A method and system for real-time monitoring of TFDS intelligent identification of railway freight cars

Abstract

This invention relates to the field of image recognition and discloses a real-time monitoring method and system for intelligent identification of railway freight cars using the TFDS (Total Frequency Detection and Recognition) mechanism. The method includes: acquiring a temporal sequence of original images and image motion vectors of a moving target as it passes; extracting multi-scale feature tensors using a convolutional neural network; constructing an anisotropic convolution operator based on the image motion vectors to analyze the directional gradient response values ​​in the multi-scale feature tensors; calculating a phase offset vector using the spatial displacement constancy of a rigid structure to achieve sub-pixel-level phase reassembly and alignment of the multi-scale feature tensors; analyzing temporal semantic consistency relationships and suppressing high-entropy noise using a gated aggregation unit; reconstructing and generating a synthetic feature map; and outputting fault identification results. This invention effectively compensates for anisotropic semantic degradation caused by high-speed motion, synergistically enhances the purity of target component feature representation, and improves the confidence level of identifying minor faults under harsh working conditions.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, and in particular relates to a method and system for real-time monitoring and intelligent identification of railway freight cars using the TFDS (Total Traffic Detection and Recognition System). Background Technology

[0002] With the current evolution of automated rail transit inspection technology, freight car fault image dynamic detection systems acquire images of key parts of trains in motion through imaging components deployed at the trackside. The computing unit uses convolutional neural network algorithms to process the pixel distribution features in the images, extracting geometric morphology information of components such as bogies, brake beams, and fasteners to determine whether there is mechanical damage or missing parts. Since train speeds are often in the high range of 120km / h to 160km / h, the imaging system is limited by the physical integration layout caused by the sensor exposure time, resulting in anisotropic motion degradation features in the acquired raw images. Specifically, edge features parallel to the travel trajectory in the image undergo directional haloing, leading to uneven distribution of high-frequency feature energy. Existing solutions typically use a single-frame pixel matrix as the recognition benchmark. Since it is impossible to distinguish the source of feature residuals in the directional degradation area, the accuracy of the recognition unit in judging the deformation of fine structures such as cotter pins is reduced.

[0003] Besides the physical limitations of hardware imaging, the image recognition processing algorithm suffers from insufficient software control, making it difficult to remove motion noise from the dynamic environment. Chinese invention patent application CN115661776A discloses an image recognition method and system for railway freight car brake beam safety chain detachment faults. By improving the SSD network loss function and introducing the DIOU-NMS algorithm, it addresses the low detection accuracy problem caused by the flexible characteristics of the safety chain behind the brake beam leading to occlusion and overlap. However, this scheme is based on static pixel matrix feature regression, lacks sensitivity discrimination of motion vector direction, and fails to utilize the constant law of train displacement. In actual working conditions, it cannot distinguish the source of feature residuals in directional degradation areas, easily misinterpreting directional ghosting. Noise is misread as structural edges, or the phase misalignment due to feature energy recombination reduces the accuracy of identifying cotter pin microstructures. To alleviate image blurring, physical adjustment methods such as shortening exposure time or increasing ambient light intensity are limited by the power redundancy and scattering interference of the trackside environment. Furthermore, conventional isotropic deblurring strategies are difficult to adapt to the imaging conditions of railway freight cars with specific motion vector deviations. Simply increasing the depth of the neural network or introducing a general semantic attention mechanism not only increases computational resource consumption but also fails to effectively suppress the activation interference of non-car-mounted targets during the feature reconstruction stage due to the lack of utilization of prior logic of motion constancy, resulting in aliasing of semantic boundaries after cross-frame feature fusion.

[0004] Therefore, how to utilize the physical prior of the constant spatial displacement during the movement of railway freight cars to achieve anisotropic feature enhancement and cross-frame semantic energy reorganization in the feature space, thereby eliminating the negative impact of motion blur on the feature recognition of fine components, has become the technical problem to be solved by this invention. Summary of the Invention

[0005] This invention discloses a real-time monitoring method for TFDS intelligent identification of railway freight cars, comprising the following steps:

[0006] Step S1: Obtain the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously obtain the image motion vector of the moving target to be monitored relative to the image acquisition array;

[0007] Step S2: Extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths.

[0008] Step S3: Based on the image motion vector, construct an anisotropic convolution operator parallel to the image motion vector in the feature space, and use the anisotropic convolution operator to perform gradient convolution operation on the multi-scale feature tensor to extract the directional gradient response value representing the distribution along the motion trajectory.

[0009] Step S4: Utilize the constant spatial displacement of the rigid structure in the moving target to be monitored between adjacent image frames to calculate the phase offset vector of the orientation gradient response value on the time axis, and perform sub-pixel-level phase reorganization alignment of the multi-scale feature tensor based on the phase offset vector.

[0010] Step S5: The temporal semantic consistency relationship between the aligned feature tensors of each frame is analyzed by the gated aggregation unit, the feature weights with pixel residual distribution higher than the preset entropy threshold are suppressed, and the retained complementary feature fragments are superimposed with energy weights in the channel dimension to reconstruct and generate a synthetic feature map.

[0011] Step S6: Input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored.

[0012] Preferably, the gradient convolution processing of the multi-scale feature tensor using an anisotropic operator includes: mapping the image motion vector to a two-dimensional feature coordinate system, constructing a weight distribution matrix of the anisotropic convolution operator based on the direction of the image motion vector; performing sliding window convolution on the multi-scale feature tensor using the weight distribution matrix, and calculating the gradient partial derivatives of the eigenvalues ​​in the multi-scale feature tensor in the direction of the image motion vector to obtain the directional gradient response value reflecting the intensity of the photon integral tail.

[0013] Preferably, in step S4, the calculation of the phase offset vector includes: identifying feature anchor points representing the rigid parts of the moving target to be monitored in the multi-scale feature tensor; calculating the spatial coordinate difference of the same semantic feature anchor points between adjacent frames to generate an initial phase offset; and performing kinematic reference correction on the initial phase offset based on the real-time moving speed of the moving target to be monitored to generate a phase offset vector.

[0014] Preferably, the calculation of the phase offset vector further includes: in response to the nonlinear displacement deviation of the moving target to be monitored perpendicular to its direction of travel, calculating the gradient centroid coordinates of the feature anchor point cluster of the rigid part of the target in the multi-scale feature tensor; and performing secondary phase calibration on the phase offset vector based on the gradient centroid coordinates; wherein, the gradient centroid coordinates The calculation formula is: ,in, Let be the spatial coordinates of the i-th pixel in the feature anchor cluster. Let be the gradient energy weight of the i-th pixel, and n be the total number of pixels in the feature anchor cluster.

[0015] Preferably, in step S5, the feature weights for suppressing pixel residual distribution above a preset entropy threshold include: calculating the feature residual map between adjacent frame feature tensors after alignment; statistically analyzing the spatial distribution entropy value of the feature residual map within a preset local window, and calculating the deviation of the spatial distribution entropy value from the global average entropy value; when the deviation exceeds the preset entropy threshold, determining the local window as a high-entropy noise region, and reducing the channel activation gain of the corresponding region in the multi-scale feature tensor.

[0016] Preferably, the method further includes: extracting the semantic residual tensor generated during the phase recombination and alignment process; analyzing the high-frequency energy spectrum features of the semantic residual tensor in the frequency domain; and calculating the reconstruction confidence of the synthesized feature map by establishing a monotonic mapping relationship between the high-frequency energy spectrum features and the reconstruction error.

[0017] Preferably, the method further includes: when the reconstruction confidence is lower than a preset threshold, retrieving the energy features of adjacent background reference channels on the time axis; and performing physical boundary consistency verification on the identification results of the fault identification model based on the signal-to-noise ratio ratio of the background reference channel energy features and the synthetic feature map.

[0018] Preferably, in step S6, the output fault identification result includes: extracting a semantic vector from the synthetic feature map that represents the edge sharpness of the moving target structure to be monitored; matching the semantic vector with a preset fault feature library; and outputting a monitoring instruction containing the fault type, fault location, and identification confidence level.

[0019] Preferably, in step S2, the multi-scale feature tensor includes: a high-resolution feature layer for preserving high-frequency details of cracks; and a high-semantic feature layer for locking the overall structural contour of the moving target to be monitored. The method further includes: monitoring real-time ambient light intensity data and rain / fog concentration data while performing step S1; calculating the scattered light attenuation coefficient of the detection environment based on the ambient light intensity data and rain / fog concentration data; and performing dynamic linear gain compensation on the activation weights of each feature channel in the convolutional neural network based on the scattered light attenuation coefficient to enhance the representation intensity of edge details in the synthesized feature map.

[0020] A real-time monitoring system for TFDS intelligent identification of railway freight cars includes:

[0021] The image acquisition unit is used to acquire the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously acquire the image motion vector of the moving target to be monitored relative to the image acquisition array;

[0022] The feature extraction unit is used to extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths.

[0023] The directional gradient analysis unit is used to construct an anisotropic convolution operator in the feature space based on the motion vector of the image, and extract the directional gradient response values ​​distributed along the motion trajectory axis in the multi-scale feature tensor.

[0024] The phase reorganization and alignment unit is used to calculate the phase offset vector of the directional gradient response value on the temporal axis by utilizing the spatial displacement constancy of the rigid structure in the moving target under monitoring between adjacent image frames, and to complete the sub-pixel-level phase reorganization and alignment of the multi-scale feature tensor.

[0025] The gated aggregation reconstruction unit is used to parse the temporal semantic consistency relationship between the aligned feature tensors of each frame, and reconstruct the synthetic feature map by suppressing the feature weights whose pixel residual distribution is higher than the preset entropy threshold.

[0026] The fault identification unit is used to input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored.

[0027] An environmental compensation unit is used to monitor real-time ambient light intensity data and rain and fog concentration data, and dynamically compensate the feature activation weights in the feature extraction unit according to the attenuation coefficient of scattered light in the detected environment, so as to enhance the representation intensity of edge details in the synthesized feature map.

[0028] Compared with existing technologies, the TFDS intelligent identification and real-time monitoring method for railway freight cars of the present invention has the following advantages:

[0029] 1. In the TFDS intelligent recognition of railway freight cars, by establishing the inner product mapping relationship between the two-dimensional gradient vector and the phase offset vector in the feature space, the system obtains the logical immunity to the photon integral trailing generated by unidirectional high-speed motion. Since the gradient of the component edge perpendicular to the direction of train motion is orthogonal to the direction of motion and the inner product approaches zero, while the gradient of the trailing image in the parallel direction is collinear with the direction of motion and the inner product reaches a maximum value, this mechanism achieves accurate stripping of directional fuzzy noise and faithful extraction of rigid structure edges. It eliminates the directional feature degradation caused by high-speed photography from the underlying semantic level, and provides a pure feature field of view for the subsequent discriminator that is not contaminated by motion.

[0030] 2. By utilizing the spatial displacement constancy of truck motion, shallow feature tensors in time-series image sequences are aligned and recombined based on phase offset vectors. Complementary feature fragments with minimal environmental interference between frames are extracted through gated aggregation units. This incremental injection method of cross-frame feature fragments enables the superposition of effective semantic information from multiple frames in the feature space, reconstructing a synthetic feature map with a high signal-to-noise ratio that exceeds the physical imaging limit of a single frame, thereby improving the confidence level of identifying faults in small components under high-speed operating conditions.

[0031] 3. To address the nonlinear displacement deviation caused by the high-frequency vibration of heavy-load trains, the gradient centroid coordinates of the feature anchor point clusters of rigid parts in the reference feature map and the initial feature map are calculated. Based on the centroid offset vector, a second sub-pixel-level phase calibration is performed on the reference feature map to eliminate feature phase aliasing caused by mechanical vibration, prevent edge blurring or logical ghosting in the feature reconstruction stage, and ensure that high-frequency detail features such as cotter pins and micro-cracks maintain extremely high edge sharpness when energy is superimposed. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the workflow of the real-time monitoring method for TFDS intelligent identification of railway freight cars of the present invention;

[0033] Figure 2 This is the core operational architecture diagram of the TFDS intelligent identification real-time monitoring system for railway freight cars of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0035] It should be noted that all directional and positional terms used in this invention, such as: up, down, left, right, front, back, vertical, horizontal, inner, outer, top, bottom, transverse, longitudinal, center, etc., are only used to explain the relative positional relationship and connection between components in a specific state (as shown in the accompanying drawings). They are only for the convenience of describing this invention and do not require that this invention be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting this invention. In addition, the descriptions of "first," "second," etc., in this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated.

[0036] In the description of this invention, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections; they can refer to direct connections or indirect connections through an intermediate medium; they can refer to the internal connection of two components. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0037] In the description of this specification, references to the terms "an embodiment," "some embodiments," "illustrative embodiments," "examples," "specific examples," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example, and the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0038] A method for real-time monitoring of railway freight cars using TFDS intelligent identification includes the following steps:

[0039] Step S1: Obtain the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously obtain the image motion vector of the moving target to be monitored relative to the image acquisition array;

[0040] Step S2: Extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths.

[0041] Step S3: Based on the image motion vector, construct an anisotropic convolution operator parallel to the image motion vector in the feature space, and use the anisotropic convolution operator to perform gradient convolution operation on the multi-scale feature tensor to extract the directional gradient response value representing the distribution along the motion trajectory.

[0042] Step S4: Utilize the constant spatial displacement of the rigid structure in the moving target to be monitored between adjacent image frames to calculate the phase offset vector of the orientation gradient response value on the time axis, and perform sub-pixel-level phase reorganization alignment of the multi-scale feature tensor based on the phase offset vector.

[0043] Step S5: The temporal semantic consistency relationship between the aligned feature tensors of each frame is analyzed by the gated aggregation unit, the feature weights with pixel residual distribution higher than the preset entropy threshold are suppressed, and the retained complementary feature fragments are superimposed with energy weights in the channel dimension to reconstruct and generate a synthetic feature map.

[0044] Step S6: Input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored.

[0045] Preferably, the gradient convolution processing of the multi-scale feature tensor using an anisotropic operator includes: mapping the image motion vector to a two-dimensional feature coordinate system, constructing a weight distribution matrix of the anisotropic convolution operator based on the direction of the image motion vector; performing sliding window convolution on the multi-scale feature tensor using the weight distribution matrix, and calculating the gradient partial derivatives of the eigenvalues ​​in the multi-scale feature tensor in the direction of the image motion vector to obtain the directional gradient response value reflecting the intensity of the photon integral tail.

[0046] Preferably, in step S4, the calculation of the phase offset vector includes: identifying feature anchor points representing the rigid parts of the moving target to be monitored in the multi-scale feature tensor; calculating the spatial coordinate difference of the same semantic feature anchor points between adjacent frames to generate an initial phase offset; and performing kinematic reference correction on the initial phase offset based on the real-time moving speed of the moving target to be monitored to generate a phase offset vector.

[0047] Preferably, the calculation of the phase offset vector further includes: in response to the nonlinear displacement deviation of the moving target to be monitored perpendicular to its direction of travel, calculating the gradient centroid coordinates of the feature anchor point cluster of the rigid part of the target in the multi-scale feature tensor; and performing secondary phase calibration on the phase offset vector based on the gradient centroid coordinates; wherein, the gradient centroid coordinates The calculation formula is: ,in, Let be the spatial coordinates of the i-th pixel in the feature anchor cluster. Let be the gradient energy weight of the i-th pixel, and n be the total number of pixels in the feature anchor cluster.

[0048] Preferably, in step S5, the feature weights for suppressing pixel residual distribution above a preset entropy threshold include: calculating the feature residual map between adjacent frame feature tensors after alignment; statistically analyzing the spatial distribution entropy value of the feature residual map within a preset local window, and calculating the deviation of the spatial distribution entropy value from the global average entropy value; when the deviation exceeds the preset entropy threshold, determining the local window as a high-entropy noise region, and reducing the channel activation gain of the corresponding region in the multi-scale feature tensor.

[0049] Preferably, the method further includes: extracting the semantic residual tensor generated during the phase recombination and alignment process; analyzing the high-frequency energy spectrum features of the semantic residual tensor in the frequency domain; and calculating the reconstruction confidence of the synthesized feature map by establishing a monotonic mapping relationship between the high-frequency energy spectrum features and the reconstruction error.

[0050] Preferably, the method further includes: when the reconstruction confidence is lower than a preset threshold, retrieving the energy features of adjacent background reference channels on the time axis; and performing physical boundary consistency verification on the identification results of the fault identification model based on the signal-to-noise ratio ratio of the background reference channel energy features and the synthetic feature map.

[0051] Preferably, in step S6, the output fault identification result includes: extracting a semantic vector from the synthetic feature map that represents the edge sharpness of the moving target structure to be monitored; matching the semantic vector with a preset fault feature library; and outputting a monitoring instruction containing the fault type, fault location, and identification confidence level.

[0052] Preferably, in step S2, the multi-scale feature tensor includes: a high-resolution feature layer for preserving high-frequency details of cracks; and a high-semantic feature layer for locking the overall structural contour of the moving target to be monitored. The method further includes: monitoring real-time ambient light intensity data and rain / fog concentration data while performing step S1; calculating the scattered light attenuation coefficient of the detection environment based on the ambient light intensity data and rain / fog concentration data; and performing dynamic linear gain compensation on the activation weights of each feature channel in the convolutional neural network based on the scattered light attenuation coefficient to enhance the representation intensity of edge details in the synthesized feature map.

[0053] A real-time monitoring system for TFDS intelligent identification of railway freight cars includes:

[0054] The image acquisition unit is used to acquire the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously acquire the image motion vector of the moving target to be monitored relative to the image acquisition array;

[0055] The feature extraction unit is used to extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths.

[0056] The directional gradient analysis unit is used to construct an anisotropic convolution operator in the feature space based on the motion vector of the image, and extract the directional gradient response values ​​distributed along the motion trajectory axis in the multi-scale feature tensor.

[0057] The phase reorganization and alignment unit is used to calculate the phase offset vector of the directional gradient response value on the temporal axis by utilizing the spatial displacement constancy of the rigid structure in the moving target under monitoring between adjacent image frames, and to complete the sub-pixel-level phase reorganization and alignment of the multi-scale feature tensor.

[0058] The gated aggregation reconstruction unit is used to parse the temporal semantic consistency relationship between the aligned feature tensors of each frame, and reconstruct the synthetic feature map by suppressing the feature weights whose pixel residual distribution is higher than the preset entropy threshold.

[0059] The fault identification unit is used to input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored.

[0060] An environmental compensation unit is used to monitor real-time ambient light intensity data and rain and fog concentration data, and dynamically compensate the feature activation weights in the feature extraction unit according to the attenuation coefficient of scattered light in the detected environment, so as to enhance the representation intensity of edge details in the synthesized feature map.

[0061] Example 1: In a heavy-haul railway freight monitoring scenario, the instantaneous speed of a freight train passing through the image acquisition array is 125 km / h. The time-series raw image sequence acquired by the image acquisition array contains anisotropic degradation features caused by high-speed motion. Specifically, the edges of fine components such as bogie cotter pins generate directional photon integral tails in a direction parallel to the travel trajectory, resulting in uneven distribution of high-frequency feature energy and overlapping semantic boundaries in the image. The processor uses a convolutional neural network to extract the multi-scale feature tensor corresponding to the time-series raw image sequence and uses the synchronously acquired image motion vector of the moving target to be monitored relative to the image acquisition array. Constructing image motion vectors within the feature space Parallel anisotropic convolution operators are used to perform gradient convolution operations on multi-scale feature tensors, calculating the eigenvalues ​​of the multi-scale feature tensors in the image motion vector. The partial derivatives of the gradient in the direction are used to obtain the directional gradient response value reflecting the intensity of the photon integral tail; the processor calculates the two-dimensional spatial gradient vector of the multi-scale feature tensor. With image motion vectors absolute value of inner product When the absolute value of the inner product When the value approaches its maximum, the corresponding region is identified as a trailing region contaminated by motion and an anisotropic suppression mask matrix is ​​generated. This anisotropic suppression mask matrix is ​​then used to forcibly suppress the channel activation gain of the corresponding region in the multi-scale feature tensor. For the two-dimensional spatial gradient vector... With image motion vectors In regions approaching orthogonality, the system, based on the highest injection weight, strips away directional fuzzy noise in the feature space and extracts rigid structural edges with high spatial fidelity.

[0062] When extracting features, the processor divides the multi-scale feature tensor into feature subspaces with different physical properties. The high-resolution feature layer uses a smaller convolution stride and shallower pooling depth to preserve high-frequency texture components representing crack-like defects within the pixel space. Meanwhile, the high-semantic feature layer extracts low-frequency contour information representing the overall topology of the bogie and brake beam by increasing the number of convolutional layers and the global receptive field. This achieves decoupling representation of microscopic details and global targets during the feature extraction stage. To address the optical degradation caused by severe weather in physical imaging, the system simultaneously monitors real-time ambient light intensity data while acquiring the original image sequence. and rain and fog concentration data The system calculates the attenuation coefficient of scattered light in the detection environment based on ambient light intensity data and rain / fog concentration data. The specific calculation formula is set as follows: ,in, The preset medium scattering cross section constant, This is the illumination compensation adjustment coefficient. Using the natural logarithm as the base, this formula physically characterizes the strong scattering effect of high-concentration rain and fog on photons, as well as the nonlinear suppression characteristics of optical attenuation caused by high background light intensity. It is based on the calculated scattered light attenuation coefficient. The processor performs dynamic linear gain compensation on the activation weights of each feature channel in the convolutional neural network, thereby specifically amplifying the weights of high-frequency edge features that are severely attenuated by rain and fog, thus effectively enhancing the representation intensity of edge details in the synthesized feature map. The system utilizes the constant spatial displacement of rigid structures in the moving target between adjacent image frames to identify feature anchor points representing rigid parts in the multi-scale feature tensor. It calculates the spatial coordinate difference of the same semantic feature anchor points between adjacent frames to generate an initial phase offset, and combines this with real-time movement speed to perform kinematic benchmark correction on the initial phase offset, generating a phase offset vector to represent temporal axis semantic drift. For the nonlinear displacement deviation caused by the high-frequency jerking of heavy-load trains, the system extracts the feature anchor point clusters of preset rigid parts from the reference feature map and the initial feature map, and calculates their gradient centroid coordinates Gc. The formula for calculating the gradient centroid coordinates Gc is as follows: ,in, Let be the spatial coordinates of the i-th pixel in the feature anchor cluster. The gradient energy weight for the i-th pixel is defined as follows, and its value range is limited to 1. This weight It is obtained by calculating the ratio of the local directional gradient magnitude of the pixel to the maximum gradient magnitude within the entire feature anchor cluster, where n is the total number of pixels in the feature anchor cluster; in the above formula, the numerator... It represents the sum of the spatial distributions of combined gradient energy intensity, expressed as the sum of the weights of all pixels. The physical significance of weighted normalization in the denominator lies in the weighted normalization of spatial coordinates. This normalization operation can eliminate the interference caused by drastic changes in illumination or local absolute energy differences in coordinate calculation, ensuring that the calculated gradient centroid coordinates are consistent regardless of the overall gradient response strength of the feature anchor point cluster. It always converges strictly to its true geometric and physical position. The centroid offset vector determined by the gradient centroid coordinate Gc is used to perform a second subpixel-level phase calibration on the reference feature map to eliminate feature phase aliasing caused by mechanical vibration and ensure that the multi-scale feature tensor achieves accurate phase recombination and alignment on the time axis.

[0063] Construct the anisotropic convolution operator described in step S3, based on synchronously acquired image motion vectors. The geometric topology of the weight distribution matrix is ​​determined, and a 1×11 pixel matrix is ​​selected as the convolution kernel template, aligning the major axis with the image motion vector v. The weight values ​​are set to conform to a discrete Gaussian distribution along the major axis, with zeros padded at other positions. A sliding window convolution is performed to calculate the partial derivatives of the gradient values ​​of the eigenvalues ​​in the direction of the image motion vector v, obtaining the directional gradient response value reflecting the intensity of the photon integral tail. Step S4 is executed, where feature anchor points are determined by calculating the Hessian matrix of each pixel in the multi-scale feature tensor. The coordinates of the point with the largest eigenvalue exceeding three times the global mean and belonging to the local response peak are selected as the rigid part anchor points. Based on this, the spatial coordinate difference of the same semantic anchor points between adjacent frames is calculated to generate the initial phase offset. For the nonlinear displacement deviation caused by the high-frequency vibration of heavy-load trains, the gradient centroid coordinates of the feature anchor point cluster are calculated using the formula in the embodiment. ,variable The variable represents the two-dimensional spatial coordinates of the i-th pixel in the feature anchor cluster. The gradient energy weights of pixels are represented by the variable n, which represents the total number of pixels. The gradient centroid coordinates are compared between the reference feature map and the current feature map. The offset vector performs secondary subpixel-level phase calibration to eliminate phase aliasing caused by mechanical vibration; step S5 is executed, and the entropy threshold is preset. Based on the statistical distribution analysis of the feature residual map of the fault-free sample set, the sum of the global average entropy value μ and three times the standard deviation σ is selected as the judgment threshold. When the measured spatial distribution entropy value E of the local window exceeds the threshold, the gated aggregation unit generates an attenuation factor with a value range between 0.05 and 0.15. The attenuation factor is used to suppress the channel activation weights of the corresponding region of the multi-scale feature tensor, suppress high-entropy noise, and reconstruct and generate a synthetic feature map.

[0064] Example 2: A simulation test platform for dynamic image detection of railway freight car faults was established. This platform includes an image acquisition array consisting of six linear image sensors with a sampling frequency set to 100kHz to support high-frequency spatial sampling of the moving target at a speed of 150km / h. The original image sequence used in the test was derived from freight car passing data collected on-site at the detection station. This data included background noise with a signal-to-noise ratio of 15.2dB due to insufficient ambient light. The test also considered the stride Δs of the anisotropic convolution operator and the weight distribution of the convolution kernel. The setting is based on a technical consideration of balancing image edge sharpness enhancement with computational load; when the image motion vector As the modulus increases, to ensure that the directional gradient response value can cover the complete integral tail region, the effective receptive field of the anisotropic convolution operator is shifted to the upper limit of its value range, increasing the number of convolution sampling points along the motion direction; for a driving speed of 125 km / h, the kernel size of the anisotropic convolution operator is selected as 1×15 pixels, and the stride Δs is 1 pixel, where Δs is the spatial interval between adjacent sampling points; the experimental group uses the method claimed in this invention, control group one removes the anisotropic convolution operator in step S3 and replaces it with the isotropic Sobel operator, control group two removes the sub-pixel level phase calibration in step S4 and only uses integer pixel level feature translation alignment; test loop The test group simulated high-frequency mechanical vibration of a heavy-haul train by superimposing random displacement disturbances with a frequency of 25 Hz and an amplitude of 3.2 pixels onto the original image sequence. Measurement data showed that under the condition of 125 km / h, the test group had an accuracy rate of 98.24% in identifying bogie cotter pin loss faults, with a mean confidence level of 0.951. However, the semantic boundary aliasing in the feature tensor extracted by the control group 1 caused the fault identification accuracy to drop to 81.56%, and the false alarm rate to increase by 12.63%. The control group 2 lacked sub-pixel phase calibration driven by gradient centroid coordinate Gc, resulting in feature overlap artifacts during feature recombination, which caused the accuracy rate of identifying micro-cracks to drop to 74.38%.

[0065] Analysis of the control group data shows that the directional gradient response value generated in step S3 provides the high-contrast feature anchor point in step S4. Without the directional aggregation of feature energy by anisotropic convolution, the calculated gradient centroid coordinates Gc will be biased due to the scattered distribution of pixel grayscale gradients, leading to inaccurate phase reconstruction. The experimental group, through the synergistic effect of anisotropic convolution and sub-pixel alignment, improved the edge sharpness of structural components in the synthesized feature map by 45.67%, demonstrating a positive feedback relationship between feature purification and phase compensation. A gradient test group was constructed for driving speeds, testing at 50 km / h and 100 km / h respectively. Performance at speeds of 50 km / h, 150 km / h, and beyond 180 km / h; data shows that within the range of 50 km / h to 150 km / h, the reconstruction confidence of the synthesized feature map remains above 0.923; when the speed reaches 180 km / h, the recognition accuracy decreases to 65.41% due to the tail length exceeding the preset anisotropic convolution kernel size, confirming 150 km / h as the performance inflection point under the parameter configuration of this embodiment. This also confirms that the speed adaptation range defined by this invention has definite engineering boundaries for the preset entropy threshold in the gated aggregation unit. The calibration was performed by statistically modeling the feature residual maps of 500 sets of samples, calculating the average distribution entropy value of the local window under normal conditions and superimposing three times the standard deviation as the judgment boundary; when the local window was affected by speckle interference caused by wheel rim flicker, the measured entropy value reached 4.85, which was higher than the preset entropy threshold of 2.15. At this time, the system suppressed the channel activation gain of the region to less than 0.1 times the initial value according to the deviation, suppressed the injection of false features, and improved the signal-to-noise ratio of the synthesized feature map by 18.42dB compared with the original image.

[0066] Example 3: In a real-time monitoring scenario of heavy-haul railway freight, the moving target to be monitored passes through the image acquisition array at a speed of 125 km / h. The system is subjected to high-frequency mechanical vibration interference, resulting in local nonlinear geometric deformation of the image. This deformation causes non-uniform displacement of semantic features between adjacent frames in the pixel space, resulting in a large number of invalid residual components representing deformation noise in the generated feature residual map. In order to suppress the interference of invalid residual components on the recognition results, the system enables the gating aggregation logic of the feature space after performing anisotropic feature enhancement. The processor performs element-wise subtraction of the aligned adjacent frame feature tensors to obtain the feature residual map R. A sampling window with a size of 8×8 pixels and a sliding step size of 4 pixels is selected to traverse the feature residual map R. The gray-level probability distribution of the pixels in each sampling window area is statistically analyzed, and the corresponding spatial distribution entropy value is calculated according to the definition of information entropy.

[0067] The system uses a pre-stored feature library of fault-free samples in memory to determine the judgment criteria, calculates the global average entropy value μ and standard deviation σ under fault-free conditions, and measures the spatial distribution entropy value. Relative to the global average entropy deviation The calculation formula is defined as follows: Meanwhile, based on outlier detection criteria using mathematical statistics, the system sets a preset entropy threshold. The absolute boundary is set as When the measured spatial distribution entropy value of the sliding window Deviation generated Too large, causing its actual value Exceeded the preset entropy threshold The upper limit, that is, satisfying the mathematical relation , equivalent to If the deviation exceeds the controlled range, it indicates that the feature dispersion of the window exceeds the controlled deviation range. The local space containing the window is determined to be a high-entropy noise region. The gating aggregation unit generates an attenuation factor matrix based on the deviation and uses it to suppress the channel activation weights at corresponding positions in the multi-scale feature tensor, thus reducing the deviation to a preset entropy threshold. The complementary structural details are injected into the main feature stream through energy weighting to reconstruct the synthetic feature map. The processor extracts the semantic residual tensor generated during phase recombination and alignment, and transforms the semantic residual tensor from the spatial domain to the frequency domain by performing a fast Fourier transform. The high-frequency component weights of the semantic residual tensor in the frequency domain are analyzed and their energy amplitudes are accumulated to obtain high-frequency energy spectrum features for characterizing spatial abrupt changes. The system establishes a negatively correlated monotonic mapping relationship between the high-frequency energy spectrum features and the reconstruction error, and the reconstruction confidence of the synthetic feature map is determined. The calculation formula is set as follows: ,in, The measured energy amplitude of the high-frequency energy spectrum characteristics obtained by accumulation, The system uses a preset empirical decay penalty coefficient (value greater than 0), where e is the base of the natural logarithm. The system employs a negative exponential function model. The residual energy of the interval is smoothly mapped to Within the confidence interval; when the high-frequency residual energy in the frequency domain A higher value indicates a greater amount of deformation artifacts and reconstruction errors introduced during the alignment process, resulting in a higher calculated reconstruction confidence level. The corresponding energy exhibits a monotonically exponential decay, and the reconstruction confidence of the synthetic feature map is calculated online based on the measured energy amplitude.

[0068] The synthesized feature map is input into the fault identification model, which adopts a deep residual network architecture containing four residual blocks. It receives a three-dimensional tensor with dimensions C×H×W, where C is the number of channels, H is the spatial height, and W is the spatial width. Each residual block maps local spatial features into semantic vectors through convolutional layers and skip connections. After passing through a global average pooling layer, it outputs a topological feature vector with a length of 512, which is used to characterize the geometric constraint relationship of the target structural component. The system calculates the cosine similarity between this topological feature vector and the feature vectors in the preset fault feature library, and outputs a monitoring command when the cosine similarity exceeds 0.85.

[0069] Example 4: When the system faces the condition of physical environment migration of the detection station, the processor acquires the original image sequence of the moving target under no-load conditions through the image acquisition array, statistically analyzes the spatial distribution entropy value E of the M consecutive frames of original images under different time periods and illumination conditions, and calculates the initial average entropy value corresponding to each sampling window region. Compared with the initial standard deviation The system uses the obtained data to set a preset entropy threshold in the gating aggregation unit. Dynamic gain compensation is performed to eliminate static component interference caused by strong track reflections and ballast shadows on the feature residual map R, so that the output synthetic feature map has the physical boundary adaptability for the current site. When the reconstruction confidence of the synthetic feature map is lower than the preset confidence threshold of 0.8, the processor retrieves the background features of the non-target coverage area in the adjacent frames on the time axis, calculates the mean spatial variance of the non-target coverage area and defines it as the background reference channel energy feature. The processor calculates the ratio of the background reference channel energy feature to the measured signal-to-noise ratio of the synthetic feature map and uses it as the input criterion for physical boundary consistency verification. When the ratio deviates from the preset safety range, the system performs physical constraint correction of the recognition result.

[0070] In real-time monitoring scenarios involving mixed operation of heavy-haul and general-purpose railway freight cars, the processor collects sample image data of bogie cotter pins and brake beam structural components under different wear levels, and extracts the initial topological feature vectors corresponding to each fault sample through a feature extraction network. Establish initial topological feature vectors A unique correspondence with fault classification codes; after the system receives the image sequence of the moving target to be monitored and extracts the real-time topological feature vector, it uses the cosine distance calculation logic to complete the semantic similarity retrieval with the preset fault feature library in the 512-dimensional feature tensor space. The judgment result includes the corresponding fault classification index and structural component maintenance priority, enabling the system to enter a state of accurate capture of structural details of different vehicle models.

[0071] Example 5: Under the condition that the camera at the detection station is installed at an elevation angle of 30 degrees and the lens focal length is 12mm, the system performs the size calibration of the anisotropic convolution operator and calculates the length parameter of the convolution kernel using the motion degradation model pre-stored in the memory. The specific calculation formula is as follows: Where V is the speed at which the truck passes. For the camera's exposure time, The kernel length parameter is calculated based on the actual spatial scale corresponding to a single pixel, a speed of 125 km / h, and an exposure time of 10 μs. For a length of 11 pixels and a width of 1 pixel, the processor selects a matrix with a length of 11 pixels and a width of 1 pixel as the initial weight template of the anisotropic convolution operator, and performs spatial geometric correction on the weight template according to the lens radial distortion coefficient to eliminate edge projection distortion caused by camera installation deviation, so that the feature enhancement logic adapts to the physical imaging optical path.

[0072] To address the structural differences in bogies across different heavy-duty truck models, the processor performs feature anchor point identification and sample library construction. It uses a spatial variance statistical model to extract the mean of the second derivative of the spatial gradient in the original image sequence. Regions where this mean exceeds a preset curvature threshold and whose relative displacement remains constant within five consecutive frames are marked as rigid feature anchor points. The starting index position of the feature vector is determined by calculating the local response peak of the rigid feature anchor point in the feature tensor. For cosine similarity matching in the fault identification model, the system collects sample data including missing pins, broken pins, and pins with excessive displacement and extracts the corresponding topological feature vectors. The minimum Manhattan distance between positive and negative samples in the feature space is calculated. The midpoint of this distance is selected as the benchmark for matching confidence, eliminating the drift of recognition results caused by light intensity fluctuations, and keeping the fault detection rate of the system within the preset error range at different times.

[0073] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. A method for real-time monitoring of railway freight cars using TFDS intelligent identification, characterized in that, Includes the following steps: Step S1: Obtain the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously obtain the image motion vector of the moving target to be monitored relative to the image acquisition array; Step S2: Extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths. Step S3: Based on the image motion vector, construct an anisotropic convolution operator parallel to the image motion vector in the feature space, and use the anisotropic convolution operator to perform gradient convolution operation on the multi-scale feature tensor to extract the directional gradient response value representing the distribution along the motion trajectory. Step S4: Utilizing the constant spatial displacement of the rigid structure in the moving target between adjacent image frames, calculate the phase offset vector of the directional gradient response value on the temporal axis. The calculation of the phase offset vector includes: identifying feature anchor points representing the rigid parts of the moving target in the multi-scale feature tensor; calculating the spatial coordinate difference of the same semantic feature anchor points between adjacent frames to generate an initial phase offset; performing kinematic reference correction on the initial phase offset based on the real-time movement speed of the moving target to generate the phase offset vector; and, in response to the nonlinear displacement deviation of the moving target perpendicular to its direction of travel, calculating the gradient centroid coordinates of the feature anchor point cluster of the rigid part of the target in the multi-scale feature tensor; performing secondary phase calibration on the phase offset vector based on the gradient centroid coordinates; and performing sub-pixel-level phase re-alignment on the multi-scale feature tensor based on the phase offset vector after secondary phase calibration. The gradient centroid coordinates... The calculation formula is: ,in, For the first feature anchor point in the cluster The spatial coordinates of each pixel is the gradient energy weight of the i-th pixel, and n is the total number of pixels in the feature anchor cluster; Step S5: The temporal semantic consistency relationship between the aligned feature tensors of each frame is analyzed by the gated aggregation unit, the feature weights with pixel residual distribution higher than the preset entropy threshold are suppressed, and the retained complementary feature fragments are superimposed with energy weights in the channel dimension to reconstruct and generate a synthetic feature map. Step S6: Input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored.

2. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 1, characterized in that, Gradient convolution processing of multi-scale feature tensors using anisotropic operators includes: mapping image motion vectors to a two-dimensional feature coordinate system, constructing a weight distribution matrix of anisotropic convolution operators based on the direction of image motion vectors; performing sliding window convolution on multi-scale feature tensors using the weight distribution matrix, and calculating the gradient partial derivatives of the eigenvalues ​​in the multi-scale feature tensors in the direction of image motion vectors to obtain directional gradient response values ​​that reflect the intensity of photon integral tailing.

3. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 1, characterized in that, In step S5, the feature weights for suppressing pixel residual distribution above a preset entropy threshold include: calculating the feature residual map between adjacent frame feature tensors after alignment; statistically analyzing the spatial distribution entropy value of the feature residual map within a preset local window, and calculating the deviation of the spatial distribution entropy value from the global average entropy value; when the deviation exceeds the preset entropy threshold, determining the local window as a high-entropy noise region, and reducing the channel activation gain of the corresponding region in the multi-scale feature tensor.

4. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 1, characterized in that, The method also includes: extracting the semantic residual tensor generated during the phase recombination and alignment process; analyzing the high-frequency energy spectrum features of the semantic residual tensor in the frequency domain; and calculating the reconstruction confidence of the synthesized feature map by establishing a monotonic mapping relationship between the high-frequency energy spectrum features and the reconstruction error.

5. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 4, characterized in that, The method also includes: when the reconstruction confidence is lower than a preset threshold, retrieving the energy features of adjacent background reference channels on the time axis; and performing physical boundary consistency verification on the identification results of the fault identification model based on the signal-to-noise ratio ratio of the background reference channel energy features and the synthetic feature map.

6. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 1, characterized in that, In step S6, the output fault identification results include: extracting semantic vectors representing the edge sharpness of the moving target structure to be monitored from the synthetic feature map; matching the semantic vectors with the preset fault feature library; and outputting monitoring instructions containing fault type, fault location, and identification confidence.

7. The method for real-time monitoring of railway freight cars using TFDS intelligent identification according to claim 1, characterized in that, In step S2, the multi-scale feature tensor includes: a high-resolution feature layer for preserving high-frequency details of crack-like structures; A high semantic feature layer is used to lock the overall structural contour of the moving target to be monitored. The method also includes: monitoring real-time ambient light intensity data and rain and fog concentration data while performing step S1; calculating the scattered light attenuation coefficient of the detection environment based on the ambient light intensity data and rain and fog concentration data; and performing dynamic linear gain compensation on the activation weights of each feature channel in the convolutional neural network based on the scattered light attenuation coefficient to enhance the representation intensity of edge details in the synthesized feature map.

8. A real-time monitoring system for intelligent identification of TFDS (Traffic Freight Vehicle Data System) for railway freight cars, used to implement the real-time monitoring method for intelligent identification of TFDS for railway freight cars as described in claim 1, characterized in that, include: The image acquisition unit is used to acquire the time-series raw image sequence generated when the moving target to be monitored passes through the image acquisition array, and simultaneously acquire the image motion vector of the moving target to be monitored relative to the image acquisition array; The feature extraction unit is used to extract the multi-scale feature tensor corresponding to the original temporal image sequence through a convolutional neural network. The multi-scale feature tensor contains semantic feature matrices with different spatial resolutions and channel depths. The directional gradient analysis unit is used to construct an anisotropic convolution operator in the feature space based on the motion vector of the image, and extract the directional gradient response values ​​distributed along the motion trajectory axis in the multi-scale feature tensor. The phase reconfiguration and alignment unit is used to calculate the phase offset vector of the directional gradient response value on the temporal axis by utilizing the spatial displacement constancy of the rigid structure in the monitored moving target between adjacent image frames. The calculation of the phase offset vector includes: identifying feature anchor points representing the rigid parts of the monitored moving target in the multi-scale feature tensor; calculating the spatial coordinate difference of the same semantic feature anchor points between adjacent frames to generate an initial phase offset; performing kinematic reference correction on the initial phase offset based on the real-time movement speed of the monitored moving target to generate a phase offset vector; and calculating the gradient centroid coordinates of the feature anchor point cluster of the target's rigid parts in the multi-scale feature tensor in response to the nonlinear displacement deviation of the monitored moving target perpendicular to its direction of travel; and performing secondary phase calibration on the phase offset vector based on the gradient centroid coordinates to complete the sub-pixel-level phase reconfiguration and alignment of the multi-scale feature tensor. The gradient centroid coordinates... The calculation formula is: ,in, For the first feature anchor point in the cluster The spatial coordinates of each pixel is the gradient energy weight of the i-th pixel, and n is the total number of pixels in the feature anchor cluster; The gated aggregation reconstruction unit is used to parse the temporal semantic consistency relationship between the aligned feature tensors of each frame, and reconstruct the synthetic feature map by suppressing the feature weights whose pixel residual distribution is higher than the preset entropy threshold. The fault identification unit is used to input the synthesized feature map into the fault identification model, extract the topological feature vector of the target structural component, and output the fault identification result for the moving target to be monitored. An environmental compensation unit is used to monitor real-time ambient light intensity data and rain and fog concentration data, and dynamically compensate the feature activation weights in the feature extraction unit according to the attenuation coefficient of scattered light in the detected environment, so as to enhance the representation intensity of edge details in the synthesized feature map.

Images