An intelligent brake detection system for railway freight cars

Abstract

The present invention discloses an intelligent brake detection system for railway freight cars, which relates to the technical field of intelligent monitoring of railway freight cars and includes collecting thermal imaging and brake state data for preprocessing; calculating the gradient of the entire image to generate a high-temperature gradient region; calculating the total contact area and the area of the closed polygon, calculating the local temperature rise and the temperature field, and calculating the predicted temperature field after local heat conduction based on the local temperature rise and the temperature field; calculating the brake risk index and judging the brake state. The present invention preprocesses by collecting thermal imaging and brake state data, calculates the predicted temperature field after local heat conduction based on the local temperature rise and the temperature field, calculates the brake risk index, and judges the brake state; enhances the applicability and generation quality of complex image generation tasks, improves the efficiency and intelligent level of railway freight car brake monitoring, and ensures the stability and generalization ability of perception loss in different tasks.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent monitoring of railway freight cars, and particularly to an intelligent brake detection system for railway freight cars. Background Art

[0002] Railway freight cars are an important part of the modern logistics transportation system. Their operation safety and efficiency are crucial for transportation and economic development. The brake device is the core component of the railway freight car braking system, and its working state directly affects the operation safety of the train. In traditional technologies, the state monitoring of the brake device mainly relies on manual inspection and simple sensor monitoring based on mechanical performance. This method is not only inefficient but also easily affected by human factors, and cannot meet the high requirements of modern railway transportation for real-time and accuracy. With the development of sensor technology, computer vision, and artificial intelligence, the brake state detection system based on multi-source data fusion and intelligent analysis has gradually become a research hotspot. In particular, by collecting key data through technologies such as thermal imaging and pressure sensing, and combining deep learning algorithms for data processing and fault prediction, the automation level and detection accuracy of the monitoring system can be significantly improved. However, the current technical solutions still have many deficiencies in terms of the depth of data processing and fusion, the degree of refinement, and the real-time performance, and cannot fully meet the intelligent detection requirements of the railway freight car brake device.

[0003] The deficiencies of the existing technologies in the field of brake detection are mainly reflected in the following aspects. The existing signal analysis methods lack intelligence and real-time performance, and fail to make full use of machine learning or deep learning technologies to deeply analyze and optimize the detection data. In terms of image processing and data fusion, the existing technologies are difficult to balance real-time performance and accuracy, resulting in a lag in the monitoring of the brake system state. Summary of the Invention

[0004] In view of the problems existing in the above-mentioned existing intelligent brake detection system for railway freight cars, the present invention is proposed.

[0005] Therefore, the problems to be solved by the present invention are that the existing signal analysis methods lack intelligence and real-time performance, and fail to make full use of machine learning or deep learning technologies to deeply analyze and optimize the detection data. In terms of image processing and data fusion, the existing technologies are difficult to balance real-time performance and accuracy, resulting in a lag in the monitoring of the brake system state.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: An intelligent brake detection system for railway freight cars, which includes a collection and processing module for collecting thermal imaging and brake state data and preprocessing the collected thermal imaging and brake state data; a filtering and calculation module for calculating the gradient of the entire image to generate a high-temperature gradient region; a calculation and prediction module for calculating the total contact area and the area of the closed polygon, calculating the local temperature rise and temperature field, and calculating the predicted temperature field after local heat conduction based on the local temperature rise and temperature field; an index judgment module for calculating the brake risk index and judging the brake state; a visualization and storage module for constructing a visualization interface to display the brake state result and storing the thermal imaging and brake state data collected and analyzed.

[0007] As a preferred solution of the intelligent brake detection system for railway freight cars of the present invention, wherein: the collection of thermal imaging and brake state data means installing a thermal imaging sensor at the front end of the brake device to collect thermal imaging data, installing a pressure sensor at the joint of the brake hydraulic system, and installing an optical slip speed sensor on the wheel side edge to collect brake state data;

[0008] Synchronize the collected thermal imaging and brake state data using the Network Time Protocol.

[0009] The brake state data includes pressure data and slip speed data.

[0010] As a preferred solution of the intelligent brake detection system for railway freight cars of the present invention, wherein: the preprocessing of the collected thermal imaging and brake state data includes,

[0011] Using the moving average method to smooth and denoise the brake state data, using the interquartile range method to screen and delete outliers in the brake state data, using the linear interpolation method to fill in the missing data in the brake state data, and normalizing the smoothed brake state data;

[0012] Using Gaussian filtering to denoise the collected thermal imaging data, using the Jet mapping to convert the denoised thermal imaging data into a pseudo-color image, using the dynamic window adjustment method to set the segmentation window, and segmenting the pseudo-color image according to the segmentation window to obtain a segmented image;

[0013] Calculate the gradient amplitude of the segmented image Use a pre-trained CycleGAN model to generate a blurred mask from the segmented image Extract the blurred region

[0014] For the blurred region Perform a fast Fourier transform to obtain the spectrum F blur (u,v), calculate the spectrum F blur(u, v) high-frequency energy distribution F high (u, v);

[0015] Calculate the fuzzy region gradient mean μ G and standard deviation σ G , based on the gradient mean μ G and standard deviation σ G calculate the fuzzy kernel parameter θ k ;

[0016] Based on the fuzzy kernel parameter θ k , construct the two-dimensional Gaussian fuzzy kernel K(u, v);

[0017] Based on the high-frequency energy distribution F high (u, v) optimize the two-dimensional Gaussian fuzzy kernel K optim (u, v), the formula is:

[0018] K optim (u, v) = K(u, v) · (1 - β · |F high (u, v)|),

[0019] where β is the optimization coefficient;

[0020] Use the optimized two-dimensional Gaussian fuzzy kernel K optim (u, v) to compensate the spectrum F blur (u, v) to obtain the compensated frequency-domain data F com (u, v);

[0021] For the compensated frequency-domain data F com (u, v) use a band-pass Gaussian filter for smoothing to obtain the smoothed frequency-domain data F filtered (u, v);

[0022] Use to convert the smoothed frequency-domain data back to the spatial domain to obtain the compensated fuzzy region image (i, j, t);

[0023] Replace the compensated fuzzy region image into the original segmented image to obtain the initial segmentation compensation image

[0024] Stitch the initial segmentation compensation image and the fuzzy mask into an input tensor

[0025]

[0026] Collect thermal images with historical labels for preprocessing and extract historical depth features to generate a training set;

[0027] The clear image of the target is obtained by using the clear image mapping of thermal imaging based on historical tags

[0028] Extract tensors using a pre-trained ResNet-50 model The deep feature F deep ;

[0029] Construct a generative adversarial network, including a generator and a discriminator;

[0030] Set the input format of the generative adversarial network to the deep feature F deep ;

[0031] Input the deep feature into the generator to generate an optimized compensation image Use the discriminator to receive the generated optimized compensation image and the clear target image, and output the scores of the compensation image and the clear target;

[0032] Use a pre-trained VGG network to extract the feature representations of the compensation image and the clear target image;

[0033] Train the generative adversarial network using the training set;

[0034] Calculate the perceptual loss using spatial error Calculate the adversarial loss using cross-entropy loss

[0035] Based on the perceptual loss And the adversarial loss Construct the loss function of the generator The formula is:

[0036]

[0037] Where Is the feature representation of the clear target image at the l-th layer, and l is the feature layer;

[0038] Construct the loss function of the discriminator using cross-entropy loss

[0039] Use the loss change monitoring method to set the training stop condition. When the loss function of the generator And the loss function of the discriminator Meet the training stop condition, stop training;

[0040] Input the real-time deep feature F deep Into the trained generative adversarial network to obtain the final segmentation compensation image;

[0041] Use overlapping weight fusion for the final segmentation compensation image to obtain the corrected thermal imaging data T mer (i,j,t).

[0042] As a preferred solution of the intelligent brake detection system for railway freight cars according to the present invention, wherein: calculating the global gradient to generate a high-temperature gradient region means calculating the temperature gradients in the horizontal and vertical directions based on the corrected thermal imaging data using the Sobel operator;

[0043] Calculating the global gradient using the gradient magnitude method based on the temperature gradients in the horizontal and vertical directions;

[0044] Setting a gradient threshold using the mean standard deviation method, marking the global gradients greater than the gradient threshold, and generating a high-temperature gradient region.

[0045] As a preferred solution of the intelligent brake detection system for railway freight cars according to the present invention, wherein: calculating the total contact area and the area of the closed polygon means setting a neighborhood radius using the k-nearest neighbor distance curve method and setting a minimum number of points using the data dimension method;

[0046] Calculating the set of points within the neighborhood of each pixel point in the high-temperature gradient region. When the neighborhood point set is greater than or equal to the minimum number of points, it is determined as a core point and marked as visited. When the neighborhood point set is less than the minimum number of points, it is determined as a boundary point, and all the pixel points within the neighborhood are added to the core points until all the pixel points in the high-temperature gradient region are visited and the expansion stops;

[0047] Using the contact area point set composed of core points and boundary points, calculating the total contact area A using discrete integration total ;

[0048] Using the convex hull algorithm to sort the boundary points in geometric order to generate a set of closed boundary polygon points;

[0049] Calculating the area A(t) of the closed polygon using the Shoelace Formula.

[0050] As a preferred solution of the intelligent brake detection system for railway freight cars according to the present invention, wherein: calculating the local temperature rise and temperature field, and calculating the predicted temperature field after local heat conduction based on the local temperature rise and temperature field means obtaining the material parameters of the friction block from the material database and performing preprocessing;

[0051] Calculating the local heat power P(t) of the high-temperature region according to the frictional force and slip velocity;

[0052] Calculating the local temperature rise ΔT(t) at time t based on the local heat power P(t) and material parameters;

[0053] Using the explicit difference method to discretize the heat conduction equation and calculating the temperature field T predict (i,j,t);

[0054] According to the current local temperature rise ΔT(t) and the temperature field T predict (i,j,t), calculate the predicted temperature field T predict (i,j,t + 1);

[0055] The material parameters include specific heat capacity, density, thickness, and the temperature threshold allowed for the material.

[0056] As a preferred solution of the intelligent brake detection system for railway freight cars described in the present invention, wherein: the calculated brake risk index refers to calculating the temperature difference ΔT predict (i,j,t + 1) based on the predicted temperature rise T predict (i,j,t + 1);

[0057] Compare the temperature difference with the temperature threshold allowed for the material to generate over-temperature points M risk (i,j,t + 1), and calculate the area of the over-temperature region;

[0058] Based on the average temperature rise ΔT avg and the area A of the over-temperature region risk Calculate the brake risk index R(t) at time t.

[0059] As a preferred solution of the intelligent brake detection system for railway freight cars described in the present invention, wherein: the determination of the brake state refers to setting a risk threshold using a statistical distribution;

[0060] Compare the brake risk index R(t) at time t with the risk threshold. If the brake risk index is greater than or equal to the risk threshold, it is determined as abnormal, trigger an alarm signal and notify the maintenance personnel via email;

[0061] If the brake risk index is less than the risk threshold, it is determined as normal, and continue to continuously monitor the thermal imaging and brake state data.

[0062] As a preferred solution of the intelligent brake detection system for railway freight cars described in the present invention, wherein: the construction of a visual interface to display the brake state result refers to using React.js to construct a visual interface to display the brake risk index and the brake state;

[0063] Allow users who have passed real-name verification to view.

[0064] As a preferred solution of the intelligent brake detection system for railway freight cars according to the present invention, wherein: the thermal imaging and brake state data generated by storage, collection and analysis refer to storing the collected thermal imaging and brake state data, and the brake risk indicators and brake states generated by analysis in a central database. The central database is sorted in chronological order and corresponding tags are marked. Meanwhile, the collected thermal imaging and brake state data, and the brake risk indicators and brake states generated by analysis are backed up to the cloud, and the integrity of the backup data is detected regularly, significantly improving the data quality.

[0065] The beneficial effects of the present invention are as follows: by collecting thermal imaging and brake state data, preprocessing the collected thermal imaging and brake state data, calculating the gradient of the whole image to generate a high thermal gradient region, calculating the total contact area and the area of the closed polygon, calculating the local temperature rise and temperature field, calculating the predicted temperature field after local heat conduction based on the local temperature rise and temperature field, calculating the brake risk indicators, and judging the brake state; it enhances the applicability and generation quality of complex image generation tasks, improves the efficiency and intelligent level of railway freight car brake monitoring, and ensures the stability and generalization ability of perception loss in different tasks. Brief Description of the Drawings

[0066] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0067] Figure 1 It is a schematic structural diagram of the intelligent brake detection system for railway freight cars.

[0068] Figure 2 It is a schematic flow diagram of the intelligent brake detection system for railway freight cars. Detailed Embodiments

[0069] To make the above objects, features and advantages of the present invention more obvious and understandable, the following will describe the detailed embodiments of the present invention in conjunction with the drawings in the specification.

[0070] Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar promotions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0071] Second, the "one embodiment" or "embodiment" referred to herein means a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The appearances of "in one embodiment" in different places in this specification do not all refer to the same embodiment, nor are they separate or selectively exclusive embodiments from other embodiments.

[0072] Embodiment 1, referring to Figure 1 and Figure 2 , is the first embodiment of the present invention. This embodiment provides an intelligent brake detection system for railway freight cars. The intelligent brake detection system for railway freight cars includes

[0073] S1. A collection and processing module, which is used to collect thermal imaging and brake state data and preprocess the collected thermal imaging and brake state data;

[0074] Specifically, collecting thermal imaging and brake state data means installing a thermal imaging sensor at the front end of the brake device to collect thermal imaging data, installing pressure sensors at the joints of the brake hydraulic system, and installing optical slip speed sensors on the wheel flanges to collect brake state data;

[0075] Use the Network Time Protocol to synchronize the collected thermal imaging and brake state data in terms of time;

[0076] The brake state data includes pressure data and slip speed data.

[0077] By combining multiple data sources such as thermal imaging, pressure, and slip speed, the present invention can achieve comprehensive monitoring of the brake state. In particular, the introduction of thermal imaging data provides a new dimension for the abnormal detection of the brake state, greatly improving the detection accuracy and reliability. By detecting the slip speed between the wheel and the track during braking, it is possible to analyze whether the brake performance is normal and prevent situations of excessive braking or skidding out of control. Using the Network Time Protocol to synchronize and process the data enables the system to perform fusion analysis of multi-source data under transient working conditions. Compared with traditional detection methods that rely on a single data source, this system significantly improves the real-time performance and intelligence level.

[0078] Furthermore, preprocessing the collected thermal imaging and brake state data includes

[0079] Using the moving average method to perform smoothing and denoising processing on the brake state data, using the interquartile range method to screen and delete outliers in the brake state data, using the linear interpolation method to fill in missing data in the brake state data, and performing normalization processing on the smoothed brake state data;

[0080] Denoise the collected thermal imaging data using Gaussian filtering, convert the denoised thermal imaging data into a pseudo-color image using Jet mapping, set a segmentation window using the dynamic window adjustment method, and segment the pseudo-color image according to the segmentation window to obtain a segmented image;

[0081] Calculate the gradient magnitude of the segmented image Generate a blur mask from the segmented image using a pre-trained CycleGAN model 1 indicates that the pixel belongs to the blurred area, 0 indicates the non-blurred area, and extract the blurred area

[0082] For the blurred area Perform a fast Fourier transform to obtain the spectrum F blur (u, v), calculate the spectrum F blur (u, v)'s high-frequency energy distribution F high (u, v);

[0083] Calculate the gradient mean μ of the blurred area G and the standard deviation σ G , based on the gradient mean μ G and the standard deviation σ G calculate the blur kernel parameter θ k , the formula is:

[0084]

[0085] Based on the blur kernel parameter θ k , construct a two-dimensional Gaussian blur kernel K(u, v), the formula is:

[0086]

[0087] where u and v are the frequency domain coordinates after Fourier transform, corresponding to the frequency components of the image in the horizontal and vertical directions respectively;

[0088] Based on the high-frequency energy distribution F high (u, v) optimize the two-dimensional Gaussian blur kernel K optim (u, v), the formula is:

[0089] K optim (u, v) = K(u, v) · (1 - β · |F high (u, v)|),

[0090] where β is the optimization coefficient;

[0091] High-frequency information usually represents texture or detail information in image processing. For the optimization of the blur kernel, these high-frequency characteristics need to be precisely considered. By introducing the high-frequency energy distribution, the processing intensity of the blur kernel on the high-frequency region can be effectively controlled. This targeted method enables the blur kernel optimization to avoid over-smoothening details and maintain the image texture features, and adaptively adjusts the optimization coefficient according to different high-frequency energy distributions. The traditional two-dimensional Gaussian blur kernel optimization method is usually adjusted based on global characteristics, while this method adjusts the blur kernel through the local high-frequency energy distribution, thus more precisely optimizing different regions. The two-dimensional Gaussian blur kernel is usually adjusted based on fixed parameters or simple global optimization, without fully considering the influence of high-frequency characteristics on the blur effect. The present invention introduces F high (u, v), and directly integrates the high-frequency distribution information into the blur kernel optimization, significantly improving the ability to retain high-frequency characteristics. The optimization formula of the present invention is particularly suitable for image processing scenarios with high resolution and complex textures, such as medical images, remote sensing images, and industrial inspections.

[0092] Use the optimized two-dimensional Gaussian blur kernel K optim (u, v) to compensate the spectrum F blur (u, v), and obtain the compensated frequency-domain data F com (u, v). The formula is:

[0093]

[0094] where ∈ is the regularization parameter to prevent the denominator from being zero;

[0095] Perform smoothing processing on the compensated frequency-domain data F com (u, v) using a band-pass Gaussian filter to obtain the smoothed frequency-domain data F filtered (u, v). The formula is:

[0096]

[0097] where H filter (u, v) is the value of the Gaussian filter at the frequency-domain position (u, v), and θ f is the control frequency range, which is dynamically adjusted according to the image resolution and blur degree;

[0098] Convert the smoothed frequency-domain data back to the spatial domain to obtain the compensated blurred region image The formula is:

[0099]

[0100] where k is the dimension size of the segmented image, t is the time, g is the imaginary unit, H is the height of the segmented image, and W is the width of the segmented image. is the contribution of the vertical direction frequency component. is the contribution of the horizontal direction frequency component, where i and j are the pixel coordinates in the spatial domain, representing the vertical and horizontal positions of the image respectively;

[0101] Replace the compensated blurred region image into the original segmented image to obtain the initial segmented compensation image

[0102] Concatenate the initial segmented compensation image and the blurred mask into an input tensor

[0103] Collect thermal images with historical labels for preprocessing and extract historical depth features to generate a training set;

[0104] Use the thermal image based on the historical label to map with a clear image to obtain the target clear image

[0105] Use a pre-trained ResNet-50 model to extract the depth feature F of the tensor of the tensor deep ;

[0106] Build a generative adversarial network, including a generator and a discriminator;

[0107] Set the input format of the generative adversarial network as the depth feature F deep ;

[0108] Input the depth feature into the generator to generate an optimized compensation image Use the discriminator to receive the generated optimized compensation image and the target clear image, and output the scores of the compensation image and the target clear image;

[0109] Use the training set to train the generative adversarial network;

[0110] Use a pre-trained VGG network to extract the feature representations of the compensation image and the target clear image;

[0111] Use the spatial error to calculate the perceptual loss The formula is:

[0112]

[0113] where is the feature representation of the compensation image at the l-th layer, is the feature representation of the target clear image at the l-th layer, and l is the feature layer;

[0114] Use the cross-entropy loss to calculate the adversarial loss The formula is:

[0115]

[0116] wherein is the score of the discriminator for the compensated image;

[0117] Based on the perceptual loss and the adversarial loss construct the loss function of the generator The formula is:

[0118]

[0119] The perceptual loss can measure the similarity between the generated image and the target clear image in the high-level feature space in the image generation task. Such high-level features usually correspond to the texture and content information of the image. It is difficult to capture these high-level features directly using pixel-level losses (such as mean squared error). The perceptual loss is based on the feature representation of a specific layer in a pre-trained network (such as the VGG network), which can better optimize the generator to generate images with more natural textures. The adversarial loss, through the feedback of the discriminator, prompts the generator to generate results closer to the real data distribution. Using only the perceptual loss may result in the generated image being close to the target image in the feature space but lacking in the overall visual effect (such as image sharpness). The normalization process of the perceptual loss ensures that the calculation of the loss will not be imbalanced due to the scale difference of the target feature values. By the weighted combination of the perceptual loss and the adversarial loss, two optimization objectives of feature similarity and image realism are comprehensively considered. This joint optimization strategy has unique advantages in high-resolution image generation tasks and can achieve a better balance between the quality of the generated image and feature retention. In many applications of the perceptual loss, the target features are not normalized, which may lead to uneven loss weights between different image datasets or feature layers. By normalizing the feature values of the target clear image, the stability and generalization ability of the perceptual loss in different tasks are ensured. Existing methods often use the perceptual loss or the adversarial loss alone, resulting in the generator optimization being too biased towards a specific target. This formula, through the combination of the perceptual loss and the adversarial loss, takes into account both the matching of high-level features and the optimization of overall realism, achieving high-quality output of the generated image. The present invention uses the feature representation of a specific layer to ensure that the optimization process is adjusted for the key feature levels, improving the generation ability of the generator. The perceptual loss of this formula is based on high-level features, which can capture complex textures and semantic information. At the same time, combined with the adversarial loss, it further enhances the realism and detail expressiveness of the generated image.

[0120] Use the cross-entropy loss to construct the loss function of the discriminator The formula is:

[0121]

[0122] wherein is the score of the discriminator for the target clear image, is the discriminator's score for the compensated image;

[0123] The training stop condition is set using the loss change monitoring method. When the loss function of the generator and the loss function of the discriminator meet the training stop condition, the training is stopped;

[0124] Input the real-time depth feature F deep into the trained generative adversarial network to obtain the final segmentation compensated image;

[0125] Use overlapping weight fusion on the final segmentation compensated image to obtain the corrected thermal imaging data T mer (i, j, t).

[0126] Through the combined use of the moving average method, interquartile range method, and linear interpolation method, the brake state data can be efficiently cleaned and repaired, ensuring its high stability and accuracy, improving the reliability of subsequent analysis, avoiding result deviations caused by data anomalies or missing data. Through Gaussian filtering denoising and Jet mapping pseudocolor processing, the noise in the thermal imaging data is significantly reduced, and the visual effect of the image is more intuitive. Using the CycleGAN model to generate a blur mask and extract the blurred area, analyzing the high-frequency energy distribution through fast Fourier transform, and optimizing the two-dimensional Gaussian blur kernel, the blurred area in the thermal imaging data can be accurately identified and effectively compensated. The compensated image is clearer and the information expression is more complete. The constructed generative adversarial network can generate a compensated image close to the target clear image through depth feature extraction and perceptual loss optimization. By correcting the final segmentation compensated image through overlapping weight fusion, the overall accuracy and reliability of the thermal imaging data can be significantly improved.

[0127] S2. A filtering calculation module for calculating the full-image gradient to generate a high-heat gradient region;

[0128] Specifically, calculating the full-image gradient to generate a high-heat gradient region means calculating the temperature gradients in the horizontal and vertical directions based on the corrected thermal imaging data using the Sobel operator;

[0129] Calculate the full-image gradient using the gradient magnitude method based on the temperature gradients in the horizontal and vertical directions;

[0130] Set the gradient threshold using the mean standard deviation method, mark the full-image gradients greater than the gradient threshold, and generate a high-heat gradient region.

[0131] The Sobel operator can efficiently calculate the temperature gradients in the horizontal and vertical directions of the corrected thermal imaging data, thereby precisely capturing the directional information of the surface temperature change of the brake device, helping the system identify possible overheating edges and abnormal heat source positions. By using the gradient magnitude method, the temperature gradients in the horizontal and vertical directions are synthesized into a scalar value, which can intuitively reflect the gradient intensity of each pixel point. The generation of the full-image gradient can effectively display the areas with the most significant temperature changes in the thermal imaging map. By setting the gradient threshold using the mean standard deviation method, the threshold range can be dynamically adjusted according to the data distribution, thereby improving the extraction accuracy of high-temperature gradient regions. Marking the high-temperature gradient regions can significantly highlight the overheating parts that need attention in the thermal imaging map and intuitively display the areas where abnormalities may exist. This dynamic adjustment mechanism significantly improves the accuracy of high-temperature gradient region extraction and avoids missed detections or false detections caused by fixed thresholds. Compared with traditional overall thermal map analysis methods, this region marking method can more accurately lock the problem areas and improve the diagnostic efficiency. By analyzing the thermal imaging data in real time, the system can identify possible overheating problems of the equipment in advance, thereby preventing safety hazards caused by brake failure or wear.

[0132] S3. A calculation and prediction module, which is used to calculate the total contact area and the area of the closed polygon, calculate the local temperature rise and the temperature field, and calculate the predicted temperature field after local heat conduction based on the local temperature rise and the temperature field;

[0133] Specifically, calculating the total contact area and the area of the closed polygon means setting the neighborhood radius using the k-nearest neighbor distance curve method and setting the minimum number of points using the data dimension method;

[0134] For each pixel point in the high-temperature gradient region, calculate the set of points within its neighborhood. When the neighborhood point set is greater than or equal to the minimum number of points, it is determined as a core point and marked as visited. When the neighborhood point set is less than the minimum number of points, it is determined as a boundary point, and all the pixel points within the neighborhood are added to the core points until all the pixel points in the high-temperature gradient region are visited and the expansion stops;

[0135] For the set of contact area points composed of core points and boundary points, calculate the total contact area A using discrete integration total ;

[0136] Use the convex hull algorithm to sort the boundary points in geometric order to generate a set of closed boundary polygon points;

[0137] Use the Shoelace Formula to calculate the area A(t) of the closed polygon.

[0138] Set the neighborhood radius through the k-nearest neighbor distance curve method, and combine with the data dimension method to set the minimum number of points, so that the pixel clustering in the high heat gradient region has higher robustness and adaptability, effectively distinguish the core points and boundary points, reduce the interference of noise on region extraction. By recursively expanding the neighborhood set of each pixel point in the high heat gradient region, all relevant pixel points can be completely covered to form a set of contact region points. The generation method ensures the coherence and integrity of the high heat region extraction. Use the discrete integral method to calculate the area of the set of contact region points, which can accurately quantify the total contact area of the high heat region, providing an intuitive basis for evaluating the braking effect of the brake device and the equipment status. Through the convex hull algorithm to geometrically sort the boundary points, a closed set of boundary polygon points is generated to ensure a high geometric accuracy of the boundary definition of the high heat gradient region. This boundary information is convenient for subsequent area calculation and shape analysis. By using the Shoelace Formula to quickly calculate the area of the closed polygon, the shape characteristics and area size of the high heat region can be effectively quantified. Combining with the total contact area for analysis can more comprehensively evaluate the performance of the brake device.

[0139] Furthermore, calculate the local temperature rise and temperature field. Based on the local temperature rise and temperature field, calculate the predicted temperature field after local heat conduction, which means obtaining the material parameters of the friction block from the material database and performing preprocessing.

[0140] According to the friction force and slip speed, calculate the local heat power P(t) of the high heat region. The formula is:

[0141]

[0142] where F(t) is the pressure at time t, and v(t) is the slip speed at time t.

[0143] Based on the local heat power P(t) and material parameters, calculate the local temperature rise ΔT(t) at time t. The formula is:

[0144]

[0145] where Δt is the time interval, C is the specific heat capacity, ρ is the density, and h is the thickness of the friction block.

[0146] Use the explicit difference method to discretize the heat conduction equation and calculate the temperature field T predict (i,j,t);

[0147] According to the current local temperature rise ΔT(t) and temperature field T predict (i,j,t), calculate the predicted temperature field T predict (i,j,t + 1). The formula is:

[0148] T predict(i,j,t+1) = T predict (i,j,t) + ΔT(t),

[0149] The material parameters include specific heat capacity, density, thickness, and the temperature threshold allowed for the material.

[0150] Obtain the material parameters of the friction block from the material database and perform preprocessing to make the calculation more accurate and efficient. Material parameters such as specific heat capacity and density directly affect the rate of heat conduction and the amplitude of local temperature rise, while the temperature threshold helps to determine whether the material is within the safe operating range. Calculate the local heat power based on the frictional force and slip velocity, which can quantitatively describe the rate of heat generation in the high-temperature region, providing the initial conditions for subsequent local temperature rise and temperature field calculations, ensuring the accuracy of heat conduction analysis. By combining the local heat power with the material parameters, the change in local temperature rise can be accurately predicted. The local temperature rise data can be used to identify potential overheating regions in the brake device, helping to detect potential material failure risks in advance. The explicit difference method discretizes the continuous heat conduction equation, enabling rapid calculation of the dynamic changes of temperature in time and space. By combining the local temperature rise and the temperature distribution after heat conduction, predicting the temperature field can intuitively present the diffusion state of heat after time t. This predictive ability can identify the heat conduction characteristics and potential failure risks of high-temperature regions in advance.

[0151] S4. An index judgment module, used to calculate the brake risk index and judge the brake state;

[0152] Specifically, calculating the brake risk index means calculating the temperature difference ΔT predict (i,j,t+1) based on the predicted temperature rise T predict (i,j,t+1);

[0153] Compare the temperature difference with the temperature threshold allowed for the material to generate the over-temperature point M risk (i,j,t+1), and calculate the area A of the over-temperature region risk , and the formula is:

[0154] A risk = Δx·Δy·∑ (i,j) M risk (i,j,t+1),

[0155] where Δx is the physical distance between pixel points in the horizontal direction, and Δy is the physical distance between pixel points in the vertical direction;

[0156] Calculate the average temperature rise ΔT of the over-temperature region avg ;

[0157] Based on the average temperature rise ΔT avg and the area A of the over-temperature region riskCalculate the braking risk index R(t) at time t, and the formula is:

[0158]

[0159] where A total is the total contact area.

[0160] Predicting the temperature rise can intuitively reflect the local thermal response of the braking device. By calculating the temperature difference with the initial temperature, the change in thermal stress borne by the material during operation can be clearly identified, which helps to detect potential thermal runaway risks in advance. Predicting the temperature rise can intuitively reflect the local thermal response of the braking device. By calculating the temperature difference with the initial temperature, the change in thermal stress borne by the material during operation can be clearly identified, which helps to detect potential thermal runaway risks in advance. It can quickly screen out high-temperature points that may exceed the material's tolerance limit, providing guidance for subsequent thermal management measures. By performing spatial integration on the over-temperature points to calculate the over-temperature area, the spatial distribution characteristics of the high-temperature risk can be quantified. The larger the over-temperature area, the higher the potential risk of thermal runaway of the equipment. Combining the average temperature rise and the over-temperature area to calculate the braking risk index can comprehensively evaluate the thermodynamic performance and risk status of the equipment. The risk index is a comprehensive parameter that can quantify the safety margin during equipment operation.

[0161] Furthermore, judging the braking state means setting a risk threshold using statistical distribution;

[0162] Compare the braking risk index R(t) at time t with the risk threshold. If the braking risk index is greater than or equal to the risk threshold, it is judged as abnormal, an alarm signal is triggered and the maintenance personnel are notified by email;

[0163] If the braking risk index is less than the risk threshold, it is judged as normal, and the thermal imaging and braking state data continue to be monitored continuously.

[0164] By analyzing the historical statistical distribution of the braking risk index, a risk threshold suitable for the current working condition can be dynamically set. Compared with the fixed threshold method, this dynamic adjustment method can better adapt to different environmental conditions and equipment states, significantly improving the accuracy of the monitoring system. Comparing the braking risk index calculated in real time with the dynamically set risk threshold can quickly judge the operating state of the equipment. This process can efficiently identify potential high-temperature or abnormal states and trigger corresponding response mechanisms according to the results. When the braking risk index is greater than or equal to the risk threshold, the system automatically triggers an alarm signal and notifies the maintenance personnel by email. The notification method shortens the equipment failure response time and reduces the risk caused by delayed failures. When the braking risk index is lower than the risk threshold, the system continues to monitor the thermal imaging data and the braking state data, thereby realizing continuous equipment operation health management. Through continuous monitoring, the system can timely capture the change trend of potential risks.

[0165] S5, a visualization storage module, is used to build a visualization interface to display the brake status results and store the thermal imaging and brake status data generated by collection and analysis.

[0166] Specifically, building a visualization interface to display the brake status results means using React.js to build a visualization interface to display the brake risk indicators and brake status.

[0167] It allows users who have passed real-name verification to access.

[0168] By using React.js to build a visualization interface, the dynamic update and response of the page can be quickly realized, ensuring that the brake risk indicators and status information can be presented to users in real time. The component-based design of React.js also makes the interface development more efficient and facilitates subsequent function expansion. By visualizing the display of brake risk indicators and status information, users can intuitively understand the operation of the device. Through the real-name verification mechanism, the system can effectively restrict data access rights to ensure that only authorized users can view the brake status information.

[0169] Furthermore, storing the thermal imaging and brake status data generated by collection and analysis means storing the collected thermal imaging and brake status data, as well as the brake risk indicators and brake status generated by analysis, into the central database. The central database is sorted in chronological order and marked with corresponding tags. At the same time, the collected thermal imaging and brake status data, as well as the brake risk indicators and brake status generated by analysis, are backed up to the cloud, and the integrity of the backup data is detected regularly.

[0170] By storing the collected thermal imaging and brake status data, as well as the brake risk indicators generated by analysis, into the central database, centralized management of data can be achieved. The chronological sorting ensures the storage logic and retrieval efficiency of the data, while the tag marking enhances the flexibility of data classification. The cloud backup function provides additional data security protection and can quickly restore data in case of a central database failure. By regularly detecting the integrity of the backup data, it can be ensured that the data has not been damaged or tampered with, providing accurate and reliable basic data for subsequent analysis and decision-making.

Claims

1. An intelligent brake detection system for railway freight cars, characterized in that: include, A collection and processing module, used for collecting thermal imaging and brake status data, and pre-processing the collected thermal imaging and brake status data; The filtering calculation module is used to calculate the full-image gradient to generate high thermal gradient areas; A calculation prediction module is used to calculate the total contact area and the area of ​​the closed polygon, calculate the local temperature rise and temperature field, and calculate the predicted temperature field after local heat conduction based on the local temperature rise and temperature field; Index judgment module, used to calculate the brake risk index and judge the brake status; Visual storage module, used to build a visual interface to display the brake status results, store, collect and analyze the generated thermal imaging and brake status data; The collecting of thermal imaging and brake status data refers to installing a thermal imaging sensor at the front end of the brake device to collect thermal imaging data, installing a pressure sensor at the joint of the brake hydraulic system, and installing an optical slip speed sensor at the side edge of the wheel to collect brake status data; Use the time protocol to synchronize the collected thermal imaging and brake status data; The brake state data includes pressure data and slip speed data.

2. The intelligent brake detection system for railway freight cars according to claim 1, characterized in that: The preprocessing of the collected thermal imaging and brake status data includes: The brake status data is smoothed and denoised using the sliding average method, the outliers of the brake status data are screened and deleted using the interquartile range method, the missing data of the brake status data are filled using the linear interpolation method, and the smoothed brake status data is normalized; The collected thermal imaging data is denoised using Gaussian filtering, the denoised thermal imaging data is converted into a pseudo-color image using Jet mapping, the segmentation window is set using a dynamic window adjustment method, and the pseudo-color image is segmented according to the segmentation window to obtain a segmented image; Calculate the gradient magnitude of the segmented image Use the pre-trained CycleGAN model to generate a blurred mask from the segmented image Extracting Blurry Areas For fuzzy areas Perform fast Fourier transform to get the spectrum F blur (u,v), calculate the spectrum F blur The high-frequency energy distribution F of (u,v) high (u,v); Calculate the fuzzy area The mean gradient μ G and standard deviation σ G , based on the gradient mean μ G and standard deviation σ G Calculate the blur kernel parameter θ k ; Based on the fuzzy kernel parameter θ k , construct a two-dimensional Gaussian blur kernel K(u,v); Based on high frequency energy distribution F high (u,v) optimized two-dimensional Gaussian blur kernel K optim (u,v), the formula is: K optim (u,v)=K(u,v)·(1-β·|F high (u,v)|), Where β is the optimization coefficient; Use optimized 2D Gaussian blur kernel K optim (u,v) spectrum F blur (u,v) is compensated to obtain the compensated frequency domain data F com (u,v); For the compensated frequency domain data F com (u,v) is smoothed using a bandpass Gaussian filter to obtain the smoothed frequency domain data F filtered (u,v); The smoothed frequency domain data is converted back to the spatial domain to obtain the compensated blurred area image. (i, j, t); The compensated blurred area image Replace it with the original segmented image to get the initial segmented compensation image Concatenate the initial segmentation compensated image and blur mask into the input tensor Collect thermal images with historical labels for preprocessing and extraction of historical deep features to generate a training set; Thermal imaging based on historical tags uses clear image mapping to obtain a clear image of the target Extract tensors using the pre-trained ResNet-50 model The deep feature F deep ; Build a generative adversarial network, including a generator and a discriminator; Set the input format of the generative adversarial network to the deep feature F deep ; Input the deep features into the generator to generate optimized compensated images Using a discriminator to receive the generated optimized compensated image and the target clear image, and output the scores of the compensated image and the target clearness; Use the pre-trained VGG network to extract feature representations of the compensated image and the target clear image; Train the GAN using the training set; Calculate perceptual loss using spatial error Calculate adversarial loss using cross entropy loss Based on perceptual loss and combat loss Constructing the loss function of the generator The formula is: in is the feature representation of the target clear image at the lth layer, where l is the feature layer; Use cross entropy loss to construct the discriminator loss function Use the loss change monitoring method to set the training stop condition when the generator's loss function And the loss function of the discriminator Stop training when the training stop condition is met; The real-time deep feature F deep Input into the trained generative adversarial network to obtain the final segmentation compensation image; The final segmented and compensated image is fused using overlapping weights to obtain the corrected thermal imaging data T mer (i,j,t).

3. The intelligent brake detection system for railway freight cars according to claim 2, characterized in that: The calculation of the full image gradient to generate the high thermal gradient area refers to calculating the temperature gradients in the horizontal and vertical directions using the Sobel operator based on the corrected thermal imaging data; The full-image gradient is calculated using the gradient modulus method based on the temperature gradients in the horizontal and vertical directions; The mean standard deviation method is used to set the gradient threshold, and the full-image gradient greater than the gradient threshold is marked to generate a high thermal gradient area.

4. The intelligent brake detection system for railway freight cars as claimed in claim 3, characterized in that: The calculation of the total contact area and the area of ​​the closed polygon refers to setting the neighborhood radius using the k-nearest neighbor distance curve method and setting the minimum number of points using the data dimension method; For each pixel in the high thermal gradient area, the point set in its neighborhood is calculated. When the neighborhood point set is greater than or equal to the minimum number of points, it is determined to be a core point and marked as visited. When the neighborhood point set is less than the minimum number of points, it is determined to be a boundary point. All pixel points in the neighborhood are added to the core point until all pixel points in the high thermal gradient area are visited and the expansion stops. The contact area point set composed of core points and boundary points is used to calculate the total contact area A using discrete integration total ; Use the convex hull algorithm to sort the boundary points in geometric order to generate a closed boundary polygon point set; Use the Shoelace Formula to calculate the area A(t) of the closed polygon.

5. The intelligent brake detection system for railway freight cars according to claim 4, characterized in that: The calculation of the local temperature rise and temperature field, based on the local temperature rise and temperature field, calculating the predicted temperature field after local heat conduction means obtaining material parameters of the friction block from a material database and performing preprocessing; According to the friction force and the sliding speed, the local heat power P(t) of the high heat area is calculated; Calculate the local temperature rise ΔT(t) at time t based on the local thermal power P(t) and material parameters; Use the explicit difference method to separate the heat conduction equation and calculate the temperature field T predict (i, j, t); According to the current local temperature rise ΔT(t) and temperature field T predict (i, j, t), calculate the predicted temperature field T after local heat conduction predict (i,j,t+1); The material parameters include specific heat capacity, density, thickness and a temperature threshold allowed by the material.

6. The intelligent brake detection system for railway freight cars according to claim 5, characterized in that: The calculated brake risk index refers to the predicted temperature rise T predict (i,j,t+1) calculate the temperature difference ΔT predict (i,j,t+1); Compare the temperature difference with the temperature threshold allowed by the material to generate the over-temperature point M risk (i,j,t+1), calculate the area of ​​over-temperature region; Based on the average temperature rise ΔT avg and the over-temperature area A risk Calculate the brake risk index R(t) at time t.

7. The intelligent brake detection system for railway freight cars according to claim 6, characterized in that: Determining the brake state refers to setting a risk threshold using statistical distribution; Compare the brake risk index R(t) at time t with the risk threshold. If the brake risk index is greater than or equal to the risk threshold, it is judged as abnormal, triggering an alarm signal and notifying the maintenance personnel via email. If the brake risk index is less than the risk threshold, it is judged to be normal and the thermal imaging and brake status data continue to be monitored.

8. The intelligent brake detection system for railway freight cars according to claim 7, characterized in that: The constructing a visual interface to display the brake status result refers to using React.js to construct a visual interface to display the brake risk index and the brake status; Users who have passed real-name verification are allowed to view the information.

9. The intelligent brake detection system for railway freight cars according to claim 8, characterized in that: The storing, collecting and analyzing the thermal imaging and brake status data refers to storing the collected thermal imaging and brake status data and the brake risk indicators and brake status generated by the analysis in a central database, sorting the central database in chronological order and marking corresponding tags, synchronously backing up the collected thermal imaging and brake status data and the brake risk indicators and brake status generated by the analysis in the cloud, and regularly performing integrity checks on the backup data.

Images