Automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit

Abstract

This invention discloses an automatic identification and state assessment method for UAV inspection of rail transit, specifically in the field of UAV inspection of rail transit. It addresses the problem that existing UAV inspection methods suffer from incomplete track defect identification and localization due to changes in track attitude and environmental interference. The method constructs an attitude temporal matrix by controlling a UAV swarm to fly along the track direction, and obtains analysis units using spatiotemporal alignment and sliding window segmentation. Principal component analysis is used to extract attitude cooperative oscillation features, constructing an attitude reference surface and calculating individual attitude anomaly indices. Dynamic thresholding is used to identify abnormal segments and locate track defects. Based on the defect distribution density, the swarm is controlled to perform matrix scanning and cross-validation flights. Finally, a continuous track smoothness assessment map is output. This invention maintains stable inspection performance in complex environments, improves the accuracy and coverage of track anomaly identification, and has high engineering application value.

Description

Technical Field

[0001] This invention relates to the field of rail transit drone inspection technology, and more specifically, to an automatic identification and condition assessment method for rail transit drone inspection. Background Technology

[0002] Rail transit systems are typically widely distributed and involve complex inspection scenarios. Traditional track inspection methods rely heavily on manual patrols or the installation of fixed detection equipment, which not only suffers from limited coverage and slow response times, but also poses significant safety risks to inspection personnel in high-risk areas. Drones, as an emerging intelligent inspection tool, can perform routine track inspection tasks with high flexibility and rapid deployment capabilities without disrupting normal train operations.

[0003] During actual inspections, to maintain the optimal imaging angle and focusing distance for the track structure, drones automatically adjust their attitude based on changes in the observation perspective caused by track smoothness, local gradient differences, and rail twisting. Furthermore, sudden changes in wind pressure and airflow deflections caused by large facilities around the track can also disturb the drone's flight attitude. If a structural anomaly exists at a certain track location, such as track warping, sleeper misalignment, or track bed subsidence, different drones in the same formation will exhibit consistent attitude shifts when flying over that area to maintain stable imaging and navigation reference.

[0004] Therefore, the impact of track geometry deformation or environmental disturbances on UAV attitude exhibits significant spatial coupling characteristics. How to utilize the dynamic attitude characteristics of multiple UAVs during collaborative inspections to automatically identify track conditions and locate defects, thereby constructing a higher-dimensional and more robust inspection and evaluation method, is a pressing issue in the field of intelligent rail transit inspection. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, embodiments of the present invention provide an automatic identification and condition assessment method for unmanned aerial vehicle (UAV) inspection of rail transit to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An automatic identification and condition assessment method for unmanned aerial vehicle (UAV) inspection of rail transit includes the following steps:

[0008] S1. Control the drone swarm to fly along the orbital direction. Each drone synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency to construct the original attitude time matrix of the swarm.

[0009] S2. Perform spatiotemporal alignment on the attitude data of each UAV and use a sliding window mechanism to divide the continuous attitude sequence into equal-length analysis units.

[0010] S3. Perform cooperative mode extraction on all analysis units, calculate the first principal component load of each UAV attitude sequence in the principal component space, and identify cooperative oscillation segments;

[0011] S4. Establish an attitude reference surface based on the cooperative oscillation segment, calculate the deviation between the actual attitude of each UAV and the reference surface, and generate an individual attitude alienation index by combining spatial coordinates.

[0012] S5. Use dynamic thresholds to identify abnormal segments in the individual attitude distortion sequence. When an abnormal segment is identified, trigger track defect localization and record the defect spatial coordinates and distortion intensity.

[0013] S6. Control the UAV cluster to perform formation reorganization based on the defect distribution density, switch to matrix scanning mode in areas with dense defects, and carry out cross-verification flight by adjusting the relative height and spacing between UAVs to output a smoothness evaluation map of rail transit UAV inspection.

[0014] In a preferred embodiment, in step S1, controlling the UAV swarm to fly along the orbital direction, with each UAV synchronously collecting its own pitch angle, roll angle, and spatial coordinate data at a fixed sampling frequency, and constructing the swarm's original attitude time-series matrix specifically includes:

[0015] The drone swarm is controlled to fly along the orbit in a parallel strip formation, with each drone synchronously collecting its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency.

[0016] Construct the original attitude time-series matrix of the cluster, and organize the synchronously collected data into a multi-dimensional array according to the UAV number and time sequence. The array rows correspond to the UAV identifier, the columns correspond to the sampling time points, and each cell stores the pitch angle, roll angle and three-dimensional coordinate values.

[0017] Data gap compensation is performed on the original attitude time series matrix of the cluster, and spatial proximity weighted interpolation is performed on the missing data points based on the spatial coordinates of surrounding UAVs.

[0018] In a preferred embodiment, step S2, which involves performing a spatiotemporal alignment operation on the attitude data of each UAV and using a sliding window mechanism to divide the continuous attitude sequence into equal-length analysis units, specifically includes:

[0019] Based on the timestamps and spatial coordinates in the original attitude time series matrix of the cluster, the attitude data of each UAV is synchronized to a unified time reference point using the time axis normalization method.

[0020] A sliding window mechanism is used to divide the aligned continuous attitude sequence into equal-length analysis units. The window length is dynamically adjusted based on the product of flight speed and sampling frequency, and the sliding step size is set to a fixed proportion of the unit length.

[0021] In a preferred embodiment, step S3 involves performing cooperative mode extraction on all analysis units, calculating the first principal component load of each UAV attitude sequence in the principal component space, and identifying cooperative oscillation segments, specifically including:

[0022] Based on the segmented analysis units, the three-dimensional coordinate values ​​in all analysis units are removed to construct an attitude data matrix that contains only the pitch and roll angles of the UAV.

[0023] Calculate the covariance matrix of the attitude data matrix, perform spectral decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors, and select the eigenvector corresponding to the largest eigenvalue as the first principal component load vector;

[0024] Based on the statistical distribution of the principal component load vector magnitude of the historical data analysis unit, a specific quantile is set as the dynamic threshold.

[0025] The load vector magnitude of the current analysis unit is compared with the dynamic threshold. When the magnitude exceeds the threshold and the duration meets the minimum window length, the analysis unit is marked as a cooperative oscillation segment.

[0026] In a preferred embodiment, step S4, which involves establishing an attitude reference surface based on the cooperative oscillation segment, calculating the deviation between the actual attitude of each UAV and the reference surface, and generating an individual attitude anomaly index by combining spatial coordinates, specifically includes:

[0027] The input data is based on the spatial coordinates of all UAVs within the cooperative oscillation segment and the corresponding first principal component payload vector.

[0028] A continuous attitude reference surface is generated in the spatial coordinate system using a surface fitting algorithm. The function value of each point on the surface represents the theoretical attitude value of that spatial position in the swarm cooperative motion.

[0029] Calculate the residuals of the actual pitch and roll angles of each UAV relative to the theoretical attitude values ​​of the corresponding spatial positions on the attitude reference surface at each time point;

[0030] By combining the residual sequences at all time points within each analysis unit, the root mean square of the sequence is calculated as the amplitude dissimilarity component, and the phase difference between the actual attitude sequence and the reference attitude sequence is calculated as the phase dissimilarity component.

[0031] The amplitude anisotropy component and the phase anisotropy component are combined according to a preset weight to generate an individual attitude anisotropy index.

[0032] In a preferred embodiment, in step S5, anomaly segment identification is performed on the individual attitude anomaly sequence using a dynamic threshold. When an anomaly segment is identified, track defect localization is triggered, and the defect spatial coordinates and anomaly intensity are recorded. Specifically, this includes:

[0033] Calculate the statistical distribution characteristics of the individual posture alienation index sequence of historical analysis units, and set a dynamic threshold based on the statistical distribution characteristics;

[0034] A sliding window scan is performed on the individual attitude alienation sequence within the current analysis unit. When the alienation index of the sampling point corresponding to the three-dimensional coordinate value exceeds the dynamic threshold at the same time, the point is marked as a candidate anomaly.

[0035] If a set number of consecutive sampling points are all marked as candidate outliers, verify whether the alienation index shows a monotonic trend of continuous decay or enhancement.

[0036] If the verification results show a monotonic trend, the orbital defect location is triggered, and the largest individual attitude deviation index value is taken as the deviation intensity of the defect.

[0037] In a preferred embodiment, the track defect location method is to use the center of the spatial coordinates of continuous candidate anomaly points as the defect center coordinates.

[0038] In a preferred embodiment, in step S6, the UAV cluster is controlled to perform formation reorganization based on the defect distribution density, switching to matrix scanning mode in areas with dense defects, and cross-verification flights are implemented by adjusting the relative altitude and spacing between UAVs. The output of the smoothness evaluation map of UAV inspection for rail transit specifically includes:

[0039] The number of defect center coordinates recorded within a unit length interval corresponding to the statistical analysis unit is used as the original distribution density;

[0040] The original distribution density sequence is processed by moving average filtering to obtain a smooth distribution density curve, and the peak intervals in the curve are identified as areas with dense defects.

[0041] The drone swarm can be controlled to switch its flight formation to a multi-layered grid structure in areas with dense defects. The top-level drones maintain their original cruising altitude, while the bottom-level drones adjust their cruising altitude linearly according to the original distribution density of defects.

[0042] Cross-validation flights are performed in a multi-layered grid structure, controlling the top and bottom layer UAVs to fly along orthogonal trajectories and simultaneously collecting individual attitude variation indices as cross-validation data.

[0043] The cross-validation data is weighted and integrated to generate a continuous color gradient map on the track plane, which serves as a smoothness evaluation map for UAV inspection of rail transit.

[0044] The technical effects and advantages of the automatic identification and condition assessment method for unmanned aerial vehicle (UAV) inspection of rail transit in this invention are as follows:

[0045] By fully utilizing the attitude coordination changes of multi-UAV formations during inspection, this method achieves sensor-free identification of track smoothness, geometric anomalies, and potential defects, significantly improving inspection accuracy and stability. Based on pitch angle, roll angle, and spatial coordinate data synchronously collected along the track direction, this method implements a complete processing chain from data acquisition to automatic track defect localization through collaborative modal extraction and attitude heterogeneity construction. It no longer relies entirely on image clarity and lighting conditions, effectively compensating for the susceptibility of traditional visual inspections to environmental interference. Furthermore, this invention enhances data acquisition coverage and directional diversity in defect-dense areas through defect distribution density-driven formation reorganization and multi-layer grid cross-validation mechanisms, significantly improving the reliability of identification results and generating track smoothness assessment maps that can be used to assist maintenance decisions.

[0046] This invention has the advantages of flexible deployment, applicability to various track scenarios, high robustness, and high degree of unmanned and intelligent operation. It can reduce inspection and operation costs, improve the timeliness and comprehensiveness of defect detection, and has important engineering value and promotion prospects for promoting the transformation of rail transit inspection methods towards autonomous and data-driven intelligent operation and maintenance. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of the automatic identification and condition assessment method for unmanned aerial vehicle (UAV) inspection of rail transit according to the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention. Example 1

[0049] Figure 1 The present invention provides an automatic identification and condition assessment method for unmanned aerial vehicle (UAV) inspection of rail transit, which includes the following steps:

[0050] S1. Control the drone swarm to fly along the orbital direction. Each drone synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency to construct the original attitude time matrix of the swarm.

[0051] S2. Perform spatiotemporal alignment on the attitude data of each UAV and use a sliding window mechanism to divide the continuous attitude sequence into equal-length analysis units.

[0052] S3. Perform cooperative mode extraction on all analysis units, calculate the first principal component load of each UAV attitude sequence in the principal component space, and identify cooperative oscillation segments;

[0053] S4. Establish an attitude reference surface based on the cooperative oscillation segment, calculate the deviation between the actual attitude of each UAV and the reference surface, and generate an individual attitude alienation index by combining spatial coordinates.

[0054] S5. Use dynamic thresholds to identify abnormal segments in the individual attitude distortion sequence. When an abnormal segment is identified, trigger track defect localization and record the defect spatial coordinates and distortion intensity.

[0055] S6. Control the UAV cluster to perform formation reorganization based on the defect distribution density, switch to matrix scanning mode in areas with dense defects, and carry out cross-verification flight by adjusting the relative height and spacing between UAVs to output a smoothness evaluation map of rail transit UAV inspection.

[0056] In S1, the drone swarm is controlled to fly along the orbital direction, and each drone synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency to construct the original attitude time sequence matrix of the swarm.

[0057] The drone swarm is controlled to fly in a parallel strip formation along the track. This parallel strip formation refers to a grouping structure where multiple drones are arranged in parallel columns along the track centerline. Each column maintains the same flight speed and a fixed lateral offset, resulting in a regular rectangular array arrangement along the flight path. Specifically, a column of drones close to the track centerline is selected as the baseline column and controlled by navigation to cruise stably along the centerline. One or more columns of drones are arranged on either side of the baseline column, maintaining a uniform lateral spacing between them. This lateral spacing is typically set as an integer multiple of the standard track gauge to ensure uniform coverage of the track structure above and to the sides. A fixed longitudinal spacing is also maintained in the forward and backward directions, placing all drones in a multi-row, multi-column layout. This longitudinal spacing is set to be at least twice the maximum wingspan of the drones to strictly ensure flight safety and avoid wake turbulence interference. During flight, all UAVs maintain attitude and heading stability based on the same flight control command source. Especially in curved sections of the trajectory, automatic heading adjustment is used to maintain the parallel formation and prevent deviation, ensuring that the formation remains constant and the relative positions are consistent throughout the flight. While maintaining the formation, each UAV collects attitude and positioning information at a fixed sampling frequency, outputting pitch and roll angles in real time through the inertial measurement unit and three-dimensional coordinate values ​​through the positioning device.

[0058] When constructing the original attitude time series matrix for the cluster, the pitch angle, roll angle, and 3D coordinate data collected synchronously were sorted by UAV number and written into a multidimensional array. This array is organized using a two-dimensional index structure, with rows representing the identifiers of different UAVs and columns representing continuous sampling time indices. Each cell simultaneously stores the attitude angle and spatial position data corresponding to the sampling time of the UAV in that row and column. To ensure data integrity and traceability, this attitude time series matrix is ​​also correlated with the orbital mileage sequence, ensuring that each column of sampling time points can be uniquely associated with a specific spatial position along the orbital direction. This facilitates feature alignment and defect localization based on orbital position in subsequent data processing stages. During data organization, all data retains its original resolution and is not downsampled, ensuring that subtle spatial disturbances caused by the orbital structure are fully represented in the final analysis results.

[0059] To address potential data recording anomalies during flight, such as sudden loss of short-term attitude data or missing or interrupted positioning values ​​due to unstable satellite tracking signals, spatial proximity-weighted interpolation compensation is performed on missing data points to ensure the continuity and accuracy of the original attitude time-series matrix of the cluster. Before performing the interpolation operation, the positional relationship of the UAV to which the missing data point belongs in the parallel bar formation is first identified. If the UAV to which the missing point belongs is located in a non-edge column, the spatial coordinates of its left and right adjacent columns at the same time point are used for interpolation. If the missing point is located in an edge column, the data of the UAV in the nearest inner column is used for interpolation. The interpolation method is based on the principle of spatial distance weighting, that is, the interpolation contribution ratio is allocated according to the three-dimensional spatial distance between the reference UAV and the UAV to be compensated, so that the reference data with the smallest spatial position difference has a higher weight. When compensating for pitch and roll angles, angle change smoothing processing is performed simultaneously to ensure that there are no sudden jumps in the attitude angle compensation results. For consecutive missing data segments, linear interpolation is first performed using data from preceding and following time points. Then, a secondary spatial consistency correction is performed based on neighboring UAV data to ensure that the compensation results maintain continuity and formation correlation in both the time and spatial domains. After data compensation is completed, abrupt change detection is performed on the repaired segments to remove compensation values ​​that significantly deviate from the normal measurement range. Finally, a complete cluster attitude time series matrix that can be used for subsequent collaborative attitude analysis is formed.

[0060] In step S2, a spatiotemporal alignment operation is performed on the attitude data of each UAV, and a sliding window mechanism is used to divide the continuous attitude sequence into equal-length analysis units.

[0061] Based on the timestamps and spatial coordinates in the original attitude timing matrix of the cluster, the attitude data collected by each UAV is synchronized to a unified time reference point. First, a UAV with stable flight and complete sampling throughout its flight is selected as the time reference source. The moment of the first data acquisition by this UAV is taken as the global zero moment, and all UAVs undergo time alignment processing based on this. Specifically, the original timestamps recorded in the data collected by each UAV are read, and through a unified format conversion, the time values ​​output by the internal clocks of different devices are converted into a unified absolute time series. Based on the sampling time of the reference UAV, the corresponding timestamps of other UAVs are precisely aligned according to the sampling frequency, ensuring that the same sampling sequence number of all UAVs strictly corresponds to the same time node. To avoid the accumulation of time errors caused by minor drifts in the internal clocks of the sampling devices, a bidirectional calibration mechanism is introduced during time synchronization. That is, for records whose sampling time deviates from the reference sampling point, the time difference between the sampling points before and after the phase is compared, and correction is performed according to the magnitude of the deviation using time interpolation or time sliding, ensuring that all UAVs correspond to the same physical time point at the same sampling point number.

[0062] A sliding window mechanism is used to segment the aligned continuous attitude sequence into equal-length analysis units. The length of the sliding window is dynamically adjusted based on the product of flight speed and sampling frequency to ensure that all analysis units cover a consistent length of orbital space. For example, when inspecting a straight section of the orbit, if the UAV's flight speed is several meters per second and the sampling frequency is fixed, the window length is set to the complete integer value corresponding to the product of speed and sampling frequency, ensuring that the window covers a fixed spatial segment on the orbit. When the flight speed is slightly adjusted in areas with orbital slope or curves, the window length is automatically recalculated based on real-time speed parameters to ensure that the orbital space range corresponding to the analysis unit remains uniformly represented. The sliding step size is set to a certain proportion of the window length, for example, half of the window length is selected as the fixed sliding step size, so that there is a data overlap between adjacent analysis units, thereby ensuring that no data is missed during continuous changes in attitude characteristics. When performing this segmentation operation, only the window is translated according to the time series, without changing the internal order of the attitude data. An orbital space index is attached after each window is formed, so that each analysis unit uniquely corresponds to a continuous orbital segment in the flight path.

[0063] In S3, cooperative mode extraction is performed on all analysis units, the first principal component load of each UAV attitude sequence in the principal component space is calculated, and cooperative oscillation segments are identified.

[0064] Based on the sliding window mechanism, the analysis units are segmented, and the three-dimensional spatial coordinate data recorded in each analysis unit are removed, retaining only the pitch and roll angle data that demonstrate the attitude change characteristics of the UAV. To ensure the specificity of the collaborative mode analysis and reduce unnecessary interference of spatial positioning errors on the matrix structure, the pitch and roll angle information of each UAV at each sampling time point within the same analysis unit is arranged into an attitude data vector in a predetermined order. The vectors corresponding to all time points of all UAVs are then superimposed sequentially in time to construct a two-dimensional attitude data matrix. The rows of this matrix correspond to the different UAV identification numbers, and the columns correspond to the consecutive sampling sequence numbers in the analysis unit.

[0065] When calculating the covariance matrix of the attitude data matrix, the mean of all UAV attitude sequences is first calculated based on the total sequence values ​​of the attitude data over the time dimension. The mean is then subtracted from each data point to form a zero-centered data deviation matrix. This allows subsequent covariance calculations to accurately characterize the commonalities and differences in attitude change trends among the UAVs. Based on this, covariance calculation is performed to obtain a symmetric matrix structure describing the statistical relationships between different UAV attitude sequences. Spectral decomposition is then performed on this covariance matrix to extract its eigenvalues. The eigenvalue set of the covariance matrix is ​​calculated and transformed into scale-consistent eigenvectors. The eigenvalues ​​are then sorted by value, and the largest eigenvalue is determined based on the sorting result. Subsequently, the eigenvector corresponding to the largest eigenvalue is extracted. This eigenvector serves as the first principal component loading vector, reflecting the most significant uniform oscillation trend in the attitude changes of all UAVs. To ensure the stability of the spectral decomposition results, this embodiment adopts an iterative convergence strategy when performing feature solving, and normalizes the magnitude of the feature vector after the solution is completed, so that it only reflects the directional attribute and does not carry the magnitude accumulation, thereby ensuring that the first principal component loads extracted by different analysis units have a consistent expression form.

[0066] The dynamic threshold is determined based on the statistical distribution of the principal component load vector amplitudes of historical data analysis units. This embodiment first establishes a load vector amplitude reference library in the non-defect area inspection records or the initial baseline inspection stage, collects the first principal component load vectors of multiple historical analysis units, calculates the absolute value of their amplitudes, and forms a statistical distribution according to their numerical magnitudes. Based on this, a specific quantile value is set as the dynamic threshold. By default, the value corresponding to the upper quartile of the statistical distribution is taken as the threshold interval selection range. Finally, the amplitude corresponding to the specific quantile value adapted to the current inspection environment is determined as the dynamic threshold. The amplitude of the first principal component load vector of the current analysis unit is compared element-wise with this dynamic threshold. When the amplitudes of multiple consecutive sampling time points exceed the dynamic threshold, and the duration of the consecutive exceedance reaches the preset minimum window length requirement, the analysis unit is marked as a cooperative oscillation segment. The minimum window length is set according to the sampling frequency and track length coverage requirements. For example, half the analysis unit time length is taken as a fixed setting value to ensure that the oscillation behavior is a real cooperative behavior rather than an occasional disturbance. After marking, the cooperative oscillation segment number, corresponding time index, and associated spatial location are written into a structured record.

[0067] In S4, an attitude reference surface is established based on the cooperative oscillation segment, the deviation between the actual attitude of each UAV and the reference surface is calculated, and an individual attitude alienation index is generated by combining spatial coordinates.

[0068] Based on the identified cooperative oscillation segments, the spatial coordinates and first principal component load vectors of all UAVs within these segments at each sampling time point are selected as input data to construct an attitude reference surface. Spatial coordinates describe the actual flight positions of the UAVs above the orbit, while the first principal component load vector represents the theoretical value of the coordinated attitude change of the cluster at the corresponding position. First, the spatial coordinate points of all UAVs within the cooperative oscillation segments are sorted according to the longitudinal orbital direction to form a continuous spatial distribution point set, and a three-dimensional discrete grid is established based on the lateral and altitude relationships between adjacent UAVs. Subsequently, the corresponding first principal component load vector value is read at each spatial point and used as the theoretical attitude attribute value at that position. To convert the discrete points into a three-dimensional surface that can be used for continuous interpolation calculations, a spline surface fitting algorithm is used for surface construction. During the fitting process, the surface is divided into multiple spline segments according to the longitudinal position direction. Each segment uses six to ten surrounding data points to perform a local surface fitting operation. By constraining the first and second derivatives of the fitted surface to be continuous at the nodes, the entire surface maintains a smooth transition in both the orbital and lateral directions. In the boundary region of the surface, an extrapolation fitting method is used to extend the surface, ensuring its continuous existence within the spatial area covered by all sampling points, without any voids or breaks. The resulting attitude reference surface has a unified definition throughout the entire cooperative oscillation segment, and the function value at any position represents the theoretical attitude change value at that spatial position under the cooperative motion of the cluster.

[0069] The residuals of the actual pitch and roll angles of each UAV relative to the theoretical values ​​on the attitude reference surface are calculated at each time point. Specifically, at each sampling time point, the UAV's mapped position on the reference surface is first located based on its spatial coordinates, and the surface function value at that position is read as the theoretical attitude reference value. To reduce computational interference caused by differences in attitude data components, the difference between pitch and roll angles is uniformly represented as a signed angle error. For the residual calculation at each time point, the attitude residual sequence is obtained by subtracting the theoretical value from the actual collected value, and the complete residual change trajectory is saved within each analysis unit. To avoid unnecessary impact on the overall results due to instantaneous errors caused by short-term wind disturbances, an amplitude rationality check is added after the residual sequence calculation is completed. That is, if the residual value at a certain time point suddenly deviates from the residual value of the adjacent sampling point by more than the preset error growth limit, the result of smooth interpolation of the adjacent point is used to replace the original residual value, so as to ensure the continuity and consistency of the residual sequence over time.

[0070] After obtaining the residual sequence, the amplitude anisotropy component is calculated by integrating the values ​​of all sampling time points in each analysis unit. The root mean square method is used for calculation, which involves summing the squares of the residuals at all sampling points, dividing by the number of sampling points, and then taking the square root. This value reflects the average intensity of the UAV attitude change deviating from the theoretical attitude of the reference surface within the corresponding analysis unit. Simultaneously, the phase anisotropy component is calculated to measure the temporal difference between the actual attitude sequence and the reference attitude sequence. By comparing the fluctuation patterns of the two sequences in the time dimension, the two sequences are standardized, and the differences in the positions of peaks and troughs are statistically analyzed and quantified into angular phase differences using a unified unit. For the two anisotropy components mentioned above, a weighted combination method is used to generate an individual attitude anisotropy index. The weighting coefficients are specifically set based on the flight speed and turning frequency of the UAV formation, ensuring that the overall index simultaneously includes information on both the degree of attitude deviation and the temporal difference of the deviation.

[0071] In step S5, an abnormal segment is identified by using a dynamic threshold on the individual attitude anomaly sequence. When an abnormal segment is identified, the track defect is located and the spatial coordinates and anomaly intensity of the defect are recorded.

[0072] The statistical distribution characteristics of the historical individual attitude anomaly index sequence of the calculation and analysis unit are used to set a dynamic threshold. Individual attitude anomaly index values ​​calculated by all analysis units are selected from multiple orbital baseline inspection tasks performed during the initial equipment deployment phase to construct a historical anomaly statistical database. All values ​​are arranged in ascending order to form a continuous historical statistical distribution curve. Subsequently, the distribution characteristics of this curve are analyzed to extract key positional features describing its overall distribution pattern, including the changes in the middle, high, and low segments of the distribution. Simultaneously, to ensure the dynamic threshold has practical filtering capabilities, quantile selection principles are set based on the concentration of high anomaly regions in the distribution curve. The upper quartile is defaulted as the anomaly value corresponding to the threshold, serving as the dynamic threshold. After determining the threshold, this value is saved and used as the current reference threshold in all subsequent inspection data, ensuring that the dynamic threshold remains updated and adaptable as historical inspection data accumulates.

[0073] A sliding window scan is performed on the individual attitude anomaly sequence within the current analysis unit. Specifically, at each sampling time point, the anomaly index value is compared with a dynamic threshold. If the anomaly index value exceeds the dynamic threshold simultaneously, the point is marked as a candidate anomaly, and its current spatial coordinate index and sampling time index are appended. Subsequently, a fixed sliding window slides forward point by point along the time series, continuously tracking the continuous occurrence of candidate anomalies. A condition is set for three to five consecutive sampling points being in a candidate anomaly state. The number of consecutive points is determined based on the actual sampling frequency and the local track length coverage requirements, ensuring that only anomalies with continuous deviation characteristics are marked, excluding isolated data points caused by random disturbances. For candidate regions that meet the continuous marking requirement, a trend verification operation is further performed, sorting the anomaly index values ​​of all sampling points within the region and calculating the monotonic trend of the values ​​over time. Continuous increases or decreases are both considered monotonic trends, both indicating a correlation between UAV attitude deviation and orbital geometric anomalies.

[0074] If the verification results show a monotonic trend, the track defect localization operation is triggered. During localization, firstly, the anomaly index values ​​of all sampling points within the continuous candidate region are globally scanned, and the maximum value is selected as the anomaly intensity of the corresponding track defect in that region, characterizing the severity of the track anomaly at the attitude migration level. Subsequently, the geometric center of the three-dimensional spatial coordinates of all candidate anomaly points within the region is calculated. The center coordinates of the region are obtained by averaging the spatial positions of all points, and these center coordinates are taken as the center location of the track defect. This method achieves automatic track defect localization based on attitude migration.

[0075] In step S6, the UAV cluster is controlled to perform formation reorganization based on the defect distribution density. In areas with dense defects, it switches to matrix scanning mode. Cross-verification flight is carried out by adjusting the relative height and spacing between UAVs, and a smoothness evaluation map of rail transit UAV inspection is output.

[0076] Based on the identified track defect center coordinate records, the number of defect center coordinates recorded within a unit length interval corresponding to the statistical analysis unit is used as the original defect distribution density for the corresponding segment. To ensure that this density information has a smooth and continuously interpretable numerical change trend, a moving average filter is used to process the original distribution density sequence. Specifically, a fixed number of adjacent unit length intervals are used as a sliding window. By arithmetically averaging the original densities of each segment within the window, independent abrupt values ​​are filtered out and non-continuous defect signals are suppressed, resulting in a smooth change characteristic of the density sequence. The processed smooth distribution density curve can display the degree of defect distribution concentration in the direction of track length as a continuous curve. In this embodiment, during curve analysis, intervals with significant increases and forming local peaks are identified, and such peak intervals are defined as densely populated track defect regions.

[0077] Within the identified areas of high defect density, the drone swarm is switched from its original parallel strip formation to a multi-layered grid structure to improve the spatial integrity of the data acquisition trajectory. Before the formation switch, the number of layers and horizontal columns of the multi-layered grid structure is determined based on the defect density fluctuation trend. Specifically, the top-level drones maintain their original cruising altitude to ensure comparability with data from the previous inspection phase; the bottom-level drones adjust their cruising altitude according to the changes in smooth distribution density using a linear function. For example, when the density increases to a certain threshold, the altitude of the bottom-level drones decreases accordingly to collect data closer to the orbital plane and enhance data sensitivity. The grid structure layout follows the principle of equal distribution in both the longitudinal and lateral directions to ensure complementary perspectives between the top and bottom layers. After the formation switch is completed, all drones maintain their relative positions without disrupting the stable grid structure and continuously cover the entire area of ​​high defect density at a constant speed.

[0078] In this embodiment, cross-validation flights are performed in a multi-layered grid formation. The top and bottom layer UAVs perform synchronous sampling along mutually orthogonal trajectory directions. For example, the top layer UAV maintains a straight line along the trajectory direction, while the bottom layer UAV performs periodic reciprocating scans along a lateral direction perpendicular to the trajectory direction. This orthogonal sampling structure forms a cross-dataset composed of longitudinal and lateral viewpoint data. During sampling, all UAVs record individual attitude anisotropy index values ​​at the same sampling frequency as before, and simultaneously record spatial coordinates to achieve spatial mapping of attitude changes. After data acquisition, the cross-validation data is weighted and integrated. The weights are determined based on the UAV's position relative to the trajectory center at the time of sampling; for example, the weights of the top and bottom layer data are set to the same ratio to ensure a balanced expression of top-down angle and lateral perception capabilities. Finally, the weighted results are mapped onto the trajectory plane to form a continuous color gradient distribution map, where color changes correspond to the degree to which the overall cluster attitude deviates from the reference surface. This map has a continuous representation on the trajectory plane, fully presenting the trajectory smoothness change structure and facilitating intuitive observation of trajectory status information. Once the map is generated, it is incorporated into the inspection report storage structure as the final evaluation result of the rail transit drone inspection output.

[0079] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0080] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0081] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0082] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0083] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0084] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0085] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0086] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0087] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0088] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit, characterized in that, The method comprises the following steps: S1, control the formation of the UAV cluster to fly along the track direction, and each UAV synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency to construct a cluster original attitude time sequence matrix; S2, perform a space-time alignment operation on the attitude data of each UAV, and divide the continuous attitude sequence into equal-length analysis units using a sliding window mechanism; S3, perform cooperative mode extraction on all analysis units, calculate the first principal component load of the attitude sequence of each UAV in the principal component space, and identify the cooperative oscillation segment; S4, establish an attitude reference surface based on the cooperative oscillation segment, calculate the deviation of the actual attitude of each UAV from the reference surface, and generate an individual attitude alienation index in combination with the spatial coordinates; S5, identify an abnormal segment in the individual attitude alienation sequence using a dynamic threshold, trigger track defect positioning when an abnormal segment is identified, and record the defect spatial coordinates and alienation intensity; S6, control the UAV cluster to perform formation reorganization according to the defect distribution density, switch to a matrix scanning mode in the defect dense area, and implement cross-validation flight by adjusting the relative height and distance between the UAVs, and output a smoothness evaluation atlas of the track traffic UAV patrol.

2. The automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S1, the control of the formation of the UAV cluster to fly along the track direction, and each UAV synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency to construct a cluster original attitude time sequence matrix specifically comprises: controlling the formation of the UAV cluster to fly along the track direction in a parallel strip formation mode, and each UAV synchronously collects its own pitch angle, roll angle and spatial coordinate data at a fixed sampling frequency; constructing a cluster original attitude time sequence matrix, organizing the synchronously collected data into a multi-dimensional array according to the UAV number and time sequence, with the array rows corresponding to the UAV identifiers, the columns corresponding to the sampling time points, and each unit storing the pitch angle, roll angle and three-dimensional coordinate values; performing data gap compensation on the cluster original attitude time sequence matrix, and performing spatial proximity weighted interpolation on the missing data points based on the spatial coordinates of the surrounding UAVs. 3.The automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S2, the space-time alignment operation is performed on the attitude data of each UAV, and the continuous attitude sequence is divided into equal-length analysis units using a sliding window mechanism, specifically comprising: synchronizing the attitude data of each UAV to a unified time reference point using a time axis normalization method based on the time stamp and spatial coordinates in the cluster original attitude time sequence matrix; dividing the aligned continuous attitude sequence into equal-length analysis units using a sliding window mechanism, with the window length dynamically adjusted according to the product of the flight speed and the sampling frequency, and the sliding step set to a fixed proportion of the unit length.

4. The automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S3, the cooperative mode extraction is performed on all analysis units, the first principal component load of the attitude sequence of each UAV in the principal component space is calculated, and the cooperative oscillation segment is identified, specifically comprising: based on the divided analysis units, removing all three-dimensional coordinate value data in the analysis units to construct an attitude data matrix containing only the pitch angle and roll angle of the UAV; calculating the covariance matrix of the attitude data matrix, performing spectral decomposition on the covariance matrix to obtain eigenvalues and eigenvectors, and selecting the eigenvector corresponding to the maximum eigenvalue as the first principal component load vector; Based on the principal component load vector amplitude statistical distribution of the historical data analysis unit, a certain percentile is set as a dynamic threshold value; The load vector amplitude of the current analysis unit is compared with the dynamic threshold value. When the amplitude exceeds the threshold value and the duration meets the minimum window length, the analysis unit is marked as a cooperative oscillation segment.

5. The automatic identification and condition assessment method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S4, a posture reference surface is established based on the cooperative oscillation segment, the deviation of the actual posture of each UAV from the reference surface is calculated, and the individual posture alienation degree index is generated in combination with the spatial coordinates, specifically including: Taking the spatial coordinates and the corresponding first principal component load vector of all UAVs in the cooperative oscillation segment as input data; A continuous posture reference surface is generated in the spatial coordinate system by using a surface fitting algorithm. The function value of each point on the surface represents the theoretical posture value of the space position in the cluster cooperative motion; The actual pitch angle and roll angle of each UAV are calculated point by point with respect to the theoretical posture value of the corresponding spatial position on the posture reference surface; The residual sequences of all time points in each analysis unit are integrated to calculate the root mean square as the amplitude alienation degree component, and the phase difference between the actual posture sequence and the reference posture sequence is calculated as the phase alienation degree component; The amplitude alienation degree component and the phase alienation degree component are combined according to the preset weight to generate the individual posture alienation degree index.

6. The automatic identification and condition assessment method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S5, the individual posture alienation degree sequence is subjected to abnormal segment identification by using a dynamic threshold value. When an abnormal segment is identified, track defect positioning is triggered, and the defect spatial coordinates and alienation degree intensity are recorded, specifically including: The statistical distribution characteristics of the individual posture alienation degree index sequence of the historical analysis unit are calculated, and a dynamic threshold value is set based on the statistical distribution characteristics; The individual posture alienation degree sequence in the current analysis unit is subjected to sliding window scanning. When the alienation degree index of the sampling point corresponding to the three-dimensional coordinate value simultaneously exceeds the dynamic threshold value, the point is marked as a candidate abnormal point; If a continuous number of sampling points are marked as candidate abnormal points, it is verified whether the alienation degree index presents a monotonic change trend of continuous attenuation or enhancement; If the verification result presents a monotonic change trend, track defect positioning is triggered, and the largest individual posture alienation degree index value is taken as the alienation degree intensity of the defect.

7. The automatic identification and condition assessment method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, The track defect positioning mode is to take the center of the continuous candidate abnormal point spatial coordinates as the defect center coordinates. 8.The automatic identification and state evaluation method for unmanned aerial vehicle inspection of rail transit according to claim 1, characterized in that, In the S6, the UAV cluster is controlled to perform formation reorganization according to the defect distribution density, the matrix scanning mode is switched in the defect dense area, the relative height and distance between the UAVs are adjusted to implement cross-verification flight, and the smoothness evaluation atlas of the track transportation UAV patrol is output, specifically including: The number of defect center coordinates recorded in the unit length interval corresponding to the analysis unit is taken as the original distribution density; The original distribution density sequence is subjected to sliding average filtering to obtain a smooth distribution density curve, and the peak interval in the curve is identified as a defect dense area; The flight formation of the UAV cluster is switched to a multi-layer grid structure in the defect dense area. The top layer of UAVs maintains the original cruising altitude, and the bottom layer of UAVs adjusts the cruising altitude linearly according to the defect original distribution density change. Cross-validation flight is performed under the multi-layer grid structure, and the top layer and the bottom layer unmanned aerial vehicles fly along the orthogonal trajectories and synchronously collect individual attitude alienation indicators as cross-validation data; The cross-validation data is weighted and integrated, and a continuous color gradient atlas is generated in the orbital plane as a smoothness evaluation atlas for orbital transportation unmanned aerial vehicle inspection.

Images