Railway track crack identification method based on machine identification

Abstract

The invention discloses a railway track crack identification method based on machine identification, and relates to the technical field of track crack identification, and the method comprises the steps: extracting a sample signal from a multi-dimensional sample data set, decomposing the sample signal, calculating the energy entropy of the decomposed sample signal, and calculating the energy entropy of the decomposed sample signal; dividing the sample signal into K sample clusters with similar energy entropy characteristics; inputting the sample signal of each sample cluster into a crack identification model, and calculating the identification accuracy of each sample cluster; calculating the weight of each dimension sample signal feature according to the accuracy of each dimension sample signal, fusing the weighted sample signal features to form a comprehensive feature vector, inputting the comprehensive feature vector into a crack recognition model for recognition, and performing classification recognition on the crack according to the output result of the classification recognition model. Whether the current railway track has cracks or not is judged, so that possible misjudgment caused by single dimension and interference of low-quality data on model judgment are avoided.

Description

Technical Field

[0001] The invention relates to the technical field of track crack identification, and in particular to a railway track crack identification method based on machine identification. Background Art

[0002] During the long-term operation of railway tracks, tiny cracks will form due to factors such as material fatigue and stress concentration. If these cracks are not discovered and repaired in time, they will gradually expand and eventually lead to rail breakage, causing serious safety accidents. Traditional rail crack detection methods mainly rely on physical methods such as manual inspection and ultrasonic detection. These methods have problems such as low detection efficiency and susceptibility to human factors. In recent years, with the rapid development of computer vision and deep learning technology, rail crack detection methods based on image recognition have gradually become a research hotspot.

[0003] In a Chinese invention application with application publication number CN118799245A, a method and device for identifying sleeper cracks are disclosed, wherein the method comprises: obtaining a sleeper image; inputting the sleeper image into a trained sleeper crack semantic segmentation model to obtain a mask image containing sleeper cracks; the sleeper crack semantic segmentation model is obtained by training a deep learning semantic segmentation network based on a sleeper image with cracks that has been pixel-annotated; extracting the contour pixel position and contour outer frame of the sleeper crack from the mask image, and obtaining the number and position of the sleeper cracks based on the contour pixel position and contour outer frame; performing skeleton extraction on the mask image to obtain the skeleton line of the sleeper crack, and calculating the length of the sleeper crack based on the skeleton line of the sleeper crack; extracting the normal of the skeleton line of the sleeper crack, and calculating the width of the sleeper crack based on the normal of the skeleton line of the sleeper crack.

[0004] In a Chinese invention application with application publication number CN111301489A, a method for online monitoring of track cracks is disclosed, including: judging whether a train is in the process of running, if so, starting the system, filtering the acceleration signal collected by the acceleration sensor and the acoustic emission signal collected by the acoustic emission sensor, and extracting characteristic values ​​respectively, the characteristic values ​​including the peak value and main frequency information of the acceleration signal and the number of events, the number of ringings, the event duration and the acoustic emission rate of the acoustic emission signal; if the peak value, the main frequency information, the number of events, the number of ringings, the event duration and the acoustic emission rate meet the preset conditions, it is determined that a crack has occurred, and the characteristic value is stored and sent.

[0005] In view of the above invention, the prior art still has the following deficiencies:

[0006] 1. In the current railway track crack recognition technology, crack detection mainly relies on a single signal source (such as image data). This singleness not only limits the flexibility and adaptability of the recognition system, but also may lead to one-sidedness or errors in the recognition results due to the limitations of the signal source. For example, relying solely on image data may not fully capture key information such as the depth and shape of the crack, thereby affecting the accuracy and reliability of recognition;

[0007] 2. In actual applications, due to factors such as weather changes, differences in lighting conditions, and the performance of collection equipment, the collected railway track data often has uneven quality. However, the existing technology has not fully taken this into consideration. There is no quality assessment of data in different dimensions, nor is there any classification and processing of data of different qualities. As a result, during the recognition process, low-quality data may interfere with the judgment of the model and reduce the accuracy of recognition. Summary of the invention

[0008] 1. Technical issues to be resolved

[0009] In view of the shortcomings of the prior art, the present invention provides a railway track crack recognition method based on machine recognition, which extracts sample signals from a multi-dimensional sample data set, decomposes the sample signals, calculates the energy entropy of the decomposed sample signals, and uses the energy entropy as a classification basis to divide the samples into K sample clusters with similar energy entropy characteristics; inputs the sample signals of each sample cluster into a crack recognition model respectively, and calculates the recognition accuracy of each sample cluster; calculates the weight of the sample signal features of each dimension according to the accuracy of the sample signals of each dimension, fuses the weighted sample signal features to form a comprehensive feature vector, inputs the comprehensive feature vector into the crack recognition model, performs crack recognition, and judges whether there are cracks in the current railway track according to the output result of the classification recognition model, thereby solving the problems mentioned in the background technology.

[0010] (II) Technical solution

[0011] To achieve the above objectives, the present invention is implemented by the following technical solutions: a railway track crack identification method based on machine identification, comprising the following steps:

[0012] Select a section of railway track, collect multi-dimensional railway track sample data, including sound signals, image signals and vibration signals, and construct a multi-dimensional sample data set;

[0013] Extract sample signals from a multi-dimensional sample data set, match the corresponding signal decomposition method for the sample signals of each dimension, decompose the sample signals, and calculate the energy entropy of the decomposed sample signals;

[0014] Taking energy entropy as the classification basis, the K-mean++ algorithm is used to classify the sample signals at the same latitude and divide them into K sample clusters with similar energy entropy characteristics;

[0015] Select a machine learning model, obtain a crack recognition model after training, input the sample signals of each sample cluster into the crack recognition model, calculate the recognition accuracy of each sample cluster, determine the energy entropy range of each sample cluster and the corresponding accuracy, and establish an energy entropy mapping relationship;

[0016] Multi-dimensional railway track sample signals are collected in real time. According to the accuracy of the sample signals in each dimension, the weights of the sample signal features in each dimension are calculated. The weighted sample signal features are fused to form a comprehensive feature vector. The comprehensive feature vector is input into the crack recognition model for crack recognition. According to the output results of the classification recognition model, it is determined whether there are cracks in the current railway track.

[0017] Further, based on the known crack characteristics, the multi-dimensional railway track sample data is labeled, and the sample data is labeled as "containing cracks" or "not containing cracks";

[0018] The labeled sample data are integrated according to timestamps and location information to construct a multi-dimensional sample data set, including but not limited to an image signal sample data set, a sound signal sample data set, and a vibration signal sample data set. The sample data set of each dimension is divided into a training set, a validation set, and a test set.

[0019] Furthermore, decomposing the sample signal includes:

[0020] For image signals, image decomposition operations are performed to obtain a series of sub-band images. For sound signals, spectrum analysis is performed to obtain the spectrum diagram or spectrum characteristics of the sound signals. For vibration signals, vibration signal decomposition is performed to obtain a series of intrinsic mode functions or local mean functions.

[0021] Furthermore, the energy entropy of the sample signal of each dimension after decomposition is calculated including:

[0022] For image signals, the energy distribution of the pixel value of each subband image is calculated, and the energy entropy is calculated based on the energy distribution. For sound signals, the energy of each frequency component is calculated, and the energy entropy is calculated based on the energy distribution. For vibration signals, the energy of each intrinsic mode function or local mean function is calculated, and the energy entropy is calculated based on the energy distribution.

[0023] Furthermore, the energy entropy calculation formula is as follows:

[0024]

[0025] Where H represents energy entropy, pi represents the probability of occurrence of the i-th energy value, i = 1, 2, …, N, N represents the total number of energy values, ε represents the correction parameter, 0<ε<<1.

[0026] Furthermore, the process of using the K-mean++ algorithm to classify sample signals at the same latitude is as follows:

[0027] For each dimension of sample signal, the number of clusters K is pre-set, and the K-mean++ algorithm is used to initialize the cluster center; according to the energy entropy of each sample signal, the distance from each sample signal to all cluster centers is calculated, and according to the calculated distance, each sample is assigned to the cluster center closest to it to form a sample cluster;

[0028] For each sample cluster, the energy entropy mean of all sample signals in each sample cluster is calculated, and the sample signal closest to the energy entropy mean is used as the new cluster center.

[0029] Further, check whether the cluster center changes. If the cluster center does not change or the change is less than the preset convergence threshold, the algorithm is determined to have converged and the iteration is stopped; otherwise, the iteration is continued until the algorithm converges.

[0030] When the algorithm converges, the sample signals of each dimension are divided into K different sample clusters according to the final clustering center, and each sample cluster represents sample signals with similar energy entropy characteristics.

[0031] Furthermore, for each sample cluster, the prediction results of the crack recognition model are compared with the sample data annotations to calculate the recognition accuracy of each sample cluster: recognition accuracy = number of correctly recognized samples / total number of samples in each sample cluster.

[0032] Furthermore, for the sample signal of each dimension, the energy entropy of the sample signal of each dimension is calculated, and the correct rate corresponding to the sample signal of each dimension is found according to the established energy entropy mapping relationship, and the weight of the sample signal feature of each dimension is calculated. The weight calculation formula is:

[0033]

[0034] Among them, ω j represents the weight of the sample signal feature of the jth dimension, Ac j It represents the accuracy of the sample signal of the jth dimension, j = 1, 2, …, M, and M represents the number of dimensions.

[0035] Furthermore, according to the output result of the classification recognition model, it is judged whether there is crack in the current railway track, including:

[0036] If the recognition result shows that there is a crack, an alarm will be issued and relevant personnel will be prompted to repair it. If the recognition result shows that there is no crack, sample signals will continue to be collected in real time for monitoring.

[0037] (III) Beneficial effects

[0038] The present invention provides a railway track crack identification method based on machine identification, which has the following beneficial effects:

[0039] (1) The signal decomposition method is used to decompose the complex signal into simpler and easier to analyze parts, and the energy entropy of the decomposed signal is calculated, which realizes the effective extraction of rail crack characteristics. Energy entropy, as an indicator to measure the uniformity of signal energy distribution, can reflect the energy distribution characteristics of different frequency or scale components in the signal, which helps to improve the sensitivity of crack identification and provide a data basis for subsequent crack identification.

[0040] (2) Energy entropy, as an indicator to measure signal complexity and information content, can reflect the energy distribution of the signal at different frequencies or scales, and to a certain extent determines the quality of the signal, thereby highlighting the differences between different sample signals. The K-mean++ algorithm is a partition-based clustering algorithm that iteratively divides the sample signal into K sample clusters with similar characteristics. It can effectively classify the sample signals according to the energy entropy characteristics, so that sample signals with similar energy entropy characteristics are divided into the same sample cluster, thereby improving the accuracy of crack identification.

[0041] (3) By inputting the sample signals of each sample cluster into the trained crack recognition model and calculating the recognition accuracy of each sample cluster, the accuracy of sample signals in different energy entropy ranges for crack recognition can be accurately evaluated, which helps to screen out those signal features that are most sensitive and accurate for crack recognition, thereby improving the overall recognition accuracy. The energy entropy range and corresponding accuracy of each sample cluster in each dimension of sample signal are determined, and the energy entropy mapping relationship is established, which can provide an important reference basis for subsequent crack recognition work.

[0042] (4) By collecting multi-dimensional railway track sample signals (images, sounds, vibrations) in real time and calculating the energy entropy of each sample signal, the track status can be more comprehensively reflected, thereby improving the accuracy of crack identification. Since multiple signal sources are used, even if a certain signal source is interfered with or restricted, other signal sources can still provide useful information, thereby enhancing the robustness of identification;

[0043] (5) By quickly determining the sample cluster to which the detected signal belongs, the recognition accuracy information associated with the sample cluster can be quickly called. The accuracy of the sample signal in each dimension reflects the reliability of the dimension in crack recognition. Through weighted fusion, the information of different dimensions can be more accurately integrated to avoid possible misjudgment of a single dimension, thereby improving the overall recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the steps of a railway track crack identification method based on machine identification according to the present invention;

[0045] Figure 2 The figure is a flow chart of the railway track crack identification method based on machine identification according to the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] See also Figure 1-Figure 2 The present invention provides a railway track crack identification method based on machine identification, comprising the following steps:

[0048] Step 1: Select a section of railway track, collect multi-dimensional railway track sample data, including but not limited to sound signals, image signals and vibration signals, and construct a multi-dimensional sample data set;

[0049] The step 1 includes the following contents:

[0050] Step 101: Select a section of railway track and collect multi-dimensional railway track sample data, including but not limited to sound signals, image signals and vibration signals. For example, use a high-resolution camera or a camera mounted on a drone to collect images of the railway track surface, use acoustic emission technology to collect sound signals of the railway track, and use a vibration sensor to collect vibration signals of the railway track.

[0051] When selecting railway tracks, representative railway sections should be selected as data collection points based on the importance of the railway line, frequency of use, historical crack records and other factors. The length of the railway track can be 100 meters, 200 meters, 500 meters, etc. When collecting sample data, in order to increase the diversity and accuracy of the samples, the data should be collected repeatedly in different time periods and under different weather conditions.

[0052] Step 102: labeling the multi-dimensional railway track sample data according to the known crack characteristics, and labeling the sample data as "including cracks" or "excluding cracks". The labeling can be completed by manual labeling or automatic labeling tools;

[0053] Step 103: Integrate the labeled sample data according to timestamps and location information to construct a multi-dimensional sample data set, including but not limited to an image signal sample data set, a sound signal sample data set, and a vibration signal sample data set. According to the data labeling results, divide the sample data set of each dimension into a training set, a validation set, and a test set;

[0054] When using, combine the contents of step 101 to step 103:

[0055] By collecting multi-dimensional sample data such as image signals, sound signals and vibration signals, the status of the railway track can be reflected more comprehensively. Data of different dimensions can reveal the existence and characteristics of cracks from different angles and complement each other, thereby improving the accuracy of crack identification.

[0056] Step 2: Extract sample signals from the sample data sets of each dimension respectively, match the corresponding signal decomposition method for the sample signals of each dimension, decompose the sample signals, and calculate the energy entropy of the decomposed sample signals;

[0057] The step 2 includes the following contents:

[0058] Step 201: obtaining a multi-dimensional sample data set, extracting sample signals from the sample data set of each dimension, and matching the corresponding signal decomposition method to the sample signals of each dimension;

[0059] For example, for image signals, wavelet transform, Fourier transform or discrete cosine transform are used to decompose the image signals into sub-bands of different frequencies or scales; for sound signals, wavelet transform, Fourier transform and other methods are used to perform spectrum analysis to decompose the sound signals into different frequency components; for vibration signals, empirical mode decomposition, local mean decomposition and other methods are used to decompose the vibration signals into a series of intrinsic mode functions or local mean functions;

[0060] Step 202: Decompose the sample signal of each dimension using a matching signal decomposition method. For image signals, perform image decomposition operation to obtain a series of sub-band images. For sound signals, perform spectrum analysis to obtain a spectrum diagram or spectrum feature of the sound signal. For vibration signals, perform vibration signal decomposition to obtain a series of intrinsic mode functions or local mean functions.

[0061] Step 203: Calculate the energy entropy of the sample signal of each dimension after decomposition. For image signals, calculate the energy distribution of the pixel value of each sub-band image, and calculate the energy entropy based on the energy distribution. For sound signals, calculate the energy of each frequency component, and calculate the energy entropy based on the energy distribution. For vibration signals, calculate the energy of each intrinsic mode function or local mean function, and calculate the energy entropy based on the energy distribution. The energy entropy calculation formula is as follows:

[0062]

[0063] Where H represents energy entropy, p i represents the probability of the i-th energy value appearing, i = 1, 2, ..., N, N represents the total number of energy values, ε represents the correction parameter used to avoid mathematical errors, 0 < ε < < 1;

[0064] It should be noted that energy entropy is an indicator used to measure the uniformity of energy distribution in a signal. The calculation of energy entropy is based on the probability distribution of each energy value in the signal. If the energy distribution in the signal is relatively uniform, then the value of energy entropy will be relatively large. On the contrary, if the energy distribution is relatively concentrated, then the value of energy entropy will be relatively small, which is used to reflect the change of track state and the existence of cracks.

[0065] When used, combine the contents of step 201 to step 203:

[0066] The signal decomposition method is used to decompose complex signals into simpler and easier to analyze parts, and the energy entropy of the decomposed signal is calculated, which realizes the effective extraction of rail crack characteristics. Energy entropy, as an indicator to measure the uniformity of signal energy distribution, can reflect the energy distribution characteristics of different frequency or scale components in the signal, which helps to improve the sensitivity of crack identification and provide a data basis for subsequent crack identification.

[0067] Step 3: Use energy entropy as the basis for classification and use the K-mean++ algorithm to classify sample signals at the same latitude into K sample clusters with similar energy entropy characteristics;

[0068] The step three includes the following contents:

[0069] Step 301: For each dimension of sample signal, the number of clusters K (i.e., the number of sample clusters) is preset, and the K-mean++ algorithm is used to initialize the cluster center;

[0070] It should be noted that the number of clusters can be determined based on actual conditions and prior knowledge. It is usually necessary to select the best K value through experiments or experience. The K-mean++ algorithm is an improved K-means algorithm that accelerates convergence and reduces errors by selecting the initial cluster center.

[0071] Specifically, first randomly select a sample as the first cluster center, then calculate the shortest distance between each sample and the selected cluster center, and select the next cluster center according to the probability distribution of the sum of the squares of these distances to ensure that the distance between the initial cluster centers is as far as possible;

[0072] Step 302: for the sample signals of each dimension, the distance from each sample signal to all cluster centers is calculated according to the energy entropy of each sample signal, and each sample is assigned to the cluster center closest to it according to the calculated distance to form a sample cluster, wherein the distance calculation can use Euclidean distance, Manhattan distance or other suitable distance measurement methods, and use energy entropy as a variable in the distance formula for calculation;

[0073] Step 303: For each sample cluster, calculate the energy entropy mean of all sample signals in each sample cluster, and use the sample signal closest to the energy entropy mean as a new cluster center;

[0074] Step 304: Check whether the cluster center has changed. If the cluster center has not changed or the change is less than the preset convergence threshold, it is determined that the algorithm has converged and the iteration is stopped; otherwise, return to step 302 to continue iteration;

[0075] It should be noted that we check whether the cluster center has changed. Here, "change" refers to whether the cluster center has moved or adjusted during the iteration process. This change is compared with the preset convergence threshold. The convergence threshold is a standard for judging whether the change of the cluster center is small enough to consider that the algorithm has converged.

[0076] The setting of the convergence threshold can be based on experience to set a reasonable threshold, set a smaller threshold (such as 0.001 or less) to ensure that the cluster center can be stabilized during the iteration process, or adjust the convergence threshold through experiments. By running the algorithm multiple times and observing the changes in the clustering results, a suitable threshold range can be gradually found;

[0077] If the change of the cluster center is less than the convergence threshold, it is considered to have converged, that is, the position of the cluster center has stabilized and no longer changes significantly. In this case, the iteration is stopped; if the change of the cluster center is greater than or equal to the convergence threshold, it is considered to have not converged yet, and it is necessary to continue iteration to further adjust the position of the cluster center;

[0078] Step 305: When the algorithm converges, the sample signals of each dimension are divided into K different sample clusters according to the final clustering center, and each sample cluster represents sample signals with similar energy entropy characteristics;

[0079] When used, combine the contents of step 301 to step 305:

[0080] Energy entropy, as an indicator to measure signal complexity and information content, can reflect the energy distribution of signals at different frequencies or scales, and to a certain extent determines the quality of the signal, thereby highlighting the differences between different sample signals. The K-mean++ algorithm is a partition-based clustering algorithm that iteratively divides sample signals into K sample clusters with similar characteristics. It can effectively classify sample signals according to energy entropy characteristics, so that sample signals with similar energy entropy characteristics are divided into the same sample cluster, thereby improving the accuracy of crack identification.

[0081] Step 4: Select a machine learning model, obtain a crack recognition model after training, input the sample signals of each sample cluster into the crack recognition model, calculate the recognition accuracy of each sample cluster, determine the energy entropy range of each sample cluster in each dimension sample signal, and the corresponding accuracy, and establish an energy entropy mapping relationship;

[0082] The step 4 includes the following contents:

[0083] Step 401: Select a machine learning model, such as a convolutional neural network (CNN), a support vector machine (SVM), or a random forest, and train it using a training set. After verification with a validation set and testing with a test set, a crack recognition model is obtained to ensure that the model can process and recognize sample signals from different dimensions (images, sounds, vibrations).

[0084] For example, a standard CNN architecture, including convolutional layers, pooling layers, and fully connected layers, is used to extract image features, convert sound signals into features such as spectrograms or Mel-frequency cepstral coefficients (MFCCs), and then use CNN for feature extraction, preprocess the vibration signal (such as denoising and normalization), and then use a one-dimensional CNN for feature extraction;

[0085] The features from different branches are concatenated, added or fused through the attention mechanism, and classified using a fully connected layer and a softmax function to output the probability of whether a crack exists or not;

[0086] Get the training set, validation set, and test set from the multi-dimensional sample data set, select the cross entropy loss function as the loss function for the classification task, select optimizers such as Adam or SGD to update the weights, use the training set data to train the model, update the model parameters through the back propagation algorithm, monitor the performance of the model on the validation set, and stop training when the performance no longer improves to prevent overfitting;

[0087] Evaluate the performance of the model on the test set, including indicators such as accuracy, recall, and F1 score. Adjust the model's hyperparameters (such as learning rate, batch size, convolution kernel size, etc.) through grid search, random search, or Bayesian optimization to improve model performance and obtain the final crack recognition model.

[0088] Step 402: for each dimension of sample signal, the sample signal of each sample cluster is respectively input into the crack recognition model to perform crack recognition test. For each sample cluster, the prediction result of the crack recognition model is compared with the actual annotation result (including cracks or not), and the recognition accuracy of each sample cluster is calculated: recognition accuracy = number of correctly recognized samples / total number of samples in each sample cluster;

[0089] Step 403: Determine the energy entropy range of each sample cluster in each dimension sample signal and the corresponding accuracy, and establish an energy entropy mapping relationship;

[0090] When used, combine the contents of step 401 to step 403:

[0091] By inputting the sample signals of each sample cluster into the trained crack recognition model and calculating the recognition accuracy of each sample cluster, the accuracy of sample signals in different energy entropy ranges for crack recognition can be accurately evaluated, which helps to screen out those signal features that are most sensitive and accurate for crack recognition, thereby improving the overall recognition accuracy. Determining the energy entropy range and corresponding accuracy of each sample cluster in each dimension of sample signal, and establishing an energy entropy mapping relationship can provide an important reference basis for subsequent crack recognition work.

[0092] Step 5: Collect multi-dimensional railway track sample signals in real time, calculate the weight of the sample signal features of each dimension according to the accuracy of the sample signal of each dimension, fuse the weighted sample signal features to form a comprehensive feature vector, input the comprehensive feature vector into the crack recognition model to perform crack recognition, and judge whether there are cracks in the current railway track based on the output results of the classification recognition model.

[0093] The step five includes the following contents:

[0094] Step 501: Collect multi-dimensional railway track sample signals in real time, and for the sample signals of each dimension, use the same signal decomposition method and energy entropy calculation method as in step 2 to calculate the energy entropy of the sample signals of each dimension;

[0095] Step 502: According to the energy entropy mapping relationship established in step 4, find the correct rate corresponding to each dimension sample signal, calculate the weight of each dimension sample signal feature after normalization, assign weights to each dimension sample signal feature, fuse the weighted features, and form a comprehensive feature vector. The weight calculation formula is:

[0096]

[0097] Among them, ω j represents the weight of the sample signal feature of the jth dimension, Ac j represents the accuracy of the sample signal of the jth dimension, j = 1, 2, ..., M, M represents the number of dimensions;

[0098] It should be noted that, according to the calculated accuracy, a weight is assigned to the sample features of each dimension, and the size of the weight is proportional to the accuracy, that is, the dimension with a higher accuracy has a larger weight. For example, if the accuracy corresponding to the image signal is A, the accuracy corresponding to the sound signal is B, and the accuracy corresponding to the vibration signal is C, then the weight of the image signal is A / A+B+C, the weight of the sound signal is B / A+B+C, and the weight of the vibration signal is C / A+B+C;

[0099] Step 503: Input the fused comprehensive feature vector into the crack recognition model to perform crack recognition. According to the output result of the classification recognition model, determine whether there are cracks in the current railway track. If the recognition result is that there are cracks, an alarm is issued and relevant personnel are prompted to repair. If the recognition result is that there are no cracks, continue to collect sample signals in real time for monitoring.

[0100] When used, combine the contents of step 501 to step 503:

[0101] By collecting multi-dimensional railway track sample signals (images, sounds, vibrations) in real time and calculating the energy entropy of each sample signal, the track status can be more comprehensively reflected, thereby improving the accuracy of crack identification. Due to the use of multiple signal sources, even if a certain signal source is interfered or restricted, other signal sources can still provide useful information, thereby enhancing the robustness of identification;

[0102] By quickly determining the sample cluster to which the signal to be detected belongs, the recognition accuracy information associated with the sample cluster can be quickly called up. The accuracy of the sample signal in each dimension reflects the reliability of that dimension in crack recognition. Through weighted fusion, information from different dimensions can be more accurately integrated to avoid possible misjudgments in a single dimension, thereby improving the overall recognition accuracy.

[0103] In the application, the several formulas involved are all calculated by removing dimensions and taking their numerical values. The formula is a formula of the most recent real situation obtained by collecting a large amount of data and performing software simulation. The coefficients in the formula are set by technicians in this field according to actual conditions.

[0104] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in combination with electronic hardware, computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0105] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0106] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A railway track crack identification method based on machine recognition, characterized in that: The following steps are involved: Select a section of railway track, collect multi-dimensional railway track sample data, including sound signals, image signals and vibration signals, and construct a multi-dimensional sample data set; Extract sample signals from a multi-dimensional sample data set, match the corresponding signal decomposition method for the sample signals of each dimension, decompose the sample signals, and calculate the energy entropy of the decomposed sample signals; Taking energy entropy as the classification basis, the K-mean++ algorithm is used to classify the sample signals at the same latitude and divide them into K sample clusters with similar energy entropy characteristics; Select a machine learning model, obtain a crack recognition model after training, input the sample signals of each sample cluster into the crack recognition model, calculate the recognition accuracy of each sample cluster, determine the energy entropy range of each sample cluster and the corresponding accuracy, and establish an energy entropy mapping relationship; Multi-dimensional railway track sample signals are collected in real time. According to the accuracy of the sample signals in each dimension, the weights of the sample signal features in each dimension are calculated. The weighted sample signal features are fused to form a comprehensive feature vector. The comprehensive feature vector is input into the crack recognition model for crack recognition. According to the output results of the classification recognition model, it is determined whether there are cracks in the current railway track.

2. The railway track crack identification method based on machine recognition according to claim 1 is characterized in that: According to known crack characteristics, the multi-dimensional railway track sample data is labeled, and the sample data is labeled as containing cracks or not containing cracks; The labeled sample data are integrated according to timestamps and location information to construct a multi-dimensional sample data set, including but not limited to an image signal sample data set, a sound signal sample data set, and a vibration signal sample data set. The sample data set of each dimension is divided into a training set, a validation set, and a test set.

3. The railway track crack identification method based on machine recognition according to claim 1 is characterized in that: Decomposing the sample signal includes: For image signals, image decomposition operations are performed to obtain a series of sub-band images. For sound signals, spectrum analysis is performed to obtain the spectrum diagram or spectrum characteristics of the sound signals. For vibration signals, vibration signal decomposition is performed to obtain a series of intrinsic mode functions or local mean functions.

4. The railway track crack identification method based on machine recognition according to claim 3 is characterized in that: Calculating the energy entropy of each dimension sample signal after decomposition includes: For image signals, the energy distribution of pixel values ​​of each subband image is calculated, and the energy entropy is calculated based on the energy distribution. For sound signals, the energy of each frequency component is calculated, and the energy entropy is calculated based on the energy distribution. For vibration signals, the energy of each intrinsic mode function or local mean function is calculated, and the energy entropy is calculated based on the energy distribution.

5. The railway track crack identification method based on machine identification according to claim 4 is characterized in that: The energy entropy calculation formula is as follows: Where H represents energy entropy, p i represents the probability of occurrence of the i-th energy value, i = 1, 2, …, N, N represents the total number of energy values, ε represents the correction parameter, 0<ε<<1.

6. The railway track crack identification method based on machine identification according to claim 1 is characterized in that: The process of using the K-mean++ algorithm to classify sample signals at the same latitude is as follows: For each dimension of sample signal, the number of clusters K is pre-set, and the K-mean++ algorithm is used to initialize the cluster center; according to the energy entropy of each sample signal, the distance from each sample signal to all cluster centers is calculated, and according to the calculated distance, each sample is assigned to the cluster center closest to it to form a sample cluster; For each sample cluster, the energy entropy mean of all sample signals in each sample cluster is calculated, and the sample signal closest to the energy entropy mean is used as the new cluster center.

7. The railway track crack identification method based on machine identification according to claim 6 is characterized in that: Check whether the cluster center has changed. If the cluster center has not changed or the change is less than the preset convergence threshold, the algorithm is considered to have converged and the iteration is stopped; otherwise, the iteration continues until the algorithm converges; When the algorithm converges, the sample signals of each dimension are divided into K different sample clusters according to the final clustering center, and each sample cluster represents sample signals with similar energy entropy characteristics.

8. The railway track crack identification method based on machine recognition according to claim 1, characterized in that: For each sample cluster, the prediction results of the crack recognition model are compared with the sample data annotations to calculate the recognition accuracy of each sample cluster: recognition accuracy = number of correctly recognized samples / total number of samples in each sample cluster.

9. The railway track crack identification method based on machine recognition according to claim 1, characterized in that: For the sample signal of each dimension, the energy entropy of the sample signal of each dimension is calculated. According to the established energy entropy mapping relationship, the correct rate corresponding to the sample signal of each dimension is found, and the weight of the sample signal feature of each dimension is calculated. The weight calculation formula is: Among them, ω j represents the weight of the sample signal feature of the jth dimension, Ac j It represents the accuracy of the sample signal of the jth dimension, j = 1, 2, …, M, and M represents the number of dimensions.

10. The railway track crack identification method based on machine identification according to claim 9, characterized in that: According to the output results of the classification and recognition model, it is judged whether there are cracks in the current railway track, including: If the recognition result shows that there is a crack, an alarm will be issued and relevant personnel will be prompted to repair it. If the recognition result shows that there is no crack, sample signals will continue to be collected in real time for monitoring.

Images