AI wear and tear and loosening identification system for catenary of rail transit rotating equipment

Abstract

The application relates to the technical field of contact network state recognition, in particular to an AI abrasion and loosening recognition system for a contact network of a rotating device of track transportation, which comprises a data time sequence arrangement module, a rotating motion flow form analysis module, a frequency domain energy extraction module, an abrasion gradient determination module and a state risk assessment module.In the application, a local tangent plane is constructed, and a geodesic line projection distance to a standard trajectory surface is calculated, so that constrained motion and abnormal deviation can be effectively stripped, nonlinear space anomalies hidden in vibration noise are captured by using flow form geometry analysis to identify slight loosening, texture energy and morphological energy are separated by filtering to realize accurate quantification of material loss, the spatial gradient value of deformation energy proportion is calculated to effectively distinguish surface shallow interference and structural deep mutation, and the quantitative indexes of time and space alignment are comprehensively used to realize accurate identification of equipment operation stability and health state in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of overhead contact line condition recognition technology, and in particular to an AI wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit. Background Technology

[0002] The field of overhead contact line condition recognition technology mainly involves monitoring and evaluating the physical form, geometric parameters, and mechanical connection status of overhead contact line facilities in rail transit power supply networks. This field encompasses the systematic inspection of the spatial position of the contact suspension device, the integrity of its components, and the surface condition of the contact wire using sensor technology and data acquisition terminals. The aim is to maintain a stable current collection relationship between the pantograph and the overhead contact line and ensure train operation safety. Traditional AI wear and loosening recognition systems for overhead contact lines of rotating equipment in rail transit utilize laser scanners or linear array cameras mounted on inspection vehicles to scan and photograph the rotating equipment and contact wire. Image preprocessing techniques are used to remove background noise and enhance contrast. Edge detection operators or Hough transforms are employed to extract the contour features of the contact wire and fasteners. By calculating the pixel size or relative position of feature points of the contour and comparing these values ​​with standard design parameters, the wear depth of the contact wire or the loosening and displacement of the fasteners can be determined.

[0003] Existing technologies rely on edge detection and pixel size comparison to treat complex rotational dynamics as static geometric parameters. Ignoring the spatial depth constraints of rotating parts makes it impossible to distinguish between normal mechanical vibration and abnormal axial displacement. Simple contour comparison lacks sensitivity to nonlinear trajectory anomalies, resulting in early loosening and missed detection. Relying on global image contour analysis cannot effectively distinguish between surface texture interference and substantial structural defects. Failure to decouple frequency domain features leads to misjudging surface scratches or oil stains as deep wear, reducing the accuracy of fault diagnosis. Relying solely on geometric parameter comparison makes it difficult to keenly capture minute deformations in dynamic high-frequency vibration environments, thus leading to the risk of false alarms. Summary of the Invention

[0004] To address the technical problems of existing technologies, such as ignoring the spatial depth constraints of rotating components, which makes it impossible to distinguish between normal mechanical vibration and abnormal axial displacement; the lack of sensitivity to nonlinear trajectory anomalies in simple contour comparison leading to missed early loosening detection; the inability of global image contour analysis to effectively distinguish between surface texture interference and substantial structural defects; the failure to decouple frequency domain features, which leads to misjudging surface scratches or oil stains as deep wear and reducing fault diagnosis accuracy; and the difficulty of sensitively capturing minute deformations in dynamic high-frequency vibration environments by relying solely on geometric parameter comparison, thus leading to false alarm risks, this invention provides an AI wear and loosening identification system for the contact wire of rotating equipment in rail transit. The technical solution is as follows:

[0005] On the one hand, an AI-based wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit is provided. This system includes:

[0006] The data time sequence arrangement module acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment of rail transit, quantizes the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, and serializes the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the acquisition timestamp to generate a spatiotemporal fusion point cloud image sequence.

[0007] The rotational motion manifold analysis module calculates the covariance matrix of the three-dimensional spatial point vectors of the ratchet and the wrist arm base based on the spatiotemporal fusion point cloud image sequence, calculates the dispersion and growth trend of the geodesic length over time, and constructs the trajectory geodesic deviation and grayscale dataset.

[0008] The frequency domain energy extraction module calls the trajectory geodesic deviation and grayscale dataset, performs convolution operation on the two-dimensional grayscale pixel array, calculates the proportion of the contour shape energy amplitude in the total energy, and generates a set of deformation energy proportion and deviation.

[0009] The wear gradient determination module calculates the spatial gradient value of the deformation energy ratio along the contact line extension direction based on the deformation energy ratio and deviation set, compares the spatial gradient value with the preset abrupt change threshold, divides the wear level, and establishes wear level and loosening quantification index.

[0010] As a further aspect of the present invention, the spatiotemporal fusion point cloud image sequence includes serialized grayscale image frames, a three-dimensional spatial coordinate sequence, and a synchronization time index; the trajectory geodesic deviation and grayscale dataset includes geodesic projection distance values, trajectory dispersion feature values, deviation growth trend coefficients, and associated grayscale matrices; the deformation energy proportion and deviation set includes texture detail energy amplitude, contour morphology energy amplitude, and deformation energy proportion coefficients; and the wear level and loosening quantification index includes energy proportion spatial gradient values, wear level classification identifiers, and loosening risk quantification values.

[0011] As a further aspect of the present invention, the data timing arrangement module includes:

[0012] The multimodal data acquisition submodule acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment in rail transit, drives the acquisition sensor to quantize the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, captures the hardware clock signal at the moment of sensor triggering, generates a timestamp, and encapsulates the two-dimensional grayscale pixel array, the three-dimensional spatial point vector and the timestamp into independent data units to generate a discrete grayscale coordinate raw dataset.

[0013] The sequence synchronization alignment submodule calls the original discrete grayscale coordinate dataset, traverses and reads the timestamp values ​​in the data units, performs an ascending sort operation on the data units according to the size of the timestamp values, constructs a time index sequence with monotonically increasing characteristics, performs one-to-one time alignment between the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the time index sequence, checks and filters abnormal data units with missing or mismatched timestamps, and generates a time index serialized data frame.

[0014] The spatiotemporal data fusion submodule defines a four-dimensional data structure including horizontal, vertical, and axial coordinates and grayscale intensity values ​​based on the time-indexed serialized data frame. It fills the coordinate field of the structure with the aligned three-dimensional spatial point vectors, maps the two-dimensional grayscale pixel array to the corresponding texture intensity field, and chains together multiple four-dimensional data structures according to the arrangement order of the time index sequence to establish a spatiotemporal fused point cloud image sequence.

[0015] As a further aspect of the present invention, the rotational motion manifold analysis module includes:

[0016] The tangent plane feature extraction submodule analyzes the three-dimensional spatial point vectors of the ratchet and wrist arm base in continuous time steps based on the spatiotemporal fusion point cloud image sequence. It performs centering processing on the three-dimensional spatial point vectors and calculates the covariance matrix. It performs eigenvalue decomposition operation on the covariance matrix and selects feature vectors representing the main motion direction according to the magnitude of the eigenvalues. The local spatial plane that can characterize the real-time motion state by the feature vectors is calculated. The normal vector parameters perpendicular to the local spatial plane are calculated to generate a local tangent plane feature vector group.

[0017] The manifold projection measurement submodule calls the local tangent plane feature vector group, loads the preset standard trajectory equation describing the ideal rotation path of the device, constructs the corresponding manifold surface geometric constraints based on the standard trajectory equation, projects the measured three-dimensional spatial coordinates onto the manifold surface geometric constraint surface along the direction of the local tangent plane normal vector, determines the position coordinates of the projection point, calculates the geodesic path length of the measured three-dimensional spatial coordinates and the position coordinates of the projection point on the manifold surface, and generates a manifold space geodesic distance sequence.

[0018] The deviation trend fusion submodule calculates the statistical variance of the geodesic path length values ​​according to the time index order for the geodesic distance sequence in the manifold space, calculates the first difference of the geodesic path length values ​​as the time changes, extracts the corresponding two-dimensional gray matrix from the spatiotemporal fusion point cloud image sequence, establishes a corresponding storage relationship between the statistical variance values, the first difference values ​​and the two-dimensional gray matrix, and establishes a trajectory geodesic deviation and gray dataset.

[0019] As a further aspect of the present invention, the centralization process of the three-dimensional spatial point vector is specifically defined as follows: the arithmetic mean vector of the three-dimensional spatial point vectors within a continuous time step is used as the central reference vector to form a centralized point set.

[0020] The calculation of the covariance matrix is ​​specifically defined as constructing a symmetric real matrix based on a centralized point set, and the matrix elements are obtained by multiplying the dimension components pairwise and normalizing the number of time steps; the selection of feature vectors representing the main motion direction based on the magnitude of the feature values ​​is specifically defined as selecting feature vectors with corresponding peak feature values ​​and whose feature values ​​account for a proportion of the total number of feature values ​​not less than a preset proportion threshold.

[0021] The local spatial plane that can characterize the real-time motion state by feature vectors is specifically defined as a set of planar basis vectors determined by no less than two mutually orthogonal feature vectors; the calculation of the normal vector parameter perpendicular to the local spatial plane is specifically defined as performing a vector cross product on the set of planar basis vectors and performing normalization processing.

[0022] As a further aspect of the present invention, the frequency domain energy extraction module includes:

[0023] The frequency domain convolution filtering submodule calls the trajectory geodesic deviation and grayscale dataset, extracts the included two-dimensional grayscale matrix, loads the preset high-pass filter matrix for extracting image edge sharpness and the low-pass filter matrix for smoothing image background, uses the high-pass filter matrix to perform a sliding convolution operation with a stride of one on the two-dimensional grayscale matrix, and uses the low-pass filter matrix to perform weighted average convolution processing on the two-dimensional grayscale matrix to generate high-frequency texture feature maps and low-frequency contour feature maps, and establishes high-frequency texture coefficient matrices and low-frequency contour coefficient matrices;

[0024] The frequency band energy quantization submodule, based on the high-frequency texture coefficient matrix and the low-frequency contour coefficient matrix, traverses each frequency response value in the matrix, performs a square operation on the frequency response value to obtain the signal power spectral density, accumulates and sums the squared values ​​in the high-frequency texture coefficient matrix to obtain the signal energy value, and quantifies the micro-texture intensity and macro-geometric morphology intensity of the contact line surface according to the accumulation result to generate the texture detail energy amplitude and contour morphology energy amplitude.

[0025] The energy proportion aggregation submodule performs an addition operation on the energy amplitude of the texture details and the energy amplitude of the contour shape to obtain the total energy value of the entire frequency band. It calculates the ratio between the energy amplitude of the contour shape and the total energy value of the entire frequency band. Combining the corresponding geodesic distance statistical variance and first-order difference data, it associates and binds the calculated ratio value with the extracted trajectory deviation data to generate a deformation energy proportion and deviation set.

[0026] As a further aspect of the present invention, the wear gradient determination module includes:

[0027] The spatial gradient calculation submodule extracts a numerical sequence representing the energy proportion of the contour shape based on the deformation energy proportion and deviation set. It sorts and positions the numerical sequence along the spatial direction of the physical extension of the contact line, calculates the difference in energy proportion between adjacent acquisition points, quantifies the rate of change of frequency domain energy characteristics in the spatial dimension, and generates the spatial gradient value of energy proportion.

[0028] The wear level classification submodule calls the energy proportion spatial gradient value, loads the preset gradient threshold parameter used to define structural abrupt changes, performs a comparison operation between the gradient value and the gradient threshold parameter, determines whether there is an abnormal jump in energy distribution based on the comparison result, and maps the surface state of the contact line to a predefined physical loss category in combination with the size range of the original energy proportion value to generate a wear type classification identifier.

[0029] The index construction submodule classifies the wear type and, in conjunction with the trajectory geodesic deviation data, performs multi-dimensional data correlation between the wear classification results and the trajectory deviation values ​​to construct an evaluation data structure that includes wear level and the degree of mechanical loosening of rotating parts. It outputs quantified equipment health status parameters and establishes quantitative indicators for wear level and loosening.

[0030] As a further aspect of the present invention, the sorting and positioning of the numerical sequence along the spatial direction of the physical extension of the contact line is specifically defined as arranging them in ascending order according to the spatial coordinate projection distance of the collection points on the contact line.

[0031] The calculation of the difference in energy percentage between adjacent collection points is specifically limited to performing a difference operation on the sorted adjacent energy percentage values ​​and dividing by the spatial distance between the corresponding points to obtain the change within a unit length.

[0032] The gradient threshold parameter is specifically limited to the upper limit of a fixed value range obtained statistically from stable operating conditions.

[0033] As a further aspect of the present invention, the system also includes a state risk assessment module:

[0034] The status risk assessment module aligns the wear level and loosening quantification index in time and space according to the timestamp, maps the aligned index data to the preset contact network maintenance urgency range, and generates an early warning result for the operating status of the contact network equipment.

[0035] The early warning results of the overhead contact line equipment operation status include spatiotemporal alignment status data, maintenance urgency level code, and abnormal alarm signals.

[0036] As a further aspect of the present invention, the state risk assessment module includes:

[0037] The spatiotemporal index alignment submodule analyzes the wear level identification data and loosening deviation numerical data included in the wear level and loosening quantification index, extracts the collection timestamp embedded in the data packet, uses the timestamp as the reference index key, and performs row-dimensional splicing operation on the wear level identification data and loosening deviation numerical data at the same time to generate a spatiotemporal synchronization state feature matrix.

[0038] The urgency mapping and determination submodule calls the spatiotemporal synchronization state feature matrix, loads the preset contact network maintenance urgency interval table that defines differentiated maintenance priorities, maps the wear level value and loosening deviation value to the corresponding value range in the contact network maintenance urgency interval table, retrieves the risk weight coefficient associated with the value range, calculates the score representing the severity of real-time equipment failure based on the risk weight coefficient, and generates the equipment maintenance urgency rating coefficient.

[0039] The status early warning submodule sets a risk warning threshold for triggering the alarm mechanism based on the equipment maintenance urgency rating coefficient. It performs a comparison operation between the equipment maintenance urgency rating coefficient and the risk warning threshold. For coefficient points that exceed the risk warning threshold, it extracts the corresponding equipment spatial coordinates and anomaly type codes, encapsulates alarm data packets including location, type and level according to a predefined communication protocol format, and establishes an early warning result for the operating status of the overhead contact line equipment.

[0040] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0041] By constructing local tangent planes and calculating the geodesic projection distance onto the standard trajectory surface, constrained motion and abnormal deviations can be effectively separated. Manifold geometry analysis is used to capture nonlinear spatial anomalies hidden in vibration noise to identify minor loosening. By filtering and separating texture energy and morphological energy, the material loss can be accurately quantified. The spatial gradient value of the deformation energy ratio can be calculated to effectively distinguish between shallow surface interference and structural depth abrupt changes. Wear types are classified and judged according to energy distribution characteristics to eliminate false alarms caused by scratches. The quantitative indicators of spatiotemporal alignment can be used to accurately identify the stability and health status of equipment operation in dynamic environments. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A schematic diagram of the system provided by the present invention;

[0044] Figure 2 This is a schematic diagram of the system framework of the present invention;

[0045] Figure 3 This is a flowchart of the data timing arrangement module in this invention;

[0046] Figure 4 This is a flowchart of the rotational motion manifold analysis module in this invention;

[0047] Figure 5 This is a flowchart of the mid-frequency domain energy extraction module of the present invention;

[0048] Figure 6 This is a flowchart of the wear gradient determination module in this invention;

[0049] Figure 7 This is a flowchart of the state risk assessment module in this invention. Detailed Implementation

[0050] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0051] This invention provides an AI-based wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit, such as... Figure 1-2 The diagram shown illustrates an AI wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit. The system includes:

[0052] The data time sequence arrangement module acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment of rail transit, quantizes the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, and serializes the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the acquisition timestamp to generate a spatiotemporal fusion point cloud image sequence.

[0053] The rotational motion manifold analysis module is based on spatiotemporal fusion point cloud image sequences. It calculates the covariance matrix of the three-dimensional spatial point vectors of the ratchet and the wrist arm base, extracts the feature vectors of the corresponding peak feature values ​​to construct a local tangent plane, projects the measured coordinate points along the normal vector of the local tangent plane onto the surface defined by the preset standard trajectory equation, measures the geodesic length between the measured coordinate points and the projected points, calculates the dispersion and growth trend of the geodesic length over time, and constructs the trajectory geodesic deviation and grayscale dataset.

[0054] The frequency domain energy extraction module calls the trajectory geodesic deviation and grayscale dataset, and uses the preset high-pass filter matrix and low-pass filter matrix to perform convolution operation on the two-dimensional grayscale pixel array to calculate the texture energy amplitude and contour shape energy amplitude respectively, calculate the proportion of contour shape energy amplitude in the total energy, and generate the deformation energy proportion and deviation set.

[0055] The wear gradient determination module calculates the spatial gradient value of the deformation energy ratio along the contact line extension direction based on the deformation energy ratio and deviation set, compares the spatial gradient value with the preset abrupt change threshold, divides the wear level according to the energy distribution characteristics, and establishes wear level and loosening quantification index.

[0056] The status risk assessment module aligns the wear level and loosening quantification indicators in time and space based on the timestamp, maps the aligned indicator data to the preset contact network maintenance urgency range, and generates early warning results for the operating status of contact network equipment.

[0057] The spatiotemporal fusion point cloud image sequence includes serialized grayscale image frames, three-dimensional spatial coordinate sequences, and synchronization time indexes. The trajectory geodesic deviation and grayscale dataset includes geodesic projection distance values, trajectory dispersion feature values, deviation growth trend coefficients, and associated grayscale matrices. The deformation energy proportion and deviation set includes texture detail energy amplitude, contour morphology energy amplitude, and deformation energy proportion coefficients. The wear level and loosening quantification indicators include energy proportion spatial gradient values, wear level classification identifiers, and loosening risk quantification values. The overhead contact line equipment operation status early warning results include spatiotemporal alignment status data, maintenance urgency level codes, and abnormal alarm signals.

[0058] Specifically, such as Figure 2 , 3 As shown, the data time-series arrangement module includes:

[0059] The multimodal data acquisition submodule acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment in rail transit, drives the acquisition sensor to quantize the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, captures the hardware clock signal at the moment of sensor triggering, generates a timestamp, and encapsulates the two-dimensional grayscale pixel array, the three-dimensional spatial point vector and the timestamp into independent data units to generate a discrete grayscale coordinate raw dataset.

[0060] An industrial-grade linear array camera and a phase-detection lidar scanner, deployed on top of the track inspection vehicle, are activated. The linear array camera, equipped with a 4096-pixel-width CMOS sensor, continuously captures photon energy from the contact wire ratchet compensation device and the cantilever arm base surface using a line scan method. An analog-to-digital converter inside the sensor discretizes the received analog voltage signal into grayscale levels ranging from 0 to 255, forming a two-dimensional grayscale pixel array, where 0 represents pure black and 255 represents pure white. This array faithfully records the rust texture and reflective properties of the metal surface. Simultaneously, the phase-detection lidar emits laser pulses with a wavelength of 1550 nanometers, calculating the straight-line distance between the target point and the sensor by measuring the phase difference between the emitted and reflected waves. Combined with the deflection angle of the scanning mirror of the internally integrated microelectromechanical system The horizontal coordinate is obtained by using the transformation logic from polar coordinates to Cartesian coordinates, that is, by multiplying the distance value by the cosine of the scanning angle. The vertical coordinate is obtained by multiplying the distance value by the sine of the scanning angle. The geometric contour information is solved into a three-dimensional spatial point vector containing values ​​of the horizontal axis X, vertical axis Y, and vertical axis Z (estimated from the vehicle's travel distance). At the moment of each triggering acquisition action, the submodule captures the hardware interrupt signal issued by the synchronization controller through the field programmable gate array (FPGA) and reads the current high-precision crystal counter value. The counter accumulates with a 10 nanosecond base pulse. The system converts the read count value into an absolute time value in microseconds and generates a unique timestamp. Subsequently, the submodule allocates a structured storage space in memory and encapsulates the memory address pointer of the two-dimensional grayscale pixel array, the floating-point value group of the three-dimensional spatial point vector, and the long integer value of the timestamp into an independent data packet, which is continuously stored in the high-speed cache queue to generate a discrete grayscale coordinate raw dataset. As shown in Table 1, the specific configuration parameters of the acquisition sensor at this stage are displayed.

[0061] Table 1: Configuration Parameters of Multimodal Sensor

[0062] Parameter name Parameter values unit Remark Linear scan camera resolution 4096 Pixels Horizontal coverage LiDAR scanning frequency 100 hertz Harness refresh rate Distance measurement accuracy 2 millimeters Within 100 meters Time synchronization error 1 microseconds FPGA hardware triggering

[0063] The sequence synchronization and alignment submodule calls the original dataset of discrete grayscale coordinates, iterates through and reads the timestamp values ​​in the data units, performs an ascending sort operation on the data units according to the size of the timestamp values, constructs a time index sequence with monotonically increasing characteristics, performs one-to-one time alignment between the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the time index sequence, checks and filters abnormal data units with missing or mismatched timestamps, and generates a time index serialized data frame.

[0064] The quicksort algorithm is used to process the data units in the cache queue. This algorithm selects the first timestamp in the queue as the pivot, moves all data units with timestamp values ​​less than the pivot to the left, and all data units with timestamp values ​​greater than the pivot to the right, and recursively repeats this process on both sides until all data units are arranged strictly in ascending order of timestamp values, thus constructing a monotonically increasing time index sequence. Based on this, the system sets a time tolerance window, for example, setting the tolerance value to [value missing]. In microseconds, the system iterates through the sorted sequence, checking the timestamp difference between adjacent data units. If the timestamp of a row of scanned data is a grayscale image... Corresponding radar point cloud timestamp The absolute value of the difference satisfies If the difference is within microseconds, it is determined that the data was collected at the same time and data association is performed. If the difference exceeds the tolerance range or a certain type of data is missing, the data set is marked as an invalid frame and removed from the sequence, thereby generating a time-indexed serialized data frame.

[0065] The spatiotemporal data fusion submodule is based on time-indexed serialized data frames. It defines a four-dimensional data structure including horizontal, vertical, and axial coordinates and grayscale intensity values. The aligned three-dimensional spatial point vectors are filled into the coordinate fields of the structure, and the two-dimensional grayscale pixel array is mapped to the corresponding texture intensity field. Multiple four-dimensional data structures are chained together according to the arrangement order of the time index sequence to establish a spatiotemporal fused point cloud image sequence.

[0066] A four-dimensional data structure is dynamically allocated in the memory heap area. This structure contains four floating-point member variables, named respectively. , , and The system directly assigns the three coordinate components of the aligned 3D spatial point vector to the first three member variables. Simultaneously, it reads the grayscale value of the corresponding spatial position in the 2D grayscale pixel array, normalizes it to a floating-point number between 0.0 and 1.0, and then assigns it to... The system utilizes a doubly linked list data structure, following the time index order, to connect the scattered four-dimensional data structures in memory through pointers, forming a continuous spatiotemporal fusion point cloud image sequence. For example, in a certain actual acquisition, the lidar measured the three-dimensional coordinates of a point on the ratchet surface as the X-axis. millimeters, Y-axis millimeters, Z-axis Millimeters, at the same moment, the linear scan camera captured a grayscale level of 128 at this location. The system verified that the timestamp of this data point deviated from the camera frame timestamp by only 120 microseconds, which is less than the threshold of 500 microseconds. Therefore, the alignment was deemed successful, and the system created a four-dimensional structure. Assigned value , Assigned value , Assigned value For grayscale values, the system performs normalization calculation, which involves dividing the acquired value of 128 by the maximum range of 255. The calculation formula is as follows: Assign it to Field, the four-dimensional description of this point This processing logic ensures that geometric deformation information and surface texture information can be used simultaneously in the subsequent process, avoiding the limitations of single-modal data, and establishing a spatiotemporal fusion point cloud image sequence.

[0067] Specifically, such as Figure 2 , 4 As shown, the rotational motion manifold analysis module includes:

[0068] The tangent plane feature extraction submodule is based on the spatiotemporal fusion point cloud image sequence. It analyzes the three-dimensional spatial point vectors of the ratchet and the wrist arm base in continuous time steps, performs centering processing on the three-dimensional spatial point vectors and calculates the covariance matrix. It performs eigenvalue decomposition operation on the covariance matrix, selects feature vectors representing the main motion direction according to the magnitude of the eigenvalues, and calculates the normal vector parameters perpendicular to the local spatial plane to generate a local tangent plane feature vector group.

[0069] Define a sliding window containing a set of neighborhood points, for example, selecting 50 points before and after the current sampling point, for a total of 101 three-dimensional spatial point vectors. Calculate the geometric center coordinates of these 101 points, that is, sum the X, Y, and Z axis coordinates and divide by the number of points 101 to obtain the mean vector. Centering is performed on the vector of each point within the window, that is, the coordinate vector of each point is... Subtract the above mean vector, that is To eliminate the influence of the origin offset, based on the centered point set, construct... Covariance matrix of dimension The diagonal elements of this matrix represent the degree of dispersion along each coordinate axis, while the off-diagonal elements represent the correlation between axes. The system calls the Jacobi iterative algorithm to perform eigenvalue decomposition on the covariance matrix, iterating until the off-diagonal elements converge to zero, thereby resolving the three real eigenvalues. and the corresponding three feature vectors The system compares the numerical values ​​of these three features (assuming...). Select the eigenvalue with the smallest value. The corresponding eigenvector Define it as a normal vector Because the data variance is smallest in this direction, it is perpendicular to the main plane of data distribution; select the eigenvectors corresponding to the two eigenvalues ​​with larger values ​​to form a local tangent plane, which is the tangential space of the current rotational motion of the device, and generate a set of eigenvectors of the local tangent plane.

[0070] The manifold projection measurement submodule calls the local tangent plane feature vector group, loads the preset standard trajectory equation describing the ideal rotation path of the device, constructs the corresponding manifold surface geometric constraints based on the standard trajectory equation, projects the measured three-dimensional spatial coordinates onto the manifold surface geometric constraint surface along the direction of the local tangent plane normal vector, determines the position coordinates of the projection point, calculates the geodesic path length of the measured three-dimensional spatial coordinates and the position coordinates of the projection point on the manifold surface, and generates a manifold space geodesic distance sequence.

[0071] The standard rotational trajectory equation for a loading ratchet mechanism, for example, is defined as a circle in the XY plane with radius r. The system constructs geometric constraints for the manifold surface based on this equation. Then, it performs projection calculations. Starting from the measured three-dimensional coordinate point, a ray is drawn along the direction of the previously calculated normal vector. The intersection point of this ray with the standard manifold surface is calculated, and this intersection point is the projection point. For a rotating manifold in Euclidean space, this distance can be approximated as the shortest Euclidean distance from the measured point to the ideal trajectory curve. However, in manifold analysis, the system uses Riemannian metric logic, that is, it calculates the difference in arc length between two points on the curved manifold surface and generates a sequence of geodesic distances in manifold space.

[0072] The deviation trend fusion submodule calculates the statistical variance of the geodesic path length values ​​according to the time index order for the geodesic distance sequence in the manifold space, calculates the first difference of the geodesic path length values ​​as the time changes, extracts the corresponding two-dimensional gray matrix from the spatiotemporal fusion point cloud image sequence, establishes a corresponding storage relationship between the statistical variance values, the first difference values ​​and the two-dimensional gray matrix, and establishes a trajectory geodesic deviation and gray dataset.

[0073] To calculate the statistical variance of the sequence, iterate through each distance value in the sequence, calculate the square of the difference between that distance and the sequence mean, sum all the squared differences, and divide by the number of data points to obtain the variance value that characterizes the stability of the motion. Simultaneously, the first-order difference is calculated, which involves subtracting the distance from the previous time step from the current distance value and dividing by the sampling time interval to obtain the gradient value representing the rate of change of the deviation. The system associates and stores the calculated kinematic features with the original two-dimensional grayscale matrix through a time index. For example, after calculating the covariance matrix of the three-dimensional point set within the sliding window, the normal vector corresponding to the smallest eigenvalue is set to... This indicates that the local plane is parallel to the XY plane, and the coordinates of the current measured point are... The preset standard trajectory equation is the radius. A circle of millimeters, centered at the origin. The system calculates the projected distance of the measured point on the XY plane, i.e., from the origin to... The modulus is calculated as follows: If the distance is millimeters, then the geodesic distance of that point from the standard manifold surface (radius 500 millimeters) is... millimeters, if the geodesic distance at the previous moment was millimeters, sampling interval is If the time is seconds, then the first difference is calculated as follows: Millimeters per second, assuming the variance of the distance sequence within the statistical window is calculated as follows: Square millimeters, the system will have variance ,gradient The calculation logic is beneficial because it binds the grayscale matrix of the current point and stores it in the dataset. The advantage of this logic is that it transforms the complex three-dimensional spatial error into a single-dimensional geodesic distance through manifold projection. Combined with gradient and variance, it can accurately quantify the mechanical loosening and wear of rotating equipment and establish trajectory geodesic deviation and grayscale dataset.

[0074] Specifically, such as Figure 2 , 5 As shown, the frequency domain energy extraction module includes:

[0075] The frequency domain convolution filtering submodule calls the trajectory geodesic deviation and grayscale dataset, extracts the included two-dimensional grayscale matrix, loads the preset high-pass filter matrix for extracting image edge sharpness and the low-pass filter matrix for smoothing image background, uses the high-pass filter matrix to perform a sliding convolution operation with a stride of one on the two-dimensional grayscale matrix, and uses the low-pass filter matrix to perform weighted average convolution processing on the two-dimensional grayscale matrix to generate high-frequency texture feature maps and low-frequency contour feature maps, and establishes high-frequency texture coefficient matrix and low-frequency contour coefficient matrix;

[0076] The decompression yields a two-dimensional grayscale matrix, which is... The numerical array is loaded with preset convolution kernels from read-only memory, wherein the high-pass filter matrix is ​​selected as... The Laplacian operator, with a central element of 8 and eight neighboring elements of -1, is designed to suppress flat regions and enhance signals from abrupt edge changes; the low-pass filter matrix is ​​selected... The Gaussian smoothing kernel, whose element values ​​follow a two-dimensional Gaussian distribution and whose sum of all elements is normalized to 1, aims to filter out noise and preserve the overall contour. A sliding convolution operation is performed, aligning the center of the filter matrix with each pixel in the grayscale matrix. Each weight coefficient within the filter is multiplied by the corresponding pixel value in the grayscale matrix, and all products are summed to obtain the new value for that pixel. The convolution stride is set to 1, meaning the filter moves pixel by pixel, covering the entire image area. The system performs a two-dimensional discrete Fourier transform (DFT) on the convolution output matrix, converting the grayscale distribution in the spatial domain into spectral coefficients in the frequency domain, resulting in a high-frequency texture coefficient matrix and a low-frequency contour coefficient matrix.

[0077] The frequency band energy quantization submodule is based on the high-frequency texture coefficient matrix and the low-frequency contour coefficient matrix. It traverses each frequency response value in the matrix, performs a square operation on the frequency response value to obtain the signal power spectral density, accumulates and sums the square values ​​in the high-frequency texture coefficient matrix to obtain the signal energy value, and quantifies the micro-texture intensity and macro-geometric morphology intensity of the contact line surface based on the accumulation result to generate the texture detail energy amplitude and contour morphology energy amplitude.

[0078] For each complex form of the frequency response value in the matrix The system calculates the square of its modulus. In a numerical physical sense, this represents the power spectral density of the corresponding frequency component. The system sets two accumulator variables, both initially set to zero. For the high-frequency texture coefficient matrix, the system accumulates the squared values ​​of the modulus of each element into the first accumulator to obtain the texture detail energy amplitude. This numerical value quantifies the microscopic wear and wire-drawing degree of the contact line surface; for the low-frequency profile coefficient matrix, the system accumulates the squared modulus values ​​of the elements to a second accumulator to obtain the profile shape energy amplitude. The macroscopic geometric deformation of the numerical quantization device generates texture detail energy amplitude and contour shape energy amplitude.

[0079] The energy proportion aggregation submodule performs an addition operation on the energy amplitude of texture details and the energy amplitude of contour shape to obtain the total energy value of the entire frequency band. It calculates the ratio between the energy amplitude of contour shape and the total energy value of the entire frequency band. Combining the corresponding geodesic distance statistical variance and first-order difference data, it associates and binds the calculated ratio value with the extracted trajectory deviation data to generate a deformation energy proportion and deviation set.

[0080] The energy amplitude of texture detail and the energy amplitude of contour shape are added together to obtain the total energy value across the entire frequency band. By using division, the percentage of the contour shape energy amplitude to the total energy value across the entire frequency band is calculated, i.e., the energy percentage. Finally, the system queries the statistical variance and first difference of the geodesic distance generated by the previous module, and combines and encapsulates the energy percentage value with these two kinematic parameters. For example, in a simplified... In a local grayscale region, the pixel value is Using the Qualcomm Laplacian operator (center 8, perimeter -1) to convolve the center point, the calculation logic is as follows: Assuming that after Fourier transform and full-image summation, the obtained high-frequency texture detail energy amplitude for Unit: Low-frequency profile energy amplitude for The unit is the total energy value calculated by the system across the entire frequency band. Calculate the ratio, that is At this point, the system extracts the corresponding trajectory deviation data, assuming the variance is... square millimeters, gradient is The system generates a dataset of millimeters per second. The calculation logic intuitively reveals the equipment status through frequency domain energy distribution. When the ratio is 0.75, which is within the normal range, it indicates that the equipment mainly exhibits contour features and the surface wear is relatively light. If the ratio drops significantly, it means that the high-frequency texture energy increases, the surface roughness increases, and the deformation energy ratio and deviation set are generated.

[0081] Specifically, such as Figure 2 , 6 As shown, the wear gradient determination module includes:

[0082] The spatial gradient calculation submodule extracts a numerical sequence representing the energy proportion of the contour shape based on the deformation energy proportion and deviation set. It sorts and locates the numerical sequence along the spatial direction of the physical extension of the contact line, calculates the difference in energy proportion between adjacent acquisition points, quantifies the rate of change of frequency domain energy characteristics in the spatial dimension, and generates the spatial gradient value of energy proportion.

[0083] Based on the odometer markers or GPS coordinates at the time of data collection, the data points in the deformation energy percentage and deviation set are linearly sorted from the starting point to the ending point according to spatial distance. The system sets a spatial step interval, for example, every [number] [times]. Mi selects a sampling point, traverses the sorted sequence, and performs a difference operation: reads the first sampling point... The energy percentage value of each point, minus the value of the first point. The energy percentage value of each point is used to obtain the energy percentage increment; the energy percentage value of the first point is read. The mileage value of the first point, minus the mileage value of the second point. The system calculates the distance increment by dividing the energy percentage increment by the distance increment and taking the absolute value to obtain the energy percentage spatial gradient value.

[0084] The wear level classification submodule calls the energy proportion spatial gradient value, loads the preset gradient threshold parameter used to define structural abrupt changes, performs a comparison operation between the gradient value and the gradient threshold parameter, determines whether there is an abnormal jump in energy distribution based on the comparison result, and maps the surface state of the contact line to the predefined physical loss category by combining the size range of the original energy proportion value, and generates a wear type classification identifier.

[0085] A pre-defined multi-level threshold table is established, based on experimental determinations of the material fatigue characteristics of copper-magnesium alloy contact wires. The first-level normal threshold is set as follows: The second-level warning threshold is per meter. For each meter, the calculated spatial gradient value is compared with a threshold value one by one. If the gradient is less than a threshold value... If the gradient is between 0 and 1, it is considered a smooth transition; if the gradient is between 0 and 1, it is considered a smooth transition. and Between these values, it is determined to be a minor mutation; if the gradient is greater than... If the system determines the wear type as a severe mutation, it combines the original energy percentage value (e.g., between 0 and 1) with a two-dimensional lookup table method. For example, when the energy percentage is lower than... And the gradient is greater than When the energy percentage is higher than 10%, it is judged as "wave wear"; when the energy percentage is higher than 10%, it is judged as "wave wear". And the gradient is less than When the wear is normal, it is determined to be "normal wear" and a wear type classification label is generated.

[0086] The indicator construction submodule identifies wear type classification and combines it with trajectory geodesic deviation data to perform multi-dimensional data correlation between wear classification results and trajectory deviation values. It constructs an evaluation data structure that includes wear level and mechanical loosening degree of rotating parts, outputs quantified equipment health status parameters, and establishes quantitative indicators for wear level and loosening.

[0087] Create a system that includes a "depth loss level" ( ) and "mechanical loosening" ( The evaluation data structure directly maps the depth of wear level to the wear type classification identifier, for example, mapping "wave wear" to level 3 and "normal wear" to level 1. Mechanical loosening is quantified using the statistical variance of the trajectory geodesic deviation data calculated by the aforementioned module. The system multiplies the variance value by a preset amplification factor (e.g., ...). (), and round down to obtain a quantitative indicator, for example, at kilometer marker K100+000, the energy percentage is At kilometer marker K100+001 (1 meter interval), the energy percentage is: The system calculates the energy percentage increment as follows: per meter, the distance increment is If the distance is meters, then the spatial gradient value is... per meter, this value Greater than the first threshold And less than the second-level threshold This is determined to be a slight mutation; assuming the corresponding trajectory variance at this time is... The system determines the wear type to be level two based on the square millimeter, i.e. Assigning a value of 2, for looseness, the system will use variance Multiply by a coefficient Calculated ,Right now The value is assigned to 4, and the health status parameter is... Establish wear levels and quantitative indicators for loosening.

[0088] Specifically, such as Figure 2 , 7 As shown, the status risk assessment module includes:

[0089] The spatiotemporal index alignment submodule analyzes the wear level identification data and loosening deviation numerical data for wear level and loosening quantification indicators, extracts the collection timestamp embedded in the data packet, uses the timestamp as the reference index key, and performs row-dimensional splicing operation on the wear level identification data and loosening deviation numerical data at the same time to generate a spatiotemporal synchronization state feature matrix.

[0090] Extract the timestamp field (accurate to milliseconds) and geographic location information (latitude, longitude, and odometer) from the header of each data unit, and construct a hash table structure, using the timestamp as the unique key. The system iterates through the wear level data stream, storing the value of each level into the hash bucket corresponding to the timestamp; it then iterates through the loosening quantification index data stream, appending the loosening value to the corresponding hash bucket by searching for the same timestamp key. After all data streams have been traversed, the system exports the data in the hash table into a matrix format, where each row of the matrix contains: Generate a spatiotemporal synchronization state feature matrix.

[0091] The urgency mapping and determination submodule calls the spatiotemporal synchronization state feature matrix, loads the preset contact network maintenance urgency interval table that defines differentiated maintenance priorities, maps the wear level value and loosening deviation value to the corresponding value range in the contact network maintenance urgency interval table, retrieves the risk weight coefficient associated with the value range, calculates the score representing the severity of real-time equipment failure based on the risk weight coefficient, and generates the equipment maintenance urgency rating coefficient.

[0092] The maintenance urgency interval table is read from the configuration database. This table defines the risk weights for different indicator intervals, as shown in Table 2, which displays the system's preset risk weight configuration.

[0093] Table 2: Risk Weighting Table for the Urgency of Overhead Contact Line Maintenance

[0094] Indicator Type numerical range Risk weighting coefficient Remark Wear levels [1 2) 0.2 Slight wear Wear levels [2 3) 0.6 Moderate wear Wear levels [3 5] 1 Severe wear Loosening deviation [0 5) 0.3 Normal vibration Loosening deviation [5 10) 0.7 slight vibration Loosening deviation [10 ∞) 1.2 Severe vibration

[0095] Referring to Table 2, the system performs weighted scoring calculations based on this table. The system reads the wear level value of a certain row in the matrix, determines the interval it falls into, and retrieves the corresponding wear risk weight (denoted as ). Read the loosening deviation value and retrieve the corresponding loosening risk weight (denoted as ). The system performs a linear weighted summation operation: Equipment maintenance urgency rating coefficient. This scoring logic can comprehensively reflect the multidimensional health status of the equipment.

[0096] The status early warning submodule sets a risk warning threshold for triggering the alarm mechanism based on the equipment maintenance urgency rating coefficient. It performs a comparison operation between the equipment maintenance urgency rating coefficient and the risk warning threshold. For coefficient points that exceed the risk warning threshold, it extracts the corresponding equipment spatial coordinates and anomaly type codes, encapsulates alarm data packets including location, type and level according to a predefined communication protocol format, and establishes an early warning result for the operating status of the contact network equipment.

[0097] Set a global risk warning threshold, for example, set the threshold to be [value]. Iterate through each calculated rating coefficient, perform a comparison operation, and if the rating coefficient at a certain point is greater than... The system immediately triggers the alarm process. It extracts the odometer information (e.g., "K120+500") and the specific anomaly type code (e.g., "ERR_WEAR_LOOSE") for that point, and encapsulates it according to ModbusTCP or a custom industrial Ethernet protocol format. The encapsulation process involves converting the start character, device ID, alarm level (e.g., "High"), odometer string, and rating coefficient value into a hexadecimal byte stream, calculating a CRC checksum, and appending it to the end to create an alarm data packet. For example, if the system processes a row of data with a wear level value of 3 (belonging to the severe wear range, weight 1.0) and a loosening deviation value of 6.0 (belonging to the slight loosening range, weight 0.7), the system calculates the rating coefficient based on the logic in Table 2. Assuming the set risk warning threshold is The system comparison found Therefore, it was determined that the location had not yet reached the emergency alarm standard, and the data was only recorded in the database without sending an immediate alarm. If another set of data showed a different result... greater than The system will then generate an alarm data packet containing "Location: K120+500, Coefficient: 9.5" and send it through the network port to establish an early warning result for the operating status of the contact network equipment.

[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. An AI-based wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit, characterized in that, The system includes: The data time sequence arrangement module acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment of rail transit, quantizes the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, and serializes the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the acquisition timestamp to generate a spatiotemporal fusion point cloud image sequence. The rotational motion manifold analysis module calculates the covariance matrix of the three-dimensional spatial point vectors of the ratchet and the wrist arm base based on the spatiotemporal fusion point cloud image sequence, calculates the dispersion and growth trend of the geodesic length over time, and constructs the trajectory geodesic deviation and grayscale dataset. The frequency domain energy extraction module calls the trajectory geodesic deviation and grayscale dataset, performs convolution operation on the two-dimensional grayscale pixel array, calculates the proportion of the contour shape energy amplitude in the total energy, and generates a set of deformation energy proportion and deviation. The wear gradient determination module calculates the spatial gradient value of the deformation energy ratio along the contact line extension direction based on the deformation energy ratio and deviation set, compares the spatial gradient value with the preset abrupt change threshold, divides the wear level, and establishes wear level and loosening quantification index.

2. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 1, characterized in that, The spatiotemporal fusion point cloud image sequence includes serialized grayscale image frames, a three-dimensional spatial coordinate sequence, and a synchronization time index. The trajectory geodesic deviation and grayscale dataset includes geodesic projection distance values, trajectory dispersion feature values, deviation growth trend coefficients, and associated grayscale matrices. The deformation energy proportion and deviation set includes texture detail energy amplitude, contour shape energy amplitude, and deformation energy proportion coefficients. The wear level and loosening quantification index includes energy proportion spatial gradient values, wear level classification identifiers, and loosening risk quantification values.

3. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 1, characterized in that, The data time-series arrangement module includes: The multimodal data acquisition submodule acquires the optical reflection features and geometric contour information of the ratchet, cantilever base and contact line surface of the rotating equipment in rail transit, drives the acquisition sensor to quantize the optical reflection features into a two-dimensional grayscale pixel array, solves the geometric contour information into a three-dimensional spatial point vector, captures the hardware clock signal at the moment of sensor triggering, generates a timestamp, and encapsulates the two-dimensional grayscale pixel array, the three-dimensional spatial point vector and the timestamp into independent data units to generate a discrete grayscale coordinate raw dataset. The sequence synchronization alignment submodule calls the original discrete grayscale coordinate dataset, traverses and reads the timestamp values ​​in the data units, performs an ascending sort operation on the data units according to the size of the timestamp values, constructs a time index sequence with monotonically increasing characteristics, performs one-to-one time alignment between the two-dimensional grayscale pixel array and the three-dimensional spatial point vector according to the time index sequence, checks and filters abnormal data units with missing or mismatched timestamps, and generates a time index serialized data frame. The spatiotemporal data fusion submodule defines a four-dimensional data structure including horizontal, vertical, and axial coordinates and grayscale intensity values ​​based on the time-indexed serialized data frame. It fills the coordinate field of the structure with the aligned three-dimensional spatial point vectors, maps the two-dimensional grayscale pixel array to the corresponding texture intensity field, and chains together multiple four-dimensional data structures according to the arrangement order of the time index sequence to establish a spatiotemporal fused point cloud image sequence.

4. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 3, characterized in that, The rotational motion manifold analysis module includes: The tangent plane feature extraction submodule analyzes the three-dimensional spatial point vectors of the ratchet and wrist arm base in continuous time steps based on the spatiotemporal fusion point cloud image sequence. It performs centering processing on the three-dimensional spatial point vectors and calculates the covariance matrix. It performs eigenvalue decomposition operation on the covariance matrix and selects feature vectors representing the main motion direction according to the magnitude of the eigenvalues. The local spatial plane that can characterize the real-time motion state by the feature vectors is calculated. The normal vector parameters perpendicular to the local spatial plane are calculated to generate a local tangent plane feature vector group. The manifold projection measurement submodule calls the local tangent plane feature vector group, loads the preset standard trajectory equation describing the ideal rotation path of the device, constructs the corresponding manifold surface geometric constraints based on the standard trajectory equation, projects the measured three-dimensional spatial coordinates onto the manifold surface geometric constraint surface along the direction of the local tangent plane normal vector, determines the position coordinates of the projection point, calculates the geodesic path length of the measured three-dimensional spatial coordinates and the position coordinates of the projection point on the manifold surface, and generates a manifold space geodesic distance sequence. The deviation trend fusion submodule calculates the statistical variance of the geodesic path length values ​​according to the time index order for the geodesic distance sequence in the manifold space, calculates the first difference of the geodesic path length values ​​as the time changes, extracts the corresponding two-dimensional gray matrix from the spatiotemporal fusion point cloud image sequence, establishes a corresponding storage relationship between the statistical variance values, the first difference values ​​and the two-dimensional gray matrix, and establishes a trajectory geodesic deviation and gray dataset.

5. The AI ​​wear and loosening recognition system for contact wires of rotating equipment in rail transit according to claim 4, characterized in that, The specific definition of performing centering processing on the three-dimensional spatial point vector is as follows: using the arithmetic mean vector of the three-dimensional spatial point vector within a continuous time step as the central reference vector to form a centralized point set; The calculation of the covariance matrix is ​​specifically defined as constructing a symmetric real matrix based on a centralized point set, and the matrix elements are obtained by multiplying the dimension components pairwise and normalizing the number of time steps; the selection of feature vectors representing the main motion direction based on the magnitude of the feature values ​​is specifically defined as selecting feature vectors with corresponding peak feature values ​​and whose feature values ​​account for a proportion of the total number of feature values ​​not less than a preset proportion threshold. The local spatial plane that can characterize the real-time motion state by feature vectors is specifically defined as a set of planar basis vectors determined by no less than two mutually orthogonal feature vectors; the calculation of the normal vector parameter perpendicular to the local spatial plane is specifically defined as performing a vector cross product on the set of planar basis vectors and performing normalization processing.

6. The AI ​​wear and loosening recognition system for contact wires of rotating equipment in rail transit according to claim 4, characterized in that, The frequency domain energy extraction module includes: The frequency domain convolution filtering submodule calls the trajectory geodesic deviation and grayscale dataset, extracts the included two-dimensional grayscale matrix, loads the preset high-pass filter matrix for extracting image edge sharpness and the low-pass filter matrix for smoothing image background, uses the high-pass filter matrix to perform a sliding convolution operation with a stride of one on the two-dimensional grayscale matrix, and uses the low-pass filter matrix to perform weighted average convolution processing on the two-dimensional grayscale matrix to generate high-frequency texture feature maps and low-frequency contour feature maps, and establishes high-frequency texture coefficient matrices and low-frequency contour coefficient matrices; The frequency band energy quantization submodule, based on the high-frequency texture coefficient matrix and the low-frequency contour coefficient matrix, traverses each frequency response value in the matrix, performs a square operation on the frequency response value to obtain the signal power spectral density, accumulates and sums the squared values ​​in the high-frequency texture coefficient matrix to obtain the signal energy value, and quantifies the micro-texture intensity and macro-geometric morphology intensity of the contact line surface according to the accumulation result to generate the texture detail energy amplitude and contour morphology energy amplitude. The energy proportion aggregation submodule performs an addition operation on the energy amplitude of the texture details and the energy amplitude of the contour shape to obtain the total energy value of the entire frequency band. It calculates the ratio between the energy amplitude of the contour shape and the total energy value of the entire frequency band. Combining the corresponding geodesic distance statistical variance and first-order difference data, it associates and binds the calculated ratio value with the extracted trajectory deviation data to generate a deformation energy proportion and deviation set.

7. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 6, characterized in that, The wear gradient determination module includes: The spatial gradient calculation submodule extracts a numerical sequence representing the energy proportion of the contour shape based on the deformation energy proportion and deviation set. It sorts and positions the numerical sequence along the spatial direction of the physical extension of the contact line, calculates the difference in energy proportion between adjacent acquisition points, quantifies the rate of change of frequency domain energy characteristics in the spatial dimension, and generates the spatial gradient value of energy proportion. The wear level classification submodule calls the energy proportion spatial gradient value, loads the preset gradient threshold parameter used to define structural abrupt changes, performs a comparison operation between the gradient value and the gradient threshold parameter, determines whether there is an abnormal jump in energy distribution based on the comparison result, and maps the surface state of the contact line to a predefined physical loss category in combination with the size range of the original energy proportion value to generate a wear type classification identifier. The index construction submodule classifies the wear type and, in conjunction with the trajectory geodesic deviation data, performs multi-dimensional data correlation between the wear classification results and the trajectory deviation values ​​to construct an evaluation data structure that includes wear level and the degree of mechanical loosening of rotating parts. It outputs quantified equipment health status parameters and establishes quantitative indicators for wear level and loosening.

8. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 7, characterized in that, The sorting and positioning of the numerical sequence along the spatial direction of the physical extension of the contact line is specifically defined as arranging them in ascending order according to the spatial coordinate projection distance of the collection points on the contact line; The calculation of the difference in energy percentage between adjacent collection points is specifically limited to performing a difference operation on the sorted adjacent energy percentage values ​​and dividing by the spatial distance between the corresponding points to obtain the change within a unit length. The gradient threshold parameter is specifically limited to the upper limit of a fixed value range obtained statistically from stable operating conditions.

9. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 1, characterized in that, The system also includes a state risk assessment module: The status risk assessment module aligns the wear level and loosening quantification index in time and space according to the timestamp, maps the aligned index data to the preset contact network maintenance urgency range, and generates an early warning result for the operating status of the contact network equipment. The early warning results of the overhead contact line equipment operation status include spatiotemporal alignment status data, maintenance urgency level code, and abnormal alarm signals.

10. The AI ​​wear and loosening recognition system for overhead contact lines of rotating equipment in rail transit according to claim 9, characterized in that, The status risk assessment module includes: The spatiotemporal index alignment submodule analyzes the wear level identification data and loosening deviation numerical data included in the wear level and loosening quantification index, extracts the collection timestamp embedded in the data packet, uses the timestamp as the reference index key, and performs row-dimensional splicing operation on the wear level identification data and loosening deviation numerical data at the same time to generate a spatiotemporal synchronization state feature matrix. The urgency mapping and determination submodule calls the spatiotemporal synchronization state feature matrix, loads the preset contact network maintenance urgency interval table that defines differentiated maintenance priorities, maps the wear level value and loosening deviation value to the corresponding value range in the contact network maintenance urgency interval table, retrieves the risk weight coefficient associated with the value range, calculates the score representing the severity of real-time equipment failure based on the risk weight coefficient, and generates the equipment maintenance urgency rating coefficient. The status early warning submodule sets a risk warning threshold for triggering the alarm mechanism based on the equipment maintenance urgency rating coefficient. It performs a comparison operation between the equipment maintenance urgency rating coefficient and the risk warning threshold. For coefficient points that exceed the risk warning threshold, it extracts the corresponding equipment spatial coordinates and anomaly type codes, encapsulates alarm data packets including location, type and level according to a predefined communication protocol format, and establishes an early warning result for the operating status of the overhead contact line equipment.

Images