A bridge deformation real-time prediction method based on reinforcement learning

Abstract

The application discloses a bridge deformation real-time prediction method based on reinforcement learning, relates to the technical field of bridge structure health monitoring and intelligent prediction, and comprises the following steps: S1, collecting bridge real-time monitoring responses according to structural parts, and converging homologous load deformation segments; S2, determining a load removal boundary, and forming a bridge deformation response window chain; S3, verifying a falling back closed state, and packaging a deformation prediction object with a boundary marker; S4, connecting an improved Pathformer model; S5, dividing layers from patch segments, rewriting fixed patch segment size continuous indexes to generate paths, limiting index remaining relations through reinforcement learning actions, and forming candidate patch segments; S6, extracting changes in the segments, checking changes between the segments, and reading out a prediction result; and S7, pasting structural parts and rewriting archives. The application realizes deformation real-time prediction under a load removal boundary.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring and intelligent prediction technology, and in particular to a method for real-time prediction of bridge deformation based on reinforcement learning. Background Technology

[0002] Existing bridge deformation monitoring typically relies on displacement, deflection, tilt, or strain sensors to obtain real-time responses, and then uses threshold judgment, statistical analysis, or conventional time series models to predict deformation trends. While these methods can reflect overall changes, they often incorporate vehicle load, temperature change, and wind load responses into the same continuous sequence, making it difficult to distinguish between synchronous fluctuations under load, normal settling after unloading, and residual deformation that does not disappear with unloading.

[0003] With the development of deep learning technology, Transformer-type models and multi-scale time series models are increasingly being used for bridge deformation prediction. Existing models typically divide monitoring sequences into fixed time windows or fixed segment sizes, then extract features and extrapolate trends. However, the two sides of the load removal boundary have different physical meanings, and fixed segment division can easily compress the load application stage and the unloading rebound stage into the same segment, resulting in the continued prediction of reboundable deformation and the weakening of non-rebound residuals by the fallback process.

[0004] In addition, existing prediction methods focus more on error reduction or threshold adjustment, and rarely modify the internal fragment index generation path of the model. It is difficult to truncate, rearrange and mask the index connection across the load removal boundary, which makes the real-time prediction results prone to drift and lag.

[0005] Therefore, how to provide a real-time bridge deformation prediction method based on reinforcement learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a real-time bridge deformation prediction method based on reinforcement learning. This invention employs a combination of load unloading boundary identification, non-rebound residual retention, and an improved Pathformer model to organize the real-time bridge monitoring response into a bridge deformation response window chain and encapsulate deformation prediction objects with boundary markers. Based on this, the continuous index generation path of fixed patch segment size is rewritten within the patch segment division layer. The placement direction of the boundary adjacent index and the non-rebound residual index is determined by reinforcement learning actions, avoiding mixed modeling of the load application stage and the unloading and rebound stage within the same patch segment. This method has the advantages of high prediction accuracy, strong real-time performance, and applicability to long-term online bridge monitoring.

[0007] A method for real-time prediction of bridge deformation based on reinforcement learning according to an embodiment of the present invention includes the following steps: S1. Collect the real-time monitoring response of the bridge according to the structural parts, and consolidate the synchronous fluctuations of deformation under the same load into deformation segments of the same load. S2. Determine the load removal boundary along the deformation segment of the same load source, and compress the deformation bearing relationship on both sides of the boundary into a bridge deformation response window chain. S3. Verify the closed state of the bridge deformation response window chain after unloading and retain the changes that do not follow the unloading and fall as non-rebound residuals, and encapsulate them as deformation prediction objects with boundary markers. S4. Integrate the deformation prediction object into the improved Pathformer model, and the rewriting position of the path generated by the fixed patch fragment size continuous index falls on the patch fragment partitioning layer. S5, the patch segmentation layer takes over the continuous index generation path of fixed patch segment size, arranges the starting point and position within the segment around the candidate patch segment size, and cuts off the continuous index connection when the position within the segment crosses the load unloading boundary. The reinforcement learning action limits the retention relationship between the adjacent index of the boundary and the non-bounced residual index. The position that crosses the boundary is supplemented by the same side boundary index or the masking index to form candidate patch segments that maintain the alignment of the segment size. S6. Extract intra-fragment changes and verify inter-fragment changes for candidate patch segments, retain deformation changes on the same side of the boundary as effective multi-scale features, mask changes across opposite deformation stages, and read out the real-time prediction results of bridge deformation. S7. Paste the real-time prediction results of bridge deformation back to the structural parts corresponding to the bridge deformation response window chain, merge them into the real-time prediction record of bridge deformation and write them back to the bridge monitoring archive.

[0008] Optionally, S1 specifically includes: S11. Verify the installation attribution of the bridge real-time monitoring response and the bridge structural parts, and group the deformation responses with the same attribution and continuous sampling time into the same part response group. S12. Align the sampling time axis of response groups in different parts, cut off isolated responses with time breaks, and retain the deformed responses that can be aligned within the same time range as the response groups to be verified. S13. Compare the undulation direction, undulation starting point and duration within the response group to be verified. Classify the deformation changes that rise synchronously, fall synchronously, or stop synchronously into the same load response group, and return the deformation changes that have opposite undulation directions or discontinuous duration to the corresponding part response group. S14. The deformation changes that continuously participate in synchronous fluctuations in the response group of the same source load are gathered together, and the static changes before and after the load are removed to form the deformation segment of the same source load.

[0009] Optionally, S2 specifically includes: S21. Compare the deformation changes at adjacent sampling times along the time sequence of deformation segments of the same load source to identify candidate unloading sampling points where the deformation changes change from continuous expansion to continuous convergence. S22. Verify the adjacency relationship between the candidate unloading sampling point and the previous peak response, and check the fall direction and fall duration after the candidate unloading sampling point to determine the load unloading boundary. S23. Use the load removal boundary to fit the deformation segments of the same source load before and after, retain the deformation changes that are continuously connected in time between the deformation pushing direction before the boundary and the falling direction after the boundary, and return the deformation changes that cannot be fitted due to time breaks or abrupt changes in direction to the deformation segments of the same source load. S24. Gather the deformation changes that are attached front and back, and compress the continuous deformation changes at the boundary front and back into a bridge deformation response window chain.

[0010] Optionally, S3 specifically includes: S31. Read the load removal boundary in the bridge deformation response window chain, and extract the adjacent response before the load removal boundary and the fallback response after the load removal boundary. S32. Using the adjacent response on the front side of the load removal boundary as the reference for the fall-off closure, verify the fall-off direction, fall-off endpoint, and duration of the fall-off response on the rear side of the load removal boundary. S33. Deformation changes in which the fall direction is toward the fall closing reference and the fall endpoint is in contact with the fall closing reference are calibrated as fall closing changes. Deformation changes in which the fall direction deviates from the fall closing reference or the fall endpoint is not in contact with the fall closing reference are calibrated as non-rebound residuals. S34. Strip the fallback closure change in the bridge deformation response window chain, pair and encapsulate the load removal boundary, non-rebound residual, structural part attribution and time bearing position to form a deformation prediction object with boundary markings.

[0011] Optionally, the improved Pathformer model includes an instance normalization layer, a patch segmentation layer, a dual attention layer, a multi-scale path selector, a path weight propagation path, an aggregator, and a predictor.

[0012] Optionally, S5 specifically includes: S51. Read the deformation prediction object with boundary markers by dividing the patch fragment into layers. Take the sampling order of the bridge deformation response window chain as the reference continuous index arrangement corresponding to the fixed patch fragment size. The fragment starting point in the reference continuous index arrangement is expanded in sequence according to the fixed patch fragment size. The position in the fragment is attached to the continuous sampling index according to the sampling order. S52. Align the load removal boundary to the reference continuous index arrangement. When the load removal boundary falls between adjacent patch segments, preserve the continuous index connection of adjacent patch segments. When the load removal boundary falls inside the same patch segment, cut the continuous index arrangement within the same patch segment into a boundary front index arrangement and a boundary back index arrangement. S53. Rearrange the cut indexes around the size of the candidate patch segment. Fill the indexes before the boundary into the positions of the segments before the boundary according to the original sampling order. Fill the indexes after the boundary into the positions of the segments after the boundary according to the original sampling order. The indexes of the non-bounced residuals are retained in the positions of the segments on the side of the boundary that have not closed after falling back. S54. The retention destination of the boundary adjacent index and the non-bounced residual index of the reinforcement learning action decision boundary is determined. The boundary same-side index corresponding to the fallback closure change is retained along the boundary same-side sampling order. The non-bounced residual index is retained along the residual continuity direction. The boundary same-side sampling index of adjacent structural parts is connected to the vacancy position under the same candidate patch segment size. S55. Perform size alignment on the position within the cut segment. If there is a sampling index on the same side of the boundary, supplement it with the sampling index on the same side of the boundary. If the sampling index on the same side of the boundary is insufficient, backfill it with the sampling index near the load unloading boundary. If neither the sampling index on the same side of the boundary nor the sampling index near the load unloading boundary is hit, fill in the mask index. S56. Replace the baseline continuous index arrangement with the rearranged index arrangement, retain the number of positions within the fragment corresponding to the candidate patch fragment size, and form a candidate patch fragment that does not cross the load removal boundary and maintains the fragment size alignment.

[0013] Optionally, S54 specifically includes: S541, the patch segmentation layer reads the baseline continuous index arrangement, the index arrangement before the boundary, the index arrangement after the boundary, the index of the non-bounced residual, and the position within the missing segment under the same candidate patch segment size, and arranges them as the index arrangement to be decided along the segment start point to the segment end; S542. Improve the path weight passing path in the Pathformer model. Read the path weight corresponding to the size of the same candidate patch fragment, remove the direct connection between the path weight and the baseline continuous index arrangement that crosses the load removal boundary, and hand the path weight over to the reinforcement learning action to reconnect it within the index arrangement to be decided. S543. When the fallback direction of the reinforcement learning action check boundary back index is met, and the fallback endpoint is aligned with the deformation position carried by the boundary front index, the path weight is attached to the boundary back index arrangement, and the non-bounced residual index returns to the segment position corresponding to the residual continuation direction. S544. When the reinforcement learning action checks the continuation direction of the non-bounced residual index, and the non-bounced residual index continues in the same direction along the unloaded time sequence, the path weight is attached to the non-bounced residual index, and the fallback index behind the boundary returns to the same sampling sequence on the same side of the boundary. S545. When the boundary same-side sampling index of adjacent structural parts is checked in reinforcement learning, and the boundary same-side sampling index of adjacent structural parts is close to the position in the missing segment in time and the deformation direction is continuous, the path weight is attached to the boundary same-side sampling index of adjacent structural parts, and the boundary same-side sampling index of adjacent structural parts is filled into the position in the missing segment. S546 When the boundary back-side index, the non-bounced residual index, and the boundary same-side sampling index of adjacent structural parts compete for the same segment position, the index with attached path weight is retained, and the index that is not retained is returned to the segment position corresponding to the boundary same-side sampling order or the residual continuity direction. S547. When the path weight has not been reattached, the boundary adjacent index is retained in the original segment position, the non-bounced residual index is retained along the residual continuity direction, and the position in the missing segment is filled by the same side boundary index, the adjacent sampling index or the masking index, forming an index retention arrangement limited by the path weight reattachment.

[0014] Optionally, S6 specifically includes: S61. The multi-scale path selector reads the candidate patch fragment size, path weight, and mask index of the candidate patch fragment, retains the candidate patch fragment size that has completed path weight reattachment, and pairs the candidate patch fragment size with the mask index that has not completed path weight reattachment. S62. The dual attention layer reads the position within the fragment according to the size of the candidate patch fragment, extracts the deformation changes on the same side of the boundary, and sets the position within the fragment paired with the mask index as the mask position in the attention weight calculation to form the fragment change feature. S63. The dual attention layer verifies the segment start point, load removal boundary and non-bounced residual index of adjacent candidate patch segments. The intra-segment variation features of segments with the same boundary and continuous deformation direction are retained as inter-segment variation features, and the intra-segment variation features of segments that cross the load removal boundary and have opposite directions are paired with the masking index. S64. The aggregator reads the intra-fragment change features, inter-fragment change features, and path weights for reattaching the fragments. It then merges the deformation changes on the same side of the boundary along the candidate patch fragment size to form effective multi-scale features. S65. The predictor reads the segment positions of the effective multi-scale features and the masked index pairings, retains the prediction components corresponding to the effective multi-scale features, reduces the prediction components corresponding to the segment positions of the masked index pairings, and reads out the real-time prediction results of bridge deformation.

[0015] Optionally, S7 specifically includes: S71. Retrieve the real-time prediction results of bridge deformation and the bridge deformation response window chain, and verify the corresponding status of the prediction period, predicted deformation direction, predicted deformation amplitude and structural part attribution. S72. The time connection position along the bridge deformation response window chain is aligned with the prediction period, and the predicted deformation direction and predicted deformation amplitude are back-attached to the bridge structural parts with the same affiliation. S73. Merge the posting results that are continuous in time and have connected deformation directions within the same bridge structural part to form a structural part deformation prediction segment. S74. Reply results with conflicting structural part attributions or disconnected prediction periods are retained as prediction results pending review, and the corresponding bridge deformation response window chain is attached. S75. Merge the deformation prediction segments of structural parts and the prediction results to be reviewed to form a real-time prediction record of bridge deformation, and write it back to the bridge monitoring archive.

[0016] The beneficial effects of this invention are: This invention aggregates real-time bridge monitoring responses by structural components and forms homogeneous load deformation segments around synchronous fluctuations under the same load. This allows the dispersed bridge deformation data to be organized into predictable objects with load-bearing significance. By determining the load removal boundary and verifying the fallback closure state after unloading, this invention can distinguish between normal fallback elastic deformation and non-fallback residuals that do not disappear with unloading, avoiding the problem of mixing different physical stages in traditional continuous time series prediction.

[0017] This invention introduces an improved Pathformer model, with the core improvements concentrated in the patch segmentation layer. By rewriting the continuous index generation path for fixed patch segment sizes, cross-boundary index connections can be truncated at the load removal boundary. Reinforcement learning actions are used to determine the placement of boundary-adjacent indices and the unrebounded residual indices, ensuring that candidate patch segments maintain size alignment and do not cross opposite deformation stages. Therefore, when extracting intra-segment changes and verifying inter-segment changes, the model can reduce prediction drift caused by mixing the load application stage and the unloading / springback stage within the same patch segment.

[0018] This invention also appends the real-time bridge deformation prediction results to the corresponding structural parts and merges them into a real-time bridge deformation prediction record, writing it back into the bridge monitoring archive. This ensures that the prediction results maintain a traceable correlation with the structural parts, load removal boundaries, and non-rebound residuals. This method can improve the ability of real-time bridge deformation prediction to identify residual deformation and adapt to frequent load changes. It has the advantages of high prediction stability, strong real-time performance, and suitability for long-term online bridge monitoring. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a real-time bridge deformation prediction method based on reinforcement learning proposed in this invention. Figure 2 This is a schematic diagram of the improved Pathformer model, which is a method for real-time prediction of bridge deformation based on reinforcement learning proposed in this invention. Figure 3 This is a schematic diagram of the patch index rearrangement path for a real-time bridge deformation prediction method based on reinforcement learning proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figures 1-3 A method for real-time prediction of bridge deformation based on reinforcement learning includes the following steps: S1. Collect the real-time monitoring response of the bridge according to the structural parts, and consolidate the synchronous undulating deformation changes under the same load into deformation segments of the same load. S2. Determine the load removal boundary along the deformation segment of the same load source, and compress the deformation bearing relationship on both sides of the boundary into a bridge deformation response window chain. S3. Verify the closed state of the bridge deformation response window chain after unloading and retain the changes that do not follow the unloading and fall as non-rebound residuals, and encapsulate them as deformation prediction objects with boundary markers. S4. Integrate the deformation prediction object into the improved Pathformer model, and the rewriting position of the path generated by the fixed patch fragment size continuous index falls on the patch fragment partitioning layer. S5, the patch segmentation layer takes over the continuous index generation path of fixed patch segment size, arranges the starting point and position within the segment around the candidate patch segment size, and cuts off the continuous index connection when the position within the segment crosses the load unloading boundary. The reinforcement learning action limits the retention relationship between the adjacent index of the boundary and the non-bounced residual index. The position that crosses the boundary is supplemented by the same side boundary index or the masking index to form candidate patch segments that maintain the alignment of the segment size. S6. Extract intra-fragment changes and verify inter-fragment changes for candidate patch segments, retain deformation changes on the same side of the boundary as effective multi-scale features, mask changes across opposite deformation stages, and read out the real-time prediction results of bridge deformation. S7. Paste the real-time prediction results of bridge deformation back to the structural parts corresponding to the bridge deformation response window chain, merge them into the real-time prediction record of bridge deformation and write them back to the bridge monitoring archive.

[0022] In this embodiment, S1 specifically includes: S11. Verify the installation attribution of the bridge real-time monitoring response and the bridge structural parts, and group the deformation responses with the same attribution and continuous sampling time into the same part response group. S12. Align the sampling time axis of response groups in different parts, cut off isolated responses with time breaks, and retain the deformed responses that can be aligned within the same time range as the response groups to be verified. S13. Compare the undulation direction, undulation starting point and duration within the response group to be verified. Classify the deformation changes that rise synchronously, fall synchronously, or stop synchronously into the same load response group, and return the deformation changes that have opposite undulation directions or discontinuous duration to the corresponding part response group. S14. The deformation changes that continuously participate in synchronous fluctuations in the response group of the same source load are gathered together, and the static changes before and after the load are removed to form the deformation segment of the same source load.

[0023] In practice, the real-time monitoring responses of the bridge come from displacement, deflection, tilt, or strain sensors installed at supports, beam ends, connection points, mid-span of the main beam, and key sections. The installation affixation of structural components is determined by the sensor installation log, bridge structure number, and measuring point number. During verification, the structural component field in the measuring point number is read and compared with the component name in the bridge structure number. Deformation responses with matching names are assigned to the corresponding component response group, while those with inconsistent names are retained in the pending verification response pool, along with the original measuring point number. Sampling continuity is determined by the time difference between two adjacent response records. In one implementation, the sampling interval can be 1 second; adjacent time differences of no more than 1.5 seconds are considered continuous, while differences exceeding 1.5 seconds are considered discontinuous.

[0024] The sampling timelines of different response groups are aligned with a unified start time and a unified sampling interval. During alignment, the sampling times in each response group are mapped to the same time scale. Missing sampling points are only marked as missing, without being inserted as predicted values. The missing markers are used for timeline integrity verification but not for fluctuation direction calculation. Isolated responses are determined by considering both preceding and following sampling times and synchronous fluctuations in adjacent locations. If a response record cannot be continuously matched with either the preceding or following sampling time, and there are no synchronous fluctuations in adjacent locations within the same time range, the response record is removed and returned to the corresponding response group. If there are synchronous fluctuations in adjacent locations, the response record is retained and added to the response group to be verified.

[0025] The direction of fluctuation is determined by the sign of the difference between the current sampled value and the previous sampled value. A positive difference indicates an upward movement, and a negative difference indicates a downward movement. When the absolute value of the difference is lower than the sensor noise amplitude, it is considered to have stabilized. The sensor noise amplitude is calculated from the response of the same stationary section. First, the average value of the stationary section is calculated, then the square root of the square mean of the differences between each sampled value and the average value is calculated, and the resulting standard deviation is multiplied by 3 to obtain the noise amplitude. The starting point of fluctuation is the first sampling moment when continuous upward or continuous downward movement begins, and the duration is the time interval from the starting point of fluctuation to the end of the change in direction or stabilization. When multiple parts of the response have the same fluctuation direction within the same time range, and the overlap ratio of their duration reaches 60%, the deformation changes are merged into the same load response group; when the overlap ratio is less than 60% or the fluctuation directions are opposite, the deformation changes are returned to the corresponding part response group.

[0026] The convergence boundary of the response group of the same load originates from the earliest undulation point and ends when the last part participating in synchronous undulation reaches a stable state. The static changes before and after the load application are determined by the continuous stable state. When 30 sampling points are continuously stable, the stable interval is cut off as a static segment. After cutting off the static segment, the remaining response records are bound into same load deformation segments according to structural location, sampling time, deformation value, undulation direction, duration, and same load group number. The same load group number is uniformly assigned by the response group that has completed synchronous undulation verification. The same load deformation segments retain the original sampling order and common undulation boundary of the response of each part and enter the load removal boundary determination process.

[0027] In this embodiment, S2 specifically includes: S21. Compare the deformation changes at adjacent sampling times along the time sequence of deformation segments of the same load source to identify candidate unloading sampling points where the deformation changes change from continuous expansion to continuous convergence. S22. Verify the adjacency relationship between the candidate unloading sampling point and the previous peak response, and check the fall direction and fall duration after the candidate unloading sampling point to determine the load unloading boundary. S23. Use the load removal boundary to fit the deformation segments of the same source load before and after, retain the deformation changes that are continuously connected in time between the deformation pushing direction before the boundary and the falling direction after the boundary, and return the deformation changes that cannot be fitted due to time breaks or abrupt changes in direction to the deformation segments of the same source load. S24. Gather the deformation changes that are attached front and back, and compress the continuous deformation changes at the boundary front and back into a bridge deformation response window chain.

[0028] In the specific implementation process, deformation segments of the same source load are arranged in ascending order according to the sampling time. Each sampling time retains the structural location, deformation value, undulation direction, and same source load group number. The deformation change between adjacent sampling times is obtained by subtracting the previous sampling value from the current sampling value. When the difference is continuously positive, it is considered as continuous expansion; when the difference is continuously negative, it is considered as continuous convergence. In one embodiment, continuous expansion or continuous convergence can be determined by taking 3 consecutive sampling points as the effective judgment length. Reverse changes shorter than 3 sampling points are retained as noise disturbances in the original time sequence and are not included in the candidate unloading sampling point delineation.

[0029] The candidate unloading sampling point is delineated starting from the position where it transitions from the continuous expansion interval to the continuous convergence interval. The preceding peak response is obtained by comparing the absolute deformation value within the continuous expansion interval before the candidate unloading sampling point; the sampling record with the largest absolute deformation value is taken as the preceding peak response. The adjacency status between the candidate unloading sampling point and the preceding peak response is determined by the time difference. In one embodiment, adjacency is determined when the time difference does not exceed two sampling intervals; when the time difference exceeds two sampling intervals, the candidate unloading sampling point reverts to the deformation segment of the same source load and continues to participate in the scanning.

[0030] The direction of the load reversal is determined by the sign of the difference between adjacent samples after the candidate load removal sampling point. A reversal is defined as the direction of reversal occurring when the reversal direction is opposite to the expansion direction corresponding to the previous peak response. The duration of the reversal begins at the candidate load removal sampling point and ends when the deformation difference enters a stable state or re-enters an expansion state. The stable state is obtained by comparing the absolute value of the deformation difference with the sensor noise amplitude. In one embodiment, a stable state is defined as the absolute value of the deformation difference at 30 consecutive sampling points being lower than the sensor noise amplitude. When a candidate load removal sampling point simultaneously satisfies peak adjacency, directional reversal, and continuous reversal, it is determined to be the load removal boundary; if peak adjacency, directional reversal, or continuous reversal is not met, the candidate load removal sampling point reverts to the deformation segment of the same source load.

[0031] After the load removal boundary is determined, deformation segments of the same source load are fitted together before and after the boundary. The direction of deformation before the boundary is taken as the main direction of the continuous expansion interval before the boundary, and the direction of fall after the boundary is taken as the main direction of the continuous convergence interval after the boundary. When the time difference between the last sampling time before the boundary and the first sampling time after the boundary does not exceed one sampling interval, it is considered that the time is continuous; when the difference between two consecutive sampling values ​​after the boundary is opposite to the fall direction, it is considered that the direction is abrupt. Deformation changes with time break or abrupt direction change are returned to the deformation segment of the same source load and attached to the break position. The deformation changes after fitting together converge around the load removal boundary. The continuous expansion interval is retained before the boundary, and the continuous fall interval is retained after the boundary. The structural part attribution, sampling time, deformation value, peak response of the previous segment, load removal boundary, fall direction, and fall duration range together form the bridge deformation response window chain. The bridge deformation response window chain retains the deformation bearing relationship on both sides of the boundary and enters the fall closure state verification process.

[0032] In this embodiment, S3 specifically includes: S31. Read the load removal boundary in the bridge deformation response window chain, and extract the adjacent response before the load removal boundary and the fallback response after the load removal boundary. S32. Using the adjacent response on the front side of the load removal boundary as the reference for the fall-off closure, verify the fall-off direction, fall-off endpoint, and duration of the fall-off response on the rear side of the load removal boundary. S33. Deformation changes in which the fall direction is toward the fall closing reference and the fall endpoint is in contact with the fall closing reference are calibrated as fall closing changes. Deformation changes in which the fall direction deviates from the fall closing reference or the fall endpoint is not in contact with the fall closing reference are calibrated as non-rebound residuals. S34. Strip the fallback closure change in the bridge deformation response window chain, pair and encapsulate the load removal boundary, non-rebound residual, structural part attribution and time bearing position to form a deformation prediction object with boundary markings.

[0033] In the specific implementation process, the bridge deformation response window chain is centered on the load removal boundary, capturing the preceding adjacent response and the following fallback response. The preceding adjacent response is obtained from the continuous expansion interval before the load removal boundary, including the undulation start point, peak response, and the last sampling point before the boundary; the following fallback response is captured from the first sampling point after the load removal boundary until it falls into a stable state or returns to an expansion state. The capture range does not change the sampled values, but only determines the range of sampling records participating in the fallback closure verification. When sampling is insufficient, the actual sampling length is retained and marked as insufficient sampling; the closure confidence is obtained by dividing the actual number of captured sampling points by the target number of captured sampling points. The target number of captured sampling points is obtained by adding the target window lengths of the preceding adjacent response and the following fallback response. In one embodiment, when the closure confidence is less than 60%, the window chain enters a pending verification state and does not participate in the fallback closure change stripping; when the closure confidence reaches 60%, the window chain enters the fallback closure verification.

[0034] The fallback closure reference is obtained from the adjacent responses on the preceding side. The mean deformation value of continuously stable samples before the undulation initiation is used as the closure reference, and the peak response before the boundary is used as the starting point for calculating the fallback direction. The closure reference is obtained by summing the stable sample values ​​and dividing by the number of sample points. The fallback direction is determined by the sign of the difference between adjacent sample values ​​in the subsequent fallback response; when the difference direction is towards the closure reference, it is considered to be towards fallback, and when the difference direction is away from the closure reference, it is considered to be deviating from fallback. The fallback endpoint is taken from the sample point in the subsequent fallback response that first enters a continuous stable state. In one embodiment, a continuous stable state is defined as when the absolute value of the deformation difference of 30 consecutive sample points is lower than the sensor noise amplitude. The fallback duration is the time interval between the load removal boundary and the fallback endpoint.

[0035] The alignment of the fall-off endpoint is determined by the difference between the deformation value at the fall-off endpoint and the closing reference. The sensor noise amplitude is amplified along the standard deviation of the stationary section of the same structural part. If the difference between the deformation value at the fall-off endpoint and the closing reference does not exceed twice the sensor noise amplitude, the fall-off endpoint is considered to be aligned with the closing reference; if the difference exceeds twice the sensor noise amplitude, the fall-off endpoint is considered not aligned with the closing reference. Deformation changes where the fall-off direction is towards the closing reference and the fall-off endpoint aligns with the closing reference are calibrated as fall-off closure changes; deformation changes where the fall-off direction deviates from the closing reference, or where the fall-off endpoint does not align with the closing reference, are calibrated as non-rebound residuals. The residual value of the non-rebound residual is obtained by subtracting the closing reference from the deformation value at the fall-off endpoint. The direction of the residual is determined by the sign of the residual value, and the duration of the residual is retained along the unclosed interval in the subsequent fall-off response.

[0036] The stripping of fallback closure changes is performed within the bridge deformation response window chain. During stripping, the original sampling records are retained; only the prediction participation markers corresponding to fallback closure changes are removed from the deformation prediction objects. Normal fallback changes are no longer included in residual modeling. Non-bounced residuals are paired and encapsulated with the load removal boundary. The pairing includes the structural part attribution, the temporal position of the load removal boundary within the window chain, the residual value of the non-bounced residual, the residual direction, and the residual duration. The structural part attribution follows the attribution results of the part response group, and the temporal continuation position is taken from the sampling sequence range between the load removal boundary and the fallback endpoint.

[0037] The deformation prediction object with boundary markers consists of a bridge deformation response window chain, load removal boundaries, non-rebound residuals, and prediction participation markers. The deformation prediction object retains the original sampling order; load removal boundaries are converted into boundary markers, and non-rebound residuals are converted into residual markers. Boundary markers record the sampling position of the load removal boundary within the bridge deformation response window chain, and residual markers record the structural component attribution, residual direction, and residual duration corresponding to the non-rebound residuals, forming a deformation prediction object with boundary markers capable of accepting model input.

[0038] In this embodiment, the improved Pathformer model includes an instance normalization layer, a patch segmentation layer, a dual attention layer, a multi-scale path selector, a path weight propagation path, an aggregator, and a predictor. In practical implementation, when processing time series data, the ordinary Pathformer model first performs instance normalization on the input sequence, then continuously segments the time index according to a fixed patch size. The segmented patch fragments undergo dual attention, multi-scale path selection, aggregation, and prediction readout to complete the time series prediction. Bridge deformation response corresponds to the load application and unloading rebound stages on either side of the load removal boundary. Ordinary continuous segmentation would place deformation changes in opposite directions into the same patch fragment, causing the attention within the fragment to simultaneously read both rebound and residual changes, easily leading to tailing and residual weakening in the prediction results.

[0039] The improved Pathformer model consists of an instance normalization layer, a patch segmentation layer, a dual attention layer, a multi-scale path selector, a path weight propagation path, an aggregator, and a predictor connected sequentially. The instance normalization layer organizes the scale differences in the responses of different structural parts. The patch segmentation layer takes boundary markers and residual markers and rewrites the path generation path using a fixed-size continuous index. The dual attention layer extracts intra- and inter-segment variations. The multi-scale path selector assigns path weights corresponding to candidate patch sizes. The path weight propagation path stores the path weight attachment state. The aggregator merges effective multi-scale features. The predictor reads the real-time prediction results of bridge deformation. Compared to the ordinary Pathformer model, the improvements are concentrated in the patch segmentation layer and the path weight propagation path. The fixed-size continuous index no longer directly crosses the load removal boundary, and the path weights are no longer directly attached to the cross-boundary index arrangement.

[0040] The input samples are obtained by organizing the bridge deformation response window chain. Before entering the model, the bridge real-time monitoring response is uniformly resampled to the same sampling interval, which follows the interval determined in the aforementioned time axis alignment process. The data format is a two-dimensional tensor of time length multiplied by the feature channel. In one embodiment, the input tensor dimension can be 64×32, where 64 corresponds to the number of sampling points within the continuous load action window, and 32 corresponds to the feature channels formed by structural part response, normalized deformation change, boundary markers, residual markers, and sampling position encoding. Each sampling point is written to a line in chronological order, and each channel is written with the corresponding deformation value, marker value, or position value; the boundary marker corresponds to the sampling position of the load removal boundary, and the residual marker corresponds to the sampling position of the non-rebound residual. When there are fewer than 64 input sampling points, the missing positions are filled with a mask index; when there are more than 64 input sampling points, the adjacent sampling range is clipped with the load removal boundary as the center, and the clipped part is retained within the bridge deformation response window chain.

[0041] Training data comes from displacement, deflection, tilt, or strain responses collected by the long-term online bridge monitoring system, and is aligned with vehicle load records, temperature records, maintenance review records, and stable operation review records. Training samples are constructed using bridge deformation response window chains as units, including structural location attribution, sampling time, normalized deformation sequence, load removal boundary location, non-rebound residual markers, and measured deformation values ​​for the future prediction period. Prediction labels are taken from measured deformation values ​​for the future prediction period; boundary labels are taken from the sampling location where deformation changes from continuous expansion to continuous convergence; residual retention labels are determined by the difference between the fallback endpoint and the closure benchmark; a difference exceeding twice the sensor noise amplitude for more than 20 consecutive sampling points is marked as residual retention, and a difference falling back to within twice the sensor noise amplitude is marked as fallback closure.

[0042] The training loss is a weighted average of prediction error, boundary index consistency error, residual retention error, and masked index penalty. Prediction error is the mean absolute difference between the predicted deformation value and the future measured deformation value. Boundary index consistency error is the sampling distance between the boundary index position and the boundary label position within the model. Residual retention error is the difference between the position where the non-bounced residual index falls and the residual retention label. The masked index penalty is the proportion of masked indexes to the number of positions within the candidate patch segment. In one implementation, the weights of the four terms can be 0.5, 0.2, 0.2, and 0.1. When the proportion of masked indexes exceeds 30%, the masked index penalty increases to reduce ineffective masking.

[0043] The model is trained using the AdamW optimizer, with a learning rate of 0.0001, a batch size of 32, and a maximum of 200 training epochs. After each training epoch, the mean absolute error of the validation set, the load removal boundary index deviation, and the non-bounced residual recognition rate are calculated. Training is stopped when the mean absolute error decreases by less than 1% for 10 consecutive epochs and the non-bounced residual recognition rate increases by less than 0.5%. When the mean absolute error of the validation set increases for 5 consecutive epochs, the parameters are rolled back to the epoch with the lowest mean absolute error on the validation set.

[0044] The instance normalization layer only performs scale adjustment. For each structural part of each sample, the mean and standard deviation are calculated separately for each channel. First, the deformation values ​​of the 64 sampling points are summed and divided by 64 to obtain the channel mean. Then, the square root of the squared mean of the differences between each sampling point and the channel mean is calculated to obtain the channel standard deviation. During normalization, the sampled value is subtracted from the channel mean and then divided by the channel standard deviation. When the channel standard deviation is lower than the sensor noise amplitude, the channel retains its original value and a low-fluctuation marker is added. Boundary markers and residual markers are not involved in the mean and standard deviation calculation; they are only retained synchronously with the sampling index. After training, the scale adjustment method, index rewriting parameters, path weight attachment parameters, candidate size selection parameters, feature aggregation parameters, and prediction mapping parameters are collectively encapsulated into improved Pathformer model parameters.

[0045] In this embodiment, S5 specifically includes: S51. Read the deformation prediction object with boundary markers by dividing the patch fragment into layers. Take the sampling order of the bridge deformation response window chain as the reference continuous index arrangement corresponding to the fixed patch fragment size. The fragment starting point in the reference continuous index arrangement is expanded in sequence according to the fixed patch fragment size. The position in the fragment is attached to the continuous sampling index according to the sampling order. S52. Align the load removal boundary to the reference continuous index arrangement. When the load removal boundary falls between adjacent patch segments, preserve the continuous index connection of adjacent patch segments. When the load removal boundary falls inside the same patch segment, cut the continuous index arrangement within the same patch segment into a boundary front index arrangement and a boundary back index arrangement. S53. Rearrange the cut indexes around the size of the candidate patch segment. Fill the indexes before the boundary into the positions of the segments before the boundary according to the original sampling order. Fill the indexes after the boundary into the positions of the segments after the boundary according to the original sampling order. The indexes of the non-bounced residuals are retained in the positions of the segments on the side of the boundary that have not closed after falling back. S54. The retention destination of the boundary adjacent index and the non-bounced residual index of the reinforcement learning action decision boundary is determined. The boundary same-side index corresponding to the fallback closure change is retained along the boundary same-side sampling order. The non-bounced residual index is retained along the residual continuity direction. The boundary same-side sampling index of adjacent structural parts is connected to the vacancy position under the same candidate patch segment size. S55. Perform size alignment on the position within the cut segment. If there is a sampling index on the same side of the boundary, supplement it with the sampling index on the same side of the boundary. If the sampling index on the same side of the boundary is insufficient, backfill it with the sampling index near the load unloading boundary. If neither the sampling index on the same side of the boundary nor the sampling index near the load unloading boundary is hit, fill in the mask index. S56. Replace the baseline continuous index arrangement with the rearranged index arrangement, retain the number of positions within the fragment corresponding to the candidate patch fragment size, and form a candidate patch fragment that does not cross the load removal boundary and maintains the fragment size alignment.

[0046] In the specific implementation process, the patch segmentation layer, as the core modification in the improved Pathformer model, includes a baseline continuous index generation unit, a load removal boundary alignment unit, an index cutting unit, an index rearrangement unit, a size alignment unit, and a masked index writing unit. After receiving the deformation prediction object with boundary markers, the patch segmentation layer first generates a baseline continuous index arrangement corresponding to a fixed patch segment size, and then aligns the load removal boundary to the baseline continuous index arrangement. When the load removal boundary falls within the same patch segment, it cuts the cross-boundary continuous index, rearranges the indices on both sides of the boundary and the non-bounced residual index, and forms candidate patch segments through size alignment and masked index writing.

[0047] The baseline continuous index arrangement is generated by the sampling order in the deformation prediction object with boundary markers. The sampling order comes from the bridge deformation response window chain, and the sampling index is numbered continuously starting from the first valid sampling point in the window chain, with the numbering order consistent with the sampling time. In one embodiment, the input window can take 64 sampling points, and the candidate patch fragment size can take 8, 16, and 32 sampling points. The baseline continuous index generation unit generates a set of fragment starting points for each candidate patch fragment size. The fragment starting point starts from the first valid sampling index and moves backward by the number of sampling points corresponding to the current candidate patch fragment size each time. The position within the fragment is sequentially attached with continuous sampling indices starting from the fragment starting point until the current candidate patch fragment size is reached. When the valid sampling points are insufficient for a complete patch fragment, the position within the missing fragment is not filled with sampling values ​​and is reserved for processing by the size alignment unit.

[0048] The load removal boundary alignment unit reads the sampling positions recorded by the boundary markers and maps the boundary markers to the reference continuous index arrangement. When a boundary marker falls between two adjacent patch segments, the two patch segments maintain their original continuous index arrangement, with the segments before and after the boundary treated as independent segments. When a boundary marker falls within the same patch segment, the index cutting unit cuts the continuous index connection within the segment at the sampling position corresponding to the boundary marker. The sampling index before the cutting position is assigned to the index arrangement before the boundary, and the sampling index after the cutting position is assigned to the index arrangement after the boundary. The cutting action does not change the original sampling values, but only changes the assigned position of the sampling index within the patch segment, so that the same candidate patch segment no longer carries continuous indices from both the load application stage and the unloading and rebound stage simultaneously.

[0049] The index rearrangement unit rearranges the cut indexes around the current candidate patch segment size. The indexes before the boundary are filled into the segment positions of the candidate patch segments before the boundary in the original sampling order, and the indexes after the boundary are filled into the segment positions of the candidate patch segments after the boundary in the original sampling order. The non-bounced residual index comes from the sampling position recorded by the residual marker. After reading the residual direction and residual duration, the index rearrangement unit retains the non-bounced residual index in the segment position on the boundary side where the fallback is not yet closed. When a non-bounced residual spans multiple sampling points, the non-bounced residual index is filled into the segment positions sequentially according to the unloaded time order. When the non-bounced residual index and the index after the boundary point point to the same sampling point, the residual marker carried by the non-bounced residual index is retained first, and the index after the boundary is returned to the same sampling order on the boundary side and re-arranged.

[0050] Within the patch segmentation layer, reinforcement learning actions determine the placement of boundary adjacent indices and non-bounced residual indices. The states read by the reinforcement learning action include the boundary front index arrangement, boundary back index arrangement, non-bounced residual index, candidate patch segment size, and the current missing position within the segment. Boundary-side indices corresponding to fallback closure changes are placed along the boundary-side sampling order, while non-bounced residual indices are placed along the residual continuation direction. Boundary-side sampling indices of adjacent structural parts are connected to the missing positions within the same candidate patch segment size. When a reinforcement learning action fails to hit the sampling index of an adjacent structural part, the missing position is not extended to other structural parts but is directly handled by the size alignment unit to prevent unsourced sampling indices from entering the candidate patch segment.

[0051] The size alignment unit performs padding on the positions within the rearranged segment. If available sampling indices still exist on the same side of the boundary, the missing positions are filled with the same-side sampling index; if the same-side sampling index is insufficient, the missing positions are filled with the sampling index adjacent to the load removal boundary. If neither the same-side sampling index nor the adjacent-to-load-removal-boundary sampling index is matched, the mask index writing unit fills in the mask index. In one embodiment, the mask index can be an independent number that does not correspond to any valid sampling point. The mask index only occupies a position within the segment and does not carry deformation values, boundary markers, or residual markers. After padding, each candidate patch segment retains the number of positions within the segment specified by the corresponding candidate patch segment size; for example, a candidate patch segment with 16 sampling points always retains 16 positions within the segment.

[0052] Index replacement is performed at the output of the patch segment partitioning layer. The rearranged index layout covers the intra-segment indices in the baseline continuous index layout that cross the load removal boundary, while segments that do not cross the load removal boundary retain the baseline continuous index layout. After replacement, each candidate patch segment simultaneously carries the segment start point, intra-segment position, sampling index, boundary marker, residual marker, and masked index status. Candidate patch segments do not cross the load removal boundary and maintain a consistent number of intra-segment positions across different candidate patch segment sizes, enabling them to handle intra-segment change extraction and inter-segment change verification.

[0053] The improvement to the Pathformer model lies in the change to the fixed-size index generation logic within the patch segmentation layer. The standard Pathformer model uses the fixed patch size as the length of a continuous time slice, with indices within the slice extending naturally along the time sequence. The improved Pathformer model incorporates the load removal boundary into the index generation process. Continuous indices across the boundary are cut, and the indices before, after, and without rebound residuals are rearranged within the candidate patch segments. This improvement allows the patch segmentation layer to form an internal index rearrangement path capable of recognizing load removal boundaries. This prevents the dual attention layer from simultaneously reading opposite deformation changes during the load application and unloading / rebound stages in the same local segment, and allows without rebound residuals to be retained with independent index positions, reducing prediction drift from within the model itself.

[0054] In this embodiment, S54 specifically includes: S541, the patch segmentation layer reads the baseline continuous index arrangement, the index arrangement before the boundary, the index arrangement after the boundary, the index of the non-bounced residual, and the position within the missing segment under the same candidate patch segment size, and arranges them as the index arrangement to be decided along the segment start point to the segment end; S542. Improve the path weight passing path in the Pathformer model. Read the path weight corresponding to the size of the same candidate patch fragment, remove the direct connection between the path weight and the baseline continuous index arrangement that crosses the load removal boundary, and hand the path weight over to the reinforcement learning action to reconnect it within the index arrangement to be decided. S543. When the fallback direction of the reinforcement learning action check boundary back index is met, and the fallback endpoint is aligned with the deformation position carried by the boundary front index, the path weight is attached to the boundary back index arrangement, and the non-bounced residual index returns to the segment position corresponding to the residual continuation direction. S544. When the reinforcement learning action checks the continuation direction of the non-bounced residual index, and the non-bounced residual index continues in the same direction along the unloaded time sequence, the path weight is attached to the non-bounced residual index, and the fallback index behind the boundary returns to the same sampling sequence on the same side of the boundary. S545. When the boundary same-side sampling index of adjacent structural parts is checked in reinforcement learning, and the boundary same-side sampling index of adjacent structural parts is close to the position in the missing segment in time and the deformation direction is continuous, the path weight is attached to the boundary same-side sampling index of adjacent structural parts, and the boundary same-side sampling index of adjacent structural parts is filled into the position in the missing segment. S546 When the boundary back-side index, the non-bounced residual index, and the boundary same-side sampling index of adjacent structural parts compete for the same segment position, the index with attached path weight is retained, and the index that is not retained is returned to the segment position corresponding to the boundary same-side sampling order or the residual continuity direction. S547. When the path weight has not been reattached, the boundary adjacent index is retained in the original segment position, the non-bounced residual index is retained along the residual continuity direction, and the position in the missing segment is filled by the same side boundary index, the adjacent sampling index or the masking index, forming an index retention arrangement limited by the path weight reattachment.

[0055] In its implementation, the path weight transfer path receives path weights assigned by the multi-scale path selector to the same candidate patch segment size. Internally, it includes a path weight buffer, an index attachment record, an action decision entry point, and an attachment result output point. The path weight buffer stores the path weight value corresponding to each candidate patch segment size. These values ​​are output by the trained multi-scale path selector and, after normalization, fall between 0 and 1. The index attachment record stores the current segment position and sampling index corresponding to the path weight. The action decision entry point receives reinforcement learning actions. The attachment result output point outputs the index placement arrangement after reattaching. During initial path weight attachment, the path weight corresponding to the same candidate patch segment size is written to the segment position within the corresponding candidate patch segment size in the baseline continuous index arrangement. The continuous sampling index carried by the segment position is simultaneously recorded in the index attachment record. The path weight transfer path only changes the attachment object between the path weight and the index arrangement; it does not change the sampling value itself.

[0056] The index arrangement to be decided is organized by the patch segment partitioning layer under the same candidate patch segment size. The baseline continuous index arrangement provides the segment start point and intra-segment position when the original continuous segmentation was performed. The boundary front and boundary rear index arrangements are derived from the sampling indices on both sides after the load removal boundary is cut. The non-bounced residual index is derived from the sampling position recorded by the residual marker. The position within the missing segment is derived from the empty space after index cutting and size alignment. During the organization, the intra-segment positions are arranged sequentially from the segment start point to the segment end. Each intra-segment position records the sampling index, boundary lateral, residual marker, path weight attachment status, and masking status. Positions not written to a valid sampling index are only recorded as intra-segment positions within the missing segment and do not participate in the path weight attachment calculation.

[0057] After reading the path weights corresponding to the same candidate patch segment size, the path weight propagation process first checks whether the sampling index currently attached to the path weight crosses the load removal boundary. The check method compares the lateral boundary directions corresponding to two adjacent sampling indices within the segment. If the lateral boundary direction changes, the path weight is considered to be attached to a continuous index arrangement that crosses the boundary. When a direct attachment is removed, the original sampling index in the index attachment record remains within the index arrangement to be decided. The path weight is stripped from the original sampling index and written to the action decision entry. The stripped path weight retains its original value and only waits for the reinforcement learning action to reassign the attachment object.

[0058] The state of a reinforcement learning action is composed of the boundary front index arrangement, boundary back index arrangement, non-bounce residual index, candidate patch fragment size, path weight value, and position within the missing fragment. The action set includes fallback attachment actions, residual attachment actions, adjacent part attachment actions, and retention actions. The reward is obtained by adding three parts: the decrease in prediction error after path weight attachment is the main reward; the residual reward is increased when the non-bounce residual index falls into the fragment position corresponding to the residual continuity direction; and the masking penalty is deducted when the number of masking indices increases. In one embodiment, the weights of the main reward, residual reward, and masking penalty can be 0.6, 0.3, and 0.1, respectively. When updating the action value, the sum of the current reward and the next round's maximum action value multiplied by a discount factor is first calculated, and then this sum is used to make a weighted average with the current action value; the learning rate can be 0.05, and the discount factor can be 0.9. When the action value changes by less than 1% for 5 consecutive rounds, the action decision result under the current candidate patch fragment size remains stable.

[0059] The fallback and reattachment action is performed on the boundary-back index arrangement. The reinforcement learning action checks the fallback direction of the boundary-back index, which is determined by the sign of the difference between adjacent sampled values. If the difference between the fallback endpoint and the deformed position carried by the boundary-front index does not exceed twice the sensor noise amplitude, the fallback endpoint is considered to be aligned with the boundary-front index. When the fallback endpoint is aligned, the path weights are reattached to the boundary-back index arrangement. The boundary-back index occupies its corresponding position in the segment according to the original sampling order. The non-bounced residual index does not compete with the boundary-back index for the same position, but instead retreats to the position in the segment corresponding to the residual's continuity direction.

[0060] The residual attachment action is performed on the non-bounced residual index. The reinforcement learning action checks the continuity direction of the non-bounced residual index in the time sequence after unloading. If the continuous sampling differences maintain the same sign and the absolute value of the residual value is consistently higher than twice the sensor noise amplitude, the non-bounced residual index is considered to be continuing in the same direction. When the same direction continuity is satisfied, the path weight is reattached to the non-bounced residual index. The non-bounced residual index occupies the position in the continuous segment according to the sampling order within the residual continuity range. The index falling back to the boundary side returns to the sampling order on the same side of the boundary and continues to be used as the available index on the same side of the boundary, waiting to be called during the size alignment process.

[0061] The adjacent component linkage action is executed for the sampling indexes of adjacent structural components under the same load removal boundary. The reinforcement learning action first checks the temporal continuity between the sampling indexes on the same side of the boundary of the adjacent structural components and the position within the missing segment. If the time difference does not exceed the same sampling interval, the temporal continuity is considered to be consistent. Next, the deformation direction is checked; if the signs of the adjacent sampling differences of the two structural components are consistent and the overlap in their continuous range reaches 60%, the deformation direction is considered continuous. When both checks are met, the path weight is re-linked to the sampling indexes on the same side of the boundary of the adjacent structural components, and the corresponding sampling index is filled into the position within the missing segment. If either check is not met, the sampling indexes of the adjacent structural components are returned to the original structural component response group and are not written into the current candidate patch segment.

[0062] When contending for a position within the same segment, the path weight propagation process first compares whether the contending index has completed path weight reattachment. Indexes that have completed path weight reattachment are retained at the contended position, while those that haven't are returned to their corresponding sources: indices behind the boundary are returned to the sampling order on the same side of the boundary; indices with non-bounced residuals are returned to the position within the segment corresponding to the residual's continuity direction; and indices sampled on the same side of the boundary of adjacent structural parts are returned to the response group of the adjacent structural parts. When both indices have completed path weight reattachment, the index with the larger path weight value is retained, and the index with the smaller path weight value is returned to its corresponding source. When path weight values ​​are the same, the index with non-bounced residuals is retained first to ensure that continuous residual changes are not covered by boundary fallback changes.

[0063] When path weights are not reattached, the action is retained to take over the index arrangement pending adjudication. Boundary adjacent indexes are retained within their original positions within the segment, and non-bounced residual indexes are retained along the residual continuation direction. Positions within missing segments are filled in the order of same-side boundary index, nearest sampling index, and masked index. Same-side boundary indexes are from the most recent valid sampling index on the same side of the same load removal boundary; nearest sampling indexes are from the most recent valid sampling index at the same structural location that does not cross the load removal boundary; if neither type of index is matched, the masked index is used. After completion, the path weights are passed to the output index retention arrangement. The index retention arrangement records the sampling index, path weight attachment status, and masked status corresponding to each position within the segment, and returns to the candidate patch segment formation process.

[0064] The improvement to the Pathformer model in terms of path weight propagation lies in transforming path weights from scale-selected weights for fixed continuous patch segments to internal weight objects that can be re-selected and attached to objects as the load is removed. In the ordinary Pathformer model, path weights are mainly allocated according to the size of candidate patch segments, and path weights are synchronously bound to the sampling indices within continuous patch segments. The improved Pathformer model, after the cross-boundary index is cut, decouples the path weights from the direct binding to the cross-boundary sampling index and re-attaches the path weights to the fallback side index, the non-bounced residual index, or the same-side index of adjacent structural parts. With this improvement, the target of path weights is expanded from a simple time scale to the boundary index retention object, enabling the model to retain residual indices that need to be modeled continuously at the load removal boundary, release index positions that have already fallen back and closed, and improve the ability to retain non-bounced residuals and distinguish the bounce stage.

[0065] In this embodiment, S6 specifically includes: S61. The multi-scale path selector reads the candidate patch fragment size, path weight, and mask index of the candidate patch fragment, retains the candidate patch fragment size that has completed path weight reattachment, and pairs the candidate patch fragment size with the mask index that has not completed path weight reattachment. S62. The dual attention layer reads the position within the fragment according to the size of the candidate patch fragment, extracts the deformation changes on the same side of the boundary, and sets the position within the fragment paired with the mask index as the mask position in the attention weight calculation to form the fragment change feature. S63. The dual attention layer verifies the segment start point, load removal boundary and non-bounced residual index of adjacent candidate patch segments. The intra-segment variation features of segments with the same boundary and continuous deformation direction are retained as inter-segment variation features, and the intra-segment variation features of segments that cross the load removal boundary and have opposite directions are paired with the masking index. S64. The aggregator reads the intra-fragment variation features, inter-fragment variation features, and path weights for reattaching the fragments. It merges the deformation changes on the same side of the boundary along the candidate patch fragment size to form effective multi-scale features. S65. The predictor reads the segment positions of the effective multi-scale features and the masked index pairings, retains the prediction components corresponding to the effective multi-scale features, reduces the prediction components corresponding to the segment positions of the masked index pairings, and reads out the real-time prediction results of bridge deformation.

[0066] In practical implementation, the multi-scale path selector receives candidate patch fragments and internally includes a candidate size reading unit, a path weight reading unit, a mask index verification unit, and a size retention unit. The candidate size reading unit reads the patch fragment size carried by the candidate patch fragment; the path weight reading unit reads the path weight output by the path weight transmission path; the mask index verification unit checks whether a mask index exists within the candidate patch fragment; and the size retention unit outputs the size of the candidate patch fragment after path weight reattachment. In one embodiment, the candidate patch fragment size can take 8, 16, or 32 sampling points, and the path weight, after normalization, falls between 0 and 1. The size of the candidate patch fragment that has completed reattachment is retained in the current calculation batch, while the size of the candidate patch fragment that has not completed reattachment is paired with a mask index and retained as the size of the candidate patch fragment whose path weight has not completed reattachment.

[0067] The determination of whether path weight reattachment is complete is provided by the index retention arrangement. The path weight attachment status is recorded for each position within a valid segment of a candidate patch. When all positions within valid segments have completed path weight reattachment, the current candidate patch size is entered into the size retention unit. If there are unattached positions within segments, these unattached positions are written to the mask index. The sizes of candidate patch sizes whose path weights have not been reattached are not directly discarded, but paired with the mask index for masking in the attention weight calculation by the dual attention layer. This process ensures that unattached segments do not generate sourceless features, while retaining the calculation channels for valid sampling positions.

[0068] The dual attention layer internally includes an intra-segment attention calculation unit, an inter-segment attention calculation unit, and a masking position processing unit. The intra-segment attention calculation unit reads the positions within a segment according to the candidate patch segment size. It first concatenates the deformation value, boundary marker, residual marker, and sampling position encoding of each position within the segment into an intra-segment feature vector. Then, it performs three sets of linear mappings to obtain the query vector, key vector, and value vector, respectively. The linear mapping parameters are obtained through model training. In one embodiment, the intra-segment feature dimension can be 128, the number of attention heads can be 4, and the dimension of each attention head can be 32. The query vector and key vector are multiplied dimension-wise and accumulated, then divided by the square root of the attention head dimension to form the attention score. The attention scores within the same candidate patch segment are normalized and used as weighting coefficients for the value vector. The value vectors are accumulated according to the weighting coefficients to form the intra-segment variation features.

[0069] The masking position processing unit processes the intra-segment positions paired with masking indices. The masking index does not carry deformed values, boundary markers, or residual markers. Before attention score normalization, the attention score of the intra-segment position corresponding to the masking index is written into the masking value. In one embodiment, the masking value can be -10000, making the normalized attention weights close to 0; the value vector corresponding to the masking position does not participate in the weighted accumulation. The masking index only affects the attention weight calculation within the current candidate patch segment, without changing the original sampling records or the record positions of boundary markers and residual markers in the index placement arrangement.

[0070] The deformation change on the same side of the boundary is determined by the lateral boundary direction carried by the position within the segment. When the position within the segment is located on the same side of the load removal boundary and the signs of the differences between adjacent sampled values ​​are consistent, the deformation change corresponding to the current position within the segment enters the attention calculation unit within the segment; when the position within the segment crosses the load removal boundary, or the signs of the differences between adjacent sampled values ​​are opposite on both sides of the boundary, the current position within the segment is paired with the masking index. The amplitude of the deformation change on the same side of the boundary is obtained by subtracting the previous sampled value from the current sampled value, and the direction is determined by the sign of the difference. When the amplitude is lower than the sensor noise amplitude, it is treated as a stationary change.

[0071] The inter-segment attention calculation unit checks adjacent candidate patch segments. The checking order is as follows: first, read the segment start point of the adjacent candidate patch segments; then, read the load removal boundary position between the two segment start points; and finally, read whether the non-rebound residual index falls within the boundary range between the two segments. When two candidate patch segments are located on the same side of the load removal boundary and the main direction of the variation characteristics within the two segments is continuous, the inter-segment attention calculation unit retains the variation characteristics within the two segments as inter-segment variation characteristics. When two candidate patch segments cross the load removal boundary and have opposite main directions, the variation characteristics within the cross-boundary segment are paired with the masking index to avoid opposite deformation stages participating in the same inter-segment attention calculation.

[0072] The direction of deformation continuity is determined by the principal direction in the variation characteristics within a segment. The principal direction is determined by the number of signs of the sampling differences within each candidate patch segment. When the proportion of positive differences is greater than the proportion of negative differences, the principal direction is expansion; when the proportion of negative differences is greater than the proportion of positive differences, the principal direction is decline; when the difference between the proportions of the two types is less than 10%, the principal direction is considered stationary. Adjacent candidate patch segments are considered to have the same principal direction or to smoothly transition from expansion to stationary when the principal direction is the same. When one candidate patch segment is expansion and another candidate patch segment is decline, and there is a load unloading boundary between the two segments, the directions are considered to be opposite.

[0073] The aggregator internally includes an intra-segment feature receiving unit, an inter-segment feature receiving unit, a path weight multiply-accumulate unit, and a multi-scale feature output unit. The intra-segment feature receiving unit receives intra-segment variation features, the inter-segment feature receiving unit receives inter-segment variation features, and the path weight multiply-accumulate unit reads the reattached path weights. During aggregation calculation, intra-segment and inter-segment variation features of the same candidate patch segment size are first concatenated by channel and then multiplied by the corresponding path weights. Weighted features of different candidate patch segment sizes are accumulated in ascending order of size, and the path weights are normalized to a sum of 1 before accumulation. Features paired with masked indexes do not participate in path weight multiply-accumulate; only the masked position record is retained.

[0074] The effective multi-scale features are output by the aggregator. Each row of the effective multi-scale features corresponds to an effective sampling position, and each column corresponds to the aggregated feature channels. In one implementation, the effective multi-scale feature dimension can be 64×128, where 64 corresponds to the number of retained effective sampling positions, and 128 corresponds to the aggregated channel dimension. Masked positions in candidate patch fragments are not written into the effective multi-scale feature rows; only the original fragment position number is retained in the masked position record for the predictor to reduce the corresponding predicted components.

[0075] The predictor internally includes a feature reading unit, a masking component reduction unit, and a linear mapping unit. The feature reading unit reads valid multi-scale features and arranges them in the order of sampling time. The linear mapping unit multiplies the 128-dimensional feature vector corresponding to each valid sampling position with the prediction weight dimension by dimension and accumulates the results, then adds a bias term to obtain the prediction component corresponding to the current valid sampling position. When the prediction component comes from a position within a segment paired with a masking index, the masking component reduction unit multiplies the prediction component by a reduction coefficient; in one embodiment, the reduction coefficient can be 0.1, so that the prediction component across opposite deformation stages retains only low-intensity influence.

[0076] The real-time bridge deformation prediction result is obtained by the predictor aggregating the prediction components from valid sampling locations. The predictor accumulates each valid prediction component in the order of sampling locations within the prediction period, and retains the structural location attribution, prediction period, predicted deformation direction, and predicted deformation amplitude. The predicted deformation direction is determined by the sign of the prediction component, and the predicted deformation amplitude is determined by the maximum absolute value of the prediction component within the prediction period. The absolute value of the last prediction component is added to the real-time bridge deformation prediction result as an additional record of the end position of the prediction period. The position within the segment paired with the masked index only participates in the calculation of the reduced prediction component and does not participate in the main direction statistics of the predicted deformation direction. The output real-time bridge deformation prediction result inherits the structural location attribution and time continuity position in the bridge deformation response window chain.

[0077] In this embodiment, S7 specifically includes: S71. Retrieve the real-time prediction results of bridge deformation and the bridge deformation response window chain, and verify the corresponding status of the prediction period, predicted deformation direction, predicted deformation amplitude and structural part attribution. S72. The time connection position along the bridge deformation response window chain is aligned with the prediction period, and the predicted deformation direction and predicted deformation amplitude are back-attached to the bridge structural parts with the same affiliation. S73. Merge the posting results that are continuous in time and have connected deformation directions within the same bridge structural part to form a structural part deformation prediction segment. S74. Reply results with conflicting structural part attributions or disconnected prediction periods are retained as prediction results pending review, and the corresponding bridge deformation response window chain is attached. S75. Merge the deformation prediction segments of structural parts and the prediction results to be reviewed to form a real-time prediction record of bridge deformation, and write it back to the bridge monitoring archive.

[0078] In the specific implementation process, after the real-time bridge deformation prediction results are output by the predictor, they are first checked against the bridge deformation response window chain. The prediction period is determined by the start and end sampling numbers output by the predictor, the predicted deformation direction is determined by the positive and negative directions of the predicted components, and the predicted deformation amplitude is determined by the maximum absolute value of the predicted components within the prediction period. The absolute value of the end predicted component is added to the bridge deformation real-time prediction record as an additional record of the end position of the prediction period. The structural parts are assigned to the paired part response groups in the bridge deformation response window chain. During the check, the start and end sampling numbers of the prediction period are matched one by one with the time continuation positions in the window chain, and the predicted deformation direction is compared with the undulation direction of the same part in the window chain. When the sampling numbers are within the same time continuation range and the structural parts are assigned to the same location, the prediction results enter the back-matching process. When the structural parts are assigned to different locations, the prediction results enter the pending verification prediction results.

[0079] The alignment of the predicted time period with the corresponding time continuation position is performed using the difference in sampling sequence numbers. In one embodiment, the sampling interval can be 1 second. If the deviation between the start and end sampling sequence numbers of the predicted time period and the corresponding time continuation position in the bridge deformation response window chain does not exceed one sampling interval, the predicted time period is considered to be aligned. If the deviation exceeds one sampling interval, the prediction result is not re-attached to the structural part, but is retained as a prediction result to be verified. During re-attachment, the predicted deformation direction and predicted deformation amplitude are written into the record of the bridge structural part to which the same structure belongs. The written content includes the structural part number, the predicted time period, the predicted deformation direction, the predicted deformation amplitude, the corresponding load removal boundary, and the corresponding bridge deformation response window chain number. When the predicted deformation amplitude is lower than the sensor noise amplitude, it is recorded as a low-amplitude prediction result and retained in the structural part record for trend verification, and is not separately marked as an abnormal result.

[0080] The deformation prediction segment for a structural component is formed by merging consecutive back-tracking results within the same bridge structural component. During merging, the components are first grouped by structural component number, then arranged in ascending order by the starting sampling sequence number of the prediction period. The interval between two adjacent prediction periods does not exceed one sampling interval, and when the predicted deformation trends are the same or transition from expansion to stabilization, the two back-tracking results are merged into the same structural component deformation prediction segment. The merged predicted deformation amplitude is taken as the maximum absolute value within the merged range, and the prediction period is taken from the earliest starting sampling sequence number to the latest ending sampling sequence number. The load removal boundary retains the boundary position corresponding to the maximum predicted deformation amplitude. When the prediction periods are consecutive but the predicted deformation trends are opposite, the back-tracking results are not merged and are retained as two separate structural component deformation prediction segments.

[0081] The prediction results pending review are used to store results with structural component attribution conflicts, prediction time period breaks, or missing window chains. Structural component attribution conflicts refer to a mismatch between the structural component number corresponding to the prediction result and the component response group number in the bridge deformation response window chain; prediction time period breaks refer to a deviation between the prediction time period and the time transition position exceeding one sampling interval, or the presence of missing sampling markers in between; missing window chains refer to the prediction result failing to find the corresponding load removal boundary and bridge deformation response window chain number. The prediction results pending review retain the predicted deformation direction, predicted deformation amplitude, prediction time period, conflict type, and original bridge deformation response window chain number, and are entered into the review area of ​​the bridge monitoring archive, without overwriting structural component deformation prediction segments that have already been re-attached.

[0082] The real-time bridge deformation prediction record is formed by merging the structural component deformation prediction segment and the prediction results to be reviewed. During record generation, the structural component deformation prediction segment is written first, followed by the prediction results to be reviewed, and finally the structural component prediction record, load removal boundary record, and review area record in the bridge monitoring archive are updated. The bridge monitoring archive retains the association between the current round of prediction records and the previous round of prediction records. The association method is the connection between the same structural component number and the sampling sequence number of adjacent prediction periods. When the predicted deformation amplitude in two adjacent rounds of prediction records for the same structural component is higher than the predicted deformation amplitude in the previous round, the archive updates the trend marker; when the predicted deformation amplitude falls back to within the sensor noise amplitude, the trend marker is removed. After the write-back is completed, the bridge monitoring archive retains traceable prediction periods, structural components, predicted deformation amplitudes, load removal boundaries, and review status.

[0083] Example 1: To verify the feasibility of this invention in practice, it was applied to an online deformation prediction system for a long-span continuous beam bridge. This bridge is subjected to the combined effects of vehicle loads, temperature changes, and wind loads over a long period. The mid-span deflection and beam end displacement increase rapidly when vehicles pass through densely, and then decrease after the load is removed. The original prediction system uses a fixed time window to divide the monitoring sequence, which easily compresses the load application stage and the unloading and rebound stage into the same prediction segment. This results in some reboundable deformations being predicted continuously, while some non-rebound residuals are weakened by the normal decline process. In the field, this manifests as a significant tailing of the prediction curve after the peak load, and unstable identification of local residual deformations. Especially in the case of continuous vehicle passage followed by a short period of low load, the fixed window often mixes the tail of the previous load with the next unloading and rebound stage, causing the prediction curve to remain high during the decline stage. Monitoring personnel need to repeatedly check against the original curve.

[0084] In application, the deflection, displacement, and tilt responses uploaded from bridge structural components are first aggregated by structural component. Synchronous fluctuations under the same load are clustered into deformation segments of the same load source. Then, the load removal boundary is determined and compressed into a bridge deformation response window chain. The system continues to verify the fallback closure state after unloading. Changes not occurring with the fallback are retained as non-bounced residuals and encapsulated with the load removal boundary as a deformation prediction object with boundary markers. After the deformation prediction object is integrated into the improved Pathformer model, the patch segment division layer no longer continuously generates indices along a fixed patch segment size. Instead, it cuts the baseline continuous index arrangement at the load removal boundary and uses reinforcement learning actions to determine the placement direction of the boundary adjacent indexes and the non-bounced residual indexes, forming candidate patch segments that do not cross the load removal boundary and maintain segment size alignment. This is then used to complete real-time deformation prediction and structural component re-attaching.

[0085] To verify the feasibility of this invention in practice, 7200 valid prediction windows collected during continuous operation were selected as verification samples, with a sampling interval of 1 second. Each prediction window retained the corresponding structural part attribution, load removal boundary, and residual deformation state after manual verification. All three methods used the same training, validation, and test sets, with a prediction step size of 60 seconds. The hardware environment and data cleaning methods were kept consistent. The fixed-window long short-term memory network method, the original Pathformer model method, and the method of this invention were compared, and the statistical results are shown in Table 1 below.

[0086] Table 1 Comparison of Real-time Prediction Results for Bridge Deformation

[0087] As can be seen from the data in Table 1 above, the method of the present invention outperforms the fixed-window long short-term memory network method and the original Pathformer model method in terms of bridge deformation prediction accuracy, post-unloading prediction stability, and non-rebound residual identification capability. The fixed-window long short-term memory network method has an average absolute error of 1.36 mm, a root mean square error of 2.05 mm, and an average absolute percentage error of 8.70%, indicating that the fixed window lacks distinction between the load application stage and the unloading and rebound stage, resulting in a large overall prediction deviation. The original Pathformer model reduces the average absolute error to 0.94 mm and the root mean square error to 1.41 mm, indicating that multi-scale time series modeling can improve the prediction effect. However, the post-unloading prediction drift is still 0.76 mm, and the non-rebound residual identification rate is 81.30%, indicating that the continuous indexing of the fixed patch segment size still pushes the changes across the load unloading boundary into the same modeling segment. The method of this invention reduces the mean absolute error to 0.61 mm, the root mean square error to 0.92 mm, the mean absolute percentage error to 3.90%, the prediction drift after unloading to 0.34 mm, and the non-springback residual recognition rate to 91.80%, indicating that this invention can more accurately distinguish between springback deformation and non-springback residuals. Although the single-window inference time of this invention is 24.10 ms, slightly longer than the comparative method, it still meets the requirements of real-time prediction and yields more stable prediction results and stronger residual deformation recognition capabilities.

[0088] This embodiment organizes the real-time monitoring response of the bridge into deformation segments of the same load source, and forms a bridge deformation response window chain around the load removal boundary. This allows the originally scattered changes in deflection, displacement, and tilt to be organized according to structural parts and the load-bearing process. Based on this, the fallback closure state after unloading is verified, and changes that do not fall back with unloading are retained as non-bounced residuals, providing clearer boundary marker objects for the model. The improved Pathformer model rewrites the continuous index generation path of fixed patch segment size within the patch segment partitioning layer. When encountering a position within a segment that crosses the load removal boundary, it no longer maintains continuous index connection, but instead uses reinforcement learning actions to determine the placement of the boundary adjacent index and the non-bounced residual index, and uses the same-side boundary index or the masked index to supplement the cross-boundary position. After the above processing, the candidate patch segments can maintain size alignment, while avoiding the mixed modeling of the load application stage and the unloading and rebound stage in the same patch segment. This makes the extraction of changes within the segment, the verification of changes between segments, and the prediction readout more closely match the actual deformation process of the bridge, thereby improving the prediction reliability and operation and maintenance early warning support capabilities in the long-term online monitoring scenario of the bridge.

[0089] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for real-time prediction of bridge deformation based on reinforcement learning, characterized in that, Includes the following steps: S1. Collect the real-time monitoring response of the bridge according to the structural parts, and consolidate the synchronous fluctuations of deformation under the same load into deformation segments of the same load. S2. Determine the load removal boundary along the deformation segment of the same load source, and compress the deformation bearing relationship on both sides of the boundary into a bridge deformation response window chain. S3. Verify the closed state of the bridge deformation response window chain after unloading and retain the changes that do not follow the unloading and fall as non-rebound residuals, and encapsulate them as deformation prediction objects with boundary markers. S4. Integrate the deformation prediction object into the improved Pathformer model, and the rewriting position of the path generated by the fixed patch fragment size continuous index falls on the patch fragment partitioning layer. S5, the patch segmentation layer takes over the continuous index generation path of fixed patch segment size, arranges the starting point and position within the segment around the candidate patch segment size, and cuts off the continuous index connection when the position within the segment crosses the load unloading boundary. The reinforcement learning action limits the retention relationship between the adjacent index of the boundary and the non-bounced residual index. The position that crosses the boundary is supplemented by the same side boundary index or the masking index to form candidate patch segments that maintain the alignment of the segment size. S6. Extract intra-fragment changes and verify inter-fragment changes for candidate patch segments, retain deformation changes on the same side of the boundary as effective multi-scale features, mask changes across opposite deformation stages, and read out the real-time prediction results of bridge deformation. S7. Paste the real-time prediction results of bridge deformation back to the structural parts corresponding to the bridge deformation response window chain, merge them into the real-time prediction record of bridge deformation and write them back to the bridge monitoring archive.

2. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, S1 specifically includes: S11. Verify the installation attribution of the bridge real-time monitoring response and the bridge structural parts, and group the deformation responses with the same attribution and continuous sampling time into the same part response group. S12. Align the sampling time axis of response groups in different parts, cut off isolated responses with time breaks, and retain the deformed responses that can be aligned within the same time range as the response groups to be verified. S13. Compare the undulation direction, undulation starting point and duration within the response group to be verified. Classify the deformation changes that rise synchronously, fall synchronously, or stop synchronously into the same load response group, and return the deformation changes that have opposite undulation directions or discontinuous duration to the corresponding part response group. S14. The deformation changes that continuously participate in synchronous fluctuations in the response group of the same source load are gathered together, and the static changes before and after the load are removed to form the deformation segment of the same source load.

3. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, S2 specifically includes: S21. Compare the deformation changes at adjacent sampling times along the time sequence of deformation segments of the same load source to identify candidate unloading sampling points where the deformation changes change from continuous expansion to continuous convergence. S22. Verify the adjacency relationship between the candidate unloading sampling point and the previous peak response, and check the fall direction and fall duration after the candidate unloading sampling point to determine the load unloading boundary. S23. Use the load removal boundary to fit the deformation segments of the same source load before and after, retain the deformation changes that are continuously connected in time between the deformation pushing direction before the boundary and the falling direction after the boundary, and return the deformation changes that cannot be fitted due to time breaks or abrupt changes in direction to the deformation segments of the same source load. S24. Gather the deformation changes that are attached front and back, and compress the continuous deformation changes at the boundary front and back into a bridge deformation response window chain.

4. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, S3 specifically includes: S31. Read the load removal boundary in the bridge deformation response window chain, and extract the adjacent response before the load removal boundary and the fallback response after the load removal boundary. S32. Using the adjacent response on the front side of the load removal boundary as the reference for the fall-off closure, verify the fall-off direction, fall-off endpoint, and duration of the fall-off response on the rear side of the load removal boundary. S33. Deformation changes in which the fall direction is toward the fall closing reference and the fall endpoint is in contact with the fall closing reference are calibrated as fall closing changes. Deformation changes in which the fall direction deviates from the fall closing reference or the fall endpoint is not in contact with the fall closing reference are calibrated as non-rebound residuals. S34. Strip the fallback closure change in the bridge deformation response window chain, pair and encapsulate the load removal boundary, non-rebound residual, structural part attribution and time bearing position to form a deformation prediction object with boundary markings.

5. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, The improved Pathformer model includes an instance normalization layer, a patch segmentation layer, a dual attention layer, a multi-scale path selector, a path weight propagation path, an aggregator, and a predictor.

6. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, S5 specifically includes: S51. Read the deformation prediction object with boundary markers by dividing the patch fragment into layers. Take the sampling order of the bridge deformation response window chain as the reference continuous index arrangement corresponding to the fixed patch fragment size. The fragment starting point in the reference continuous index arrangement is expanded in sequence according to the fixed patch fragment size. The position in the fragment is attached to the continuous sampling index according to the sampling order. S52. Align the load removal boundary to the reference continuous index arrangement. When the load removal boundary falls between adjacent patch segments, preserve the continuous index connection of adjacent patch segments. When the load removal boundary falls inside the same patch segment, cut the continuous index arrangement within the same patch segment into a boundary front index arrangement and a boundary back index arrangement. S53. Rearrange the cut indexes around the size of the candidate patch segment. Fill the indexes before the boundary into the positions of the segments before the boundary according to the original sampling order. Fill the indexes after the boundary into the positions of the segments after the boundary according to the original sampling order. The indexes of the non-bounced residuals are retained in the positions of the segments on the side of the boundary that have not closed after falling back. S54. The retention destination of the boundary adjacent index and the non-bounced residual index of the reinforcement learning action decision boundary is determined. The boundary same-side index corresponding to the fallback closure change is retained along the boundary same-side sampling order. The non-bounced residual index is retained along the residual continuity direction. The boundary same-side sampling index of adjacent structural parts is connected to the vacancy position under the same candidate patch segment size. S55. Perform size alignment on the position within the cut segment. If there is a sampling index on the same side of the boundary, supplement it with the sampling index on the same side of the boundary. If the sampling index on the same side of the boundary is insufficient, backfill it with the sampling index near the load unloading boundary. If neither the sampling index on the same side of the boundary nor the sampling index near the load unloading boundary is hit, fill in the mask index. S56. Replace the baseline continuous index arrangement with the rearranged index arrangement, retain the number of positions within the fragment corresponding to the candidate patch fragment size, and form a candidate patch fragment that does not cross the load removal boundary and maintains the fragment size alignment.

7. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 6, characterized in that, Specifically, S54 includes: S541, the patch segmentation layer reads the baseline continuous index arrangement, the index arrangement before the boundary, the index arrangement after the boundary, the index of the non-bounced residual, and the position within the missing segment under the same candidate patch segment size, and arranges them as the index arrangement to be decided along the segment start point to the segment end; S542. Improve the path weight passing path in the Pathformer model. Read the path weight corresponding to the size of the same candidate patch fragment, remove the direct connection between the path weight and the baseline continuous index arrangement that crosses the load removal boundary, and hand the path weight over to the reinforcement learning action to reconnect it within the index arrangement to be decided. S543. When the fallback direction of the reinforcement learning action check boundary back index is met, and the fallback endpoint is aligned with the deformation position carried by the boundary front index, the path weight is attached to the boundary back index arrangement, and the non-bounced residual index returns to the segment position corresponding to the residual continuation direction. S544. When the reinforcement learning action checks the continuation direction of the non-bounced residual index, and the non-bounced residual index continues in the same direction along the unloaded time sequence, the path weight is attached to the non-bounced residual index, and the fallback index behind the boundary returns to the same sampling sequence on the same side of the boundary. S545. When the boundary same-side sampling index of adjacent structural parts is checked in reinforcement learning, and the boundary same-side sampling index of adjacent structural parts is close to the position in the missing segment in time and the deformation direction is continuous, the path weight is attached to the boundary same-side sampling index of adjacent structural parts, and the boundary same-side sampling index of adjacent structural parts is filled into the position in the missing segment. S546 When the boundary back-side index, the non-bounced residual index, and the boundary same-side sampling index of adjacent structural parts compete for the same segment position, the index with attached path weight is retained, and the index that is not retained is returned to the segment position corresponding to the boundary same-side sampling order or the residual continuity direction. S547. When the path weight has not been reattached, the boundary adjacent index is retained in the original segment position, the non-bounced residual index is retained along the residual continuity direction, and the position in the missing segment is filled by the same side boundary index, the adjacent sampling index or the masking index, forming an index retention arrangement limited by the path weight reattachment.

8. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, S6 specifically includes: S61. The multi-scale path selector reads the candidate patch fragment size, path weight, and mask index of the candidate patch fragment, retains the candidate patch fragment size that has completed path weight reattachment, and pairs the candidate patch fragment size with the mask index that has not completed path weight reattachment. S62. The dual attention layer reads the position within the fragment according to the size of the candidate patch fragment, extracts the deformation changes on the same side of the boundary, and sets the position within the fragment paired with the mask index as the mask position in the attention weight calculation to form the fragment change feature. S63. The dual attention layer verifies the segment start point, load removal boundary and non-bounced residual index of adjacent candidate patch segments. The intra-segment variation features of segments with the same boundary and continuous deformation direction are retained as inter-segment variation features, and the intra-segment variation features of segments that cross the load removal boundary and have opposite directions are paired with the masking index. S64. The aggregator reads the intra-fragment change features, inter-fragment change features, and path weights for reattaching the fragments. It then merges the deformation changes on the same side of the boundary along the candidate patch fragment size to form effective multi-scale features. S65. The predictor reads the segment positions of the effective multi-scale features and the masked index pairings, retains the prediction components corresponding to the effective multi-scale features, reduces the prediction components corresponding to the segment positions of the masked index pairings, and reads out the real-time prediction results of bridge deformation.

9. The method for real-time prediction of bridge deformation based on reinforcement learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Retrieve the real-time prediction results of bridge deformation and the bridge deformation response window chain, and verify the corresponding status of the prediction period, predicted deformation direction, predicted deformation amplitude and structural part attribution. S72. The time connection position along the bridge deformation response window chain is aligned with the prediction period, and the predicted deformation direction and predicted deformation amplitude are back-attached to the bridge structural parts with the same affiliation. S73. Merge the posting results that are continuous in time and have connected deformation directions within the same bridge structural part to form a structural part deformation prediction segment. S74. Reply results with conflicting structural part attributions or disconnected prediction periods are retained as prediction results pending review, and the corresponding bridge deformation response window chain is attached. S75. Merge the deformation prediction segments of structural parts and the prediction results to be reviewed to form a real-time prediction record of bridge deformation, and write it back to the bridge monitoring archive.

Images