Advanced early warning method for settlement trend of bridge abutment construction based on convolutional neural network

Abstract

The invention discloses a bridge abutment construction settlement trend advanced early warning method based on a convolutional neural network, and relates to the technical field of intelligent early warning, and the method comprises the steps: collecting settlement observation data of a settlement measurement point and construction process coding data, carrying out the preprocessing to obtain settlement supplementing data and mask data, and obtaining filling height data; establishing a space grid and a reversible mapping relation, selecting representative measuring points, performing rearrangement to obtain grid settlement data and a grid mask, and determining a bridge abutment reference position; generating a first branch input and a second branch input, combining the construction process coding data and the filling height data to construct a gating vector, and obtaining a sample set; and inputting the double-branch convolutional neural network to output future grid settlement increment prediction data, grading to obtain an early warning grade, and outputting measuring point positioning information for triggering the early warning grade through a reversible mapping relation. According to the method, the space interpretability and the positioning accuracy of the settlement evolution process are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent early warning technology, and in particular to an early warning method for bridge abutment construction settlement trends based on convolutional neural networks. Background Technology

[0002] Settlement monitoring during bridge abutment construction typically relies on methods such as leveling, hydrostatic leveling, or GNSS to obtain settlement time series at measuring points. This data is then combined with construction procedure records to conduct trend analysis. In engineering practice, conventional methods usually employ statistical indicators such as single-point settlement, settlement rate, and corresponding cumulative changes, combined with time series regression or empirical models for phased analysis. When considering the spatial influence range, multi-point data is expressed regionally using plane coordinates for settlement evolution assessment under conditions such as abutment backfilling and roadbed compaction.

[0003] In existing technologies, when the distribution of settlement monitoring points is irregular or there are missing measurements, single-point indicators or regional expressions are difficult to characterize the local differences and coupling relationships of the settlement field within a unified spatial framework, which reduces the quantitative characterization and location stability of abnormal areas. In addition, construction procedures and filling conditions can modulate settlement evolution. Using only the settlement time sequence as input makes it difficult to form a dynamic characteristic expression that changes with the construction conditions, thus affecting the graded early warning at future steps. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for early warning of settlement trend during bridge abutment construction based on convolutional neural networks to solve the problems of existing technologies, such as the difficulty in uniformly depicting local differences and coupling relationships in the settlement field under conditions of missing measurements and irregular locations, and the insufficient expression of dynamic features due to the lack of sufficient integration of construction procedures and filling status.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides a method for early warning of bridge abutment construction settlement trends based on convolutional neural networks. The method includes: collecting settlement observation data and construction procedure coding data from settlement monitoring points; preprocessing the settlement observation data and construction procedure coding data to obtain supplementary settlement data and mask data; and obtaining filling height data based on the collected data. Based on the plane coordinates of the settlement monitoring points, a spatial grid and reversible mapping relationship are established, and representative monitoring points are selected. The supplementary settlement data is rearranged into grid settlement data based on the representative monitoring points, and the mask data is simultaneously rearranged into a grid mask, and the bridge abutment reference position is determined. An absolute settlement sequence is extracted from the grid settlement data and concatenated with the grid mask to generate the first branch input. The input time window is then analyzed. A differential settlement dictionary is generated, and a differential mask is calculated based on the grid mask. The differential settlement dictionary and the differential mask are concatenated to generate the second branch input. A gating vector is constructed by combining construction procedure coding data and filling height data. The monitoring target is defined as the future grid settlement increment, and a corresponding target mask is constructed. Samples are screened by combining the grid mask and the target mask to obtain a sample set. A dual-branch convolutional neural network is constructed. Based on the sample set, the first branch input, the second branch input, and the gating vector are input into the dual-branch convolutional neural network to output the predicted data of the future grid settlement increment. The differential settlement index is calculated and the warning level is obtained through classification. The location information of the measuring points that trigger the warning level is output through a reversible mapping relationship.

[0007] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural networks described in this invention, the preprocessing of settlement observation data and construction procedure coding data includes: collecting settlement observation data using measurement point identifiers and timestamps as indexes, and collecting construction procedure coding data using procedure codes and procedure timestamps as indexes; performing unified time-stamp mapping on the settlement observation data and construction procedure coding data, and marking the settlement observation data for validity to generate mask data; and filling in the missing measurement positions in the settlement observation data along the time history direction according to the updated mask data to obtain the supplemented settlement data.

[0008] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the unified time-stamp mapping includes: using the nearest neighbor priority mapping method to map settlement observation data and construction procedure code data with inconsistent timestamps to a unified time point index set; when there is no settlement observation data that meets the mapping conditions at the unified time point index, a null settlement record is generated at the corresponding time point index and marked as invalid in the mask data.

[0009] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the step of establishing a spatial grid and reversible mapping relationship based on the plane coordinates of settlement measuring points includes generating a spatial grid composed of grid row index and grid column index according to the distribution range of the plane coordinates of settlement measuring points and the grid step size; and forming a reversible mapping relationship by establishing a forward mapping relationship from settlement measuring point identifiers to grid row index and grid column index, and establishing a reverse mapping relationship from grid row index and grid column index to the measuring point identifier set.

[0010] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the selection of representative measuring points includes: for each spatial grid position, selecting representative measuring points from the set of measuring point identifiers mapped to the spatial grid position according to the distance between the plane coordinates of the settlement measuring point and the center coordinates of the spatial grid position; when there are several candidate measuring points with the same distance, selecting a unique representative measuring point according to the measuring point identifier sorting rules.

[0011] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the step of synchronously rearranging the mask data into a grid mask includes: rearranging the supplementary settlement data into grid settlement data according to the grid row index, grid column index, and unified time point index based on the representative measuring point; and synchronously rearranging the mask data corresponding to the representative measuring point into a grid mask that corresponds one-to-one with the grid settlement data at the unified time point index.

[0012] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the step of generating a differential settlement dictionary in real time for the input time window includes: calculating the baseline differential settlement of each grid position relative to the bridge abutment reference position within the input time window; calculating the multi-scale adjacent differential settlement and the lateral mirror differential settlement along the line direction; forming a differential settlement dictionary based on the baseline differential settlement, adjacent differential settlement, and lateral mirror differential settlement; and generating a differential mask based on the grid mask.

[0013] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the construction of the gated vector includes: statistically analyzing the construction process coding data within a gated time window to obtain the proportion of each process code appearing within the gated time window, generating a process proportion vector; calculating the change in filling height based on the filling height data corresponding to the start and end time indexes of the gated time window; and concatenating the process proportion vector, the change in filling height, and the available flag information of the filling height data to generate the gated vector.

[0014] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural network described in this invention, the dual-branch convolutional neural network includes: performing anisotropic three-dimensional convolutional feature extraction and feature fusion on the first branch input and the second branch input respectively with spatial and temporal convolutional dimensions; inputting the gating vector to the gating mapping to generate channel scaling coefficients; scaling the convolutional features according to the channels based on the channel scaling coefficients to form a dynamic feature expression related to the construction process and filling status, and outputting predicted data of future grid settlement increments.

[0015] As a preferred embodiment of the bridge abutment construction settlement trend early warning method based on convolutional neural networks described in this invention, the step of outputting the measurement point location information that triggers the early warning level through a reversible mapping relationship includes: locating the index category with the highest level within a limited statistical area where the predicted shielding marker field is marked as valid; selecting the combination of the grid row index and grid column index where the peak value reaches its maximum value within the differential settlement prediction value grid corresponding to the highest level index category as the trigger location index; reading the measurement point identifier list corresponding to the trigger location index from the reverse mapping table based on the reversible mapping relationship, and outputting the measurement point identifier list as the measurement point location information; reading the representative measurement point identifier corresponding to the trigger location index from the representative measurement point table, and associating the representative measurement point identifier with the measurement point identifier list for output.

[0016] The beneficial effects of this invention are as follows: by establishing a spatial grid and a reversible mapping relationship and selecting representative measurement points, discrete settlement observations are expressed in a regularized manner within a unified spatial framework, enhancing the spatial interpretability and positioning accuracy of the settlement evolution process; by constructing a dual-branch convolutional neural network and introducing a gating vector to dynamically modulate the network features, the reliability of future grid settlement increment prediction and early warning level determination is improved, enabling precise positioning and advanced early warning of measurement points that trigger early warning levels. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for early warning of bridge abutment construction settlement trends based on convolutional neural networks.

[0019] Figure 2 The flowchart generated for data preprocessing and the first branch input.

[0020] Figure 3A flowchart for establishing a spatial grid and generating a differential settlement dictionary.

[0021] Figure 4 A flowchart for neural network training and early warning output.

[0022] Figure 5 The CDF plot shows the location error for the trigger position index.

[0023] Figure 6 A schematic diagram of the spatial grid overlay for triggering the location index and measuring point identifier list backtracking. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Reference Figures 1-6 This is one embodiment of the present invention, which provides a method for early warning of bridge abutment construction settlement trends based on convolutional neural networks, including the following steps: S1. Collect settlement observation data and construction procedure code data from settlement measuring points, preprocess the settlement observation data and construction procedure code data to obtain supplementary settlement data and mask data, and obtain filling height data based on the collection situation.

[0028] Settlement observation data is collected using the measuring point identifier and timestamp as indexes, and construction process code data is collected using the process code and process timestamp as indexes.

[0029] Furthermore, several settlement monitoring points are set up at the construction site, and a unique monitoring point identifier is established for each monitoring point; after the monitoring points are set up, the plane coordinates of the settlement monitoring points are recorded; settlement observation data is collected, the collection timestamp is recorded, and a collection status mark is added to the settlement observation data; the settlement observation data is stored as a quadruple of monitoring point identifier, collection timestamp, settlement reading, and collection status mark, and written to the data storage location; construction procedure code data is collected, and the construction procedure code data is stored as a triple of procedure code, procedure timestamp, and record source mark, and written to the data storage location.

[0030] Among them, the acquisition status marker is used to characterize the abnormal and missing acquisition of settlement observation data; the acquisition timestamp is used to characterize the original acquisition time of settlement observation data; and the process timestamp is used to characterize the original recording time of construction process coding data.

[0031] It should be noted that when the number of records corresponding to the same measuring point identifier and the same acquisition timestamp is greater than 1, the last record written is retained in the order of writing time, and the identifier of the overwritten record is written into the audit field. The audit field and the retained settlement observation data record are then written into the data storage location.

[0032] It should be noted that the construction process coding data includes, but is not limited to, process codes and process timestamps; where, when a process has a continuous interval, the construction process coding data also includes process start timestamps and process end timestamps.

[0033] It should be noted that the data storage locations for construction procedure coding data and settlement observation data have related fields.

[0034] A unified time-stamped mapping is performed on the settlement observation data and the construction procedure coding data, and the settlement observation data is validated to generate masked data.

[0035] Furthermore, a unified time-scale index table is established. For each settlement monitoring point, at each time point of the unified time scale, the settlement observation data records near the corresponding time point of the current settlement monitoring point are retrieved. The nearest neighbor priority mapping method is used to perform unified time-scale mapping on the settlement observation data to obtain a unified time-scale settlement observation sequence. Through rule verification, the unified time-scale settlement observation sequence is validated to generate mask data. The mask data is written into the data storage location as a triplet of monitoring point identifier, unified time point index, and validity tag, and maintains a one-to-one correspondence with the unified time-scale settlement observation sequence.

[0036] Among them, the mask data is used to characterize whether the settlement observation data of each settlement measuring point at each unified time point is valid or invalid; the unified time point index is used to characterize the time point identifier in the unified time scale index table.

[0037] It should be noted that the unified timescale index table includes continuous time point sequences.

[0038] It should be noted that the nearest neighbor priority mapping method specifically selects the record with the smallest time difference and a timestamp no later than the unified time point; when there is no record with a timestamp no later than the unified time point, the record with the smallest time difference is selected; when there are records with the same time difference, the unique record is selected according to the order in which the records were written.

[0039] Specifically, when a settlement observation data record that meets the mapping conditions cannot be retrieved at a unified time point, a record item with a corresponding unified time point index is generated in the unified time-scaled settlement observation sequence, the settlement reading is set to null, and a missing measurement marker field is generated.

[0040] It should be noted that rule validation includes, but is not limited to, field integrity validation, numeric type validation, null value validation, and collection status flag validation.

[0041] It should be noted that the validity labeling is as follows: when a settlement reading exists at a unified time point and passes the rule verification, the mask data is marked as valid; when no settlement reading exists at a unified time point, the mask data is marked as invalid; when the unified time point fails the rule verification, the mask data is marked as invalid.

[0042] Furthermore, a unified time-stamp mapping is performed on the construction procedure coding data to obtain a unified time-stamped construction procedure coding sequence.

[0043] It should be noted that a unified time-scale mapping is performed on the construction process code data. Specifically, when the construction process code data is a time-type record, the nearest neighbor priority mapping method is used to map the process code to each time point of a unified time scale. When the construction process code data is an interval-type record and the current unified time point falls within the process start and end interval, the process code corresponding to the current interval is written into the unified time-scale construction process code sequence. When the construction process code data is an interval-type record but the current unified time point does not fall within the process start and end interval, the default process code is written into the unified time-scale construction process code sequence and a process default mark is generated. When the number of times the same unified time point falls within the same process interval is greater than 1, a unique process code is selected according to the interval start time order.

[0044] It should be noted that the current unified time point falls within the process start and end interval. Specifically, the current unified time point is considered to fall within the interval if it is not less than the process start timestamp and is less than the process end timestamp.

[0045] Anomaly removal is performed on the settlement observation data, and the mask data is updated synchronously.

[0046] Furthermore, an adjacency difference discrimination method is used to perform anomaly detection on the unified time-scale settlement observation sequence to determine the set of anomaly change points. Based on the set of anomaly change points, in the unified time-scale settlement observation sequence, the settlement readings corresponding to the anomaly change points are set to null values, and the anomaly removal flag field is written to the corresponding record item. At the same time, the mask data corresponding to the anomaly change points is updated to invalid, and the mask update reason field is written to the mask data record item. When the unified time point corresponding to the anomaly change point has passed and the missing measurement flag field is set to null values, the settlement reading remains null, and the superposition state of anomaly removal and missing measurement is recorded in the anomaly removal flag field. Meanwhile, the mask data remains invalid, and the set of reasons for the corresponding invalid state is recorded in the mask update reason field.

[0047] The mask update reason field is used to indicate that the mask data is invalid due to abnormal changes and removal.

[0048] It should be noted that the adjacent difference discrimination method specifically involves setting an adjacent difference threshold, calculating the adjacent change in settlement readings at adjacent unified time points for the same settlement measuring point; when the settlement reading corresponding to any time point in adjacent unified time points is null, the discrimination of the current adjacent change is skipped, and the abnormal change judgment result corresponding to the adjacent change is recorded as undecidable; when the settlement readings at adjacent unified time points are all non-null values, the adjacent change is compared with the adjacent difference threshold; when the absolute value of the adjacent change is greater than the adjacent difference threshold, the settlement reading corresponding to the relatively later unified time point is marked as an abnormal change point, and the index of the relatively later unified time point is written into the abnormal change point set; when the absolute value of the adjacent change is not greater than the adjacent difference threshold, the abnormal change judgment result corresponding to the adjacent change is recorded as normal.

[0049] It should be noted that the adjacent difference threshold is set by statistically analyzing the adjacent changes of the same settlement measuring point during the stable construction phase, combined with the acquisition resolution and the upper limit of the field measurement error. The value range is usually [1, 10].

[0050] The missing measurement locations are filled in based on the updated mask data to obtain the supplementary settlement data.

[0051] Furthermore, based on the updated mask data, the missing measurement locations are filled in to obtain the filled settlement data. After the filling in is completed, the missing measurement marker field, the anomaly removal marker field, the initialization filling marker field, and the mask update reason field are written to the data storage location associated with the filled settlement data.

[0052] It should be noted that a completion process is performed for missing measurement locations. Specifically, for each settlement measurement point at each unified time point, when the updated mask data is marked as valid, the settlement reading corresponding to the unified time point is written into the completed settlement data; when the updated mask data is marked as invalid, the most recent unified time point marked as valid along the historical direction is retrieved, and the settlement reading corresponding to the most recent valid unified time point is written into the completed settlement data; when no mask data is marked as valid along the historical direction, the completed settlement data is written into the initial baseline value, and an initial completion mark field is generated. The completed settlement data is written into the data storage location as a triplet of measurement point identifier, unified time point index, and completed settlement value, and a one-to-one correspondence is maintained between the mask data and the unified time-scaled settlement observation sequence and the unified time point index.

[0053] The initialization completion marker field is used to indicate that the completion settlement data comes from the initialization write.

[0054] The filling height data is obtained based on the collection situation. When the filling height data is collected, it is mapped to a unified time scale and retained. When the filling height data cannot be collected, the filling height data is set to zero and corresponding flag information is generated.

[0055] Among them, the filling height data is used to characterize the height status information of the filling process during the construction period.

[0056] Furthermore, when filling height data can be collected on-site, a unified time-scale mapping is performed on the filling height data to align the filling height data on a unified time scale. The aligned filling height data is then written to the data storage location, and a filling height data availability flag is generated. When filling height data cannot be collected on-site, a filling height data sequence is generated on a unified time scale and written as zero, and a filling height data unavailable flag is generated. The flag is then associated with and stored in relation to the time index of the filling height data.

[0057] S2. Based on the plane coordinates of the settlement measuring points, establish a spatial grid and reversible mapping relationship, select representative measuring points, rearrange the supplementary settlement data into grid settlement data based on the representative measuring points, synchronously rearrange the mask data into a grid mask, and determine the abutment reference position.

[0058] Based on the plane coordinates of the settlement measuring points, grid row indexes and grid column indexes are generated using the coordinate range and grid step size, and a spatial grid is established.

[0059] Furthermore, the plane coordinates of the settlement measuring points are read, and the grid coverage area is determined based on the coordinate range of the settlement measuring points. The minimum and maximum boundary values ​​of the settlement measuring points in the direction along the line and in the lateral direction are statistically analyzed, and a spatial grid boundary is generated through boundary expansion. Discrete grids are generated within the spatial grid boundary according to the grid step size, forming a spatial grid index set composed of grid row indexes and grid column indexes. The corresponding grid position center coordinates are generated for each combination of grid row indexes and grid column indexes, and the spatial grid boundary, grid step size, grid row index range, grid column index range, and grid position center coordinates are written to the data storage location.

[0060] Among them, boundary extension is used to ensure that the spatial grid boundary covers the plane coordinates of all settlement measurement points, as well as the abutment, the back of the abutment, and the area of ​​influence.

[0061] Based on the plane coordinates and spatial grid of settlement measurement points, a reversible mapping relationship is obtained by establishing forward and reverse mapping relationships.

[0062] Furthermore, for each measuring point identifier, the corresponding settlement measuring point plane coordinates are read, and the grid row index and grid column index are calculated based on the spatial grid boundary and grid step size. The calculation results are written into the forward mapping table. The forward mapping table is traversed, and all measuring point identifiers mapped to the same combination of grid row index and grid column index are aggregated into a measuring point identifier list. The grid row index, grid column index, and measuring point identifier list are written into the reverse mapping table. The reverse mapping table and the forward mapping table form a reversible mapping relationship.

[0063] The forward mapping relationship is used to map each measurement point identifier to a unique grid row index and grid column index; the reverse mapping relationship is used to map each grid row index and grid column index combination to the set of measurement point identifiers within the corresponding grid position.

[0064] It should be noted that for the plane coordinates of settlement measuring points falling near the grid boundary line, the rounding rule is adopted to map the plane coordinates of settlement measuring points to grid row indexes and grid column indexes in the direction along the line and in the lateral direction, respectively, to ensure that the plane coordinates of the same settlement measuring point obtain unique and consistent grid row indexes and grid column indexes when performing the mapping.

[0065] It should be noted that the rounding assignment rule is adopted. Specifically, the minimum boundary value of the spatial grid boundary is used as the coordinate reference. The relative coordinate offset of the settlement measuring point plane coordinates is calculated. The ratio of the relative coordinate offset to the grid step size is calculated and rounded down to obtain the grid row index and grid column index. When the plane coordinates of the settlement measuring point are exactly on the boundary line of the adjacent grid position, the current settlement measuring point plane coordinates are assigned to the grid position to the right or above the boundary line, and the current assignment direction is written as a configuration record field to the data storage location.

[0066] For each grid location, a representative measuring point is selected. Based on the representative measuring point, the settlement data is rearranged into grid settlement data. The mask data is then rearranged into a grid mask and aligned with the grid settlement data at the same time point index.

[0067] Furthermore, for each grid location, the center coordinates of the grid location are read, and the list of measuring point identifiers corresponding to the grid location is read from the reverse mapping table. For each measuring point identifier in the list, the plane coordinates of the corresponding settlement measuring point are read and the distance to the center coordinates of the grid location is calculated. The measuring point identifier with the smallest distance is selected as the representative measuring point identifier. When there are measuring point identifiers with the same smallest distance, they are sorted according to the measuring point identifier sorting rules, and the measuring point identifier with the highest sorting position is selected as the representative measuring point identifier. The grid row index, grid column index, representative measuring point identifier, and list of spare measuring point identifiers are written into the representative measuring point table.

[0068] It should be noted that the sorting rules for measurement point identifiers are as follows: the measurement point identifiers are sorted according to their character encoding order; when the measurement point identifiers contain a mixture of numbers and letters, they are first sorted according to the character encoding order of the letter segment, and then sorted according to the numerical value of the number segment; the measurement point identifier with the highest sorting order is selected as the representative measurement point identifier.

[0069] It should be noted that the generation of the backup measuring point identifier list is specifically as follows: The list of measuring point identifiers corresponding to the current grid row index and grid column index in the reverse mapping table is deduplicated; the representative measuring point identifier is removed from the deduplicated measuring point identifier list to obtain a candidate backup set; the candidate backup set is sorted according to the measuring point identifier sorting rules; the sorted candidate backup set is written into the backup measuring point identifier list in order, and the backup measuring point identifier list and the representative measuring point identifier are stored together in the same representative measuring point table record item.

[0070] Furthermore, for each combination of grid row index and grid column index, the representative measuring point identifier is read from the representative measuring point table; the supplementary settlement value sequence corresponding to the representative measuring point identifier is retrieved in the supplementary settlement data, and the supplementary settlement value sequence is written into the corresponding grid row index and grid column index position of the grid settlement data according to the unified time point index, forming grid settlement data organized by grid row index, grid column index, and unified time point index.

[0071] It should be noted that when writing grid settlement data, the corresponding relationships of grid row index, grid column index, and the identifier of measurement point are simultaneously written into the rearranged index record field, and the rearranged index record field is stored in association with the grid settlement data.

[0072] It should be noted that when there is no list of measurement point identifiers corresponding to the combination of grid row index and grid column index in the reverse mapping table, the grid settlement value corresponding to the current grid position is set to null in the grid settlement data and a grid missing measurement mark field is generated. In the grid mask, the validity mark corresponding to the current grid position is set to invalid and a grid missing measurement mark field is generated.

[0073] Furthermore, for each combination of grid row index and grid column index, the representative measurement point identifier is read from the representative measurement point table; the validity mark sequence corresponding to the representative measurement point identifier is retrieved from the mask data, and the validity mark sequence is written into the corresponding grid row index and grid column index position of the grid mask according to the unified time point index; for grid positions where no measurement point identifier falls, a record item for the corresponding grid position is generated in the grid mask, the grid mask is marked as invalid on all unified time point indices, and a grid missing measurement mark field is generated, so that the grid mask can maintain a one-to-one correspondence with the grid settlement data on the unified time point index; after the rearrangement is completed, an alignment check is performed, and for each combination of grid row index and grid column index, the unified time point index set of the grid settlement data record item and the grid mask record item is checked to be consistent; when an inconsistency in the index set is found, a default record item is generated based on the one with the longer unified time point index set, the grid settlement value is set to null in the default record item of the grid settlement data, the validity mark is set to invalid in the default record item of the grid mask, and an index completion mark field is generated.

[0074] Based on the locations to be verified corresponding to the bridge abutment control points, the bridge abutment reference locations are determined according to availability conditions.

[0075] Furthermore, the measurement point identifiers corresponding to the bridge abutment control points are read, and the grid row index and grid column index corresponding to the measurement point identifiers are read through the forward mapping table. The combination of the grid row index and grid column index is used as the location to be verified. The validity mark sequence of the location to be verified in the grid mask is read and availability verification is performed. When the validity mark sequence meets the availability condition, the location to be verified is determined as the bridge abutment reference location. When the validity mark sequence does not meet the availability condition, the measurement point identifiers corresponding to the candidate bridge abutment control points are replaced in sequence, and the mapping and verification process is repeated until a bridge abutment reference location that meets the availability condition is determined. The bridge abutment reference location is written into the data storage location in the form of a grid row index and grid column index tuple, and an association is established with the spatial grid definition, reversible mapping relationship, and representative measurement point table.

[0076] It should be noted that the abutment control points are settlement measurement points at the abutment location. The abutment control points are selected from the settlement measurement points located at or near the abutment location after the settlement measurement points are laid out and their plane coordinates are registered at the construction site. The abutment control points include, but are not limited to, the preferred abutment control points and the alternative abutment control points.

[0077] It should be noted that the availability conditions are specifically as follows: An availability threshold is set, and within the time period covered by the unified time-scale index table, the number of valid time points marked as valid in the grid mask for the bridge abutment reference location is counted, and the proportion of valid time points to the total number of time points within the time period is calculated. When the proportion of valid time points to the total number of time points within the time period is not less than the availability threshold, the bridge abutment reference location is determined to meet the availability conditions; when the proportion of valid time points to the total number of time points within the time period is less than the availability threshold, the bridge abutment reference location is determined not to meet the availability conditions.

[0078] It should be noted that the availability threshold is set by statistically analyzing the grid mask validity markers of the bridge abutment control points during historical monitoring and combining them with the requirements for the integrity of construction records. The value range is usually [0.60, 0.95].

[0079] S3. Extract the absolute settlement sequence from the grid settlement data and concatenate it with the grid mask to generate the first branch input. Generate the differential settlement dictionary in real time by adjusting the input time window. Calculate the differential mask based on the grid mask. Concatenate the differential settlement dictionary and the differential mask to generate the second branch input.

[0080] Based on the input time window length and the gated time window length, the absolute settlement sequence is extracted from the grid settlement data and concatenated with the grid mask to generate the first branch input.

[0081] Furthermore, the current unified time point index is determined in the unified time scale index table. Based on the input time window length, the unified time point index is traced back from the current unified time point index to the historical direction to obtain the unified time point index set covered by the input time window. Based on the gated time window length, the same current unified time point index is traced back to the historical direction to obtain the unified time point index set covered by the gated time window. For each grid row index and grid column index combination, the grid settlement value sequence of the current grid position within the unified time point index set covered by the input time window is read from the grid settlement data as the absolute settlement sequence. The validity mark sequence of the current grid position within the same unified time point index set is read from the grid mask as the grid mask sequence. The first branch input is generated through channel splicing.

[0082] It should be noted that channel splicing specifically involves writing the absolute settlement sequence into the first channel and the grid mask sequence into the second channel; the channel sequence is then written into the data storage location.

[0083] It should be noted that when the grid missing test mark field indicates that the grid location is missing, the record corresponding to the current grid location is retained in the first branch input, and the grid mask sequence of the current location is kept invalid.

[0084] By calculating the baseline differential settlement relative to the bridge abutment reference position, the multi-scale adjacent differential settlement along the line direction, and the lateral mirror differential settlement within the input time window, a differential settlement dictionary is generated in real time. A differential mask is calculated based on the grid mask, and the differential settlement dictionary and the differential mask are concatenated to generate the second branch input.

[0085] Furthermore, when constructing the second branch input each time, a differential settlement dictionary is generated in real time for the unified time point index set covered by the input time window; for each unified time point index within the input time window, the difference between the grid settlement value at each grid position and the grid settlement value at the abutment reference position is calculated to obtain the reference differential settlement value; and all reference differential settlement values ​​within the input time window are arranged in the order of the unified time point index to form a reference differential settlement channel.

[0086] It should be noted that each channel in the differential settlement dictionary maintains the same spatial index shape and uniform time point index shape as the grid settlement data.

[0087] Furthermore, a multi-scale adjacent span set is defined along the route direction. For each adjacent span in the multi-scale adjacent span set, an adjacent differential settlement channel is generated. For each unified time point index within the input time window, at each grid position, the difference between the grid settlement value at the current grid position and the grid settlement value at the adjacent grid position after offset along the route direction by adjacent span is calculated to obtain the adjacent differential settlement value. When the offset adjacent grid position exceeds the grid row index range, a default record is generated at the current position, and the differential mask corresponding to the current position is set to invalid.

[0088] It should be noted that a multi-scale set of adjacent spans along the route is set. Specifically, based on the grid row index increment of the spatial grid along the route, no less than two different positive integer adjacent spans are set. When generating adjacent differential settlement channels, the corresponding channels are generated sequentially according to the adjacent span order.

[0089] Furthermore, a mirror correspondence rule is established based on the midline of the grid column index in the horizontal direction. For each unified time point index within the input time window, at each grid position, the difference between the grid settlement value at the current grid position and the grid settlement value at the corresponding horizontal mirror position is calculated to obtain the horizontal mirror differential settlement value. When the horizontal mirror position exceeds the range of the grid column index, a default record item is generated at the current position, and the differential mask corresponding to the current position is set to invalid.

[0090] It should be noted that the mirror correspondence rule is as follows: the horizontal centerline position is determined based on the grid column index range; for any grid column index, the mirror grid column index is determined according to the principle of symmetry about the horizontal centerline, and the offset of the mirror grid column index from the horizontal centerline is equal to and opposite to that of the original grid column index; when the grid column index is at the horizontal centerline position, the mirror grid column index is set to itself.

[0091] It should be noted that the channel arrangement order of the differential settlement dictionary is as follows: first, the baseline differential settlement channels are arranged; then, the adjacent differential settlement channels are arranged in the order of adjacent span sets; and finally, the transverse mirror differential settlement channels are arranged.

[0092] Furthermore, a difference mask is calculated based on the grid mask. When the difference settlement value is obtained through the difference between the grid settlement values ​​of two grid locations, the validity flag of the grid mask at the same index of the two grid locations at the same time point is read. The corresponding difference mask is marked as valid if and only if both validity flags are valid. If either of the two validity flags is invalid, the corresponding difference mask is marked as invalid. When the difference settlement value is generated using the default record item due to out-of-bounds error, the corresponding difference mask is directly marked as invalid and an out-of-bounds flag field is generated.

[0093] Furthermore, by sequentially executing the arrangement of each channel of the differential settlement dictionary and the arrangement of each channel of the differential mask, the differential settlement dictionary and the differential mask are concatenated to generate the second branch input.

[0094] It should be noted that the spliced ​​second branch input is consistent with the first branch input in terms of spatial index and unified time point index.

[0095] S4. Combine the construction process coding data and filling height data to construct a gating vector, define the monitoring target as the future grid settlement increment and construct the corresponding target mask, and combine the grid mask and target mask to filter samples and obtain a sample set.

[0096] Based on construction procedure coding data and filling height data, a gating vector is constructed by performing construction procedure coding data processing and filling height data processing.

[0097] Furthermore, within the unified time point index set covered by the gated time window, construction procedure code data processing is performed. The procedure code sequence within the gated time window is read from the unified time-scaled construction procedure code sequence. Based on the procedure code sequence, the occurrence count is performed according to the procedure code category to obtain the procedure count vector. The procedure count vector is then normalized to obtain the procedure proportion vector.

[0098] It should be noted that the normalization process is performed on the process counting vector. Specifically, the total number of time points in the unified time point index set covered by the gated time window is counted, and the process proportion vector is obtained by calculating the ratio of the occurrence frequency of each process code category in the process counting vector to the total number of time points.

[0099] Furthermore, the filling height data processing is performed. The filling height values ​​corresponding to the start and end unified time point indices of the gated time window are read from the filling height data, and the filling height change within the gated time window is calculated. When the filling height data is unavailable due to available flag information, the zeroing rule is used to obtain the filling height change. The filling height data can be written as a gated vector field to ensure that the gated vector maintains a consistent input structure under different engineering data conditions.

[0100] It should be noted that the zeroing rule specifically means that when the available information of the filling height data indicates that the filling height data is unavailable, the filling height value corresponding to the unified starting time point index of the gated time window and the filling height value corresponding to the unified ending time point index of the gated time window are both set to zero, the filling height change is set to zero, and the available information of the filling height data is retained in the gated vector.

[0101] Among them, the filling height data retained in the gating vector can be used to distinguish the source of zeroing.

[0102] Furthermore, the process proportion vector, the change in filling height, and the available flag information of the filling height data are concatenated into a gating vector.

[0103] The gating vector is used to express the construction state conditions corresponding to the current input time window.

[0104] The supervision target is defined as the future grid settlement increment and a corresponding target mask is constructed. By constructing a set of future prediction step sizes, the grid mask and the target mask are used for joint constraints to filter samples and obtain a sample set.

[0105] It should be noted that the process of constructing future grid settlement increments and target masks and filtering sample sets is used in the sample set preparation stage. The future grid settlement values ​​used are the historical records of settlement observation data collected after the current unified time point index corresponding to the sample index.

[0106] It should be noted that during the rolling prediction phase, the future grid settlement increment and target mask are not constructed, nor are the future grid settlement values ​​read. Only the first branch input, the second branch input, and the gating vector are generated and input into the dual-branch convolutional neural network. The predicted data of the future grid settlement increment is output through the dual-branch convolutional neural network.

[0107] Furthermore, the supervision target is defined as the future grid settlement increment. A set of future prediction step sizes is constructed. For the current unified time point index corresponding to each sample index, at each grid position, the future grid settlement increment is obtained by calculating the difference between the future grid settlement value corresponding to the future prediction step size and the current grid settlement value. When both the current unified time point index and the future unified time point index at the same grid position are marked as valid in the grid mask, the target mask is marked as valid. When either the current unified time point index or the future unified time point index at the same grid position is invalid in the grid mask, the target mask is marked as invalid, and a target invalidity reason field is generated.

[0108] It should be noted that the future grid settlement value will be the value of the grid settlement data at a future unified time point index.

[0109] It should be noted that the future grid settlement increment and the target mask will remain consistent in terms of grid row index, grid column index, and prediction step size.

[0110] It should be noted that the construction of the future prediction step size set is specifically as follows: in a unified time scale index table, the prediction step size is represented by a unified time point index offset, at least two different positive integer prediction step sizes are set, and the prediction step sizes are sorted in ascending order to form the future prediction step size set; the future prediction step size set is written into the data storage location and defined as the supervision target dimension.

[0111] Furthermore, using the current unified time point index as the sample index field, construct the first branch input, second branch input, gating vector, future grid settlement increment, and target mask corresponding to the current sample index field, and form a sample record item; perform input validity judgment and target validity judgment on the sample record item, and write the sample record item into the sample set if and only if both input validity judgment and target validity judgment are satisfied.

[0112] It should be noted that the input validity determination is specifically as follows: a first proportion threshold is set, and the proportion of valid labels for the grid mask sequence corresponding to the first branch input within the input time window is counted. When the proportion of valid labels for the grid mask sequence corresponding to the first branch input within the input time window is less than the first proportion threshold, the sample record is discarded, and the input validity determination is deemed not to be satisfied. When the proportion of valid labels for the grid mask sequence corresponding to the first branch input within the input time window is not less than the first proportion threshold, the current sample record is marked as valid, the input validity label field is recorded as valid, and the input validity determination is deemed to be satisfied.

[0113] It should be noted that the first proportion threshold is set by statistically analyzing the distribution of the effective label proportion of the grid mask sequence in historical samples and combining it with the proportion of invalid labels corresponding to the data missing level. The value range is usually [0.60, 0.95].

[0114] It should be noted that the data missing level, specifically, in the historical samples, takes the statistical range consisting of all grid row indices, all grid column indices, and all unified time point indices of the grid mask sequence corresponding to the first branch input within the input time window as the statistical object, counts the number of invalid grid positions, and calculates the proportion of the number of invalid grid positions to the total number of grid positions within the statistical range.

[0115] The higher the level of missing data, the higher the proportion of missing or invalid locations within the input time window.

[0116] It should be noted that the target validity determination specifically involves setting a second proportional threshold. The statistical range is defined by counting the number of target units marked as valid across all grid row indices, all grid column indices, and all prediction steps covered by the future prediction step set, using the target mask corresponding to the current sample record as the statistical range. The proportion of these valid target units to the total number of target units within the statistical range is calculated as the target validity marking ratio. When the target validity marking ratio is less than the second proportional threshold, the sample record is discarded, and the target validity determination is deemed unsatisfactory. When the target validity marking ratio is not less than the second proportional threshold, the current sample record is marked as valid, and the target validity marking field is recorded as valid, thus satisfying the target validity determination.

[0117] It should be noted that a target cell is a corresponding record of the target mask at a grid row index, a grid column index, and a prediction step size. The total number of target cells is determined by the number of rows in the grid row index range, the number of columns in the grid column index range, and the number of steps in the future prediction step size set.

[0118] It should be noted that the second proportion threshold is set by statistically analyzing the distribution of the effective target label proportion in historical samples and combining it with the coverage requirements of the future prediction step size set. The value range is usually [0.50, 0.90].

[0119] It should be noted that each sample record in the sample set includes, but is not limited to, the sample index field, the first branch input, the second branch input, the gating vector, the future grid settlement increment, and the target mask.

[0120] S5. Construct a dual-branch convolutional neural network. Based on the sample set, input the first branch input, the second branch input, and the gating vector into the dual-branch convolutional neural network. Output the predicted data of future grid settlement increment, calculate the differential settlement index and obtain the warning level through classification, and output the measurement point location information that triggers the warning level through the reversible mapping relationship.

[0121] A dual-branch convolutional neural network is constructed. Anisotropic three-dimensional convolutional feature extraction with spatial and temporal convolutional dimensions is performed on the first and second branch inputs, and feature fusion is then performed. The gating vector is input to the gating map to generate channel scaling coefficients. Based on the channel scaling coefficients, the convolutional features are scaled by channel to form dynamic feature representations related to construction procedures and filling status, and the predicted data of future grid settlement increments are output.

[0122] It should be noted that the structure of a dual-branch convolutional neural network includes an input layer, a first-branch feature extraction layer group, a second-branch feature extraction layer group, a feature fusion layer, a gating mapping layer and a channel scaling layer, a prediction output layer, and a structural consistency verification layer.

[0123] It should be noted that the input layer of a dual-branch convolutional neural network includes a first-branch input layer, a second-branch input layer, and a gated vector input layer; The first branch input layer is used to receive the first branch input, and the channel order of the first branch input is the absolute settlement sequence channel and the grid mask sequence channel.

[0124] The second branch input layer is used to receive the second branch input. The channel order of the second branch input is the differential settlement dictionary channel group and the differential mask channel group. The channel arrangement order within the differential settlement dictionary channel group is the baseline differential settlement channel, the adjacent differential settlement channels arranged in the order of multi-scale adjacent span sets, and the lateral mirror differential settlement channel. The differential mask channel group corresponds one-to-one with the differential settlement dictionary channel group and is arranged in the same order.

[0125] The gating vector input layer is used to receive gating vectors. The gating vector fields are in the following order: process proportion vector field group, filling height change field, and filling height data availability flag information field.

[0126] It should be noted that the first branch feature extraction layer group of the dual-branch convolutional neural network consists of a first convolutional layer group, a second convolutional layer group, and a third convolutional layer group connected in sequence; each convolutional layer group includes an anisotropic three-dimensional convolutional layer, a normalization layer, and a nonlinear activation layer.

[0127] In this process, the convolutional dimensions of the anisotropic 3D convolutional layer are time and space, and the kernel size and stride configuration in the time and space directions are independent of each other and written to the data storage location.

[0128] It should be noted that the second branch feature extraction layer group of the dual-branch convolutional neural network consists of the fourth, fifth and sixth convolutional layer groups connected in sequence; each convolutional layer group includes anisotropic three-dimensional convolutional layers, normalization layers and nonlinear activation layers.

[0129] The second branch feature extraction layer group is consistent with the first branch feature extraction layer group in terms of the number of layers, the order of operators within the layer, and the convolution dimension, ensuring that the features of the two branches are aligned in terms of time index and spatial index.

[0130] It should be noted that the feature fusion layer of the dual-branch convolutional neural network is used to fuse the outputs of the first branch feature extraction layer group and the second branch feature extraction layer group. The feature fusion layer includes a channel splicing layer and a fusion transformation layer.

[0131] The channel splicing layer is used to splice the features of the two branches in the channel dimension to obtain spliced ​​features; the fusion transformation layer is used to perform convolution transformation on the spliced ​​features to obtain fused features; the fusion transformation layer includes a three-dimensional convolution transformation layer, a normalization layer and a nonlinear activation layer in sequence.

[0132] It should be noted that in a dual-branch convolutional neural network, the gating mapping layer is used to map the gating vector to the channel scaling factor vector; the channel scaling layer is used to perform channel-by-channel scaling on the fused features based on the channel scaling factor vector. The gated mapping layer consists of a first fully connected mapping layer and a second fully connected mapping layer connected in sequence, with each fully connected mapping layer followed by a non-linear activation layer; the output of the gated mapping layer is a channel scaling coefficient vector, the length of which is equal to the number of fused feature channels.

[0133] The channel scaling layer is used to broadcast the channel scaling coefficient vector to the time and space dimensions of the fused feature according to the channel index, and to perform scaling on the fused feature channel by channel to obtain the gated scaled fused feature.

[0134] In this process, scaling by channel does not change the temporal and spatial indices of the fused features, but only changes the numerical amplitude of the corresponding channel.

[0135] It should be noted that the dual-branch convolutional neural network prediction output layer is used to map the gated and scaled fused features into predicted data for future grid settlement increments.

[0136] The prediction output layer includes a first prediction convolutional layer and a second prediction convolutional layer connected in sequence. The second prediction convolutional layer is the output layer. The number of output channels of the output layer is equal to the number of steps in the future prediction step set, and the one-to-one correspondence between the output channels and each prediction step in the future prediction step set is written to the data storage location.

[0137] It should be noted that the structural consistency check layer of the dual-branch convolutional neural network is used to check that the temporal index sets of the first branch input and the second branch input are consistent, the spatial index shapes of the first branch input and the second branch input are consistent, the dimension of the gated vector field is consistent with the input dimension of the gated mapping layer, and the output dimension of the gated mapping layer is consistent with the number of fused feature channels. When any check fails, the current operation of the dual-branch convolutional neural network is terminated and the structural inconsistency marker field is written.

[0138] It should be noted that the working principle of the dual-branch convolutional neural network is as follows: In the rolling prediction phase, the dual-branch convolutional neural network starts from the current unified time point index and receives the input of the first branch, the input of the second branch, and the gating vector respectively. After completing the input index consistency verification in the structural consistency verification layer, it sequentially performs the first branch feature extraction, the second branch feature extraction, dual-branch feature fusion, gating mapping to generate channel scaling coefficients, channel scaling to form dynamic feature expression, and prediction output to obtain the prediction data of future grid settlement increment. In the sample set preparation phase, the dual-branch convolutional neural network, with the forward computation link unchanged, further combines the future grid settlement increment supervision target and target mask in the sample set to perform loss calculation and parameter update to obtain the learnable parameter set used in the rolling prediction phase.

[0139] It should be noted that the first branch feature extraction layer outputs the first branch features, which retain the input time and spatial indices and only perform feature recoding in the channel dimension. The channel dimension size is determined by the number of output channels of the third convolutional layer. The second branch feature extraction layer outputs the second branch features, which retain the input time and spatial indices and only perform differential structural feature recoding in the channel dimension. The fusion transformation layer maps the concatenated features to fused features, which retain the time and spatial indices. The fusion transformation layer remixes the information from the two branches in the channel dimension, so that the fused features simultaneously express the temporal evolution pattern of absolute settlement and the spatial coupling pattern of differential settlement.

[0140] It should be noted that the gating mapping layer receives the gating vector and outputs the channel scaling coefficient vector; the channel scaling layer receives the channel scaling coefficient vector and performs scaling on the fused features channel by channel to form a dynamic feature representation related to the construction process and filling status, within the continuous time interval covered by the gating time window. Within this process, the time expansion of the gated vector field is performed to obtain the time expansion function for the gated field. For each channel of the fused feature, calculate the channel scaling factor.

[0141] It should be noted that the channel scaling factor is calculated and expressed as: ; ; ; in, Indicates channel Channel scaling factor, This represents the continuous time values ​​corresponding to the start time point of the gated time window. This represents the continuous time value corresponding to the end time of the gated time window. Indicates the dimension of the gated vector field. Indicates the index of the gated vector field. This represents the current channel weight coefficient. This represents the summation channel weight coefficient. Indicates the current channel synthesis index. This indicates a composite index for the summation channel. The total number of channels for fusion features. This represents the gating vector after uniform scaling. The fields in continuous time The time expansion value at that point, Indicates the fusion feature channel index. Indicates the channel summation index. This represents a continuous-time variable.

[0142] It should be noted that, and The set of learnable parameters derived from the gated mapping layer is obtained through gradient backpropagation and parameter updates during the training of the dual-branch convolutional neural network.

[0143] It should be noted that, The range of values ​​is And satisfy ; The range of values ​​is usually 100. .

[0144] It should be noted that, It is a piecewise constant function. The construction method is as follows: in the unified time point index set covered by the gated time window, for the time interval corresponding to each unified time point index, ... Take the gating vector at the current unified time point index as the first... The field values ​​of each field.

[0145] It should be noted that before calculating the channel scaling factor, a uniform scaling process is performed on each field of the gate vector; among them, the process proportion vector field and the filling height data can retain their original field form using the flag information field, and the filling height change field is converted into a proportional field after being processed by comparing it with the reference height value.

[0146] The reference height value and the change in filling height use the same unit of measurement, and the reference height value is a non-zero positive value.

[0147] It should be noted that the training steps of the dual-branch convolutional neural network are as follows: First, read the sample set; second, sort the sample set according to the sample index field and generate a training sample sequence; third, divide the training sample sequence to obtain training sub-sequences and validation sub-sequences; fourth, extract sample records from the training sub-sequences according to batch size to construct training batch sets; fifth, in each training batch set, stack the first branch inputs according to batch dimensions to form the first batch inputs, stack the second branch inputs according to batch dimensions to form the second batch inputs, stack the gating vectors according to batch dimensions to form batch gating vectors, stack the future grid settlement increments according to batch dimensions to form batch supervision targets, and stack the target masks according to batch dimensions to form batch targets. The training batch is labeled; structural consistency checks are performed on each training batch set. If any check fails, the current training batch set is discarded and an invalid training batch flag is written to the training batch field; for training batch sets that pass the structural consistency check, batch validity checks are performed by counting the total number of valid flags in the batch target mask. If the total number of valid flags in the batch target mask is zero, the current training batch set is discarded and a batch discard flag is written to the batch discard flag field; if the total number of valid flags is greater than zero, the current training batch set is retained and entered into the training iteration; forward computation is performed on each training batch that passes the batch validity check, and the prediction output is obtained; to uniformly encode the spatial grid position and prediction step size position within the batch, the total number of supervised positions for each sample is defined as . The batch sample size is The synthetic index uniquely encodes the sample index and the location index within the batch; the future grid settlement increment monitoring target of the batch is expanded according to the synthetic index. The batch of future grid settlement increment prediction data is expanded according to the synthetic index as follows: Expand the batch target mask according to the composite index as follows Under the constraint of the target mask, the batch normalized loss is calculated. With the batch normalized loss as the objective, gradient backpropagation is performed on the learnable parameter set of the dual-branch convolutional neural network to obtain the gradients of each learnable parameter. Using the Adam optimizer, the learnable parameter set is updated based on the learning rate, the first moment statistics of the gradients of each learnable parameter, and the second moment statistics. The updated learnable parameter set is then written to the data storage location, along with the training epoch field, batch number field, and current batch loss field. Forward computation, loss calculation, gradient backpropagation, and parameter updates are repeated for all training batches in the training subsequence to complete the training of the current training epoch. After the first training epoch, forward computation is performed on the validation subsequence, and the validation subsequence loss corresponding to the first training epoch is calculated. The validation subsequence loss corresponding to the first training epoch is recorded as the historical minimum validation subsequence loss. The current set of learnable parameters is denoted as the current optimal set of learnable parameters. After each subsequent training epoch, forward computation is performed on the validation subsequence, and the loss of the validation subsequence corresponding to the current training epoch is calculated. When the loss of the validation subsequence corresponding to the current training epoch is less than the historical minimum loss, the loss of the validation subsequence corresponding to the current training epoch is denoted as the historical minimum loss, and the current set of learnable parameters is denoted as the current optimal set of learnable parameters. When the loss of the validation subsequence corresponding to the current training epoch is not less than the historical minimum loss, the next training epoch is executed, and the validation subsequence loss is compared again. Training is terminated when the loss of the validation subsequence corresponding to two consecutive epochs is not less than the historical minimum loss. Training is terminated when the number of training epochs reaches the training epoch limit. After training is terminated, the current optimal set of learnable parameters is fixed into a parameter file and written to the data storage location.

[0148] It should be noted that, It is determined by the number of rows in the grid row index range, the number of columns in the grid column index range, and the step size of the future prediction step size set.

[0149] It should be noted that position indexing is used within each sample. Indicates the first One supervisory position, The range of values ​​is usually 100. .

[0150] It should be noted that this applies only if the target mask is marked as valid at the corresponding supervision location. When the target mask is marked as invalid at the corresponding supervision position .

[0151] It should be noted that the batch normalized loss is calculated as follows: ; ; ; in, Indicates batch normalization loss. Indicates a composite index. Indicates the sample index within the batch. Indicates the in-sample location index. Indicates the total number of monitored locations in the batch. This represents the stable term used to calculate the batch normalization loss.

[0152] It should be noted that, Example , The range of values ​​is usually 100. , The range of values ​​is .

[0153] It should be noted that each sample record in the sample set includes a sample index field, the first branch input, the second branch input, the gating vector, the future grid settlement increment, and the target mask.

[0154] The sample index field is used to indicate the index at the current unified point in time.

[0155] The predicted settlement values ​​of the future grid are reconstructed using the predicted settlement increment data of the future grid. Differential settlement indices are reconstructed within the spatial grid and the warning level is obtained through hierarchical judgment. The location information of the measuring points that trigger the warning level is output through a reversible mapping relationship.

[0156] Furthermore, based on the grid settlement data at the current unified time point index, the predicted data of future grid settlement increments are accumulated to the corresponding future prediction step size to obtain the future grid settlement prediction value sequence and generate a prediction masking marker field; at each future prediction step size, the future differential settlement sequence is calculated based on the future grid settlement prediction value sequence, and a differential settlement index set is formed.

[0157] It should be noted that the future differential settlement sequence includes the baseline differential settlement sequence, the adjacent differential settlement sequence along the route direction, and the transverse mirror differential settlement sequence.

[0158] It should be noted that the baseline differential settlement sequence is obtained by calculating the difference between the predicted future grid settlement value and the predicted future grid settlement value at the abutment baseline position for each grid location; the adjacent differential settlement sequence along the route is obtained by determining the adjacent grid locations along the route for each grid location according to the multi-scale adjacent span set, and calculating the difference between the predicted future grid settlement value of the grid location and the predicted future grid settlement value of the adjacent grid location along the route; the transverse mirror differential settlement sequence is obtained by determining the transverse mirror grid location for each grid location according to the mirror correspondence rule, and calculating the difference between the predicted future grid settlement value of the grid location and the predicted future grid settlement value of the transverse mirror grid location; the future differential settlement sequence is written to the data storage location and associated with the future prediction step size.

[0159] It should be noted that grid locations marked as invalid by the predicted masking field will not be included in the subsequent peak value statistics.

[0160] It should be noted that the differential settlement index set includes the baseline differential settlement index, the baseline differential settlement increment index, and the differential settlement index along the route direction.

[0161] It should be noted that the baseline differential settlement index is obtained by calculating the predicted differential settlement value of each grid position relative to the abutment baseline position at each future prediction step and statistically analyzing the peak value within the spatial range; the baseline differential settlement increment index is obtained by calculating the change in the baseline differential settlement index between adjacent future prediction steps and statistically analyzing the peak value within the spatial range; the differential settlement index along the route direction is obtained by calculating the predicted differential settlement value of adjacent spans along the route direction based on the multi-scale adjacent span set at each future prediction step and statistically analyzing the peak value corresponding to each span.

[0162] It should be noted that the generation of the prediction masking mark field specifically involves marking invalid grid positions at the current unified time point index of the grid mask at each future prediction step. Furthermore, when any grid position in the differential settlement sequence difference calculation is invalid in the grid mask, the corresponding grid position of the differential settlement sequence difference is marked as invalid in the prediction masking mark field, and the invalidity reason field is written.

[0163] Furthermore, after the training phase is completed, based on the samples in the training sample sequence that satisfy the target mask as valid samples, an early warning threshold interval is generated offline and written to the data storage location. For each type of peak index, a corresponding historical peak sequence is constructed, and a quantile threshold is calculated for the historical peak sequence. According to the value rules of the quantile threshold, the first threshold, the second threshold, and the third threshold are obtained, and a four-level threshold interval is defined.

[0164] It should be noted that the boundaries of the four-level threshold intervals satisfy the condition that the first threshold is less than the second threshold and the second threshold is less than the third threshold.

[0165] It should be noted that the rules for determining the quantile thresholds are as follows: the first threshold example is taken as the 0.90 quantile of the historical peak sequence, the second threshold example is taken as the 0.95 quantile of the historical peak sequence, and the third threshold example is taken as the 0.99 quantile of the historical peak sequence; the quantile calculation uses linear interpolation to obtain a unique threshold; the example values ​​are 5 mm for the first threshold example, 10 mm for the second threshold example, and 20 mm for the third threshold example.

[0166] It should be noted that the threshold intervals are defined as follows: when the peak index is less than the first threshold, it is determined to be a first-level interval; when the peak index is not less than the first threshold and is less than the second threshold, it is determined to be a second-level interval; when the peak index is not less than the second threshold and is less than the third threshold, it is determined to be a third-level interval; and when the peak index is not less than the third threshold, it is determined to be a fourth-level interval.

[0167] It should be noted that each element of the historical peak sequence is obtained by peak statistics within a defined statistical region and at a grid location marked as valid in the prediction masking field.

[0168] It should be noted that the defined statistical area is determined by the neighborhood range field, which includes the row span of the neighborhood along the line direction and the column span of the neighborhood in the horizontal direction. Taking the grid row index and grid column index corresponding to the bridge abutment reference position as the center, the rectangular neighborhood within the spatial grid is extracted according to the neighborhood range field as the defined statistical area.

[0169] Furthermore, for each future prediction step, the various peak indicators are compared with the quantile thresholds to obtain the corresponding indicator level. At each future prediction step, the threshold interval is compared for the baseline differential settlement peak, the baseline differential settlement increment peak, and the differential settlement peak along the line direction to obtain the indicator level of each of the three types of indicators. The warning level at the future prediction step is defined as the maximum value of the three types of indicator levels. The warning level is then associated with and stored in relation to the future prediction step.

[0170] Furthermore, within the spatial grid, according to the trigger location index rules, for peak indicators that reach the warning threshold, the corresponding grid row index and grid column index combination is recorded as the trigger location index; based on the reversible mapping relationship, the list of measurement point identifiers corresponding to the trigger location index is read from the reverse mapping table, and the list of measurement point identifiers is output as measurement point positioning information; the representative measurement point identifiers corresponding to the trigger location index are read from the representative measurement point table, and the representative measurement point identifiers are associated with the measurement point identifier list and output; the warning level, future prediction step size, trigger location index, representative measurement point identifier, measurement point identifier list, and corresponding peak indicator are written into the positioning output record field and written into the data storage location.

[0171] It should be noted that the trigger location index rule is as follows: within the limited statistical area and where the predicted masking marker field is valid, locate the indicator category that generates the highest level; within the differential settlement prediction value grid corresponding to the highest level indicator category, select the combination of the grid row index and grid column index where the peak value reaches the maximum value as the trigger location index; when the location where the peak value reaches the maximum value is not unique, select the trigger location index in the order of grid row index from small to large, and grid column index from small to large.

[0172] In this embodiment, to verify that the proposed gridded representation and positioning output mechanism stably provides warning points closer to the actual risk areas under conditions of missing data and construction disturbances in the scenario of bridge abutment construction settlement monitoring, and to evaluate the differences in positioning error distribution among different methods, a positioning error CDF curve is constructed as follows: Figure 5 As shown in the figure; in the experiment, a spatial grid was established based on the plane coordinates of the measuring points, and a forward mapping table and a reverse mapping table were generated to form a reversible mapping relationship; within each spatial grid position, a representative measuring point was selected from the set of measuring point identifiers, and the completed settlement observations were rearranged into grid settlement data, while a grid mask was generated simultaneously to indicate the effective position; the present invention and the control method were run under the same input conditions to obtain the trigger position index, and the corresponding measuring point identifier list and representative measuring point identifier were traced back through the reverse mapping table; the positioning distance error was defined as the distance between the grid center corresponding to the trigger position index and the true value anomaly position or the true value peak position, and the CDF was obtained by statistically analyzing all early warning events; Figure 5 The further to the left the curve is, the smaller the positioning error. The curve of this invention is shifted to the left relative to the control method, and the difference is more obvious in the tail interval. This indicates that in samples with high uncertainty, this invention can still reduce the proportion of large error positioning events, thereby improving the stability and traceability of the warning landing point.

[0173] In this embodiment, in order to verify that discrete settlement observation can form quantifiable and comparable spatial distribution results after spatial gridding under a unified spatial coordinate system, and to further verify that the grid position and measurement point identification can be consistently correlated and traced when triggering early warning and backtracking measurement points on the spatial distribution results, a simulation scenario including spatial grid, settlement measurement point identification and abnormal hotspots was constructed. Figure 6The mid-base plot shows the thermal distribution of the peak index on the spatial grid. The peak index is extracted from the cumulative result of grid settlement increment within the future prediction step and is used to characterize the concentrated areas and intensity differences of potential risks in the spatial domain. The plot marks the locations of true value anomaly hotspots as a reference for positioning accuracy. The plot also marks the trigger location index, which is the grid row index and grid column index corresponding to the peak index reaching its maximum value, used to determine the spatial grid location that triggers the warning level. After obtaining the trigger location index, the set of measurement point identifiers is obtained by backtracking at this spatial grid location through a reverse mapping relationship, forming a measurement point identifier list. The plot marks the location markers of the measurement points corresponding to the measurement point identifier list, realizing the closed-loop output of positioning from the trigger grid to the specific measurement point identifier. Through the annotation and correspondence, the spatial deviation between the trigger grid and the true value hotspot and the validity of the measurement point backtracking results can be intuitively verified.

[0174] In summary, this invention enhances the spatial interpretability and positioning accuracy of the settlement evolution process by: establishing a spatial grid and reversible mapping relationship and selecting representative measurement points, thereby enabling discrete settlement observations to form a regularized expression within a unified spatial framework; and by constructing a dual-branch convolutional neural network and introducing gating vectors to dynamically modulate network features, improving the reliability of future grid settlement increment prediction and early warning level determination, and achieving precise positioning and advanced early warning of measurement points that trigger early warning levels.

[0175] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for early warning of bridge abutment construction settlement trend based on convolutional neural networks, characterized in that, include: Settlement observation data and construction procedure code data are collected from settlement monitoring points. The settlement observation data and construction procedure code data are preprocessed to obtain supplementary settlement data and mask data. The filling height data is obtained based on the collection situation. Based on the plane coordinates of the settlement measuring points, a spatial grid and reversible mapping relationship are established, and representative measuring points are selected. Based on the representative measuring points, the settlement data is rearranged into grid settlement data, the mask data is rearranged into a grid mask, and the abutment reference position is determined. The absolute settlement sequence is extracted from the grid settlement data and concatenated with the grid mask to generate the first branch input. The differential settlement dictionary is generated in real time by the input time window. The differential mask is calculated based on the grid mask. The differential settlement dictionary and the differential mask are concatenated to generate the second branch input. By combining construction procedure coding data and filling height data, a gating vector is constructed, the monitoring target is defined as the future grid settlement increment and a corresponding target mask is constructed, and samples are screened by combining the grid mask and the target mask to obtain a sample set; A dual-branch convolutional neural network is constructed. Based on the sample set, the inputs of the first branch and the second branch, as well as the gating vector, are input into the dual-branch convolutional neural network. The network outputs the predicted data of future grid settlement increment, calculates the differential settlement index and obtains the warning level through classification, and outputs the measurement point location information that triggers the warning level through a reversible mapping relationship.

2. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The preprocessing of settlement observation data and construction procedure coding data includes: Settlement observation data were collected using measuring point identifiers and timestamps as indexes, and construction process code data were collected using process codes and process timestamps as indexes. A unified time-scale mapping is performed on the settlement observation data and construction procedure coding data, and the settlement observation data is marked with validity information to generate mask data; Based on the updated mask data, the missing locations in the settlement observation data are filled in along the time history direction to obtain the supplemented settlement data.

3. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 2, characterized in that, The unified time-scale mapping includes: The nearest neighbor priority mapping method is used to map settlement observation data and construction procedure code data with inconsistent timestamps to a unified time point index set; If no settlement observation data that meets the mapping conditions exists at the unified time point index, a null settlement record is generated at the corresponding time point index and marked as invalid in the mask data.

4. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The process of establishing a spatial grid and reversible mapping relationship based on the plane coordinates of settlement measuring points includes: A spatial grid consisting of grid row indices and grid column indices is generated based on the distribution range of the settlement measuring point plane coordinates and the grid step size; A reversible mapping relationship is formed by establishing a forward mapping relationship from settlement measuring point identifiers to grid row indices and grid column indices, and a reverse mapping relationship from grid row indices and grid column indices to the measuring point identifier set.

5. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The selection of representative measurement points includes: For each spatial grid location, a representative measuring point is selected from the set of measuring point identifiers mapped to the spatial grid location, based on the distance between the plane coordinates of the settlement measuring point and the center coordinates of the spatial grid location; When there are several candidate measuring points with the same distance, a unique representative measuring point is selected according to the measuring point identification sorting rules.

6. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 5, characterized in that, The step of synchronously rearranging the mask data into a grid mask includes: Based on representative measurement points, the supplementary settlement data will be rearranged into grid settlement data according to the grid row index, grid column index, and unified time point index; The mask data corresponding to the representative measurement points will be synchronously rearranged into a grid mask that corresponds one-to-one with the grid settlement data at a unified time point index.

7. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The real-time generation of the differential settlement dictionary for the input time window includes: Within the input time window, calculate the benchmark differential settlement of each grid position relative to the abutment benchmark position, and calculate the multi-scale adjacent differential settlement and lateral mirror differential settlement along the line direction. A differential settlement dictionary is constructed based on the baseline differential settlement, adjacent differential settlement, and lateral mirror differential settlement, and a differential mask is generated based on the grid mask.

8. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The construction of the gate vector includes: Within the gated time window, the construction process coding data is statistically analyzed to obtain the proportion of each process code appearing within the gated time window, and a process proportion vector is generated. Calculate the change in filling height based on the filling height data corresponding to the start and end time indexes of the gated time window; The process proportion vector, filling height change, and filling height data can be spliced ​​together using available marker information to generate a gating vector.

9. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 1, characterized in that, The dual-branch convolutional neural network includes: Anisotropic 3D convolutional feature extraction with spatial and temporal convolutional dimensions is performed on the first branch input and the second branch input respectively, and feature fusion is then performed. Input the gate vector into the gate mapping to generate the channel scaling factor; Based on the channel scaling factor, the convolutional features are scaled according to the channels to form a dynamic feature expression related to the construction process and filling status, and output the predicted data of future grid settlement increment.

10. The method for early warning of bridge abutment construction settlement trend based on convolutional neural networks as described in claim 9, characterized in that, The location information of the measuring points that trigger the early warning level by outputting the reversible mapping relationship includes: Within a defined statistical region, and where the predicted masking field is marked as valid, locate the indicator category that generates the highest level. Within the grid of differential settlement prediction values ​​corresponding to the highest level of indicator category, the combination of the grid row index and grid column index where the peak value reaches the maximum value is selected as the trigger location index; Based on the reversible mapping relationship, the list of measurement point identifiers corresponding to the trigger position index is read from the reverse mapping table, and the list of measurement point identifiers is output as measurement point positioning information. Read the representative measurement point identifier corresponding to the trigger position index from the representative measurement point table, and output the representative measurement point identifier along with the measurement point identifier list.

Images