A method and device for identifying damage to dirty ballast

Abstract

The present invention discloses a method and device for identifying damage to dirty ballast, and the present invention relates to the technical field of railway damage detection. The method includes the following steps: First, collect near-infrared images of the ballast surface with known hydrographic characteristic parameters, and perform preprocessing of noise reduction and enhancement to generate a training sample data set. Based on this data set, establish a neural network prediction model, with the training sample images as the input and the hydrographic characteristic parameters as the labels. Collect near-infrared images of the ballast to be detected for preprocessing to obtain predicted values of hydrographic characteristic parameters. At the same time, obtain the geometric parameters of the dirty particles. Calculate the ballast water seepage influence coefficient using the geometric parameters and the predicted values, and generate a ballast structure response coefficient in combination with the average roughness and vibration characteristic parameters of the ballast. Finally, collect environmental parameters and dynamically correct the damage judgment threshold, comprehensively generate a ballast damage index, and compare it with the corrected threshold to generate a ballast damage identification result. Efficient and accurate ballast damage identification is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of railway damage detection, and specifically to a method and device for identifying dirty ballast damage. Background Art

[0002] In railway transportation, the ballast track, as one of the main track structures, the health status of its ballast layer plays a crucial role in the safety and stability of the railway. The ballast layer not only bears the weight of the track structure but also affects the running stability and comfort of the train. However, over time, environmental factors (such as rainfall, temperature changes) and vibrations caused by train operation will lead to the dirt and damage of the ballast layer. Especially in extreme climate conditions, water accumulation and sludge accumulation will exacerbate the deterioration of the ballast, resulting in a decrease in the strength and stiffness of the ballast, and further affecting the service life and operation safety of the railway. Therefore, how to accurately identify and evaluate the damage state of the ballast has become an important issue in railway maintenance and management.

[0003] The existing ballast damage detection technologies can be roughly divided into physical detection methods and image processing methods. Physical detection methods usually rely on sensors and test equipment, such as vibration measuring instruments, ultrasonic detectors, etc. Although they can provide relatively accurate data, the high equipment cost, complex operation process, and dependence on environmental conditions limit their application. And image processing methods, especially the application of near-infrared images, provide a new idea for ballast damage identification. Near-infrared images can effectively penetrate part of the dirt and obtain the characteristic information of the ballast surface, thus overcoming the limitations of traditional methods to a certain extent. However, traditional image processing methods usually cannot be effectively combined with environmental parameters, resulting in inaccurate evaluation of the damage state. Therefore, how to combine near-infrared image technology with environmental parameters and hydrological characteristics to form an efficient and accurate damage identification scheme is an urgent problem in the current technical field.

[0004] In the prior art, the published number CN112199838A discloses a method and device for identifying damage to the ballast layer under the sleeper of a ballast track, including: establishing a two-dimensional ballast track system model for describing the response of the ballast track to be measured under the action of an excitation force, and dividing the ballast layer under the sleeper in the model into n regions; performing modal identification on the acceleration time history data of multiple data acquisition points on the sleeper of the ballast track to be measured collected in the hammering test to obtain the frequency and vibration mode of the ballast track to be measured in the damaged state; using the identified frequency and vibration mode as target data, and using the sparse Bayesian method to identify the ballast stiffness damage parameters of the model to obtain the maximum a posteriori estimate value of the ballast stiffness damage parameters in each region in the damaged state as the corresponding damage degree to determine the damage region and damage degree of the ballast layer under the sleeper of the ballast track to be measured. However, in this solution, the acceleration time history data collected through the hammering test may be affected by environmental noise or instrument errors. The data affected by noise may lead to inaccurate modal identification results. At the same time, relying solely on the acceleration time history data to judge the damage region and damage degree, the data is too single, and external parameters such as water accumulation problems caused by dirt and environmental parameters are not fully considered, thus reducing the accuracy and effectiveness of damage identification.

[0005] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and thus it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a method and device for identifying damage to dirty ballast to solve the problems raised in the above background art.

[0007] To achieve the above purpose, the present invention provides the following technical solutions:

[0008] A method for identifying damage to dirty ballast, the specific steps include:

[0009] Collect near-infrared images of the ballast surface with several known hydrological characteristic parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images and the corresponding hydrological characteristic parameters one by one to generate a training sample data set, where the hydrological characteristic parameters include the surface water accumulation area and the water accumulation volume of the ballast, and the preprocessing includes image noise reduction preprocessing and image enhancement preprocessing;

[0010] Based on the data in the training sample data set, establish a neural network prediction model, use the training sample images in the training sample data set as the input of the model, and use the hydrological characteristic parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological characteristic prediction model;

[0011] Collect the near-infrared images of the surface of the ballast to be detected. After preprocessing the images, input them into the trained hydrological feature prediction model to obtain the predicted values of the hydrological feature parameters of the ballast to be detected. At the same time, obtain the geometric parameters of the dirt particles on the surface of the ballast to be detected. The geometric parameters include the average particle size of the dirt particles, the average perimeter of the dirt particles, the average surface area of the dirt particles, and the average projected area;

[0012] According to the geometric parameters of the dirt particles on the surface of the ballast to be detected, combined with the predicted values of the hydrological feature parameters, calculate the ballast seepage influence coefficient. Obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes. Calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include the vibration frequency and amplitude;

[0013] Collect the environmental parameters of the location where the ballast to be detected is located. Dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold. Based on the obtained ballast structure response coefficient and the ballast seepage influence coefficient, comprehensively generate the ballast damage index. Compare the ballast damage index with the damage judgment correction threshold. According to different comparison results, generate the corresponding ballast damage identification results. The environmental parameters include the average rainfall, the daily average temperature, and the daily average humidity.

[0014] Furthermore, preprocess each near-infrared image of the ballast surface collected to obtain a training sample image. The preprocessing includes: unifying the image size, image enhancement, and denoising preprocessing. Among them, the wavelet transform denoising method is used to denoise each near-infrared image of the ballast surface, and bilateral filtering is used to perform image enhancement preprocessing on each near-infrared image of the ballast surface;

[0015] The generation method of the training sample data set is: map the training sample images and the corresponding hydrological feature parameters one by one to form a corresponding grid, and record the formed grid as the training sample data set;

[0016] Based on the long short-term memory network model (LSTM model), establish a neural network model, select the activation function and the optimization algorithm. Among them, select the Tanh function as the activation function and select Adam as the optimization algorithm of the LSTM model; The formula of the Tanh function is:

[0017] ;

[0018] In the formula, represents the Tanh function, and the independent variable represents the weighted sum of the inputs of the neuron, that is, the result after the inputs received by the neuron from the previous layer are weighted and summed;

[0019] Meanwhile, set the hyperparameters of the LSTM model. The hyperparameters of the LSTM model include: the number of network layers, the number of iterations, the learning rate, the batch size, the number of training times, the batch processing quantity, and the number of neurons in the hidden layer;

[0020] Among them, the number of network layers is set to a 4-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.01, the batch size is set to 32, the number of training times is set to 200, the batch processing quantity is set to 256, and the number of neurons in the hidden layer is 32;

[0021] The input of the trained hydrological feature prediction model is the preprocessed near-infrared image data of the ballast surface, and the output is the predicted values of hydrological feature parameters, including the predicted value of the accumulated water surface area of the ballast and the predicted value of the accumulated water volume.

[0022] Furthermore, according to the geometric parameters of the dirt particles on the surface of the ballast to be detected, combined with the predicted values of the hydrological feature parameters, calculate the ballast seepage influence coefficient. The specific formula for calculating the ballast seepage influence coefficient is:

[0023] ;

[0024] In the formula, is the ballast seepage influence coefficient to be detected, and are respectively the predicted value of the accumulated water surface area of the ballast to be detected and the predicted value of the accumulated water volume, is the shape factor of the dirt particle, is the average particle size of the dirt particle. The shape factor of the dirt particle is calculated according to the geometric parameters of the dirt particle. The specific formula for calculating the shape factor is:

[0025] ;

[0026] In the formula, is the average surface area of the dirt particle, is the average perimeter of the dirt particle, is the average projected area of the dirt particle.

[0027] Furthermore, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes by. Calculate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The formula for calculating the ballast structure response coefficient is:

[0028] ;

[0029] In the formula, is the ballast structure response coefficient, is the average roughness of the ballast surface to be detected, is the vibration frequency, is the vibration amplitude.

[0030] Furthermore, the average roughness of the ballast surface to be detected is obtained specifically as follows:

[0031] The ballast surface roughness measurement equipment selects a stylus profilometer. Randomly select multiple sampling areas from the ballast area to be detected. Use the stylus profilometer to measure and analyze the sampling areas to obtain the surface roughness of each sampling area, and calculate the average value of the surface roughness of all sampling areas. Take this average value as the average roughness of the ballast surface to be detected. The calculation formula is as follows:

[0032] ;

[0033] ;

[0034] In the formula, represents the sampling length of the stylus profilometer in the k-th sampling area, k is the index of the sampling area, and , K is the number of sampling areas, is the height of the q-th sampling point deviating from the center line within the sampling length of the k-th sampling area, q is the coordinate of the sampling point within the sampling length, is the surface roughness of the k-th sampling area, is the average roughness of the ballast surface to be detected, the center line and are obtained through data processing by the built-in software of the stylus profilometer.

[0035] Furthermore, based on the obtained ballast structure response coefficient and ballast water seepage influence coefficient, a ballast damage index is comprehensively generated. The formula based on which the ballast damage index is calculated is:

[0036] ;

[0037] In the formula, is the ballast damage index, and are the weight coefficients of the ballast structure response coefficient and the ballast water seepage influence coefficient to be detected respectively, where , and and are both greater than 0.

[0038] Furthermore, collect the environmental parameters of the location of the ballast to be detected, and dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold. The formula based on which the damage judgment correction threshold is calculated:

[0039] ;

[0040] In the formula, is the damage judgment correction threshold, is the preset damage judgment threshold, is the average rainfall, specifically the monthly average rainfall, is the daily average humidity, is the daily average temperature, is the reference rainfall, is the reference humidity, is the reference temperature;

[0041] Compare the ballast damage index with the damage judgment correction threshold, and generate corresponding ballast damage recognition results according to different comparison results. The specific judgment logic is as follows:

[0042] When , it is judged that the health status of the ballast to be detected is poor, indicating that the ballast to be detected is damaged and should be stopped from use immediately for repair;

[0043] When , it is judged that the health status of the ballast to be detected is good, indicating that the ballast to be detected has the risk of urban damage and should be detected and reinforced;

[0044] When , it is judged that the health status of the ballast to be detected is excellent, indicating that the ballast to be detected has no damage.

[0045] The present invention also provides a device for identifying the damage of dirty ballast. The device for identifying the damage of dirty ballast is used to execute the above-mentioned method for identifying the damage of dirty ballast, and includes:

[0046] A training sample acquisition module, which is used to collect near-infrared images of the ballast surface with several known hydrological characteristic parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images to the corresponding hydrological characteristic parameters one by one, and generate a training sample data set. The hydrological characteristic parameters include the surface water area and water volume of the ballast surface, and the preprocessing includes image noise reduction preprocessing and image enhancement preprocessing;

[0047] A prediction model training module, which is used to establish a neural network prediction model based on the data in the training sample data set, use the training sample images in the training sample data set as the input of the model, and use the hydrological characteristic parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological characteristic prediction model;

[0048] The dirt parameter acquisition module is used to acquire the near-infrared image of the surface of the ballast to be detected, preprocess the image and input it into the trained hydrological feature prediction model to obtain the predicted value of the hydrological feature parameters of the ballast to be detected. At the same time, the geometric parameters of the dirt particles on the surface of the ballast to be detected are obtained. The geometric parameters include the average particle size of the dirt particles, the average perimeter of the dirt particles, the average surface area of the dirt particles, and the average projected area.

[0049] The structural response analysis module is used to calculate the ballast seepage influence coefficient according to the geometric parameters of the dirt particles on the surface of the ballast to be detected and in combination with the predicted value of the hydrological feature parameters, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes by, and calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include the vibration frequency and amplitude.

[0050] The damage comprehensive judgment module is used to collect the environmental parameters of the location of the ballast to be detected, dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold, and comprehensively generate the ballast damage index based on the obtained ballast structure response coefficient and the ballast seepage influence coefficient. Compare the ballast damage index with the damage judgment correction threshold, and generate the corresponding ballast damage identification result according to different comparison results. The environmental parameters include the average rainfall, the average daily temperature, and the average daily humidity.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] First, by preprocessing the near-infrared image of the ballast surface, such as noise reduction and image enhancement, the image quality can be effectively improved, making the subsequent feature extraction more accurate. Second, the established neural network prediction model combines the training sample images with the hydrological feature parameters using the mapping relationship. This model can automatically extract the hydrological feature parameters of the ballast, such as the water accumulation volume and area, and then accurately predict the state of the ballast to be detected. Compared with the traditional method, the deep learning model has stronger adaptability and self-optimization ability, significantly improving the accuracy and efficiency of ballast damage identification. In addition, by combining the geometric parameters of the dirt particles on the surface of the ballast to be detected with the predicted value of the hydrological features, the ballast seepage influence coefficient is generated, providing a more comprehensive perspective for the health assessment of the ballast. Through this coefficient, the specific impacts of moisture and dirt on the ballast performance can be deeply analyzed. Finally, the dynamic correction of the collected environmental parameters (such as the average rainfall, the average daily temperature, and humidity) in the damage judgment effectively improves the accuracy of the damage judgment threshold. By real-time updating the damage judgment correction threshold, different climate conditions can be adapted, significantly reducing the risk of misjudgment caused by environmental changes. This dynamic adaptability enhances the reliability of ballast damage identification, enabling more efficient and scientific railway maintenance management. The efficient and accurate identification of dirty ballast is realized, and the monitoring and maintenance level of railway ballast is improved. Brief Description of the Drawings

[0053] Figure 1 It is a schematic diagram of the overall method flow of the present invention;

[0054] Figure 2 It is a schematic diagram of the overall device structure of the present invention. Detailed Description of the Embodiments

[0055] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with specific embodiments.

[0056] It should be noted that unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not represent any order, quantity or importance, but are only used to distinguish different components. Terms such as "including" or "comprising" mean that the elements or objects appearing before this term cover the elements or objects listed after this term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. Terms such as "upper", "lower", "left", "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0057] Embodiment:

[0058] Please refer to Figure 1 , the present invention provides a technical solution:

[0059] A method for identifying damage to dirty ballast, the specific steps include:

[0060] Step 1: Collect near-infrared images of the ballast surface with several known hydrological characteristic parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images and the corresponding hydrological characteristic parameters one by one to generate a training sample data set, where the hydrological characteristic parameters include the surface water area and the water volume of the ballast, and the preprocessing includes image noise reduction preprocessing and image enhancement preprocessing.

[0061] Preprocess each collected near-infrared image of the ballast surface to obtain a training sample image, where the preprocessing includes: unifying the image size, image enhancement and denoising preprocessing, where the denoising method using wavelet transform is used to denoise each near-infrared image of the ballast surface, and bilateral filtering is used to perform image enhancement preprocessing on each near-infrared image of the ballast surface;

[0062] The generation method of the training sample data set is as follows: Map the training sample images and the corresponding hydrological characteristic parameters one by one to form corresponding grids, and record the formed grids as the training sample data set;

[0063] The specific method for unifying the image size is as follows: Select a unified target size, such as width and height (for example, 512x512 pixels or 256x256 pixels). This size should be selected according to subsequent processing requirements and the availability of computing resources. Use a programming language (such as Python) or image processing software (such as OpenCV, PIL, etc.) to import the images to be processed, obtain the original width and height of each image for scaling processing, and select an appropriate scaling method according to specific needs. Common scaling methods include: Nearest Neighbor Interpolation: Simple and fast, but may cause jagged edges; Bilinear Interpolation: Consider the weighted average of the surrounding four pixels, with good results; Bicubic Interpolation: Consider 16 nearby pixels, usually providing smoother results.

[0064] Denoise the near-infrared image of the ballast surface using the wavelet transform denoising method. The specific steps of the wavelet transform denoising method include: Decompose the near-infrared image of the ballast surface through wavelet transform to obtain the wavelet coefficients of the image at different scales and directions; Perform threshold processing on the wavelet coefficients, set the low-amplitude wavelet coefficients to zero, and retain the high-amplitude wavelet coefficients; Perform inverse transform on the wavelet coefficients after threshold processing to reconstruct the processed coefficients into an image to complete the image denoising process;

[0065] Detail enhancement is performed on the near-infrared image of the ballast surface using bilateral filtering. The formula based on the specific filtering transformation is:

[0066] ;

[0067] In the formula, is the coordinate vector in the image coordinate system, is the coordinate vector at the gray value, is the gray value after bilateral filtering transformation, are both Gaussian functions, where The formula based on is:

[0068] ;

[0069] ;

[0070] In the formula, is the coordinate vector in the image coordinate system, is the coordinate vector at the grayscale value, and are respectively standard deviation.

[0071] Step 2: Based on the data in the training sample data set, establish a neural network prediction model. Use the training sample images in the training sample data set as the input of the model, and use the hydrological feature parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological feature prediction model.

[0072] Based on the long short-term memory network model LSTM model, establish a neural network model, select an activation function and an optimization algorithm. Among them, select the Tanh function as the activation function and select Adam as the optimization algorithm for the LSTM model; the formula of the Tanh function is:

[0073] ;

[0074] In the formula, represents the Tanh function, and the independent variable represents the input weighted sum of the neuron, that is, the result after the input received by the neuron from the previous layer is weighted and summed;

[0075] At the same time, set the hyperparameters of the LSTM model. The hyperparameters of the LSTM model include: the number of network layers, the number of iterations, the learning rate, the batch size, the number of training times, the batch processing quantity, and the number of neurons in the hidden layer;

[0076] Among them, the number of network layers is set to a 4-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.01, the batch size is set to 32, the number of training times is set to 200, the batch processing quantity is set to 256, and the number of neurons in the hidden layer is 32;

[0077] The input of the trained hydrological feature prediction model is the preprocessed near-infrared image data of the ballast surface, and the output is the predicted value of the hydrological feature parameters, including the predicted value of the ballast surface water accumulation area and the predicted value of the water accumulation volume.

[0078] In the near-infrared band (usually in the range of 700nm to 2500nm), the reflectivity of water is very low, especially near 1400nm and 1900nm. At these wavelengths, water will exhibit strong absorption characteristics. In contrast, the reflectivity of surrounding materials (such as soil or ballast) is relatively high in the same band, so it can be used to detect the ballast surface water accumulation area. Among them, the water accumulation volume can be obtained by manually pumping out the surface water and detecting it.

[0079] LSTM is particularly suitable for processing time series data, such as the prediction of hydrological characteristics. Since time series have time dependence, the data at the previous moment may have an important impact on the data at the current moment. LSTM can capture these time correlations and improve the accuracy of prediction. LSTM can learn effective features with fewer data samples, and its structure enables the model to maintain good performance in the face of data sparsity and noise. Due to its complex network structure, the LSTM model can better adapt to the changes in data when facing variable input features.

[0080] Step 3: Collect the near-infrared images of the surface of the ballast to be detected. After preprocessing the images, input them into the trained hydrological feature prediction model to obtain the predicted values of the hydrological feature parameters of the ballast to be detected. At the same time, obtain the geometric parameters of the dirt particles on the surface of the ballast to be detected. The geometric parameters include the average particle size of the dirt particles, the average perimeter of the dirt particles, the average surface area of the dirt particles, and the average projected area.

[0081] The specific method for obtaining the geometric parameters of the dirt particles on the surface of the ballast to be detected is as follows: Select several sections of the ballast to be detected for analysis. Randomly select several sampling points in each section for multiple samplings to obtain dirt particle samples. Measure the geometric parameters of the dirt particle samples through experiments. Take the average value of the geometric parameters of the dirt particle samples in the same section as the geometric parameter of this section. Finally, represent the geometric parameters of the dirt particles of the entire ballast to be detected through the average value of the geometric parameters of each section.

[0082] Among them, the sampling points are randomly selected to obtain dirt particle samples, ensuring that the collected samples can reflect the size and shape distribution of the dirt particles in this area. Use a standard sieve to physically screen particles of different particle sizes, measure the mass of the particles passing through each sieve, and calculate the particle diameter distribution according to the aperture of the sieve. Calculate the average value based on the dirt particle diameter data of each sampling point, and use this average value as the average particle size of the dirt particles in this section.

[0083] The perimeter of the dirt particles can be obtained by extracting the contour of the particles and calculating the perimeter of the particle contour through software such as OpenCV or MATLAB. Calculate the average value of the perimeter data of the particle contours at each sampling point, and use this average value as the average perimeter of the dirt particles in this section.

[0084] The projected area of the dirt particles is obtained by extracting the binary image of the particles (separating the background and the particles) in the image processing software. The software automatically calculates the projected area of the particles on the plane, or uses transparent paper to cover the particles, draw the contour and then calculate the area (suitable for small sample particles). Calculate the average value of the projected area data of the particles at each sampling point, and use this average value as the average projected area of the dirt particles in this section.

[0085] The surface area of the dirty particles can also use image processing software to calculate the actual surface area of the particles, and calculate the average value of the surface area data of the dirty particles at each sampling point, and use this average value as the surface area of the dirty particles in this section.

[0086] The preprocessing of the near-infrared image of the surface of the ballast to be detected is the same as the above method, which will not be elaborated here.

[0087] Step 4: According to the geometric parameters of the dirty particles on the surface of the ballast to be detected, combined with the predicted values of the hydrological characteristic parameters, calculate the ballast seepage influence coefficient, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes, and calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include vibration frequency and amplitude.

[0088] According to the geometric parameters of the dirty particles on the surface of the ballast to be detected, combined with the predicted values of the hydrological characteristic parameters, calculate the ballast seepage influence coefficient. The specific formula for calculating the ballast seepage influence coefficient is:

[0089] ;

[0090] In the formula, is the ballast seepage influence coefficient to be detected, and are the predicted values of the water accumulation surface area and the water accumulation volume of the ballast to be detected respectively, is the shape factor of the dirty particles, is the average particle size of the dirty particles.

[0091] It should be noted that the ballast seepage influence coefficient is used to represent the water permeability of the ballast. The larger its value, the worse the water permeability of the ballast to be detected, resulting in water accumulation and a greater probability of damage.

[0092] Among them, the predicted value of the water accumulation surface area of the ballast to be detected. The larger the water accumulation surface area, usually the more obvious the presence of water and the influence on the ballast. The more water accumulates on the ballast surface, which in turn reflects the decline in water permeability. The predicted value of the water accumulation volume of the ballast to be detected. The increase in the water accumulation volume also means that the influence of water on the penetration ability and water accumulation condition of the ballast is aggravated. Therefore, both are proportional to the ballast seepage influence coefficient to be detected. By adding these two parameters, the total influence of water accumulation is emphasized, reflecting the relationship between the water permeability of the ballast and the water accumulation condition. The more water accumulates, the worse the water permeability may be. The exponential function indicates that the predicted values of the water accumulation surface area and the water accumulation volume of the ballast to be detected significantly reflect the seepage performance of the ballast. When it increases, it will significantly increase the ballast seepage influence coefficient.

[0093] Shape factor of fouling particles , the larger the shape factor, usually indicating an increase in the surface irregularity and porosity of the particles, which may thus improve the water flow. Therefore, the shape factor of fouling particles is inversely proportional to the influence coefficient of the ballast water seepage to be detected in the form of a square emphasizes the influence of shape on the water permeability performance and increases the non-linear relationship

[0094] Average particle size of fouling particles The larger the particle size, the larger the gaps between particles may be, improving the water fluidity and resulting in less water accumulation. Therefore, the average particle size of fouling particles is inversely proportional to the influence coefficient of the ballast water seepage to be detected in the form of a logarithmic function represents the change in the influence of the smoothed particle size on the water seepage, reduces the over-sensitivity in the case of small particle sizes, and allows the change in particle size to be reflected in a relatively smooth form in the influence coefficient

[0095] Shape factor of fouling particles is calculated based on the geometric parameters of the fouling particles, where the shape factor of the fouling particles The specific formula for the calculation is as follows:

[0096] ;

[0097] In the formula, is the average surface area of the fouling particles, is the average perimeter of the fouling particles, is the average projected area of the fouling particles

[0098] Shape factor of fouling particles The larger it is, the more complex and irregular the shape of the fouling particles. The first part represents the sphericity of the fouling particles. The larger its value and the closer it is to 1, the closer the fouling particles are to the standard spherical shape. Therefore, the first part is inversely proportional to the shape factor of the fouling particles

[0099] The second part ​Represents the structure of the dirt particles, the average perimeter of the dirt particles. The perimeter is related to the boundary characteristics of the particles. A larger perimeter may imply the complexity and irregularity of the particles. The average projected area of the dirt particles. The projected area is the "shadow" of the particles in a specific direction and has a direct relationship with the shape characteristics of the particles. By normalizing the ratio of the perimeter to the projected area, the morphological characteristics of the particles in the flowing state can be further reflected. The larger the value, the more complex and irregular the structure of the dirt particles. Therefore, the second part is proportional to the shape factor of the dirt particles .

[0100] Obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes by. Calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The formula for calculating the ballast structure response coefficient is as follows:

[0101] ;

[0102] In the formula, is the ballast structure response coefficient, is the average roughness of the ballast surface to be detected, is the vibration frequency, is the vibration amplitude.

[0103] It should be noted that the ballast structure response coefficient is used to characterize the vibration situation and the average roughness of the ballast surface to be detected during normal operation. The larger the value, the rougher the ballast surface to be detected, resulting in more uneven stress, greater vibration, and a greater risk of accidents. The probability of damage to the ballast to be detected is greater.

[0104] Among them, the average roughness of the ballast surface to be detected . Roughness is an important parameter that determines the contact characteristics between the ballast and the train wheel-rail. It reflects the irregularity and friction characteristics of the ballast surface. The average roughness of the ballast surface to be detected is larger, resulting in more uneven stress on the ballast to be detected and a greater probability of damage. Therefore, the average roughness of the ballast surface to be detected is proportional to the ballast structure response coefficient . Through the square root, the influence of roughness on the ballast structure response coefficient is emphasized. A larger roughness will result in a greater response.

[0105] The vibration frequency refers to the frequency of the vibration generated when the train passes over the ballast. The higher the vibration frequency, the greater the dynamic response may be, affecting the stability of the ballast. The vibration amplitude represents the intensity or displacement of the vibration. A larger vibration amplitude will also cause an increase in the response of the ballast structure. Therefore, both the vibration frequency and the vibration amplitude are related to the ballast structure response coefficient is directly proportional, where is a comprehensive index combining vibration frequency and vibration amplitude, reflecting the total intensity of vibration. By taking the square root, it ensures that the effects of frequency and amplitude can be comprehensively obtained rather than linearly superimposed. Through the exponential function represents the attenuation characteristics of the structural response coefficient with the increase of vibration frequency and amplitude. As the vibration intensity increases, the factors affecting the stability and response of ballast will be significantly enhanced.

[0106] Among them, the vibration frequency and vibration amplitude are detected by vibration sensors set on both sides of the track, and data is obtained. The specific acquisition time period of the data is from when the train head arrives at the vibration sensor to when the train tail leaves the vibration sensor. The acquisition interval is set to 0.3 seconds. The average value of the collected vibration frequency data is calculated, and this average value is used as the vibration frequency , and the maximum value of the vibration frequency data during the acquisition process is used as the vibration amplitude .

[0107] Among them, the average surface roughness of the ballast to be detected The specific acquisition method is as follows:

[0108] A stylus profilometer is selected as the ballast surface roughness measurement device. Multiple sampling areas are randomly selected from the area of the ballast to be detected. The stylus profilometer is used to measure and analyze the sampling areas to obtain the surface roughness of each sampling area, and the average value of the surface roughness of all sampling areas is calculated. This average value is used as the average surface roughness of the ballast to be detected. The calculation formula is as follows:

[0109] ;

[0110] ;

[0111] In the formula, represents the sampling length of the stylus profilometer in the kth sampling area, k is the index of the sampling area, and , K is the number of sampling areas, is the height of the qth sampling point deviating from the center line within the sampling length of the kth sampling area, q is the coordinate of the sampling point within the sampling length, is the surface roughness of the kth sampling area, is the average surface roughness of the ballast to be detected, the center line and are obtained through data processing by the built-in software of the stylus profilometer.

[0112] Step 5: Collect the environmental parameters of the location where the ballast to be detected is located, dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain a damage judgment correction threshold, and comprehensively generate a ballast damage index based on the obtained ballast structure response coefficient and ballast water seepage influence coefficient. Compare the ballast damage index with the damage judgment correction threshold, and generate corresponding ballast damage identification results according to different comparison results. The environmental parameters include average rainfall, average daily temperature, and average daily humidity.

[0113] Based on the obtained ballast structure response coefficient and ballast water seepage influence coefficient, comprehensively generate a ballast damage index, and the formula for calculating the ballast damage index is:

[0114] ;

[0115] In the formula, is the ballast damage index, and are the weight coefficients of the ballast structure response coefficient and the ballast water seepage influence coefficient of the ballast to be detected respectively, where , and and are both greater than 0.

[0116] Among them, the larger the ballast damage index , the greater the probability that the ballast to be detected has damage. Since the proportional relationship between the ballast structure response coefficient and the ballast water seepage influence coefficient of the ballast to be detected and the ballast damage index has been described above, it will not be elaborated here. The exponential function indicates that the increase in the ballast structure response coefficient will significantly increase the ballast damage index, and the logarithmic function can effectively handle the increase in the influence of water seepage. Especially when the influence of water seepage is small, the increase of the logarithmic function is relatively gentle, which is suitable for reflecting the influence of small changes on the overall damage.

[0117] Since the ballast structure response has higher importance in the damage index, reflecting the dominant role of dynamic load on ballast damage in actual engineering, while the influence of water seepage may be affected by the environment, so , and and are both greater than 0.

[0118] Collect the environmental parameters of the location where the ballast to be detected is located, and dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain a damage judgment correction threshold. The formula for calculating the damage judgment correction threshold is:

[0119] ;

[0120] In the formula, is the damage judgment correction threshold, is the preset damage judgment threshold, is the average rainfall, specifically the monthly average rainfall, is the daily average humidity, is the daily average temperature, is the reference rainfall, is the reference humidity, is the reference temperature;

[0121] Among them, the daily average humidity is specifically the average value of the humidity detected every hour within 24 hours of the day before the detection. Similarly, the daily average temperature is the average value of the rainfall within the dates before the detection day in the month where the detection day is located.

[0122] Among them, since the greater the rainfall and humidity, the greater the probability of water accumulation in the ballast, rather than the water accumulation caused by the damage of the non-detected ballast, so the probability of the system judging that the ballast to be detected is damaged increases, resulting in misjudgment. Therefore, the rainfall and humidity are proportional to the damage judgment correction threshold and the greater the rainfall and humidity, the threshold should be increased to avoid misjudgment.

[0123] At the same time, the greater the temperature, the more likely it is to cause a decrease in the strength of the track material. Therefore, under the same vibration intensity, the probability of the detected ballast being damaged increases. Therefore, the threshold should be reduced to avoid missed judgment.

[0124] Among them, the monthly average rainfall and the reference rainfall can be obtained and set through the local meteorological station. The reference humidity is generally set to to between. The reference temperature is generally set to .

[0125] Compare the ballast damage index with the damage judgment correction threshold, and generate corresponding ballast damage recognition results according to different comparison results. The specific judgment logic is as follows:

[0126] When , it is judged that the health status of the ballast to be detected is poor, indicating that the ballast to be detected is damaged and should be stopped for use and repaired in time;

[0127] When , it is judged that the health status of the ballast to be detected is good, indicating that there is a risk of urban damage to the ballast to be detected, and detection and reinforcement should be carried out;

[0128] When , it is judged that the health status of the ballast to be detected is excellent, indicating that the ballast to be detected has no damage.

[0129] Please refer to Figure 2 , the present invention also provides a device for identifying damage to dirty ballast. The device for identifying damage to dirty ballast is used to execute the above method for identifying damage to dirty ballast, and includes:

[0130] A training sample acquisition module, which is used to acquire near-infrared images of the ballast surface with several known hydrological characteristic parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images and the corresponding hydrological characteristic parameters one by one, and generate a training sample data set. The hydrological characteristic parameters include the surface water accumulation area and the water accumulation volume of the ballast, and the preprocessing includes image noise reduction preprocessing and image enhancement preprocessing;

[0131] A prediction model training module, which is used to establish a neural network prediction model based on the data in the training sample data set, use the training sample images in the training sample data set as the input of the model, and use the hydrological characteristic parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological characteristic prediction model;

[0132] A dirty parameter acquisition module, which is used to acquire near-infrared images of the surface of the ballast to be detected, input the images into the trained hydrological characteristic prediction model after preprocessing to obtain the predicted values of the hydrological characteristic parameters of the ballast to be detected, and simultaneously obtain the geometric parameters of the dirty particles on the surface of the ballast to be detected. The geometric parameters include the average particle size of the dirty particles, the average perimeter of the dirty particles, the average surface area of the dirty particles, and the average projected area;

[0133] A structural response analysis module, which is used to calculate the ballast seepage influence coefficient according to the geometric parameters of the dirty particles on the surface of the ballast to be detected and in combination with the predicted values of the hydrological characteristic parameters, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes, and calculate and generate a ballast structural response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include the vibration frequency and amplitude;

[0134] An integrated damage judgment module, which is used to collect the environmental parameters of the location of the ballast to be detected, dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain a damage judgment correction threshold, generate a ballast damage index based on the obtained ballast structural response coefficient and the ballast seepage influence coefficient, compare the ballast damage index with the damage judgment correction threshold, and generate corresponding ballast damage identification results according to different comparison results. The environmental parameters include the average rainfall, the daily average temperature, and the daily average humidity.

[0135] The above formulas are all dimensionless and take their numerical values for calculation. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0136] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will realize that the units and algorithm steps of each example described in connection with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0137] 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. They may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0138] As mentioned above, the above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed in the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application.

Claims

1. A method for identifying damage to dirty ballast, characterized in that, The specific steps include: Collect near-infrared images of the ballast surface with several known hydrological characteristic parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images to the corresponding hydrological characteristic parameters one by one to generate a training sample data set. The hydrological characteristic parameters include the accumulated water surface area and the accumulated water volume of the ballast surface. The preprocessing includes image noise reduction preprocessing and image enhancement preprocessing; Based on the data in the training sample data set, establish a neural network prediction model. Use the training sample images in the training sample data set as the input of the model, and use the hydrological characteristic parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological characteristic prediction model; Collect the near-infrared image of the ballast surface to be detected, preprocess the image and input it into the trained hydrological characteristic prediction model to obtain the predicted value of the hydrological characteristic parameters of the ballast to be detected. At the same time, obtain the geometric parameters of the dirty particles on the surface of the ballast to be detected. The geometric parameters include the average particle size of the dirty particles, the average perimeter of the dirty particles, the average surface area of the dirty particles and the average projected area; According to the geometric parameters of the dirty particles on the surface of the ballast to be detected, combined with the predicted value of the hydrological characteristic parameters, calculate the ballast seepage influence coefficient, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes. Calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include the vibration frequency and amplitude; Collect the environmental parameters of the location of the ballast to be detected, dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold. Based on the obtained ballast structure response coefficient and the ballast seepage influence coefficient, comprehensively generate the ballast damage index. Compare the ballast damage index with the damage judgment correction threshold, and generate the corresponding ballast damage identification result according to different comparison results. The environmental parameters include the average rainfall, the daily average temperature and the daily average humidity.

2. The method for identifying damage to dirty ballast according to claim 1, wherein: Preprocess each collected near-infrared image of the ballast surface to obtain a training sample image. The preprocessing includes: unifying the image size, image enhancement and denoising preprocessing. Among them, the wavelet transform denoising method is used to denoise each near-infrared image of the ballast surface, and bilateral filtering is used to perform image enhancement preprocessing on each near-infrared image of the ballast surface; The generation method of the training sample data set is: map the training sample images to the corresponding hydrological characteristic parameters one by one to form a corresponding grid, and record the formed grid as the training sample data set; Based on the long short-term memory network model LSTM model, establish a neural network model, select an activation function and an optimization algorithm. Among them, select the Tanh function as the activation function and select Adam as the optimization algorithm of the LSTM model; The formula of the Tanh function is: ； In the formula, represents the Tanh function, and the independent variable represents the weighted sum of the inputs of the neuron, that is, the result after the inputs received by the neuron from the previous layer are weighted and summed; At the same time, set the hyperparameters of the LSTM model. The hyperparameters of the LSTM model include: the number of network layers, the number of iterations, the learning rate, the batch size, the number of training times, the batch processing quantity and the number of neurons in the hidden layer; Among them, the number of network layers is set to a 4-layer network structure, the number of iterations is set to 200, the learning rate is set to 0.01, the batch size is set to 32, the number of training times is set to 200, the batch processing quantity is set to 256, and the number of neurons in the hidden layer is 32; The input of the trained hydrological feature prediction model is the preprocessed near-infrared image data of the ballast surface, and the output is the predicted values of hydrological feature parameters, including the predicted value of the water accumulation area on the ballast surface and the predicted value of the water accumulation volume.

3. The method for identifying damage of dirty ballast according to claim 1, characterized in that: According to the geometric parameters of the dirt particles on the ballast surface to be detected, combined with the predicted values of hydrological feature parameters, the ballast seepage influence coefficient is calculated. The specific formula based on which the ballast seepage influence coefficient is calculated is as follows: ； In the formula, is the influence coefficient of ballast seepage to be detected, and are the predicted values of the surface area and volume of the ballast water accumulation to be detected, respectively, is the shape factor of the dirt particles, is the average particle size of the dirt particles, and the shape factor of the dirt particles is calculated according to the geometric parameters of the dirt particles, where the shape factor of the dirt particles The specific formula for the calculation is: ； In the formula, is the average surface area of the fouling particles, is the average perimeter of the fouling particles, is the average projected area of the fouling particles.

4. The method for identifying damage to dirty ballast according to claim 3, characterized in that: Obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes. Calculate and generate the ballast structure response coefficient based on the vibration characteristic parameters and the average roughness. The formula based on which the ballast structure response coefficient is calculated is as follows: ； In the formula, is the response coefficient of the ballast structure, is the average surface roughness of the ballast to be detected, is the vibration frequency, is the vibration amplitude.

5. The method for identifying damage of dirty ballast according to claim 4, characterized in that: wherein the average surface roughness of the ballast to be detected is obtained by the following specific method: The ballast surface roughness measurement device selects a stylus profilometer. Randomly select multiple sampling areas from the ballast area to be detected, and use the stylus profilometer to measure and analyze the sampling areas to obtain the surface roughness of each sampling area, and calculate the average value of the surface roughness of all sampling areas. Take this average value as the average roughness of the ballast surface to be detected. The calculation formula is as follows: ； ； In the formula, represents the sampling length of the stylus profilometer in the k-th sampling area, where k is the index of the sampling area, and , and K is the number of sampling areas. is the height by which the sampling point at the q-th position deviates from the center line within the sampling length of the k-th sampling area, and q is the coordinate of the sampling point within the sampling length. is the surface roughness of the k-th sampling area. is the average surface roughness of the ballast to be detected. The center line and are obtained through data processing by the built-in software of the stylus profilometer.

6. The method for identifying damage of dirty ballast according to claim 4, characterized in that: Based on the obtained ballast structure response coefficient and ballast seepage influence coefficient, a ballast damage index is comprehensively generated. The formula based on which the ballast damage index is calculated is as follows: ； In the formula, is the ballast damage index, and are the weight coefficients of the ballast structure response coefficient and the influence coefficient of the water seepage of the ballast to be detected respectively, where , and and are both greater than 0.

7. The method for identifying the damage of dirty ballast according to claim 6, wherein: Collect the environmental parameters of the location of the ballast to be detected, and dynamically correct the pre-set damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold. The formula based on which the damage judgment correction threshold is calculated: ； In the formula, is the damage judgment correction threshold, is the preset damage judgment threshold, is the average rainfall, specifically the monthly average rainfall, is the daily average humidity, is the daily average temperature, is the reference rainfall, is the reference humidity, is the reference temperature; Compare the ballast damage index with the damage judgment correction threshold, and generate corresponding ballast damage identification results according to different comparison results. The specific judgment logic is as follows: When it is determined that the health status of the ballast to be detected is poor, indicating that the ballast to be detected is damaged, and its use should be stopped immediately for repair; When it is determined that the health status of the ballast to be detected is good, indicating that there is a risk of urban damage to the ballast to be detected, and inspection and reinforcement should be carried out; When it is determined that the health status of the ballast to be detected is excellent, indicating that there is no damage to the ballast to be detected.

8. A device for identifying damage to dirty ballast, characterized in that: The described device for identifying the damage of dirty ballast is used to execute the method for identifying the damage of dirty ballast according to any one of claims 1-7, and includes: A training sample collection module, which is used to collect near-infrared images of the ballast surface with several known hydrological feature parameters, preprocess the near-infrared images of the ballast surface to obtain training sample images, map the training sample images and the corresponding hydrological feature parameters one by one to generate a training sample data set. The hydrological feature parameters include the water accumulation area on the ballast surface and the water accumulation volume. The preprocessing includes image noise reduction preprocessing and image enhancement preprocessing; A prediction model training module, which is used to establish a neural network prediction model based on the data in the training sample data set, use the training sample images in the training sample data set as the input of the model, and use the hydrological feature parameters in the training sample data set as labels to train the neural network prediction model to obtain a hydrological feature prediction model; A dirt parameter acquisition module is used to collect the near-infrared image of the surface of the ballast to be detected, preprocess the image and input it into the trained hydrological feature prediction model to obtain the predicted value of the hydrological feature parameters of the ballast to be detected. At the same time, the geometric parameters of the dirt particles on the surface of the ballast to be detected are obtained. The geometric parameters include the average particle size of the dirt particles, the average perimeter of the dirt particles, the average surface area of the dirt particles, and the average projected area. A structural response analysis module is used to calculate the ballast water seepage influence coefficient according to the geometric parameters of the dirt particles on the surface of the ballast to be detected and in combination with the predicted value of the hydrological feature parameters, obtain the average roughness of the ballast surface and the vibration characteristic parameters of the ballast when the train passes by, and calculate and generate the ballast structure response coefficient according to the vibration characteristic parameters and the average roughness. The vibration characteristic parameters include the vibration frequency and amplitude. A damage comprehensive judgment module is used to collect the environmental parameters of the location of the ballast to be detected, dynamically correct the preset damage judgment threshold according to the obtained environmental parameters to obtain the damage judgment correction threshold, comprehensively generate the ballast damage index based on the obtained ballast structure response coefficient and the ballast water seepage influence coefficient, compare the ballast damage index with the damage judgment correction threshold, and generate the corresponding ballast damage identification result according to different comparison results. The environmental parameters include the average rainfall, the daily average temperature, and the daily average humidity.

Images