Multi-source data analysis method and device for railway track structure
By constructing a machine learning model to perform anomaly cleaning and analysis on multi-source detection data of railway track structures, the problem of low intelligence level in existing technologies has been solved, realizing efficient and intelligent detection of track structure service status and improving analysis efficiency and accuracy.
Patent Information
- Application Number
- CN202411212060.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2024-11-12
AI Technical Summary
In existing technologies, the analysis of multi-source detection data for railway track structures suffers from low intelligence, low efficiency, high reliance on human subjective experience, and low accuracy, making it difficult to meet practical needs.
Machine learning models are used to clean and analyze anomalies in multi-source detection data. By constructing a data anomaly cleaning model and a track structure analysis model, and using Transformer models and interactive learning models to extract features from multi-source data, efficient and intelligent detection of track structure service status is achieved.
It improves the efficiency and accuracy of multi-source track detection data analysis, enables autonomous learning of multi-source data characteristics, and enhances the intelligence level of track structure health status assessment.
Smart Images

Figure CN118939945A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of railway infrastructure detection, and in particular to a railway track structure multi-source data analysis method and device. Background Art
[0002] This section is intended to provide a background or context for embodiments of the present invention. No description herein is admitted to be prior art by virtue of its inclusion in this section.
[0003] Track structure service status detection is a key part of ensuring the safe operation of railway lines. Carrying out analysis and mining of track multi-source dynamic detection data is an important way to evaluate the health status of track structures. Track structure multi-source dynamic detection data is obtained by large-scale comprehensive detection trains that regularly conduct periodic dynamic inspections of high-speed railway track structures. It is mainly composed of multi-dimensional time series data collected by multiple types and multiple sensors, and has the characteristics of multi-source, multi-dimensional, heterogeneous, and complex influencing factors.
[0004] Abnormal inspection data cleaning and track structure disease information mining are necessary links and important goals for track structure multi-source inspection data analysis. However, at present, dynamic inspection data analysis has the disadvantages of low intelligence, low efficiency, high reliance on human subjective experience, and low accuracy, which makes it difficult to meet actual needs. Summary of the invention
[0005] The embodiment of the present invention provides a railway track structure multi-source data analysis method, which is used to improve the efficiency of track multi-source detection data analysis and improve the accuracy of track state analysis. The method includes:
[0006] The multi-source detection data to be analyzed is obtained, and the multi-source detection data to be analyzed is cleaned of abnormalities by a data abnormality cleaning model. The multi-source detection data to be analyzed after abnormality cleaning is analyzed by a track structure analysis model to obtain the analysis results of the multi-source data of the railway track structure; wherein:
[0007] The data anomaly cleaning model is obtained as follows:
[0008] Collect historical track structure data and build a track structure multi-source detection data set based on the historical track structure data;
[0009] The track structure multi-source detection data set is processed to extract a multi-source detection training set and a multi-source detection test set, which include the track structure multi-source detection data and corresponding data anomaly labels;
[0010] Construct a first machine learning model, train the first machine learning model according to the multi-source detection training set, verify the first machine learning model according to the multi-source detection test set, adjust model parameters according to the verification result, and obtain a data anomaly cleaning model when the accuracy of the verification result reaches a preset standard;
[0011] The track structure analysis model is obtained as follows:
[0012] The track structure multi-source detection dataset is cleaned by the data anomaly cleaning model to obtain the track structure multi-source detection dataset after anomaly cleaning;
[0013] The track structure multi-source detection data set after anomaly cleaning is processed to extract a multi-source detection training set after anomaly cleaning and a multi-source detection test set after anomaly cleaning. The multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels;
[0014] Construct a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification results reaches the preset standard.
[0015] The embodiment of the present invention further provides a railway track structure multi-source data analysis device, which is used to improve the efficiency of track multi-source detection data analysis and improve the accuracy of track state analysis. The device includes:
[0016] The track structure analysis module is used to obtain the multi-source detection data to be analyzed, clean the multi-source detection data to be analyzed by using the data anomaly cleaning model, and analyze the multi-source detection data to be analyzed after the anomaly cleaning by using the track structure analysis model to obtain the railway track structure multi-source data analysis results;
[0017] A cleaning model building module is used to collect historical track structure data, build a track structure multi-source detection data set based on the historical track structure data; process the track structure multi-source detection data set to extract a multi-source detection training set and a multi-source detection test set, wherein the multi-source detection training set and the multi-source detection test set include the track structure multi-source detection data and corresponding data anomaly labels; build a first machine learning model, train the first machine learning model based on the multi-source detection training set, verify the first machine learning model based on the multi-source detection test set, adjust the model parameters based on the verification results, and obtain a data anomaly cleaning model when the accuracy of the verification results reaches a preset standard;
[0018] The analysis model building module is used to clean the track structure multi-source detection data set through the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; process the track structure multi-source detection data set after anomaly cleaning, and extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels; build a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification result reaches the preset standard.
[0019] An embodiment of the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned railway track structure multi-source data analysis method when executing the computer program.
[0020] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned railway track structure multi-source data analysis method is implemented.
[0021] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned railway track structure multi-source data analysis method is implemented.
[0022] In an embodiment of the present invention, multi-source detection data to be analyzed is obtained, anomalies are cleaned for the multi-source detection data to be analyzed by a data anomaly cleaning model, and the multi-source detection data to be analyzed after anomaly cleaning is analyzed by a track structure analysis model to obtain a railway track structure multi-source data analysis result; wherein: the data anomaly cleaning model is obtained in the following manner: historical track structure data is collected, and a track structure multi-source detection data set is constructed based on the historical track structure data; the track structure multi-source detection data set is processed to extract a multi-source detection training set and a multi-source detection test set, and the multi-source detection training set and the multi-source detection test set include track structure multi-source detection data and corresponding data anomaly labels; a first machine learning model is constructed, the first machine learning model is trained according to the multi-source detection training set, the first machine learning model is verified according to the multi-source detection test set, the model parameters are adjusted according to the verification results, and when the verification results are accurate When the data anomaly cleaning model reaches the preset standard, the data anomaly cleaning model is obtained; the track structure analysis model is obtained as follows: the track structure multi-source detection data set is cleaned by the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; the track structure multi-source detection data set after anomaly cleaning is processed to extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis label; a second machine learning model is constructed, the second machine learning model is trained according to the multi-source detection training set after anomaly cleaning, the second machine learning model is verified according to the multi-source detection test set after anomaly cleaning, the model parameters are adjusted according to the verification results, and the track structure analysis model is obtained when the accuracy of the verification results reaches the preset standard. In this way, by introducing the neural network model and the interactive model, the multi-source data features can be learned autonomously to achieve efficient and intelligent track structure service status detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0024] Figure 1 A flowchart of a railway track structure multi-source data analysis method provided in an embodiment of the present invention;
[0025] Figure 2 A schematic diagram of a data anomaly cleaning model provided in an embodiment of the present invention;
[0026] Figure 3A schematic diagram of a track structure analysis model provided in an embodiment of the present invention;
[0027] Figure 4 This is an abnormal waveform diagram of multi-source dynamic inspection data of the track structure provided in an embodiment of the present invention. Figure 4 (a) is the waveform of the gauge camera failure. Figure 4 (b) is the waveform of the string pulling fault. Figure 4 (c) is the sunlight interference waveform. Figure 4 (d) is the abnormal waveform of structural defects. Figure 4 Middle (e) is other interference abnormal waveform;
[0028] Figure 5 A geometric waveform diagram of an abnormal track state provided in an embodiment of the present invention;
[0029] Figure 6 A schematic diagram of a railway track structure multi-source data analysis device provided in an embodiment of the present invention;
[0030] Figure 7 It is a structural block diagram of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0031] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings. Here, the exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.
[0032] The term "and / or" herein only describes an association relationship, indicating that three relationships may exist. For example, A and / or B may represent the following three situations: A exists alone, A and B exist at the same time, and B exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of A, B, and C may represent including any one or more elements selected from the set consisting of A, B, and C.
[0033] In the description of this specification, the terms "include", "including", "have", "contain", etc. are all open terms, which mean including but not limited to. The descriptions with reference to the terms "one embodiment", "a specific embodiment", "some embodiments", "for example", etc. mean that the specific features, structures or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures or characteristics described may be combined in a suitable manner in any one or more embodiments or examples. The order of steps involved in each embodiment is used to schematically illustrate the implementation of the present application, and the order of steps is not limited and can be appropriately adjusted as needed.
[0034] The embodiment of the present invention provides a method for analyzing multi-source data of railway track structure. Figure 1 As shown, including:
[0035] Step 101: obtaining multi-source detection data to be analyzed, cleaning the multi-source detection data to be analyzed by using a data anomaly cleaning model, analyzing the multi-source detection data to be analyzed after cleaning the anomalies by using a track structure analysis model, and obtaining a railway track structure multi-source data analysis result; wherein:
[0036] Step 1011: collect historical track structure data, and construct a track structure multi-source detection data set based on the historical track structure data; process the track structure multi-source detection data set to extract a multi-source detection training set and a multi-source detection test set, wherein the multi-source detection training set and the multi-source detection test set include the track structure multi-source detection data and corresponding data anomaly labels; construct a first machine learning model, train the first machine learning model based on the multi-source detection training set, verify the first machine learning model based on the multi-source detection test set, adjust the model parameters based on the verification result, and obtain a data anomaly cleaning model when the accuracy of the verification result reaches a preset standard;
[0037] Step 1012: Clean the track structure multi-source detection data set through the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; process the track structure multi-source detection data set after anomaly cleaning to extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels; construct a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification result reaches the preset standard.
[0038] The railway track structure multi-source data analysis method proposed in the embodiment of the present invention reduces the dimension of the original track structure multi-source dynamic detection data, captures the local and global information of the dynamic detection data based on the attention mechanism of the Transformer model, amplifies the reconstruction error of the abnormal point, and obtains the detection results of the track invalid data and performs data cleaning by judging whether the abnormal score is higher than the threshold. Then, feature extraction is performed based on the dual sequence interactive learning model, data is reconstructed, and the error score between the reconstructed data and the original data is calculated, and the area with an error score greater than the threshold is determined as a suspected disease area.
[0039] In one embodiment, a track structure multi-source detection dataset is constructed based on historical track structure data, including:
[0040] The track gauge, left height and low, right height and low, left track direction, right track direction, level, triangular pit, superelevation, curvature, long wave left height and low, long wave right height and low, long wave left track direction, long wave right track direction, horizontal acceleration and vertical acceleration are extracted from the historical track structure data in turn to construct a multi-source detection dataset of the track structure.
[0041] The specific implementation includes:
[0042] Step S11: A large-scale inspection train collects multi-source dynamic inspection data of the track structure;
[0043] Step S12: extract 15 dimensional data in sequence, including track gauge, left height, right height, left track direction, right track direction, level, triangular pit, superelevation, curvature, long wave left height, long wave right height, long wave left track direction, long wave right track direction, horizontal acceleration and vertical acceleration, and construct a multi-source dynamic detection dataset of track structure after feature deletion.
[0044] In one embodiment, the track structure multi-source detection data set is processed to extract a multi-source detection training set and a multi-source detection test set, including:
[0045] A multi-source detection training set is extracted from the track structure multi-source detection dataset;
[0046] Some data are extracted from the track structure multi-source detection dataset, and the extracted data are cleaned for anomalies. The data before and after anomaly cleaning are used as the multi-source detection test set.
[0047] In one embodiment, constructing a first machine learning model, and training the first machine learning model according to a multi-source detection training set includes:
[0048] Convert the multi-source detection training set into a multi-dimensional feature vector sequence;
[0049] Build a Transformer-based neural network model;
[0050] According to the multi-dimensional feature vector sequence of the multi-source detection training set, the loss is calculated using the associated difference and loss function, and the model parameters are updated according to the loss calculation results.
[0051] The specific implementation includes:
[0052] Step S21: converting the multi-source dynamic detection data of the track structure to be cleaned into a multi-dimensional feature vector sequence to obtain a feature representation for data anomaly detection.
[0053] Step S22: An example of constructing a track structure multi-source detection data anomaly cleaning model is as follows: Figure 2 As shown. The multi-dimensional feature vector sequence is used to establish a track structure multi-source detection data anomaly cleaning model through an L-layer Transformer encoder. The Transformer encoder is a module composed of multiple attention layers and feedforward neural network layers to capture the association between different positions in the sequence; the attention mechanism allows the model to automatically focus on the relevant information at different positions in the sequence to establish global and local context representations, as shown in Formula 1, the prior association According to the learnable Gaussian kernel Generated, the i-th element σ i Corresponding to the i-th time point, For sequence association, is the hidden representation after the attention layer, Represents the query, key and value of self-attention respectively, d model is the number of channels.
[0054]
[0055] Step S23: In the training phase, the training samples are input into the track structure multi-source detection data anomaly cleaning model, and the association difference loss function is introduced to measure the association difference between normal sequences and abnormal sequences. As shown in Formula 2, KL(·||·) corresponds to each layer and The KL divergence calculated between two discrete distributions is is the correlation difference between each time point in the input time series X from the prior correlation P and sequence correlation S of the L layer. The correlation difference loss function compares the similarities and differences between normal sequences and abnormal sequences at the encoder output, as shown in Formula 3, represents the reconstruction of X, ||·|| F and ||·|| k represents the Frobenius norm and k-norm, and λ is a trade-off for the loss term. The loss is calculated using the association difference and MSE loss function, and the model parameters are updated through back propagation.
[0056]
[0057] Step S24: In order to better control the process of association learning, the Minimax strategy is introduced. As shown in Formula 4, in the minimization stage, the prior association is optimized let Approximate the serial correlation learned from the original time series This process will adapt the prior association to different temporal patterns; in the maximization stage, the sequence association is optimized This process makes the global information pay more attention to non-neighboring points and makes it more difficult to reconstruct outliers.
[0058]
[0059] In one embodiment, the track structure multi-source detection dataset is cleaned by a data anomaly cleaning model to obtain the track structure multi-source detection dataset after anomaly cleaning, including:
[0060] The track structure multi-source detection dataset is cleaned through the data anomaly cleaning model;
[0061] After the multi-dimensional feature vector sequence passes through the encoder, it is determined whether there is an abnormality by calculating the abnormal reference value and determining the abnormal point;
[0062] Remove isolated outliers through smoothing operations to obtain processed outlier data;
[0063] The processed outlier point data is removed from the track structure multi-source detection dataset to obtain the track structure multi-source detection dataset after anomaly cleaning.
[0064] The specific implementation includes:
[0065] Step S31: Use the trained railway track structure multi-source detection data anomaly cleaning model to clean the constructed track structure multi-source detection data set.
[0066] Step S32: After the input sequence passes through the encoder, as shown in Formula 5, the correlation difference score is calculated to evaluate whether there is an anomaly in the sequence. A high correlation difference score indicates that there is an anomaly in the sequence, and a low correlation difference score indicates that the sequence is normal.
[0067]
[0068] Step S33: Use a smoothing operation based on neighboring points to remove the isolated outliers in step S32 to obtain post-processed outlier data.
[0069] Step S34: Remove the abnormal data selected in step S33 from the data set constructed in step S1. Use the low-temperature monthly dynamic inspection data after data cleaning as the training set, and perform normalization and denoising after data preprocessing; use the high-temperature monthly dynamic inspection data after data cleaning as the test set, and insert the arch disease samples into it to speed up the model to extract features, as shown in formula 6, by inputting data x ij The average value μ can be obtained j and variance σ j , and then get the normalized result A multi-source detection dataset of track structure with data anomaly cleaning has been constructed.
[0070]
[0071] In one embodiment, a second machine learning model is constructed, and the second machine learning model is trained according to the multi-source detection training set after anomaly cleaning, including:
[0072] Constructing an interactive model, wherein the interactive model is stacked by interactive learning modules, and the interactive module is composed of a downsampling layer, a fully connected layer, an interactive learning layer, an upsampling layer, and two associated attention layers;
[0073] The multi-source detection training set after anomaly cleaning is input into the interactive model, the correlation features between the attributes of the data in the multi-source detection training set are extracted, the abnormal features are amplified through interactive learning, and the reconstructed data is determined;
[0074] The loss value is calculated based on the reconstructed data and the multi-source detection training set, and the model parameters are updated based on the loss value calculation result.
[0075] The specific implementation includes:
[0076] Step S41: An example of a track structure fault intelligent analysis model is constructed. Figure 3 The model is composed of K layers of interactive learning modules. Inside the interactive module, it consists of a downsampling layer, a fully connected layer, two associated attention layers, an upsampling layer, and an interactive learning mechanism. In the downsampling layer, the preprocessed data is Downsample to odd sequence Even sequence Two sequences, as shown in Formula 7.
[0077]
[0078] As shown in Formula 8, the sequence obtained by downsampling and Through the fully connected network, it is restored to a shape consistent with the input sequence, that is, Where L represents the original sequence length and C represents the data dimension.
[0079]
[0080] As shown in formula 9, the two sequences and Divide them into several patches of the same length, and get the shapes of (P, N, C) respectively. and shape (N, P, C) P represents the Patch length and C represents the data dimension.
[0081]
[0082] Step S42: In the training phase, the obtained track structure multi-source detection data samples after data anomaly cleaning are input into the model, and the sequences obtained in S41 are respectively input into the inter-block correlation attention and intra-block correlation attention, and the dual attention mechanism is used to extract and The intra-block correlation features and inter-block correlation features are as shown in formula 10 to formula
[0083] Formula 14.
[0084]
[0085] Similarly, inter-block correlation features can be obtained.
[0086]
[0087] The two attention weights are restored to the time series length by upsampling, as shown in Formula 15. and Interactive learning is performed and after passing through the fully connected layer, it is mapped into a feature vector consistent with the dimension of the input data, as shown in Formula 15.
[0088]
[0089] The two obtained sequences are weighted and added to obtain the final reconstructed data, as shown in Formula 17. The error between the reconstructed data and the original data is calculated, as shown in Formula 18.
[0090]
[0091] Step S43: Calculate the loss value of the reconstructed data and the input original data, as shown in Formula 18, and update the model parameters in reverse.
[0092]
[0093] Step S44: In the testing phase, the samples in the test set are input into the model, the anomaly score of the test set is calculated, the threshold is determined by the threshold selection algorithm, and the anomaly score greater than the threshold is determined as a diseased interval, as shown in Formula 19. When the model accuracy does not increase, stop training.
[0094] predict=(score>threshold) Formula 19
[0095] In one embodiment, the interaction model is verified based on the multi-source detection test set after anomaly cleaning, including:
[0096] Input the multi-source detection test set after anomaly cleaning into the interactive model and calculate the anomaly score of each data in the test set;
[0097] The anomaly score threshold is determined by the threshold selection algorithm, and the data in the test set with anomaly scores greater than the anomaly score threshold is judged as an abnormal interval, and the test results are output.
[0098] The specific implementation includes:
[0099] include:
[0100] Step S51: For the multi-source detection data to be analyzed newly collected by the comprehensive detection vehicle, multi-source detection data of the track structure are extracted therefrom, and 15 dimensional data including gauge, left height and low, right height and low, left track direction, right track direction, level, triangular pit, superelevation, curvature, long wave left height and low, long wave right height and low, long wave left track direction, long wave right track direction, horizontal acceleration and vertical acceleration are extracted in sequence to construct the multi-source detection data of the track structure to be analyzed.
[0101] Step S52: using the anomaly cleaning model trained in step S2, the original track structure multi-source dynamic inspection data anomaly cleaning work is performed on the track structure multi-source detection data to be analyzed constructed in step S51, and the track structure multi-source detection data after data anomaly cleaning is obtained.
[0102] Step S53: Use the intelligent disease analysis model trained in step S5 to carry out intelligent analysis and information mining of the railway track structure status on the multi-source detection data of the track structure after cleaning in step S52.
[0103] Take a specific example, Figure 4 As shown in the figure, several typical data anomaly (invalid data) waveforms are shown, including: Figure 4 (a) Rain interference waveform. Figure 4 (b) Sunlight interference waveform. Figure 4 (c) Displacement meter fault waveform. Figure 4 (d) Gauge machine fault waveform diagram, Figure 4 (e) Server failure waveform.
[0104] like Figure 5 As shown in the figure, the waveform evolution of the arching disease in the cold months (January, February, and March) and the hot months (July) is shown. In the hot months, the arching area has obvious quasi-sinusoidal waveform changes.
[0105] The present invention also provides a railway track structure multi-source data analysis device, as described in the following embodiments. Since the principle of solving the problem by the device is similar to that of the railway track structure multi-source data analysis method, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.
[0106] Figure 6 Schematic diagram of a railway track structure multi-source data analysis device provided in an embodiment of the present invention, such as Figure 6 As shown, the device comprises:
[0107] The track structure analysis module 601 is used to obtain multi-source detection data to be analyzed, clean the multi-source detection data to be analyzed by using a data anomaly cleaning model, and analyze the multi-source detection data to be analyzed after the anomaly cleaning by using the track structure analysis model to obtain the railway track structure multi-source data analysis results;
[0108] The cleaning model building module 602 is used to collect historical track structure data, build a track structure multi-source detection data set based on the historical track structure data; process the track structure multi-source detection data set to extract a multi-source detection training set and a multi-source detection test set, wherein the multi-source detection training set and the multi-source detection test set include the track structure multi-source detection data and corresponding data anomaly labels; build a first machine learning model, train the first machine learning model based on the multi-source detection training set, verify the first machine learning model based on the multi-source detection test set, adjust the model parameters based on the verification result, and obtain a data anomaly cleaning model when the accuracy of the verification result reaches a preset standard;
[0109] The analysis model construction module 603 is used to clean the track structure multi-source detection data set through the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; process the track structure multi-source detection data set after anomaly cleaning to extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels; construct a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification result reaches the preset standard.
[0110] In one embodiment, the cleaning model building module 602 is specifically used to:
[0111] The track gauge, left height and low, right height and low, left track direction, right track direction, level, triangular pit, superelevation, curvature, long wave left height and low, long wave right height and low, long wave left track direction, long wave right track direction, horizontal acceleration and vertical acceleration are extracted from the historical track structure data in turn to construct a multi-source detection dataset of the track structure.
[0112] In one embodiment, the cleaning model building module 602 is specifically used to:
[0113] A multi-source detection training set is extracted from the track structure multi-source detection dataset;
[0114] Some data are extracted from the track structure multi-source detection dataset, and the extracted data are cleaned for anomalies. The data before and after anomaly cleaning are used as the multi-source detection test set.
[0115] In one embodiment, the cleaning model building module 602 is specifically used to:
[0116] Convert the multi-source detection training set into a multi-dimensional feature vector sequence;
[0117] Build a Transformer-based neural network model;
[0118] According to the multi-dimensional feature vector sequence of the multi-source detection training set, the loss is calculated using the associated difference and loss function, and the model parameters are updated according to the loss calculation results.
[0119] In one embodiment, the analysis model building module 603 is specifically used to:
[0120] The track structure multi-source detection dataset is cleaned through the data anomaly cleaning model;
[0121] After the multi-dimensional feature vector sequence passes through the encoder, it is determined whether there is an abnormality by calculating the abnormal reference value and determining the abnormal point;
[0122] Remove isolated outliers through smoothing operations to obtain processed outlier data;
[0123] The processed outlier point data is removed from the track structure multi-source detection dataset to obtain the track structure multi-source detection dataset after anomaly cleaning.
[0124] In one embodiment, the analysis model building module 603 is specifically used to:
[0125] Constructing an interactive model, wherein the interactive model is stacked by interactive learning modules, and the interactive module is composed of a downsampling layer, a fully connected layer, an interactive learning layer, an upsampling layer, and two associated attention layers;
[0126] The multi-source detection training set after anomaly cleaning is input into the interactive model, the correlation features between the attributes of the data in the multi-source detection training set are extracted, the abnormal features are amplified through interactive learning, and the reconstructed data is determined;
[0127] The loss value is calculated based on the reconstructed data and the multi-source detection training set, and the model parameters are updated based on the loss value calculation result.
[0128] In one embodiment, the analysis model building module 603 is specifically used to:
[0129] Input the multi-source detection test set after anomaly cleaning into the interactive model and calculate the anomaly score of each data in the test set;
[0130] The anomaly score threshold is determined by the threshold selection algorithm, and the data in the test set with anomaly scores greater than the anomaly score threshold is judged as an abnormal interval, and the test results are output.
[0131] Based on the above invention concept, Figure 7 As shown, the present invention also proposes a computer device 700, including a memory 710, a processor 720, and a computer program 730 stored in the memory 710 and executable on the processor 720, wherein the processor 720 implements the aforementioned railway track structure multi-source data analysis method when executing the computer program 730.
[0132] An embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above-mentioned railway track structure multi-source data analysis method is implemented.
[0133] An embodiment of the present invention further provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the above-mentioned railway track structure multi-source data analysis method is implemented.
[0134] To summarize, in an embodiment of the present invention, multi-source detection data to be analyzed is obtained, anomalies are cleaned for the multi-source detection data to be analyzed by a data anomaly cleaning model, and the multi-source detection data to be analyzed after anomaly cleaning is analyzed by a track structure analysis model to obtain a railway track structure multi-source data analysis result; wherein: the data anomaly cleaning model is obtained in the following manner: historical track structure data is collected, and a track structure multi-source detection data set is constructed based on the historical track structure data; the track structure multi-source detection data set is processed to extract a multi-source detection training set and a multi-source detection test set, wherein the multi-source detection training set and the multi-source detection test set include track structure multi-source detection data and corresponding data anomaly labels; a first machine learning model is constructed, the first machine learning model is trained based on the multi-source detection training set, the first machine learning model is verified based on the multi-source detection test set, the model parameters are adjusted based on the verification results, and the verification results are used to verify the model parameters. When the accuracy reaches the preset standard, the data anomaly cleaning model is obtained; the track structure analysis model is obtained as follows: the track structure multi-source detection data set is cleaned by the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; the track structure multi-source detection data set after anomaly cleaning is processed to extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis label; a second machine learning model is constructed, the second machine learning model is trained according to the multi-source detection training set after anomaly cleaning, the second machine learning model is verified according to the multi-source detection test set after anomaly cleaning, the model parameters are adjusted according to the verification results, and the track structure analysis model is obtained when the accuracy of the verification results reaches the preset standard. In this way, by introducing the neural network model and the interactive model, the multi-source data features can be learned autonomously to achieve efficient and intelligent track structure service status detection.
[0135] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0136] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0137] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0138] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0139] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A railway track structure multi-source data analysis method, characterized in that: include: The multi-source detection data to be analyzed is obtained, and the multi-source detection data to be analyzed is cleaned of abnormalities by a data abnormality cleaning model. The multi-source detection data to be analyzed after abnormality cleaning is analyzed by a track structure analysis model to obtain the analysis results of the multi-source data of the railway track structure; wherein: The data anomaly cleaning model is obtained as follows: Collect historical track structure data and build a track structure multi-source detection data set based on the historical track structure data; The track structure multi-source detection data set is processed to extract a multi-source detection training set and a multi-source detection test set, which include the track structure multi-source detection data and corresponding data anomaly labels; Construct a first machine learning model, train the first machine learning model according to the multi-source detection training set, verify the first machine learning model according to the multi-source detection test set, adjust model parameters according to the verification result, and obtain a data anomaly cleaning model when the accuracy of the verification result reaches a preset standard; The track structure analysis model is obtained as follows: The track structure multi-source detection dataset is cleaned by the data anomaly cleaning model to obtain the track structure multi-source detection dataset after anomaly cleaning; The track structure multi-source detection data set after anomaly cleaning is processed to extract the multi-source detection after anomaly cleaning and the multi-source detection test set after anomaly cleaning of the training set. The multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels; Construct a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification results reaches the preset standard.
2. The method according to claim 1, characterized in that Based on historical track structure data, a track structure multi-source detection dataset is constructed, including: The track gauge, left height and low, right height and low, left track direction, right track direction, level, triangular pit, superelevation, curvature, long wave left height and low, long wave right height and low, long wave left track direction, long wave right track direction, horizontal acceleration and vertical acceleration are extracted from the historical track structure data in turn to construct a multi-source detection dataset of the track structure.
3. The method according to claim 1, characterized in that The track structure multi-source detection dataset is processed to extract the multi-source detection training set and multi-source detection test set, including: A multi-source detection training set is extracted from the track structure multi-source detection dataset; Some data are extracted from the track structure multi-source detection dataset, and the extracted data are cleaned for anomalies. The data before and after anomaly cleaning are used as the multi-source detection test set.
4. The method according to claim 1, characterized in that Constructing a first machine learning model, and training the first machine learning model according to the multi-source detection training set, including: Convert the multi-source detection training set into a multi-dimensional feature vector sequence; Build a Transformer-based neural network model; According to the multi-dimensional feature vector sequence of the multi-source detection training set, the loss is calculated using the associated difference and loss function, and the model parameters are updated according to the loss calculation results.
5. The method according to claim 4, characterized in that The track structure multi-source detection dataset is cleaned by the data anomaly cleaning model to obtain the track structure multi-source detection dataset after anomaly cleaning, including: The track structure multi-source detection dataset is cleaned through the data anomaly cleaning model; After the multi-dimensional feature vector sequence passes through the encoder, it is determined whether there is an abnormality by calculating the abnormal reference value and determining the abnormal point; Remove isolated outliers through smoothing operations to obtain processed outlier data; The processed outlier point data is removed from the track structure multi-source detection dataset to obtain the track structure multi-source detection dataset after anomaly cleaning.
6. The method according to claim 1, characterized in that Construct a second machine learning model, and train the second machine learning model according to the multi-source detection training set after anomaly cleaning, including: Constructing an interactive model, wherein the interactive model is stacked by interactive learning modules, and the interactive module is composed of a downsampling layer, a fully connected layer, an interactive learning layer, an upsampling layer, and two associated attention layers; The multi-source detection training set after anomaly cleaning is input into the interactive model, the attribute correlation features of the data in the multi-source detection training set are extracted, the abnormal features are amplified through interactive learning, and the reconstructed data is determined; The loss value is calculated based on the reconstructed data and the multi-source detection training set, and the model parameters are updated based on the loss value calculation result.
7. The method according to claim 6, characterized in that The interaction model is verified based on the multi-source detection test set after anomaly cleaning, including: Input the multi-source detection test set after anomaly cleaning into the interactive model and calculate the anomaly score of each data in the test set; The anomaly score threshold is determined by the threshold selection algorithm, and the data in the test set with anomaly scores greater than the anomaly score threshold is judged as an abnormal interval, and the test results are output.
8. A railway track structure multi-source data analysis device, characterized in that: include: The track structure analysis module is used to obtain the multi-source detection data to be analyzed, clean the multi-source detection data to be analyzed by using the data anomaly cleaning model, and analyze the multi-source detection data to be analyzed after the anomaly cleaning by using the track structure analysis model to obtain the railway track structure multi-source data analysis results; A cleaning model building module is used to collect historical track structure data and build a track structure multi-source detection data set based on the historical track structure data; Processing the track structure multi-source detection data set, extracting a multi-source detection training set and a multi-source detection test set, wherein the multi-source detection training set and the multi-source detection test set include the track structure multi-source detection data and corresponding data anomaly labels; constructing a first machine learning model, training the first machine learning model according to the multi-source detection training set, verifying the first machine learning model according to the multi-source detection test set, adjusting the model parameters according to the verification results, and obtaining a data anomaly cleaning model when the accuracy of the verification results reaches a preset standard; The analysis model building module is used to clean the track structure multi-source detection data set through the data anomaly cleaning model to obtain the track structure multi-source detection data set after anomaly cleaning; process the track structure multi-source detection data set after anomaly cleaning, and extract the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning, the multi-source detection training set after anomaly cleaning and the multi-source detection test set after anomaly cleaning include the track structure multi-source detection after anomaly cleaning and the corresponding track structure analysis labels; build a second machine learning model, train the second machine learning model according to the multi-source detection training set after anomaly cleaning, verify the second machine learning model according to the multi-source detection test set after anomaly cleaning, adjust the model parameters according to the verification results, and obtain the track structure analysis model when the accuracy of the verification result reaches the preset standard.
9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.
11. A computer program product, characterized in that The computer program product comprises a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.
Citation Information
Cited By
Rail transit nondestructive testing equipment and testing method thereof
CN119357876A
A rail transit nondestructive testing device and a testing method thereof
CN119357876B
Model training method and device, data cleaning method and electronic equipment
CN119782831A
Railway infrastructure detection large model method and device
CN121389037A
