to form a second data set; It can be understood that supplementing the missing values ​​is beneficial to improving the data quality of the second data set.

[0029] S230: performing normalization processing on the second data set to form a basic data set; The second data set includes contact network data of different units and amplitudes. Data with smaller values ​​are easily lost. Therefore, a data normalization method is used to map all data in the second data set to [-1, 1], which is beneficial to improving the convergence speed and accuracy of model training.

[0030] S240: setting a RUL value according to a maintenance time interval of the overhead line equipment, establishing a plurality of relevant features based on the data type of the basic data set, where the data type of the basic data set is consistent with the data type of the overhead line parameter data set, extracting a plurality of feature values ​​corresponding to the relevant features from the basic data set, and calculating a Pearson correlation coefficient between the relevant features and the RUL value; Preferably, the maintenance time interval of the contact network equipment is the difference between the previous maintenance time and the next maintenance time, each of the data types corresponds to one of the related features, and the several feature values ​​corresponding to one of the related features are several data values ​​under the data type.

[0031] S250: sorting the absolute values ​​of the Pearson correlation coefficients to select a plurality of optimal features from the plurality of related features, and forming an optimal feature data set with a plurality of feature values ​​corresponding to the plurality of optimal features.

[0032] Preferably, six optimal features are selected, the absolute values ​​of the Pearson correlation coefficients are arranged from large to small, and the correlation features corresponding to the absolute values ​​of the Pearson correlation coefficients ranked in the top six are selected as the optimal features.

[0033] Step S30: extracting a dataset to be enhanced from the optimal feature dataset, constructing a generative adversarial network based on penalty gradient and Wasserstein distance, and performing data enhancement on the dataset to be enhanced by the generative adversarial network to form an expanded dataset; Preferably, 80% of the data in the optimal feature dataset constitutes the dataset to be enhanced, and the remaining data is used to test the model's effectiveness. It can be understood that by constructing the generative adversarial network and performing data augmentation on the original data to form the expanded dataset, the problem of insufficient contact network defect data is effectively resolved. Furthermore, the generative adversarial network constructed through the penalty gradient and Wasserstein distance overcomes the problems of mode collapse and vanishing gradients that are prone to occur in traditional adversarial networks, thereby ensuring the quality of the generated data.

[0034] The step S30 includes: S310: Based on the GAN model, using Wasserstein distance as a loss function in the GAN model to form a basic adversarial network; In the traditional GAN ​​model, if the two data distributions differ greatly, the JS divergence may be 0, causing the discriminator in the GAN model to stagnate in updating the generator's gradient, resulting in the vanishing gradient problem. The Wasserstein distance is used instead of the loss function in the traditional GAN ​​model. Even if the two data distributions do not overlap, it can still measure the difference between the two data distributions.

[0035] S320: Performing a Lipschitz constraint on the basic adversarial network based on the penalty gradient to construct a generative adversarial network.

[0036] The gradient penalty mechanism is used instead of weight clipping to effectively prevent gradient explosion and non-convergence.

[0037] In step S30, the generative adversarial network includes a generator and a discriminator, and the loss function of the generator is:

[0038] in, represents the loss function of the generator, represents random noise, express The sampling sample, express The expectation of the joint distribution of represents the discriminator output result, Represents the output result of the generator; The loss function of the discriminator is:

[0039] in, represents the loss function of the discriminator, represents the real contact network sample data, express of the sampling sample.

[0040] Preferably, the generator is composed of an input layer, multiple deconvolution layers and an activation layer, which upsamples random noise from one-dimensional space to the high-dimensional space of the target sample. The deconvolution layer can expand the low-dimensional input to a high-dimensional output. Several deconvolution layers are followed by an activation layer. The activation layer uses ReLU as the activation function to enhance the nonlinear expression ability of the network. The last deconvolution layer in the generator uses the Tanh activation function. The discriminator is used to distinguish between generated data and real data. It is composed of a convolution layer and an activation layer. Each convolution layer is connected to an activation layer and a batch normalization layer. The activation function uses LeakyReLU, which is beneficial to introduce nonlinearity and prevent gradient disappearance.

[0041] Step S40: constructing a graph structure based on the expanded data set and generating several adjacency matrices to construct a multi-scale graph convolution layer; As an object with complex structure and multi-scale characteristics, the contact network is poorly suited for spatial feature extraction by traditional LSTM networks and CNN networks, making it difficult to capture global structural information. The multi-scale graph convolution layer is used to extract the multi-scale spatial features of the contact network, which is beneficial to improving the accuracy of the remaining service life prediction of the contact network.

[0042] The step S40 includes: S410: Taking the catenary pillars as nodes and the connection between two nodes as edges, constructing a graph structure consisting of a plurality of nodes and a plurality of edges based on the expanded data set; S420: Based on the plurality of edges, performing calculation and analysis using a Euclidean distance formula to obtain a plurality of adjacency matrices; S430: Constructing a multi-scale graph convolution layer based on the plurality of adjacency matrices and the plurality of nodes.

[0043] For the contact network system, the degree of mutual influence between the nodes is different, and the influence range of the neighboring nodes corresponding to the nodes is also different. The traditional graph convolutional neural network determines fixed weights for all neighboring nodes according to the topological structure. The influence between the nodes will gradually weaken, which limits the expression ability of the graph convolutional neural network. It can be understood that by analyzing the spatial characteristics of the contact network through Euclidean distance and analyzing the relationship between the contact network pillars, the quality of spatial feature extraction can be improved.

[0044] The S430 further includes: S4310: Based on the plurality of adjacency matrices, obtain a plurality of neighboring nodes corresponding to each of the nodes; S4320: Constructing a plurality of subgraph convolutional layers based on the plurality of nodes and the plurality of neighboring nodes corresponding to the plurality of nodes; S4330: Based on the number of nodes and the number of neighbor nodes corresponding to the nodes, set the scale of the subgraph convolution layer to group the subgraph convolution layers into a multi-scale graph convolution layer according to the scales.

[0045] Step S50: Based on the multi-scale graph convolution layer, predict the expanded data set to obtain a remaining useful life prediction result.

[0046] Preferably, the subgraph convolution layer of each scale is a GCN graph convolution layer, and the connection structures formed by the nodes and the neighboring nodes in different subgraph convolution layers are different. Through multi-scale aggregation, it is beneficial to extract the spatial characteristics of the contact network system and effectively improve the accuracy of the prediction results.

[0047] The step S50 includes: S510: Based on the attention mechanism, assign weights to all the nodes in the subgraph convolution layer to form a weight matrix, and calculate the dynamic attention weights between all the nodes in the subgraph convolution layer and the neighboring nodes corresponding to the nodes; S520: Calculating aggregate features of a plurality of the subgraph convolutional layers based on the plurality of weight matrices, the plurality of the dynamic attention weights, and the plurality of the adjacency matrices; S530: Perform weighted summation on the plurality of aggregated features to obtain a plurality of final node outputs, perform full connection layer mapping on the plurality of final node outputs to obtain RUL prediction values, and thereby obtain a remaining useful life prediction result.

[0048] Preferably, real contact network data is obtained, and 80% of the data in the optimal feature data set in step S30 is divided for training the prediction model formed by the contact network life prediction method based on data enhancement and spatial feature extraction in this embodiment, and the remaining 20% ​​of the data is used to test the prediction model, and the RUL prediction value is calculated. The RUL true value is obtained by combining the collected fault repair report. The prediction accuracy of the prediction model can be evaluated by calculating the average absolute error, root mean square error and average absolute percentage error between the RUL true value and the RUL prediction value. The average absolute error of the prediction model is 2.291%, the root mean square error is 3.249%, and the average absolute percentage error is 4.64%. The same data is used to train and test the CNN model and the LSTM model respectively. model, DCNN model and GCN model. The average absolute error of the CNN model is 8.284%, the root mean square error is 11.806%, and the average absolute percentage error is 12.94%. The average absolute error of the LSTM model is 10.902%, the root mean square error is 14.951%, and the average absolute percentage error is 15.01%. The average absolute error of the DCNN model is 4.296%, the root mean square error is 6.045%, and the average absolute percentage error is 8.88%. The average absolute error of the GCN model is 4.649%, the root mean square error is 6.219%, and the average absolute percentage error is 8.75%. It can be understood that the contact network life prediction method based on data enhancement and spatial feature extraction effectively improves the accuracy of contact network life prediction.

[0049] See also Figure 2 A second embodiment of the present invention provides a catenary life prediction system based on data enhancement and spatial feature extraction, applying the catenary life prediction method based on data enhancement and spatial feature extraction as described in the first embodiment. The system includes: An acquisition module 10 is configured to acquire a catenary dynamic and static data set and a catenary defect record data table, and obtain a data set to be processed based on the catenary dynamic and static data set and the catenary defect record data table; The acquisition module 10 includes: The first unit is used to determine the contact network pillar corresponding to the defect data and the kilometer mark corresponding to the contact network pillar; A second unit is configured to extract a catenary parameter data set from the catenary dynamic and static data set according to the kilometer markers, wherein the data types of the catenary parameter data set include speed, lead height, pull-out value, hard point, network voltage, spark time, contact force, and span height difference, and a plurality of the kilometer markers and the catenary parameter data sets constitute a data set to be processed; The processing module 20 is used to pre-process the data set to be processed to form a basic data set, and extract the optimal feature data set from the basic data set using a filtering method; The processing module 20 includes: A third unit is configured to draw a plurality of scatter plots based on the data set to be processed to identify a plurality of outliers in the data set to be processed, and remove the outliers to form a first data set; a fourth unit, configured to fill in a plurality of missing values ​​in the first data set using a linear interpolation method to form a second data set; a fifth unit, configured to perform normalization processing on the second data set to form a basic data set; a sixth unit, configured to set a RUL value according to a maintenance time interval of the overhead line equipment, establish a plurality of relevant features based on a data type of the basic data set, where the data type of the basic data set is consistent with the data type of the overhead line parameter data set, extract a plurality of feature values ​​corresponding to the relevant features from the basic data set, and calculate a Pearson correlation coefficient between the relevant features and the RUL value; The seventh unit is used to sort the absolute values ​​of the Pearson correlation coefficients to select a number of optimal features from the number of related features, and to form an optimal feature data set with a number of feature values ​​corresponding to the number of optimal features.

[0050] An expansion module 30 is configured to extract a dataset to be enhanced from the optimal feature dataset, construct a generative adversarial network based on a penalty gradient and a Wasserstein distance, and perform data enhancement on the dataset to be enhanced using the generative adversarial network to form an expanded dataset; The expansion module 30 includes: The eighth unit is configured to use the Wasserstein distance as a loss function in the GAN model based on the GAN model to form a basic adversarial network; A ninth unit is configured to perform a Lipschitz constraint on the basic adversarial network based on a penalty gradient to construct a generative adversarial network; In the expansion module 30, the generative adversarial network includes a generator and a discriminator, and the loss function of the generator is:

[0051] in, represents the loss function of the generator, represents random noise, express The sampling sample, express The expectation of the joint distribution of Represents the discriminator output result, Represents the output result of the generator; The loss function of the discriminator is:

[0052] in, represents the loss function of the discriminator, represents the real contact network sample data, express of the sampling sample.

[0053] A construction module 40 is used to construct a graph structure based on the expanded data set and generate several adjacency matrices to construct a multi-scale graph convolution layer; The building block 40 includes: A tenth unit is configured to construct a graph structure consisting of a plurality of nodes and a plurality of edges based on the expanded data set, using the catenary support as a node and a connection relationship between two nodes as an edge; The eleventh unit is used to perform calculation and analysis based on the plurality of edges using the Euclidean distance formula to obtain a plurality of adjacency matrices; A twelfth unit is used to construct a multi-scale graph convolution layer based on the plurality of adjacency matrices and the plurality of nodes; The twelfth unit is specifically configured to obtain a plurality of neighbor nodes corresponding to each of the nodes based on the plurality of adjacency matrices; Constructing a plurality of subgraph convolutional layers based on the plurality of nodes and the plurality of neighboring nodes corresponding to the plurality of nodes; Based on the number of nodes and the number of neighbor nodes corresponding to the nodes, the scale of the subgraph convolution layer is set, so as to group the subgraph convolution layers into a multi-scale graph convolution layer according to the scales.

[0054] The prediction module 50 is configured to predict the expanded data set based on the multi-scale graph convolution layer to obtain a remaining useful life prediction result.

[0055] The prediction module 50 includes: A thirteenth unit is configured to assign weights to all the nodes in the subgraph convolution layer based on an attention mechanism to form a weight matrix, and calculate dynamic attention weights between all the nodes in the subgraph convolution layer and the neighboring nodes corresponding to the nodes; A fourteenth unit is configured to calculate aggregate features of a plurality of the subgraph convolutional layers based on the plurality of weight matrices, the plurality of dynamic attention weights, and the plurality of adjacency matrices; The fifteenth unit is used to perform weighted summation on the plurality of aggregated features to obtain a plurality of final node outputs, perform full connection layer mapping on the plurality of final node outputs to obtain RUL prediction values, and thus obtain a remaining useful life prediction result.

[0056] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0057] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for predicting the life of a contact network based on data enhancement and spatial feature extraction, characterized in that: The steps include: Collecting a dynamic and static data group of the contact network and a data table of defect records of the contact network, and obtaining a data set to be processed based on the dynamic and static data group of the contact network and the data table of defect records of the contact network; Preprocessing the data set to be processed to form a basic data set, and extracting the optimal feature data set from the basic data set using a filtering method; Extracting a dataset to be enhanced from the optimal feature dataset, constructing a generative adversarial network based on penalty gradient and Wasserstein distance, and performing data enhancement on the dataset to be enhanced through the generative adversarial network to form an expanded dataset; Based on the expanded data set, a graph structure is constructed, and several adjacency matrices are generated to construct a multi-scale graph convolution layer; Based on the multi-scale graph convolution layer, the expanded data set is predicted to obtain a remaining useful life prediction result.

2. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 1 is characterized in that: The contact network defect record data table includes a plurality of defect data. The step of obtaining a data set to be processed based on the contact network dynamic and static data group and the contact network defect record data table includes: Establishing the contact network pillar corresponding to the defect data and establishing the kilometer mark corresponding to the contact network pillar; According to the kilometer markers, a contact network parameter data set is extracted from the contact network dynamic and static data group. The data types of the contact network parameter data set include speed, guide height, pull-out value, hard point, network pressure, spark time, contact force and height difference within the span. Several kilometer markers and the contact network parameter data sets constitute a data set to be processed.

3. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 1, characterized in that: The step of preprocessing the data set to be processed to form a basic data set includes: Drawing a plurality of scatter plots based on the data set to be processed to identify a plurality of outliers in the data set to be processed, and removing the outliers to form a first data set; Filling in a number of missing values ​​in the first data set using linear interpolation to form a second data set; The second data set is normalized to form a basic data set.

4. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 2, characterized in that: The step of extracting the optimal feature data set from the basic data set by using a filtering method comprises: setting a RUL value according to a maintenance time interval of the overhead line equipment, establishing a plurality of relevant features based on a data type of the basic data set, where the data type of the basic data set is consistent with the data type of the overhead line parameter data set, extracting a plurality of feature values ​​corresponding to the relevant features from the basic data set to calculate a Pearson correlation coefficient between the relevant features and the RUL value; The absolute values ​​of the Pearson correlation coefficients are sorted by size to select a number of optimal features from the number of related features, and a number of feature values ​​corresponding to the number of optimal features are combined into an optimal feature data set.

5. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 1, characterized in that: The steps of constructing a generative adversarial network based on penalty gradient and Wasserstein distance include: Based on the GAN model, Wasserstein distance is used as the loss function in the GAN model to form a basic adversarial network; A Lipschitz constraint is performed on the basic adversarial network based on the penalty gradient to construct a generative adversarial network.

6. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 1, characterized in that: The generative adversarial network includes a generator and a discriminator, and the loss function of the generator is: in, represents the loss function of the generator, represents random noise, express The sampling sample, express The expectation of the joint distribution of Represents the discriminator output result, Represents the output result of the generator; The loss function of the discriminator is: in, represents the loss function of the discriminator, represents the real contact network sample data, express of the sampling sample.

7. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 2, characterized in that: The steps of constructing a graph structure based on the expanded data set and generating a plurality of adjacency matrices to construct a multi-scale graph convolution layer include: Taking the catenary support as a node and the connection between two nodes as an edge, constructing a graph structure consisting of a plurality of nodes and a plurality of edges based on the expanded data set; Based on the plurality of edges, a Euclidean distance formula is used to perform calculation and analysis to obtain a plurality of adjacency matrices; Based on the plurality of adjacency matrices and the plurality of nodes, a multi-scale graph convolution layer is constructed.

8. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 7 is characterized in that: The step of setting a multi-scale graph convolution layer based on the plurality of adjacency matrices and the plurality of nodes includes: Based on the plurality of adjacency matrices, obtaining a plurality of neighbor nodes corresponding to each of the nodes; Constructing a plurality of subgraph convolutional layers based on the plurality of nodes and the plurality of neighboring nodes corresponding to the plurality of nodes; Based on the number of nodes and the number of neighbor nodes corresponding to the nodes, the scale of the subgraph convolution layer is set, so as to group the subgraph convolution layers into a multi-scale graph convolution layer according to the scales.

9. The method for predicting the life of a contact network based on data enhancement and spatial feature extraction according to claim 8, characterized in that: The step of predicting the expanded data set based on the multi-scale graph convolution layer to obtain a remaining useful life prediction result includes: Based on the attention mechanism, weights are assigned to all the nodes in the subgraph convolutional layer to form a weight matrix, and dynamic attention weights are calculated between all the nodes in the subgraph convolutional layer and the neighboring nodes corresponding to the nodes; Calculating aggregate features of a plurality of the subgraph convolutional layers based on the plurality of weight matrices, the plurality of dynamic attention weights, and the plurality of adjacency matrices; A weighted sum is performed on several of the aggregated features to obtain several final node outputs, and a fully connected layer mapping is performed on the several final node outputs to obtain RUL prediction values ​​to obtain the remaining useful life prediction result.

Images