Multi-cycle double-track marker data alignment method and system under missed detection and false detection conditions

Abstract

The application discloses a kind of multi-period double-track marker data alignment method and system under missed detection and misdetected condition, obtain the double-track marker data of reference period and the period to be aligned, determine common mileage interval and segment, according to the number density and average interval of each segmented marker, determine dense section or sparse section, configure candidate offset search range, structure consistency weight and local matching neighborhood radius;Based on segmented sequence and double-track collaborative structure features, output candidate offset prediction probability and candidate matching edge confidence, determine initial offset in offset priority area, and solve one-to-one correspondence that satisfies mileage monotonicity constraint under the condition that explicit missed detection penalty and misdetected penalty are set, after double-track consistency check, backtracking recalculation, abnormal rejection and unaligned marker filling, output alignment result.The application can improve the accuracy, consistency, robustness and integrity of multi-period double-track marker data alignment.

Description

Technical Field

[0001] This invention relates to the field of track condition detection and intelligent data processing technology, and in particular to a method and system for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection. Background Technology

[0002] In the processes of track inspection, structural condition monitoring, defect evolution analysis, and maintenance decision-making, it is often necessary to align marker data collected in different inspection cycles to establish the correspondence between the same physical location in different cycles. For dual-track scenarios, marker data not only have sequential characteristics along the mileage direction but also have a coordinated correspondence between the left and right rails. Therefore, accurate alignment of multi-cycle dual-track marker data is the foundation for subsequent change identification, trend assessment, and anomaly localization.

[0003] Existing multi-period marker data alignment methods mostly employ nearest neighbor matching, fixed offset correction, dynamic time warping, global path matching, or proportional correction based on simple statistical relationships. These methods can achieve a certain degree of alignment when acquisition conditions are relatively stable and there are few missed detections and false detections. However, in actual engineering, detection equipment is easily affected by factors such as speed fluctuations, mileage drift, occlusion interference, reflective noise, local structural repetition, and independent errors between the left and right tracks. This leads to problems such as missed detections, false detections, mismatches in dense sections, missed matching in sparse sections, and inconsistencies between the two tracks, thereby reducing the reliability of the final alignment results.

[0004] Furthermore, existing methods typically employ a uniform search range, uniform structural constraint strength, and uniform local matching neighborhood, failing to adaptively configure for differences in marker distribution within different segments. Simultaneously, existing methods often treat offset estimation and local matching separately, lacking a mechanism to link offset priority regions, candidate matching edge confidence, and dual-track collaborative structural features into a constraint mechanism, leading to the introduction of numerous invalid matching edges under conditions of missed or false detections. Moreover, while some methods can obtain locally feasible alignment results for a single track, they fail to perform closed-loop verification and backtracking recalculation of the consistency between the left and right tracks, thus making it difficult to simultaneously consider the overall consistency of the dual tracks, mileage monotonicity, and result completeness.

[0005] Therefore, there is an urgent need for a data alignment method and system that can take into account segment density differences, dual-track collaborative features, offset-first search, candidate edge confidence constraints, explicit modeling of missed and false detection costs, dual-track consistency backoff, and anomaly removal and completion under the conditions of missed and false detections, so as to improve the accuracy, robustness, consistency and completeness of multi-cycle dual-track marker data alignment. Summary of the Invention

[0006] To address the problems in existing technologies, such as excessively large search range under conditions of missed detection and false detection, numerous invalid candidate matching edges, easy mismatch in dense segments, easy missed matching in sparse segments, poor consistency between the two tracks, and insufficient completeness of results, this invention provides a method and system for aligning multi-period dual-track marker data under conditions of missed detection and false detection. This method improves the accuracy, stability, consistency, and completeness of multi-period dual-track marker data alignment by employing segmented density adaptive parameter configuration, dual constraints of offset priority region and candidate matching edge confidence, explicit setting of missed detection penalties and false detection penalties through monotonic one-to-one correspondence solving, dual-track consistency verification and backtracking recalculation, result fusion, anomaly removal, and misaligned marker completion.

[0007] To achieve the above objectives, this invention provides a method for aligning multi-cycle dual-track marker data under conditions of missed and false detections, comprising: determining a common mileage interval and dividing it into segments for dual-track marker data of a reference period and a period to be aligned; extracting the reference sequence and the sequence to be aligned for each segment; calculating the marker number density and average interval for each segment; and determining the segment type, candidate offset search range, structural consistency weight, and local matching neighborhood radius accordingly; inputting the reference sequence, the sequence to be aligned, the segment type, and the structural consistency weight for each segment into an inference model; extracting dual-track cooperative structural features; outputting the predicted probability of candidate offsets and the basic probability of candidate matching edges; and combining the dual-track cooperative structural features and the structural consistency weight to obtain the confidence level of candidate matching edges and determine the offset priority region; in the... Within the intersection of the offset priority region and the candidate offset search range, candidate offsets are scored and sorted according to a unified scoring function to determine the initial offset. A set of local candidate matching edges is then constructed based on the initial offset, the local matching neighborhood radius, and the confidence level of the candidate matching edges. With explicit settings for missed detection penalties and false detection penalties, a one-to-one correspondence satisfying the mileage monotonic constraint is solved for the set of local candidate matching edges to obtain the segment alignment result. Dual-track consistency verification and monotonic violation verification are performed on the segment alignment result. If the verification fails, the process reverts to the candidate offset with the second-best score and resolves the local matching. The segment alignment results that pass the verification are merged, and a deviation model is established based on the merged alignment points. Abnormal alignment points are removed, and misaligned markers are added, resulting in the final alignment result.

[0008] To achieve the above objectives, the present invention also provides a multi-cycle dual-track marker data alignment system under conditions of missed detection and false detection, comprising: a segmentation and density determination module, used to determine the common mileage interval and segment the dual-track marker data of the reference period and the period to be aligned, extract the reference sequence and the sequence to be aligned for each segment, calculate the marker number density and average interval of each segment, and output the segment type, candidate offset search range, structural consistency weight and local matching neighborhood radius; an offset and edge inference module, used to receive the reference sequence, the sequence to be aligned, the segment type and the structural consistency weight of each segment, extract the dual-track cooperative structure features, output the candidate offset prediction probability and the candidate matching edge basic probability, and combine the dual-track cooperative structure features and the structural consistency weight to obtain the candidate matching edge confidence and determine the offset priority region; and an offset search and monotonic matching module, used to perform offset search and monotonic matching on the offset... Within the intersection of the priority region and the candidate offset search range, an initial offset is determined based on a unified scoring function. A set of local candidate matching edges is constructed based on the initial offset, the local matching neighborhood radius, and the confidence of the candidate matching edges. The candidate offset scoring sequence is retained. Under the condition of explicitly setting missed detection penalties and false detection penalties, a one-to-one correspondence satisfying the mileage monotonic constraint is solved, and the segment alignment result is output. A consistency backoff module receives the segment alignment result and the candidate offset scoring sequence, performs dual-track consistency verification and monotonic violation verification. If the verification fails, it returns the candidate offset with the second-best score to the offset search and monotonic matching module to resolve the local matching. A fusion and completion module fuses the segment alignment results that have passed the verification, establishes deviation samples and a deviation model based on the fused alignment points, removes abnormal alignment points, completes the unaligned markers, and outputs the final alignment result.

[0009] Compared with the prior art, the present invention has at least the following beneficial effects:

[0010] Firstly, the present invention determines the segment type based on the number density and average interval of markers in each segment, and configures the candidate offset search range, structural consistency weight and local matching neighborhood radius respectively. This enables the use of differentiated alignment strategies for dense and sparse segments, reducing the risk of neighbor mismatch in dense segments, while also taking into account the matchability of sparse segments under the condition of missed detection.

[0011] Secondly, this invention inputs the segmented reference sequence, the sequence to be aligned, and the dual-track collaborative structure features into the inference model, and determines the offset priority region based on the predicted probability of the candidate offset, and obtains the confidence of the candidate matching edge based on the fusion of the comprehensive distance and the basic probability of the candidate matching edge, thereby achieving dual compression of the offset search space and the size of the candidate matching edge, and reducing invalid matching interference.

[0012] Third, this invention comprehensively considers the matching rate term, relative interval consistency term, dual-track consistency penalty term, and mileage monotonicity penalty term through a unified scoring function, so that the determined initial offset not only has high local matching feasibility, but also takes into account the overall consistency and monotonicity constraints of the dual tracks, thereby improving the quality of offset estimation.

[0013] Fourth, this invention explicitly sets penalty terms for missed detection and false detection in the local matching stage, and solves the one-to-one correspondence that satisfies the mileage monotonic constraint under the confidence constraint of candidate matching edges. This enables the suppression of mismatch problems caused by pseudo-matching, reverse matching and local structural repetition under complex missed detection and false detection conditions, thereby improving the robustness of the segment alignment results.

[0014] Fifth, this invention introduces dual-track consistency verification and monotonic violation verification into the closed-loop backoff process. When the verification fails, it backoffs to the candidate offset with the second-best score and resolves the local matching. This can suppress independent mismatch between the left and right tracks and local optimal distortion, and improve the consistency and stability of the final alignment result.

[0015] Sixth, after integrating the segmented results, this invention further establishes a deviation model, performs abnormal alignment point removal and misaligned marker completion, thereby improving the continuity, physical rationality and completeness of the final output results, which is more conducive to subsequent change analysis, disease assessment and maintenance decisions.

[0016] Seventh, the method and system provided by this invention are set up around the same technical concept, and each module of the system corresponds to one of the method steps, which has good singularity, implementability and convenient engineering deployment. Attached Figure Description

[0017] Figure 1 This is a flowchart of the multi-cycle dual-track marker data alignment method under conditions of missed detection and false detection provided in the first embodiment of the present invention.

[0018] Figure 2 This is a block diagram of the multi-cycle dual-track marker data alignment system under conditions of missed detection and false detection provided in the second embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to preferred embodiments. It should be noted that in this specific embodiment, the same letters represent the same meaning, and different letters represent different meanings; the same terms represent the same technical features, and different terms represent different technical features. Unless otherwise stated, the superscript "..." " indicates the left track, superscript " " indicates the right track, superscript " "Indicates truth value, hat number" "Indicates the predicted value, asterisk " "Indicates the optimal result, indicated by a horizontal line" "Indicates average value, apostrophe " " indicates the set or result after rollback expansion; different superscripts, subscripts or additional markers of the same letter indicate different parameters, and their meanings are explained in the corresponding positions.

[0020] This invention is applicable to the automatic alignment of dual-track marker data between a reference period and a period to be aligned. The dual-track marker data includes at least the marker mileage position, track identifier, marker type, and acquisition period identifier. For ease of description, when the left and right tracks are not distinguished, the reference period marker sequence is denoted as... The sequence of periodic markers to be aligned is denoted as ;in, Indicates the reference period number Mileage location of each marker Indicates the period to be aligned. Mileage location of each marker Indicates the total number of markers in the baseline period. This indicates the total number of period markers to be aligned. When distinguishing between the left and right rails, they are respectively denoted as... , , and .

[0021] In practical implementation, the method of the present invention can be executed by a server, industrial control computer, track detection data processing terminal or other devices with computing capabilities, or by a processor calling program instructions in memory.

[0022] First Embodiment

[0023] Figure 1 This is a flowchart of the multi-cycle dual-track marker data alignment method under conditions of missed detection and false detection provided in the first embodiment of the present invention, as follows: Figure 1 As shown, the method includes the following steps.

[0024] Step S1: Determine the common mileage interval and divide it into segments based on the dual-track marker data of the reference period and the period to be aligned.

[0025] First, acquire dual-track marker data for the baseline period and the period to be aligned. Preferably, before proceeding to subsequent alignment, the original data is first sorted by track type, duplicate points are merged, obvious outliers are removed, and marker types are uniformly coded to ensure the consistency of input data and engineering usability.

[0026] Then, a common mileage interval is determined between the two periods of data, and subsequent alignment is performed only within this common mileage interval. The common mileage interval... It can be represented as: , in, Indicates the starting mileage of the common mileage range. This indicates the ending mileage of the common mileage range. Preferably, Take the larger of the starting mileages of the two cycles. Take the smaller value between the two cycle termination mileages.

[0027] After determining the common mileage interval, the common mileage interval is divided into multiple segments, the first... Each segment Represented as: , in, Indicates the first The starting mileage of each segment, Indicates the first The final mileage of each segment, This indicates the total number of segments. Preferably, an overlap length is set between adjacent segments. This is to reduce the impact of segment boundaries on the final result.

[0028] In this step, from the common mileage range Extract the reference period left orbit subsequence corresponding to each segment. Reference period right-orbit subsequence The left-rail subsequence of the period to be aligned and the right-rail subsequence of the period to be aligned The four types of segmented subsequences are then used as data inputs for subsequent steps. In other words, the segmentation results output in step S1 serve as both the input for segment density statistics in step S2 and the basic data for the dual-track collaborative structural feature construction and inference model in step S3.

[0029] The purpose and benefits of this step are as follows: by first defining the common mileage interval and then performing segmentation processing, the global alignment problem of a long mileage interval can be decomposed into multiple segmented alignment problems with more stable local structures. This reduces the interference of global cumulative error on local matching, and by setting the overlap length... Improve the continuity and stability when merging adjacent segment results.

[0030] Step S2: Extract the baseline sequence and the sequence to be aligned for each segment, calculate the marker number density and average interval for each segment, and determine the segment type, candidate offset search range, structural consistency weight, and local matching neighborhood radius accordingly.

[0031] For each segment A statistical sequence is formed based on the position of the reference marker within the segment. And sorted in ascending order of mileage, among which, Indicates the first The first segment sorted by mileage in ascending order within the segment. The location of a reference marker. Indicates the first The number of reference markers within a segment. Preferably, the positions of the reference markers can be obtained by sorting and merging the left and right rail markers of the reference cycle within the segment according to rail type.

[0032] No. Average interval of each segment With number density They respectively satisfy: , in, Indicates the first The average interval of each segment Indicates the first Number density of segments Indicates the first The length of each segment interval; , They represent the first The first segment sorted by mileage in ascending order within the segment. The, the Location of each landmark; Indicates the first The starting mileage of each segment, Indicates the first The final mileage of each segment.

[0033] when and At that time, the first The first segment is determined to be a dense segment; otherwise, the second segment is determined to be a dense segment. Each segment is determined to be a sparse segment. Among them, Indicates the number density threshold. This represents the average interval threshold.

[0034] After obtaining the segment type, configure the candidate offset search range respectively. Structural consistency weight and local matching neighborhood radius Preferably, the dense segment adopts: , The sparse segment adopts: , And satisfy: , in, Indicates the first The candidate offset search range for each segment Indicates the first Structural consistency weights for each segment Indicates the first The local matching neighborhood radius of each segment. The candidate offset search range half-width represents the dense segment. Indicates the structural consistency weight of dense segments; Represents the local matching neighborhood radius of the dense segment; The candidate offset search range half-width represents the sparse segment; Indicates the structural consistency weight of sparse segments; This represents the radius of the local matching neighborhood of the sparse segment.

[0035] In this step, the segmented subsequences output in step S1 are used for statistical analysis. and From this, we obtain , and The data will be sent to steps S3, S4, and S5 respectively. Used to adjust the weight of structural difference terms in subsequent dual-track collaborative structural features. Used to limit the allowed range of offset search. Used to construct a set of local candidate matching edges.

[0036] The role and benefits of this step are as follows: by dividing dense and sparse segments according to the number density and average interval of markers in different segments, and by using differentiated parameter configuration, the probability of nearest neighbor mismatch can be reduced in dense segments, and the matching ability under the condition of missed detection can be improved in sparse segments, thereby providing adaptive prior constraints for subsequent offset estimation and local matching.

[0037] Step S3: Input the baseline sequence, the sequence to be aligned, the segment type, and the structural consistency weights of each segment into the inference model, extract the dual-track collaborative structure features, output the predicted probability of candidate offsets and the basic probability of candidate matching edges, and combine the dual-track collaborative structure features and structural consistency weights to obtain the confidence level of candidate matching edges, and determine the offset priority region.

[0038] This process includes three stages: dual-track collaborative structural feature construction, inference model training, and online inference.

[0039] First, construct the dual-track collaborative structure features. For the baseline period... Using markers, construct local interval vectors along the same track. For the period to be aligned Using markers, construct local interval vectors along the same track. Preferably, the local interval vector along the same track consists of the mileage intervals between the current marker and several adjacent markers along the same track before and after it. Simultaneously, the coordination interval between different tracks is calculated. and ,in, Indicates the period in the reference period and the first The distance difference between adjacent markers on different tracks corresponding to each marker. Indicates the period to be aligned with the first Each marker corresponds to the mileage difference between adjacent markers on different tracks.

[0040] Based on the above characteristics, the benchmark period is... The first marker and the period to be aligned Overall distance between the landmarks satisfy: , in, Indicates the reference period number Mileage location of each marker Indicates the period to be aligned. Mileage location of each marker Indicates the first period of the reference cycle Locally spaced vectors on the same track are constructed centered on each marker. Indicates the period to be aligned Locally spaced vectors on the same track are constructed centered on each marker. Describes the norm 1. Indicates the period in the reference period and the first The distance difference between adjacent markers on different tracks corresponding to each marker. Indicates the period to be aligned with the first The distance difference between adjacent markers on different tracks corresponding to each marker. , , This represents the weight coefficient for the corresponding distance term. Preferably, the structural consistency weight output in step S2... Used for adjustment and The segmented values ​​are determined by increasing the weight of structural difference terms in dense segments and decreasing the weight of structural difference terms in sparse segments, thereby explicitly coupling segmentation type information into the comprehensive distance. The calculation process.

[0041] Secondly, the inference model is trained. Preferably, the inference model includes a shared feature extraction unit, an offset prediction head, and a matching edge prediction head. The shared feature extraction unit receives the baseline sequence of the current segment, the sequence to be aligned, and the dual-track collaborative structure features. The offset prediction head outputs the first... Prediction score corresponding to each candidate offset Matching edge prediction head output reference period The first marker and the period to be aligned Basic probability of candidate matching edges between each marker .

[0042] During the training phase, a training sample set is first constructed. For the first... The training segments are defined by manually labeled or historical high-confidence samples. The set of truth alignment relations for each training segment is as follows: Then the truth offset of that segment It can be represented as: , in, This indicates median operations.

[0043] For the candidate offset set The truth offset can be set. Convert to the first The probability of a true soft label for each candidate offset : , in, This indicates the offset soft label smoothing parameter. Indicates the first Candidate offsets Indicates the first Candidate offsets Indicates the total number of candidate offsets. and All are candidate offset indices; and All belong to the candidate offset set .

[0044] For candidate matching edge labels, define: , in, Indicates the reference period number The first marker and the period to be aligned The truth value matching edge labels between the markers.

[0045] When training the inference model, it is preferable to use a joint training method combining cross-entropy loss and analytical consistency loss. (Offset prediction loss) Matching edge prediction loss Analyzing consistency loss and total losses They respectively satisfy: , , in, The output of the inference model represents the first... Predicted probability of each candidate offset This represents the baseline period of the inference model output. The first marker and the period to be aligned The basic probability of candidate matching edges between each marker. Indicates the period from the reference period The first marker and the period to be aligned Overall distance between the landmarks Constructed analytical soft alignment probability, Indicates the reference period number The first marker and the period to be aligned The overall distance between the landmarks This represents the offset prediction loss. This represents the loss for edge prediction. This represents the parsing consistency loss. Indicates the total loss. , , This represents the weighting coefficient of the corresponding loss term.

[0046] Preferably, during training, the adaptive moment estimator (Adam) optimizer is used to update the inference model parameters, and the minimum validation set loss or no further decrease for several consecutive rounds is used as the stopping condition. After training is completed, the model parameters are frozen, and the online inference phase begins.

[0047] During the online inference phase, the inference model outputs the first... Predicted probability of each candidate offset And thereby determine the offset priority region. : , in, Indicates the inference model for the first The prediction score output by each candidate offset Indicates the inference model for the first The prediction score of each candidate offset output Indicates the offset temperature parameter. Indicates the total number of candidate offsets. Indicates the first Candidate offsets Indicates the threshold for the offset priority region. and Both represent candidate offset indices.

[0048] At the same time, based on the comprehensive distance Constructing analytical soft alignment probabilities and compared with the baseline period output of the inference model. The first marker and the period to be aligned Basic probability of candidate matching edges between each marker Fusion, to obtain the base period number The first marker and the period to be aligned Candidate matching edge confidence among markers : , in, Indicates the alignment temperature parameter. This indicates the number of candidate markers for the current segment to be aligned. Represents the fusion coefficient. Indicates the reference period number The first marker and the period to be aligned The overall distance between the landmarks Indicates the index of the benchmark period marker. Indicates the index of the periodic marker to be aligned. This represents the index of the candidate markers for the period to be aligned, and .

[0049] In this step, the segmented subsequence output in step S1 and the output in step S2 are... Joint participation in comprehensive distance Construction; inference model output Used to form offset priority region Output With analytical soft alignment probability The confidence of candidate matching edges is formed by fusion. The This will be used as the priority region for candidate offset search in step S4. This will serve as the core input for constructing local candidate matching edges and calculating matching costs in steps S4 and S5.

[0050] This step inputs segmented sequence information and dual-track collaborative structure features into the inference model, and fuses the analytical soft alignment probability with the model output probability. This allows for the simultaneous acquisition of the offset priority region and the confidence of candidate matching edges, achieving dual compression of the offset search space and the size of candidate matching edges, and reducing the perturbations introduced by invalid matching edges.

[0051] Step S4: Within the intersection of the offset priority region and the candidate offset search range, the candidate offsets are scored and sorted according to a unified scoring function to determine the initial offset. A set of local candidate matching edges is then constructed based on the initial offset, the local matching neighborhood radius, and the confidence of the candidate matching edges.

[0052] In obtaining the offset priority region After that, only Perform a candidate offset search within the range. Preferably, a combination of coarse and fine search is used, first in... A coarse search is performed first, and then a fine search is performed on the high-scoring candidate offset neighborhood obtained from the coarse search to improve search efficiency.

[0053] No. Each segment at the candidate offset The uniform rating value Passing the exam Initial offset of each segment They respectively satisfy: , , in, This represents the candidate offset value taken within the intersection of the offset priority region and the candidate offset search range. This indicates the matching rate item. Indicates relative interval consistency terms. This indicates a dual-track consistency penalty. This indicates a penalty term for the monotonicity of mileage. , , and This represents the weight coefficient of the corresponding scoring item.

[0054] Specifically, the matching rate item Based on the current candidate offset After being applied to the segmented reference sequence and the sequence to be aligned, statistical results are obtained that satisfy the neighborhood constraint and have high confidence in candidate matching edges. The candidate matching edge ratio is obtained; the relative interval consistency term Based on the current candidate offset The degree of consistency between the adjacent mileage intervals of the reference sequence and the adjacent mileage intervals of the sequence to be aligned is obtained; the dual-track consistency penalty term By left and right rails at the current candidate offset The offset statistical difference is obtained below; the mileage monotonicity penalty term is obtained. Based on the current candidate offset The possible reverse or cross relationships that may form in the next candidate path are obtained.

[0055] Determine the initial offset Then, it can be used Local matching neighborhood radius and candidate matching edge confidence Construct a set of local candidate matching edges to provide input for solving the local monotonic matching problem in step S5.

[0056] In this step, the output of step S2 is and The output of step S3 and Together determine the initial offset And the range of the subsequent local candidate matching edge set. In other words, the candidate offset prediction probability determines "which offset regions to search first," and the local matching neighborhood radius. The confidence level of candidate matching edges determines "which candidate edges are allowed to enter the matching near this offset". Determine which candidate edges are more likely to become valid matching edges.

[0057] This step avoids blindly searching the entire range by determining the initial offset within the offset-priority region and combining it with a unified scoring function. At the same time, by using the initial offset, neighborhood constraints, and candidate matching edge confidence together to construct a local candidate matching edge set, a high-quality candidate space is provided for subsequent local one-to-one matching solutions.

[0058] Step S5: Under the condition of explicitly setting the penalty for missed detection and the penalty for false detection, solve the one-to-one correspondence that satisfies the mileage monotonic constraint for the local candidate matching edge set to obtain the segment alignment result.

[0059] No. Set of local candidate matching edges for each segment The benchmark period The first marker and the period to be aligned Matching cost between markers and before the benchmark period Each marker and the period before alignment The cumulative minimum cost when matching each marker They respectively satisfy:

[0060] , in, This represents the threshold for filtering candidate matching edges. , , and This represents the weight coefficient of the corresponding cost item. This represents a missed detection indicator variable, which takes the value 1 when a missed detection occurs and 0 otherwise. This represents a false positive indicator variable, taking the value 1 when a false positive occurs and 0 otherwise. This indicates the penalty for missed inspection. Indicates penalties for false positives. This indicates the period before the base period. Each marker and the period before alignment The minimum cumulative cost when matching each marker. This indicates the period before the base period. Each marker and the period before alignment The minimum cumulative cost when matching each marker. This indicates the period before the base period. Each marker and the period before alignment The cumulative minimum cost when matching each marker; by backtracking the optimal transfer path corresponding to the cumulative minimum cost, a one-to-one correspondence satisfying the mileage monotonic constraint is obtained.

[0061] Preferably, the initial conditions for dynamic programming are set as follows: , in, This represents the cumulative minimum cost among empty sequences. This indicates that only the period before the baseline period is retained. The cumulative minimum cost when there are 1 marker. This indicates that only the period before alignment is retained. The minimum cumulative cost when there are 1 marker.

[0062] During the solution process, only edges belonging to the local candidate matching edge set are considered. Candidate edges are transitioned using matching; for edges that do not belong to... The marker pair does not enter the matching transition path. This is achieved by minimizing the cumulative cost. By backtracking, a one-to-one correspondence that satisfies the mileage monotonic constraint can be obtained, and this one-to-one correspondence can be recorded as the segment alignment result of the current segment.

[0063] In this step, the initial offset output in step S4 The local matching neighborhood radius output in step S2 The confidence level of the candidate matching edge output in step S3 Together they determine the set of local candidate matching edges ; and penalties for missed inspections Penalties for false positives Furthermore, it determines the cost of "skipping the baseline marker" or "skipping the marker to be aligned" when solving the dynamic programming problem. Therefore, the initial offset constraint, candidate edge confidence constraint, and missed detection / false detection cost modeling are uniformly coupled in this step.

[0064] This step solves for the one-to-one correspondence that satisfies the mileage monotonic constraint by explicitly setting penalty terms for missed detections and false detections. This allows missed detections and false detections to be uniformly incorporated into the optimizable framework. Furthermore, by using the confidence of candidate matching edges to suppress low-confidence candidate edges, the probability of false matching, reverse matching, and mismatch of locally repetitive structures is reduced.

[0065] Step S6: Perform dual-track consistency check and monotonic violation check on the segmented alignment result. If the check fails, backtrack to the candidate offset with the second-best score and resolve the local matching.

[0066] After obtaining the segment alignment results, perform dual-track consistency check and monotonic violation check on them.

[0067] No. Dual-track consistency evaluation quantity for each segment and the expanded candidate matching edge set during backtracking and recalculation They respectively satisfy: , , Among them, superscript Indicates the left rail, superscript Indicates the right track. Indicates the first Offset estimation results for each segment of the left rail. Indicates the first Offset estimation results for each segment of the right rail , They represent the first The matching rate of each segment's left and right rails. , They represent the first The relative spacing consistency of the left and right rails of each segment. , , This represents the weighting coefficient of the corresponding difference item. This indicates the dual-track consistency threshold. Indicates the first The second-best candidate offset for each segment's score. Indicates the first The back-off expansion radius of each segment.

[0068] Preferably, and The predicted probabilities of the candidate offsets for the left and right rails are obtained by weighted averaging or by the offset corresponding to the highest probability. By a unified scoring function The second highest score among all candidate offsets is selected.

[0069] when If the monotonicity violation check fails, it indicates that the current segment alignment result has issues with inconsistency between the left and right tracks or insufficient path monotonicity. In this case, the candidate offset with the second-best score is used. and expanding the candidate matching edge set Re-execute the local matching solution in step S5. Preferably, a maximum number of backtracking steps can be set. If the verification still fails after reaching the maximum number of backtracking steps, the segment is marked as a segment to be reviewed.

[0070] In this step, the segment alignment result output in step S5 is used to calculate... , , , , and This leads to the formation of a dual-track consistency evaluation quantity. If the verification fails, the second-best candidate offset will be provided by the ranking result of the unified scoring function in step S4. Combined with step S2 Construct an expanded candidate matching edge set Then, return to step S5 to perform local matching and solve, forming a closed loop of "matching - verification - rollback and recalculation".

[0071] This step effectively suppresses independent mismatch between the left and right tracks and local optimum distortion by adding dual-track consistency checks and monotonic violation checks to the segmented alignment results, and reverts to the candidate offset with the second-best score to solve again when the checks fail. This improves the consistency, stability and reproducibility of the alignment results.

[0072] Step S7: Merge the alignment results of each segment that have passed the verification, establish a deviation model based on the merged alignment points, remove abnormal alignment points and fill in the misaligned markers, and output the final alignment result.

[0073] For each segment alignment result verified in step S6, segment fusion is first performed. Preferably, for the alignment results of overlapping areas of adjacent segments, a strategy of "prioritizing higher unified scores and smaller regression residuals" is adopted to remove duplicates and retain them, thereby forming a fused global alignment point set.

[0074] After obtaining the fused global alignment point set, deviation samples are established and regression fitting is performed. Deviation sample of alignment points Deviation fitting value Regression residuals And the theoretical position of the misaligned marker in the alignment period. They respectively satisfy: , in, Indicates the first The reference cycle marker mileage position corresponding to each alignment point Indicates the first The mileage position of the period marker to be aligned corresponds to each alignment point. Indicates the first Mileage deviation at each alignment point Indicates the first The deviation fit value of each alignment point Indicates the regression slope. Indicates the regression intercept. Indicates the first The regression residuals of the alignment points This indicates the theoretical position of the misaligned marker within the alignment period. This represents the corresponding deviation prediction value given by the deviation model.

[0075] Preferably, the deviation model is trained using the least squares method, with the goal of minimizing the sum of squared residuals at all fused alignment points. This can be expressed as: , in, This represents the objective function for fitting the bias model. This indicates the number of global alignment points after merging.

[0076] when When this happens, the corresponding alignment point is identified as an abnormal alignment point and removed. This represents the residual threshold. For unaligned markers, their theoretical positions in the alignment period are predicted based on the bias model. and with Centered on the local matching neighborhood radius of the corresponding segment It searches for unmatched candidate markers within its preset extended range; if there are candidate markers that meet the requirements of consistent track type, consistent marker type, and distance threshold, it establishes a matching relationship; if there are no candidate markers that meet the conditions, it marks them as points to be reviewed.

[0077] In this step, the segment alignment result that passes the verification output in step S6 serves as the source of the deviation sample. The fused alignment points then train the deviation model, which is used to identify abnormal alignment points and predict the theoretical position of misaligned markers, thereby realizing a closed-loop post-processing chain of "fusion result - deviation modeling - anomaly removal - alignment recovery".

[0078] This step establishes a deviation model on the fused global alignment results, which can further eliminate physically unreasonable abnormal alignment points and provide theoretical position predictions for misaligned markers, thereby improving the continuity, physical rationality and completeness of the final output results.

[0079] Preferably, the final alignment result output includes at least: the correspondence between the baseline period marker and the period marker to be aligned, the segment to which the correspondence belongs, the segment unified score, the dual-track consistency verification result, the regression residual, and the completion status indicator.

[0080] Second Embodiment

[0081] The second embodiment of the present invention only describes the contents that are different from those of the first embodiment; the contents that are the same will not be described again.

[0082] Figure 2 This is a block diagram of a multi-cycle dual-track marker data alignment system under conditions of missed detection and false detection provided in the second embodiment of the present invention. The system can be deployed in a server, industrial control computer, track detection data processing terminal, or other devices with data processing capabilities. The system includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it implements the steps in the above method embodiments.

[0083] like Figure 2As shown, the system includes a segmentation and density determination module, used to determine the common mileage interval and segment the dual-track marker data of the reference period and the period to be aligned, extract the reference sequence and the sequence to be aligned for each segment, calculate the marker number density and average interval of each segment, and output the segment type, candidate offset search range, structural consistency weight, and local matching neighborhood radius; an offset and edge inference module, used to receive the reference sequence, the sequence to be aligned, the segment type, and the structural consistency weight of each segment, extract the dual-track cooperative structure features, output the candidate offset prediction probability and the candidate matching edge base probability, and combine the dual-track cooperative structure features and structural consistency weight to obtain the candidate matching edge confidence and determine the offset priority region; and an offset search and monotonic matching module, used to find the intersection of the offset priority region and the candidate offset search range. The system determines the initial offset based on a unified scoring function, and constructs a set of local candidate matching edges based on the initial offset, the local matching neighborhood radius, and the confidence of candidate matching edges. It retains the candidate offset scoring sequence, and solves for the one-to-one correspondence satisfying the mileage monotonic constraint under the condition of explicitly setting missed detection penalties and false detection penalties, outputting the segment alignment result. The consistency backoff module receives the segment alignment result and the candidate offset scoring sequence, performs dual-track consistency verification and monotonic violation verification, and returns the second-best scoring candidate offset to the offset search and monotonic matching module to resolve the local matching. The fusion and completion module fuses the segment alignment results that have passed the verification, establishes deviation samples and deviation models based on the fused alignment points, removes abnormal alignment points and completes the unaligned markers, and outputs the final alignment result.

[0084] The above modules preferably collaborate according to the following data flow: the segmentation and density determination module outputs segmentation results, segmentation types, and parameter configuration results to the offset and edge inference module and the offset search and monotonic matching module; the offset and edge inference module outputs the offset priority region and candidate matching edge confidence to the offset search and monotonic matching module; the offset search and monotonic matching module outputs the segment alignment results to the consistency backoff module; the consistency backoff module outputs the verified segmentation results to the fusion completion module, or issues a backoff recalculation instruction to the offset search and monotonic matching module when the verification fails; the fusion completion module outputs the final alignment results.

[0085] This system can achieve highly robust and consistent automatic alignment of multi-cycle dual-track marker data under complex conditions such as missed detections, false detections, locally densely similar structures, and independent errors between the left and right tracks.

[0086] The present invention has been further described in detail above with reference to preferred embodiments. Those skilled in the art should understand that, without departing from the technical concept of the present invention, equivalent substitutions or modifications can be made to the common interval partitioning method, density determination threshold, specific network structure of the inference model, specific form of the loss function, specific calculation method of each item in the unified scoring function, local matching solution strategy, dual-track consistency evaluation method, and specific fitting method of the deviation model. All equivalent substitutions, simple transformations, or reasonable combinations made using the technical concept disclosed in the present invention specification and claims should fall within the protection scope defined by the claims of the present invention.

[0087] experiment

[0088] To further demonstrate the improvements in accuracy, robustness, consistency, and completeness brought about by this invention through segmented density adaptive parameter configuration, offset priority region constraint, candidate matching edge confidence screening, explicit modeling of missed detection penalties and false detection penalties, dual-track consistency verification and callback, result fusion, abnormal alignment point removal, and misaligned marker completion, overall comparative experiments, key technical feature ablation experiments, and robustness experiments under different noise intensities were conducted. The experiments used dual-track marker data from a continuous 48.6km track over six consecutive detection cycles as experimental data. The markers included rail joints, welds, and insulation joints. A baseline true value alignment relationship was manually verified as the evaluation standard. To simulate complex engineering conditions, different levels of missed detection, false detection, and mileage ratio drift disturbances were further injected into the test set. The missed detection rates were set to 5%, 10%, and 15%, the false detection rates to 3%, 6%, and 10%, and the mileage ratio drift to 0.1%, 0.3%, and 0.5%. The evaluation metrics used are precision (P), recall (R), harmonic mean of precision and recall (F1), mean absolute mileage error (MAE), monotonic violation rate (MVR), dual-track inconsistency rate (DIR), candidate edge compression rate (CER), completion success rate (CSR), and mean running time (T).

[0089] I. Overall Comparative Experiment Results

[0090] Table 1 Overall Comparison Experiment Results

[0091] Comparative Example 1 84.6 80.2 82.3 1.86 4.8 6.4 24.8 Comparative Example 2 87.1 84.4 85.7 1.39 3.2 5.1 38.6 Comparative Example 3 91.5 89.2 90.3 0.91 1.9 3.0 21.9 This invention 95.3 93.8 94.5 0.42 0.6 0.9 15.0

[0092] As shown in Table 1, even with the presence of missed detections, false detections, proportional drift, and locally densely similar structures, the present invention outperforms all comparative examples in terms of precision, recall, F1 score, mean absolute mileage error, monotonic violation rate, and dual-track inconsistency rate. Specifically, compared to the closest comparative example 3, the present invention improves the F1 score from 90.3% to 94.5%, reduces the mean absolute mileage error from 0.91m to 0.42m, reduces the monotonic violation rate from 1.9% to 0.6%, reduces the dual-track inconsistency rate from 3.0% to 0.9%, and reduces the average running time from 21.9s to 15.0s. These results demonstrate that the present invention not only significantly improves the accuracy of multi-cycle dual-track marker data alignment but also simultaneously improves path monotonicity, dual-track consistency, and computational efficiency, thus proving its beneficial effects of "improving accuracy, improving consistency, and taking efficiency into account."

[0093] II. Key Technical Features and Ablation Experiment Results

[0094] To further verify the contribution of each key technical feature to the final technical effect, the following ablation groups were set up: Ablation group A removed density adaptive parameter configuration; Ablation group B removed the joint guidance of offset priority region and candidate matching edge confidence; Ablation group C removed explicit missed detection penalty and false detection penalty; Ablation group D removed dual-track consistency verification and callback; Ablation group E removed abnormal alignment point removal and misaligned marker completion.

[0095] Table 2 Ablation Experiment Results

[0096] Ablation group A (density-adaptive ablation) 92.6 0.61 1.2 1.7 62.4 89.1 16.8 Ablation group B (de-co-guided) 91.8 0.73 1.5 1.9 28.6 88.5 22.7 Ablation Group C (Removal of penalties for missed / false detections) 90.9 0.88 1.6 2.1 61.3 86.9 15.9 Ablation group D (Consistency callback) 91.4 0.79 1.8 3.6 60.7 87.4 15.6 Ablation group E (removal and completion) 93.1 0.57 0.9 1.2 61.8 72.5 14.8 This invention 94.5 0.42 0.6 0.9 62.1 93.8 15.0

[0097] Table 2 shows that each key technical feature makes an irreplaceable contribution to the final result. After removing the density adaptive parameter configuration, the F1 score decreased from 94.5% to 92.6%, and the MAE increased from 0.42m to 0.61m, indicating that the segmented density adaptive parameter configuration can take into account the different needs of dense and sparse segments, reducing mismatches in dense areas and missed matches in sparse areas. After removing the joint guidance of the offset priority region and candidate matching edge confidence, the CER decreased from 62.1% to 28.6%, and the average running time increased from 15.0s to 22.7s, indicating that the joint guidance of the offset priority region and candidate matching edge confidence can effectively reduce the offset search space and reduce redundant candidate matching edges. After removing the explicit missed detection penalty and false detection penalty, the F1 score decreased to 90.9%, and the MAE increased to 0.88m, indicating that incorporating missed detections and false detections into the optimization framework can significantly enhance alignment robustness under complex conditions. After removing the dual-track consistency check and callback, the DIR increased to 3.6%, indicating that the closed-loop mechanism can effectively suppress independent mismatches between the left and right tracks. After removing the abnormal alignment point removal and misalignment marker completion, the CSR decreased from 93.8% to 72.5%, indicating that the re-correction and completion steps after result fusion have a significant effect on improving the continuity and completeness of the output results. Therefore, it is evident that the key steps of this invention are not simply stacked in parallel, but rather form a synergistic gain in accuracy, efficiency, consistency, and completeness.

[0098] III. Robustness test results under different noise conditions

[0099] Furthermore, under different false negative and false positive rates, the present invention is compared with Comparative Example 3.

[0100] Table 3. Robustness test results under different noise conditions

[0101] The false negative rate was 5%, and the false positive rate was 3%. Comparative Example 3 92.7 0.74 2.1 The false negative rate was 5%, and the false positive rate was 3%. This invention 95.8 0.34 0.7 The false negative rate was 10%, and the false positive rate was 6%. Comparative Example 3 90.3 0.91 3.0 The false negative rate was 10%, and the false positive rate was 6%. This invention 94.5 0.42 0.9 The false negative rate was 15%, and the false positive rate was 10%. Comparative Example 3 86.9 1.25 4.4 The false negative rate was 15%, and the false positive rate was 10%. This invention 92.2 0.59 1.4

[0102] As shown in Table 3, with the continuous increase in the false negative and false positive rates, the F1 value of Comparative Example 3 continuously decreased, while the mean absolute mileage error (MAE) and dual-track inconsistency rate continuously increased. In contrast, the present invention maintained a high F1 value and low MAE and DIR under low, medium, and high noise conditions. Especially under strong noise conditions with a false negative rate of 15% and a false positive rate of 10%, the present invention still achieved an F1 value of 92.2%, an MAE of 0.59m, and a DIR of 1.4%, significantly better than Comparative Example 3's 86.9%, 1.25m, and 4.4%, respectively. This indicates that the present invention can maintain good alignment accuracy and dual-track consistency even in complex scenarios with both false negatives and false positives and strong noise, demonstrating stronger robustness and engineering adaptability.

[0103] IV. Conclusions of Comparative Experiments

[0104] As shown in Tables 1 to 3, compared with existing conventional methods, the present invention has at least the following technical advantages: First, by configuring segmented density adaptive parameters, it can take into account the different alignment requirements of dense and sparse segments, reducing local mismatches and omissions; second, by jointly guiding the offset priority region and the confidence of candidate matching edges, it can reduce the search space, compress redundant candidate edges, and improve solution efficiency; third, by explicitly setting omission and false detection penalties, it can maintain high alignment robustness under complex omission and false detection conditions; fourth, by using dual-track consistency verification and callback recalculation, it can improve the consistency and stability of the final alignment result; fifth, by using result fusion, abnormal alignment point removal, and misalignment marker completion, it can improve the continuity, physical rationality, and completeness of the output result. Therefore, it is proven that the technical solution described in this invention can effectively achieve the beneficial effects described in the specification.

[0105] The present invention has been further described in detail above with reference to preferred embodiments. Those skilled in the art should understand that, without departing from the technical concept of the present invention, equivalent substitutions or modifications can be made to the segmentation method, density determination threshold, specific network structure of the inference model, training loss form, candidate offset search strategy, local matching solution algorithm, dual-track consistency evaluation method, and regression completion method of the present invention. All equivalent substitutions, simple transformations, or reasonable combinations made using the technical concept disclosed in the present invention specification and claims should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection, characterized in that, include: The common mileage interval of the dual-track marker data of the reference period and the period to be aligned is determined and divided into segments. The reference sequence and the sequence to be aligned of each segment are extracted. The marker number density and average interval of each segment are calculated. Based on this, the segment type, candidate offset search range, structural consistency weight and local matching neighborhood radius are determined. The dual-track marker data includes at least the marker mileage position, track identifier, marker type and acquisition period identifier. Based on the reference sequence and the sequence to be aligned for each segment, a dual-track collaborative structure feature is constructed. The dual-track collaborative structure feature includes a local interval vector on the same track and a collaborative interval amount on different tracks. The local interval vector on the same track consists of the mileage interval between the current marker and several adjacent markers on the same track before and after it. The collaborative interval amount on different tracks is the mileage difference between the current marker and adjacent markers on different tracks. The reference sequence, the sequence to be aligned, the segment type, the structural consistency weight, and the dual-track collaborative structure features of each segment are input into the inference model. The predicted probability of the candidate offset and the basic probability of the candidate matching edge are output. The confidence of the candidate matching edge is obtained by combining the dual-track collaborative structure features and the structural consistency weight, and the offset priority region is determined. Within the intersection of the offset priority region and the candidate offset search range, the candidate offsets are scored and sorted according to a unified scoring function to determine the initial offset. A set of local candidate matching edges is then constructed based on the initial offset, the local matching neighborhood radius, and the confidence of the candidate matching edges. Under the condition of explicitly setting the penalty for missed detection and the penalty for false detection, the one-to-one correspondence that satisfies the mileage monotonic constraint is solved for the local candidate matching edge set to obtain the segment alignment result; Perform dual-track consistency check and monotonic violation check on the segmented alignment result. If the check fails, backtrack to the candidate offset with the second-best score and resolve the local matching. The alignment results of each segment that have passed the verification are fused together. A deviation model is established based on the fused alignment points. Abnormal alignment points are removed and misaligned markers are filled in. The final alignment result is then output.

2. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 1, characterized in that, Set overlap length between adjacent segments ;No. Average interval of each segment With number density They respectively satisfy: ， when and At that time, the first Each segment is determined to be a dense segment, otherwise it is determined to be a sparse segment; dense segments are determined using... , , Sparse segments adopt , , And satisfy , , , in, Indicates the first The number of markers in each segment , They represent the first The first segment sorted by mileage in ascending order within the segment. The, the Location of a landmark Indicates the first The starting mileage of each segment, Indicates the first The final mileage of each segment, Indicates the number density threshold. Indicates the average interval threshold. Indicates the first The candidate offset search range for each segment Indicates the first Structural consistency weights for each segment Indicates the first The local matching neighborhood radius of each segment. The candidate offset search range half-width represents the dense segment. Indicates the structural consistency weight of dense segments; Represents the local matching neighborhood radius of the dense segment; The candidate offset search range half-width represents the sparse segment; Indicates the structural consistency weight of sparse segments; This represents the radius of the local matching neighborhood of the sparse segment.

3. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 2, characterized in that, The dual-track cooperative structure features include a local interval vector on the same track and a cooperative interval amount on different tracks, with the reference period being [number missing]. The first marker and the period to be aligned Overall distance between the landmarks satisfy: ， in, Indicates the reference period number Mileage location of each marker Indicates the period to be aligned. Mileage location of each marker Indicates the reference period number Locally spaced vectors on the same track are constructed centered on each marker. Indicates the period to be aligned Locally spaced vectors on the same track are constructed centered on each marker. Describes the norm 1. Indicates the period in the reference period and the first The distance difference between adjacent markers on different tracks corresponding to each marker. Indicates the period to be aligned with the first The distance difference between adjacent markers on different tracks corresponding to each marker. , , This represents the weight coefficient for the corresponding distance term.

4. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 3, characterized in that, The inference model for the first The predicted probability of each candidate offset output and the offset priority region determined by the predicted probability. They respectively satisfy: ， in, , These respectively represent the inference model for the first... The prediction score of the r-th candidate offset output. This indicates the offset temperature parameter. Indicates the total number of candidate offsets. Indicates the first Candidate offsets This indicates the threshold for the offset priority region.

5. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 4, characterized in that, Based on the comprehensive distance The constructed reference period The first marker and the period to be aligned Analytical soft alignment probability between markers and candidate matching edge confidence They respectively satisfy: ， in, Indicates the alignment temperature parameter. This indicates the number of candidate markers for the current segment to be aligned. This indicates the reference period output by the inference model. The first marker and the period to be aligned The basic probability of candidate matching edges between each marker. Represents the fusion coefficient. Indicates the reference period number The first marker and the period to be aligned The overall distance between the landmarks This represents the index of the candidate markers for the period to be aligned, and .

6. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 5, characterized in that, No. Uniform scoring function for each segment and the Initial offset of each segment They respectively satisfy: ， in, This represents the candidate offset value taken within the intersection of the offset priority region and the candidate offset search range. Indicates the first Each segment at the candidate offset The matching rate item below, Indicates the first Each segment at the candidate offset The relative interval consistency term, Indicates the first Each segment at the candidate offset The dual-track consistency penalty item below, Indicates the first Each segment at the candidate offset The mileage monotonicity penalty term, , , and This represents the weight coefficient of the corresponding rating item.

7. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 6, characterized in that, No. Set of local candidate matching edges for each segment The benchmark period The first marker and the period to be aligned Matching cost between markers and before the benchmark period Each marker and the period before alignment The cumulative minimum cost when matching each marker They respectively satisfy: ， , in, This represents the threshold for filtering candidate matching edges. , , and This represents the weighting coefficient of the corresponding cost item. This represents a missed detection indicator variable, which takes the value 1 when a missed detection occurs and 0 otherwise. This represents a false positive indicator variable, taking the value 1 when a false positive occurs and 0 otherwise. This indicates the penalty for missed inspection. Indicates penalties for false positives. This indicates the period before the reference period. Each marker and the period before alignment The minimum cumulative cost when matching each marker. This indicates the period before the reference period. Each marker and the period before alignment The minimum cumulative cost when matching each marker. This indicates the period before the reference period. Each marker and the period before alignment The cumulative minimum cost when matching each marker; by backtracking the optimal transfer path corresponding to the cumulative minimum cost, a one-to-one correspondence satisfying the mileage monotonic constraint is obtained.

8. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 7, characterized in that, No. Dual-track consistency evaluation quantity for each segment and the expanded candidate matching edge set during backtracking and recalculation They respectively satisfy: ， when If the monotonic violation check fails, the second-best candidate offset corresponding to the unified scoring function is used. and the extended candidate matching edge set Resolve the local matching. Among them, superscript Indicates the left rail, superscript Indicates the right track. Indicates the first Offset estimation results for each segment of the left rail. Indicates the first Offset estimation results for each segment of the right rail , They represent the first The matching rate of each segment's left and right rails. , They represent the first The relative spacing consistency of the left and right rails of each segment. , and This represents the weighting coefficient of the corresponding difference item. Indicates the dual-track consistency threshold. Indicates the first The second-best candidate offset for each segment's score. Indicates the first The back-off expansion radius of each segment.

9. The method for aligning multi-cycle dual-track marker data under conditions of missed detection and false detection as described in claim 8, characterized in that, Deviation samples are established for the fused alignment points, and regression fitting is performed. Mileage deviation at each alignment point , No. The deviation fit value of each alignment point , No. The regression residuals at each alignment point And the theoretical position of the misaligned marker in the alignment period. They respectively satisfy: ， in, Indicates the first The reference cycle marker mileage position corresponding to each alignment point Indicates the first The mileage position of the period marker to be aligned corresponds to each alignment point. Indicates the regression slope. Indicates the regression intercept. This indicates the mileage position of the misaligned marker within the reference period. This represents the predicted value of the corresponding deviation given by the regression model. Indicates the alignment point index. Indicates the residual threshold; when When this happens, the corresponding alignment point is identified as an abnormal alignment point and discarded; based on the theoretical position... Search for candidate markers within a preset neighborhood centered on the target, and establish corresponding relationships or mark them as points to be verified.

10. A multi-cycle dual-track marker data alignment system under conditions of missed detection and false detection, characterized in that, include: The segmentation and density determination module is used to determine the common mileage interval and segment the dual-track marker data of the reference period and the period to be aligned, extract the reference sequence and the sequence to be aligned of each segment, calculate the marker number density and average interval of each segment, and output the segment type, candidate offset search range, structural consistency weight and local matching neighborhood radius. The dual-track marker data includes at least the marker mileage position, track identifier, marker type and acquisition period identifier. The offset and edge inference module is used to construct dual-track cooperative structure features based on the baseline sequence and the sequence to be aligned for each segment. The dual-track cooperative structure features include local interval vectors on the same track and cooperative intervals on different tracks. The local interval vector on the same track consists of the mileage interval between the current marker and several adjacent markers on the same track before and after it. The cooperative interval on different tracks is the mileage difference between the current marker and adjacent markers on different tracks. The module is used to input the baseline sequence, the sequence to be aligned, the segment type, the structural consistency weight, and the dual-track cooperative structure features of each segment into the inference model, output the predicted probability of candidate offsets and the basic probability of candidate matching edges, and combine the dual-track cooperative structure features and the structural consistency weight to obtain the confidence of candidate matching edges and determine the offset priority region. The offset search and monotonic matching module is used to determine the initial offset based on a unified scoring function within the intersection of the offset priority region and the candidate offset search range, and to construct a set of local candidate matching edges based on the initial offset, the local matching neighborhood radius and the confidence of the candidate matching edges. The module retains the candidate offset scoring sequence, solves the one-to-one correspondence that satisfies the mileage monotonic constraint under the condition of explicitly setting the missed detection penalty and the false detection penalty, and outputs the segment alignment result. The consistency rollback module is used to receive the segment alignment result and the candidate offset score sequence, perform dual-track consistency check and monotonic violation check, and return the candidate offset with the second best score to the offset search and monotonic matching module when the check fails so as to resolve the local matching. The fusion and completion module is used to fuse the alignment results of each segment that has passed the verification, establish deviation samples and deviation models based on the fused alignment points, remove abnormal alignment points and complete the unaligned markers, and output the final alignment result.

Images