Method for collaborative monitoring of horizontal displacement of deep foundation pit and underground water level based on pattern recognition

Abstract

This invention discloses a method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition, comprising the following steps: obtaining the original coordinated monitoring sequence and timestamp information; performing time alignment processing on the original coordinated monitoring sequence to obtain an aligned coordinated monitoring sequence; performing a candidate time delay set search on the aligned coordinated monitoring sequence to obtain the current time delay updated over time; performing time delay alignment on the aligned coordinated monitoring sequence to generate a coupled feature flow and an aligned residual flow; performing ChangeFinder online change point detection on the coupled feature flow and the aligned residual flow respectively to form a coordinated scoring sequence; performing coordinated consistency gating on the coordinated scoring sequence to determine change point events and record the change point times; extracting the change point window and completing the coordinated mode category determination; and outputting graded early warning results after completing anomaly attribution discrimination. This invention improves the accuracy and reliability of coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits.

Description

Technical Field

[0001] This invention relates to the technical field of deep foundation pit monitoring, and in particular to a method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition. Background Technology

[0002] During the construction of deep foundation pit projects, the horizontal displacement of the retaining structure and changes in groundwater level directly reflect the stability of the foundation pit and the response of the surrounding environment. Continuous monitoring and early warning are typically required throughout the entire construction cycle. Current technologies often employ distributed data collection from displacement and water level monitoring points, forming a data acquisition channel through a data acquisition terminal. Sample values ​​and timestamps are recorded according to a preset sampling period before being uploaded and aggregated, enabling long-term tracking of displacement and water level.

[0003] However, existing technologies still have shortcomings in the coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits. On the one hand, sampling frequency, communication delay, and missing data are common problems for both displacement and water level. Sequence alignment errors can weaken the coordination relationship, thus affecting the stability of anomaly identification. On the other hand, the response of groundwater level changes to the displacement of the retaining structure often has time-delay characteristics, and this time delay varies with the construction stage, precipitation conditions, and geological conditions. If a fixed time delay or unconstrained correlation analysis is used, time delay estimation jitter or mismatch with the actual response can easily occur, leading to distorted coordinated judgment. For online change point detection methods, if change points are directly determined by scoring thresholds without coordination consistency verification, false alarms and missed alarms are easily generated under construction noise, short-term disturbances, or observation errors. At the same time, if the update strategy of model parameters after a change point occurs is not processed in layers, state contamination or excessive resetting can easily occur, affecting the continuity and reliability of subsequent detection.

[0004] Therefore, how to provide a method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a collaborative monitoring method for horizontal displacement and groundwater level in deep foundation pits based on pattern recognition. This invention obtains the current time delay updated over time through time alignment, candidate time delay set search, and continuity constraints. It constructs coupled feature flow and aligned residual flow and performs online change point detection using dual-path ChangeFinder. The results are fused to form a collaborative scoring sequence and robustly determined through collaborative consistency gating. Then, within the change point window, the collaborative pattern category is determined and anomaly attribution is completed, and a graded early warning result is output. This reduces false alarms and missed alarms, and improves the accuracy of collaborative monitoring, the reliability of early warning, and the operability of handling.

[0006] The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to embodiments of the present invention includes the following steps: Collect monitoring data of deep foundation pits to obtain the original collaborative monitoring sequence and timestamp information; The original collaborative monitoring sequence is time-aligned based on the timestamp information to obtain an aligned collaborative monitoring sequence. Within a sliding window, a candidate time delay set search is performed on the aligned collaborative monitoring sequence, and a continuity constraint is imposed on the search results to obtain the current time delay updated over time. Based on the current time delay, the aligned collaborative monitoring sequence is aligned with the time delay to generate a coupled feature flow and an alignment residual flow; ChangeFinder online change point detection is performed on the coupled feature stream and the aligned residual stream respectively to obtain two change point score sequences. The two change point score sequences are then fused to form a collaborative score sequence. Perform collaborative consistency gating on the collaborative scoring sequence to obtain a corrected scoring sequence, and determine the change point based on the corrected scoring sequence, identify the change point event, and record the change point time; The system captures the change point window around the change point moment and determines the collaborative mode category. It then switches the warning threshold parameters and warning judgment strategy corresponding to the collaborative mode category, performs hierarchical updates on the internal state of ChangeFinder online change point detection, and outputs hierarchical warning results after completing the anomaly attribution judgment.

[0007] Optionally, obtaining the original collaborative monitoring sequence and timestamp information specifically includes: Horizontal displacement monitoring points and groundwater level monitoring points are set up in the deep foundation pit monitoring area, and data acquisition terminals are configured for each monitoring point to form a data acquisition channel; The data acquisition terminal samples the horizontal displacement monitoring point and the groundwater level monitoring point according to the preset sampling period, and generates the original sequence of horizontal displacement and the original sequence of groundwater level respectively. The sampling time under a unified clock reference is written into the sampling record to generate the horizontal displacement timestamp information and the groundwater level timestamp information. The original sequence of horizontal displacement, the original sequence of groundwater level, and the timestamp information of horizontal displacement and groundwater level are collected to form the original collaborative monitoring sequence and timestamp information.

[0008] Optionally, obtaining the aligned collaborative monitoring sequence specifically includes: Read the original collaborative monitoring sequence and timestamp information, perform unified clock reference verification, duplicate record merging, and time sequence arrangement on the timestamp information to generate the original sampling record set; A unified alignment timeline is constructed based on the original sampled record set. The start and end times of the unified alignment are determined, and the unified time step is set as the preset alignment period. Alignment time points are generated sequentially from the start time to the end time according to the unified time step. The original sampling record set is mapped to a unified alignment time axis. For each alignment time point, the adjacent previous and next sampling times in the corresponding subset are located. The alignment value of the current alignment time point is calculated based on the previous and next sampling values. If the alignment time point exceeds the sampling range of the corresponding subset, the boundary sampling value is used as the alignment value. Each alignment time point and alignment value are collected to form an alignment collaborative monitoring sequence.

[0009] Optionally, obtaining the current time delay specifically includes: Read the alignment and coordination monitoring sequence, set the sliding window length, sliding step size, number of candidate time delays and upper limit of candidate time delays, and extract window data window by window along the alignment and coordination monitoring sequence with the sliding step size. For each sliding window, obtain the displacement alignment value sequence, water level alignment value sequence and corresponding alignment time point sequence. Within each sliding window, a candidate time delay set is constructed. The candidate time delay set consists of different time delay candidate values. Each time delay candidate value is generated by a unified time step and candidate number. Within the sliding window, the time delay offset operation of the water level sequence is performed one by one according to the candidate time delay set. The water level alignment value sequence is moved forward or backward by the corresponding time delay step to form the water level lag sequence under the current candidate time delay. The water level lag sequence is then paired one by one with the unoffset displacement alignment value sequence according to the same alignment time point to form the paired sequence under the candidate time delay. For each candidate time delay, a coupling consistency index is calculated for the paired sequences. The adjacent sampling point difference is performed on the water level lag sequence and the displacement alignment value sequence to obtain the water level increment sequence and the displacement increment sequence. The cooperative change intensity is obtained by summing the products of the water level increment sequence and the displacement increment sequence at the same position in the window. The energy quantization value is calculated for the water level increment sequence and the displacement increment sequence respectively, and the cooperative change intensity is normalized to obtain the coupling consistency index of the candidate time delay. After traversing the candidate time delay set to obtain the coupling consistency index of each candidate time delay, the candidate time delay that makes the coupling consistency index reach the maximum value is selected as the search result of the current sliding window and recorded as the optimal time delay. Apply a continuity constraint to the optimal time delay of adjacent sliding windows, set the maximum allowable range of time delay variation, and read the time delay value of the previous sliding window. When the difference between the optimal time delay of the current sliding window and the time delay value of the previous sliding window does not exceed the maximum range, output the current time delay of the current sliding window updated over time. Otherwise, recursively update the time delay value of the previous sliding window with the maximum range according to the direction of the difference, and at the same time obtain the current time delay of the current sliding window updated over time.

[0010] Optionally, the generation of the coupled feature stream and the aligned residual stream specifically includes: Read the current time delay and convert it into the number of time delay steps based on a uniform time step. A time-delay offset operation is performed on the water level alignment value sequence based on the time delay step number. The water level alignment value sequence is backtracked forward according to the time delay step number to obtain the value. This makes each water level lag value in the water level lag sequence form a driving and response pairing relationship with the displacement alignment value at the same alignment time point. When the backtracking index exceeds the starting position of the water level alignment value sequence, the boundary value of the water level alignment value sequence is used to fill it, resulting in a time-delay aligned cooperative pairing sequence. Based on the time-delay aligned collaborative pairing sequence, the displacement increment and water level lag increment are calculated between adjacent sampling points according to the alignment time point order. The displacement alignment value, water level lag value, displacement increment, and water level lag increment are combined to form a coupled feature vector sequence that is updated with the alignment time point, and output as a coupled feature stream. Based on the time-delayed aligned cooperative pairing sequence, a pre-defined coupling response mapping relationship is established and the water level lag value is converted into a displacement estimate. The displacement estimate is then compared with the displacement alignment value at the same alignment time point to obtain the alignment residual. The alignment residuals at each alignment time point are then collected in chronological order to form an alignment residual sequence, which is then output as an alignment residual stream.

[0011] Optionally, the formation of the collaborative scoring sequence specifically includes: The coupled feature stream and the aligned residual stream are arranged into the first input sequence and the second input sequence respectively according to the alignment time point, and the parameters are set to update the forgetting coefficient and the fusion coefficient of the two-stage score. ChangeFinder online change point detection is performed on the first input sequence and the second input sequence respectively. At each alignment time point, a prediction update is performed on the input sequence to obtain the prediction residual and generate the first stage score value. The first stage score value is input into the second stage prediction update to obtain the second stage score value. During the parameter update process, a dual-rate mechanism of short-term update and long-term update is adopted for the first stage and the second stage respectively to form a short-term state and a long-term state. The short-term state performs fast update on the samples in the latest window, and the long-term state performs slow update on the samples in the historical window. Based on the parameter update forgetting coefficient, the parameters of the short-term state and the long-term state are updated exponentially. At the same time, a restricted migration constraint is applied to the parameter change amplitude. At each alignment time point, the residual scale parameter is updated robustly according to the magnitude of the prediction residual and the residual exceeding the preset magnitude upper limit is truncated. The second stage score value is used as the change point score value of the current alignment time point and is collected in chronological order to form two change point score sequences. The two variable point scoring sequences are fused. Within a preset fusion window, the two variable point scoring sequences are subjected to scale unification processing to obtain two normalized scoring sequences. A first fusion weight and a second fusion weight are set. Based on the fusion coefficient of the two-stage scoring, the two normalized scores are weighted and summed according to the corresponding fusion weight to obtain the collaborative scoring value. The collaborative scoring values ​​at each aligned time point are then collected in chronological order to form a collaborative scoring sequence.

[0012] Optionally, obtaining the change point event and the change point time specifically includes: Read the collaborative scoring sequence, and set collaborative consistency gating parameters and change point determination parameters. The change point determination parameters include a scoring threshold and a number of consecutive determination steps. Establish a storage structure for the corrected scoring sequence and a recording structure for change point events. Based on the alignment of the collaborative monitoring sequence and the alignment of the residual flow, the gated input is calculated. The gated input includes the water level-side driving intensity, the displacement-side response intensity, and the residual consistency index. The driving intensity value, response intensity value, and consistency index value are subjected to scale unification processing to obtain normalized values. At the same time, the normalized values ​​are weighted and combined to generate the gated weight. Based on the gating weight, the collaborative score sequence is subjected to collaborative consistency gating. For each alignment time point, the collaborative score value at the alignment time point is weighted with the gating weight at the alignment time point to obtain the corrected score value. The corrected score values ​​at each alignment time point are then aggregated in chronological order to form the corrected score sequence. The change point determination is performed based on the modified scoring sequence. The modified score value is compared with the scoring threshold point by point in the order of the alignment time point. The alignment time points that meet the condition of not being less than the scoring threshold are continuously counted. When the continuous count reaches the number of continuous determination steps, the change point event is determined, and the earliest alignment time point that meets the continuous counting condition corresponding to the change point event is recorded as the change point time.

[0013] Optionally, the output of the tiered early warning results specifically includes: Determine the change point time, set the forward and backward lengths of the change point window, determine the start and end alignment time points of the change point window on the unified alignment time axis, extract the values ​​of the alignment collaborative monitoring sequence, alignment residual sequence, and correction scoring sequence within the change point window in the order of alignment time points, and collect them to form the change point window data; Based on the variable point window data, adjacent difference is performed, and the absolute value of the difference result is taken and accumulated within the variable point window to obtain the cooperative mode discrimination quantity. Set up a set of collaborative mode categories and collaborative mode determination rules. Perform category determination on the collaborative mode discriminant according to the collaborative mode determination rules to obtain the collaborative mode category. Record the collaborative mode category as the anomaly attribution result. Establish a correspondence table between collaborative mode categories and early warning threshold parameters and early warning judgment strategies. Locate the early warning threshold parameters and early warning judgment strategies that match the collaborative mode categories in the correspondence table, and switch to the matching early warning threshold parameters and early warning judgment strategies. Based on the switched warning threshold parameters and warning judgment strategy, the internal state of ChangeFinder online change point detection is updated hierarchically. Based on the warning threshold parameters and warning judgment strategy, the corrected score sequence in the backward interval of the change point window is judged point by point in a hierarchical warning manner to obtain the hierarchical warning result. The hierarchical warning result is then associated with the anomaly attribution result and recorded.

[0014] The beneficial effects of this invention are: This invention constructs an aligned collaborative monitoring sequence under a unified aligned time axis by performing time alignment processing on the original collaborative monitoring sequence. This makes horizontal displacement and groundwater level data comparable and calculable at the same aligned time point, reducing the impact of asynchronous sampling, missing data, and interpolation errors on collaborative analysis, and providing a stable data foundation for subsequent collaborative identification and early warning judgment.

[0015] This invention performs a candidate time delay set search within a sliding window and applies a continuity constraint to the search results to obtain the current time delay updated over time. This enables the time delay relationship between groundwater level changes and the horizontal displacement response of the retaining structure to be adaptively tracked online, avoiding mismatch caused by fixed time delays and jitter caused by unconstrained estimation, thereby improving the accuracy and continuity of the cooperative coupling relationship characterization.

[0016] This invention generates a coupled feature stream and an aligned residual stream based on the current time delay, and then performs ChangeFinder online change point detection to form two change point scoring sequences before fusing them. This can simultaneously reflect two types of abnormal features: changes in coupling relationship and coupling mismatch, thereby enhancing the ability to identify and the robustness of different abnormal forms.

[0017] This invention introduces a collaborative consistency gating system to obtain a modified scoring sequence and determines the change point event and change point time accordingly. The scoring is checked and corrected through collaborative consistency constraints, which reduces false alarms and missed alarms caused by construction noise, short-term disturbances and observation errors, and improves the reliability of change point determination.

[0018] This invention focuses on capturing the change point window at the change point moment to determine the collaborative mode category, switching the warning threshold parameters and warning judgment strategy corresponding to the collaborative mode category, and performing hierarchical updates on the internal state of ChangeFinder to maintain online detection stability. After completing the anomaly attribution judgment, it outputs graded warning results, so that the warning output not only gives the anomaly degree classification, but also provides interpretable collaborative mode attribution and targeted judgment strategy support, thereby improving the accuracy of deep foundation pit collaborative monitoring, the reliability of warnings, and the operability of engineering handling. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 The flowchart shows the method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition proposed in this invention. Fig. 2 This is a schematic diagram of the ChangeFinder collaborative change point detection and hierarchical early warning mechanism of the pattern recognition-based deep foundation pit horizontal displacement and groundwater level collaborative monitoring method proposed in this invention. Detailed Implementation

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

[0021] refer to Figs. 1-2 A pattern recognition-based method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits includes the following steps: Collect monitoring data of deep foundation pits to obtain the original collaborative monitoring sequence and timestamp information; The original collaborative monitoring sequence is time-aligned based on the timestamp information to obtain an aligned collaborative monitoring sequence. Within a sliding window, a candidate time delay set search is performed on the aligned collaborative monitoring sequence, and a continuity constraint is imposed on the search results to obtain the current time delay updated over time. Based on the current time delay, the aligned collaborative monitoring sequence is aligned with the time delay to generate a coupled feature flow and an alignment residual flow; ChangeFinder online change point detection is performed on the coupled feature stream and the aligned residual stream respectively to obtain two change point score sequences. The two change point score sequences are then fused to form a collaborative score sequence. Perform collaborative consistency gating on the collaborative scoring sequence to obtain a corrected scoring sequence, and determine the change point based on the corrected scoring sequence, identify the change point event, and record the change point time; The system captures the change point window around the change point moment and determines the collaborative mode category. It then switches the warning threshold parameters and warning judgment strategy corresponding to the collaborative mode category, performs hierarchical updates on the internal state of ChangeFinder online change point detection, and outputs hierarchical warning results after completing the anomaly attribution judgment.

[0022] In this embodiment, obtaining the original collaborative monitoring sequence and timestamp information specifically includes: Horizontal displacement monitoring points and groundwater level monitoring points are set up in the deep foundation pit monitoring area. The horizontal displacement monitoring points are used to collect horizontal displacement monitoring data of the retaining structure, and the groundwater level monitoring points are used to collect groundwater level monitoring data. Data acquisition terminals are configured for each monitoring point to form a data acquisition channel. The data acquisition terminal samples the horizontal displacement monitoring point and the groundwater level monitoring point according to the preset sampling period, and generates the original sequence of horizontal displacement and the original sequence of groundwater level respectively. The sampling time under a unified clock reference is written into the sampling record to generate the horizontal displacement timestamp information and the groundwater level timestamp information. The original sequence of horizontal displacement, the original sequence of groundwater level, and the timestamp information of horizontal displacement and groundwater level are collected to form the original collaborative monitoring sequence and timestamp information.

[0023] In this embodiment, obtaining the aligned collaborative monitoring sequence specifically includes: Read the original collaborative monitoring sequence and timestamp information, perform unified clock reference verification, duplicate record merging, and time sequence organization on the timestamp information to generate the original sampling record set, which includes a subset of displacement sampling records and a subset of water level sampling records; A unified alignment timeline is constructed based on the original sampling record set. The start and end times of the unified alignment are determined. The start time is the earlier of the earliest sampling times in the displacement sampling record subset and the water level sampling record subset. The end time is the later of the latest sampling times in the displacement sampling record subset and the water level sampling record subset. The unified time step is set as the preset alignment period. The alignment time points are generated sequentially from the start time according to the unified time step until the end time. The original sampling record set is mapped to a unified alignment time axis. For each alignment time point, the adjacent previous and next sampling times in the corresponding subset are located. The alignment value of the current alignment time point is calculated based on the previous and next sampling values. If the alignment time point exceeds the sampling range of the corresponding subset, the boundary sampling value is used as the alignment value. Each alignment time point and alignment value are collected to form an alignment collaborative monitoring sequence.

[0024] In this embodiment, obtaining the current time delay specifically includes: Read the alignment and collaborative monitoring sequence, set the sliding window length, sliding step size, number of candidate time delays and upper limit of candidate time delays, and extract window data window by window along the alignment and collaborative monitoring sequence with the sliding step size. For each sliding window, obtain the displacement alignment value sequence, water level alignment value sequence and corresponding alignment time point sequence. The time delay is the time offset of the groundwater level change relative to the horizontal displacement response of the retaining structure. Within each sliding window, a candidate time delay set is constructed. The candidate time delay set consists of different time delay candidate values. Each time delay candidate value is generated by a unified time step and candidate number. Within the sliding window, the time delay offset operation of the water level sequence is performed one by one according to the candidate time delay set. The water level alignment value sequence is moved forward or backward by the corresponding time delay step to form the water level lag sequence under the current candidate time delay. The water level lag sequence is then paired with the unoffset displacement alignment value sequence one by one according to the same alignment time point to form the paired sequence under the candidate time delay. For each candidate time delay, a coupling consistency index is calculated for the paired sequences. The adjacent sampling point difference is performed on the water level lag sequence and the displacement alignment value sequence to obtain the water level increment sequence and the displacement increment sequence. The cooperative change intensity is obtained by summing the products of the water level increment sequence and the displacement increment sequence at the same position in the window. The energy quantization value is calculated for the water level increment sequence and the displacement increment sequence respectively, and the cooperative change intensity is normalized to obtain the coupling consistency index of the candidate time delay. After traversing the candidate time delay set to obtain the coupling consistency index of each candidate time delay, the candidate time delay that makes the coupling consistency index reach the maximum value is selected as the search result of the current sliding window and recorded as the optimal time delay. Apply a continuity constraint to the optimal time delay of adjacent sliding windows, set the maximum allowable range of time delay variation, and read the time delay value of the previous sliding window. When the difference between the optimal time delay of the current sliding window and the time delay value of the previous sliding window does not exceed the maximum range, output the current time delay of the current sliding window updated over time. Otherwise, recursively update the time delay value of the previous sliding window with the maximum range according to the direction of the difference, and at the same time obtain the current time delay of the current sliding window updated over time.

[0025] In this embodiment, the generation of the coupled feature flow and the aligned residual flow specifically includes: Read the current time delay and convert it into the number of time delay steps based on a uniform time step. A time-delay offset operation is performed on the water level alignment value sequence based on the time delay step number. The water level alignment value sequence is backtracked forward according to the time delay step number to obtain the value. This makes each water level lag value in the water level lag sequence form a driving and response pairing relationship with the displacement alignment value at the same alignment time point. When the backtracking index exceeds the starting position of the water level alignment value sequence, the boundary value of the water level alignment value sequence is used to fill it, resulting in a time-delay aligned cooperative pairing sequence. Based on the time-delay aligned collaborative pairing sequence, the displacement increment and water level lag increment are calculated between adjacent sampling points according to the alignment time point order. The displacement alignment value, water level lag value, displacement increment, and water level lag increment are combined to form a coupled feature vector sequence that is updated with the alignment time point, and output as a coupled feature stream. Based on the time-delayed aligned cooperative pairing sequence, a pre-defined coupling response mapping relationship is established and the water level lag value is converted into a displacement estimate. The displacement estimate is then compared with the displacement alignment value at the same alignment time point to obtain the alignment residual. The alignment residuals at each alignment time point are then collected in chronological order to form an alignment residual sequence, which is then output as an alignment residual stream.

[0026] This invention performs time-delay offset on the water level alignment value sequence based on the current time delay updated over time and establishes a driving and response pairing relationship with the displacement alignment value. Boundary filling ensures the continuous usability of the pairing sequence. On this basis, a coupled feature flow containing displacement alignment value, water level lag value and its incremental information is constructed to characterize the cooperative change law. At the same time, according to the preset coupling response mapping relationship, the water level lag value is converted into the displacement estimate value and forms an alignment residual flow with the displacement alignment value to characterize the degree of coupling mismatch. This enables parallel characterization of cooperative relationship changes and model mismatch and improves the sensitivity and robustness of subsequent online change point detection and early warning judgment.

[0027] In this embodiment, the formation of the collaborative scoring sequence specifically includes: The coupled feature stream and the aligned residual stream are arranged into the first input sequence and the second input sequence respectively according to the alignment time point, and the parameters are set to update the forgetting coefficient and the fusion coefficient of the two-stage score. ChangeFinder online change point detection is performed on the first and second input sequences respectively. At each alignment time point, a prediction update is performed on the input sequence to obtain the prediction residual and generate the first stage score value. The first stage score value is input into the second stage prediction update to obtain the second stage score value. During the parameter update process, a dual-rate mechanism of short-term update and long-term update is adopted for the first and second stages respectively to form a short-term state and a long-term state. The short-term state performs fast update on the samples in the latest window, and the long-term state performs slow update on the samples in the historical window. Based on the parameter update forgetting coefficient, the parameters of the short-term state and the long-term state are updated exponentially. At the same time, a restricted migration constraint is applied to the parameter change amplitude. At each alignment time point, the residual scale parameter is updated robustly according to the magnitude of the prediction residual and the residual exceeding the preset magnitude upper limit is truncated. The second stage score value is used as the change point score value of the current alignment time point and is collected in chronological order to form two change point score sequences. The two variable-point scoring sequences are fused. Within a preset fusion window, the two variable-point scoring sequences are subjected to scaling uniformity processing to obtain two normalized scoring sequences. The scaling uniformity processing includes mapping the current score to a range of zero to one based on the minimum and maximum scores within the fusion window, setting a first fusion weight and a second fusion weight, the sum of the first fusion weight and the second fusion weight being one, and weighting the two normalized scores according to the fusion coefficient of the two-stage scores by the corresponding fusion weight to obtain a collaborative score value. The collaborative score values ​​at each aligned time point are then aggregated in chronological order to form a collaborative score sequence.

[0028] This invention constructs two inputs—coupled feature stream and aligned residual stream—and uses ChangeFinder two-stage online change point detection to obtain two change point scores. It combines a dual-rate mechanism of short-term and long-term updates, parameter update forgetting coefficient, restricted migration constraints, and robust update and truncation processing of residual scale to enhance the ability to accommodate both abrupt and gradual changes and suppress noise and outlier propagation. Then, the two scores are scaled and weighted according to the fusion weight within the fusion window to form a collaborative score sequence, thereby improving the stability, consistency, and comparability of collaborative anomaly representation and enhancing the reliability of subsequent change point determination and early warning.

[0029] In this embodiment, obtaining the change point event and the change point time specifically includes: Read the collaborative scoring sequence and set the collaborative consistency gating parameters and change point judgment parameters. The collaborative consistency gating parameters include the gating window length, the range of gating weight values, and the gating combination coefficient. The change point judgment parameters include the scoring threshold and the number of consecutive judgment steps. Establish the storage structure of the corrected scoring sequence and the recording structure of the change point event. Based on the alignment collaborative monitoring sequence and the alignment residual flow, the gating input is calculated. The gating input includes the water level-side driving intensity, the displacement-side response intensity, and the residual consistency index. The calculation of the water level-side driving intensity, the displacement-side response intensity, and the residual consistency index is carried out within the gating window. The difference between the groundwater level alignment value, the horizontal displacement alignment value, and the alignment residual value at adjacent alignment time points is obtained to obtain the water level change, driving intensity value, and displacement change value, respectively. At the same time, they are aggregated to obtain the driving intensity value, response intensity value, and consistency index value. The driving intensity value, response intensity value, and consistency index value are scaled and normalized to obtain normalized values. The normalized values ​​are then weighted and combined to generate the gating weight. Based on the gating weight, the collaborative score sequence is subjected to collaborative consistency gating. For each alignment time point, the collaborative score value at the alignment time point is weighted with the gating weight at the alignment time point to obtain the corrected score value. The corrected score values ​​at each alignment time point are then aggregated in chronological order to form the corrected score sequence. The change point determination is performed based on the modified scoring sequence. The modified score value is compared with the scoring threshold point by point in the order of the alignment time point. The alignment time points that meet the condition of not being less than the scoring threshold are continuously counted. When the continuous count reaches the number of continuous determination steps, the change point event is determined, and the earliest alignment time point that meets the continuous counting condition corresponding to the change point event is recorded as the change point time.

[0030] This invention introduces a collaborative consistency gating mechanism into the collaborative scoring sequence. It uses the water level-side driving strength, displacement-side response strength, and residual consistency index (composed of groundwater level alignment value, horizontal displacement alignment value, and alignment residual value) within the gating window to generate gating weights and perform weighted correction on the scores, forming a more collaborative and robust corrected scoring sequence. Combined with the scoring threshold and the number of consecutive judgment steps, it achieves stable determination of change point events and change point moments, thereby suppressing false triggers caused by short-term disturbances and noise, reducing false alarms and missed alarms, and improving the accuracy of change point location and the reliability of early warning judgment in deep foundation pit collaborative monitoring.

[0031] In this embodiment, the output of the graded early warning results specifically includes: Determine the change point time, set the forward and backward lengths of the change point window, determine the start and end alignment time points of the change point window on the unified alignment time axis, extract the values ​​of the alignment collaborative monitoring sequence, alignment residual sequence, and correction scoring sequence within the change point window in the order of alignment time points, and collect them to form the change point window data; Based on the variable point window data, adjacent difference is performed, and the absolute value of the difference result is taken and accumulated within the variable point window to obtain the cooperative mode discrimination quantity. Set up a set of collaborative mode categories and collaborative mode determination rules. Perform category determination on the collaborative mode discriminant according to the collaborative mode determination rules to obtain the collaborative mode category. Record the collaborative mode category as the anomaly attribution result. Establish a correspondence table between collaborative mode categories and early warning threshold parameters and early warning judgment strategies. Locate the early warning threshold parameters and early warning judgment strategies that match the collaborative mode categories in the correspondence table, and switch to the matching early warning threshold parameters and early warning judgment strategies. Based on the switched warning threshold parameters and warning judgment strategy, the internal state of ChangeFinder online change point detection is updated in a hierarchical manner. The hierarchical update includes resetting or quickly updating the short-term state and slowly updating the long-term state under restricted migration constraints. Based on the warning threshold parameters and warning judgment strategy, the corrected score sequence within the backward interval of the change point window is judged point by point in a hierarchical manner to obtain hierarchical warning results. The hierarchical warning results are associated with the anomaly attribution results and recorded. The hierarchical warning results express the degree of anomaly of the corrected score sequence in different grades. By combining the collaborative mode category and the anomaly attribution results, different anomaly types are distinguished and the corresponding warning threshold parameters and warning judgment strategies are matched to reduce false alarms and false negatives in collaborative monitoring and improve the operability of warnings. This enables monitoring personnel to take corresponding inspection, verification and disposal measures according to the warning level and anomaly attribution category.

[0032] This invention implements time alignment and time-delay adaptive tracking on horizontal displacement and groundwater level monitoring sequences, constructs coupled feature flow and aligned residual flow, and performs dual-path online change point detection and collaborative consistency gating. This enables robust change point determination and change point window location for collaborative anomalies. Furthermore, based on collaborative mode discriminant, it completes the determination of collaborative mode categories and anomaly attribution, and switches the warning threshold parameters and warning judgment strategies according to mode categories. At the same time, it implements hierarchical updates to the internal state of ChangeFinder to maintain online detection stability, thereby outputting interpretable hierarchical warning results, reducing false alarms and false negatives, improving the accuracy of deep foundation pit collaborative monitoring, the reliability of warnings, and the operability of engineering handling.

[0033] Example 1: To verify the feasibility of the present invention in practice, it was applied to the construction monitoring scenario of a typical deep foundation pit project. This project is located in a condition where groundwater has a significant impact and the surrounding environment is sensitive. The horizontal displacement of the retaining structure and the groundwater level show alternating characteristics with the construction stage. On-site monitoring suffers from problems such as inconsistent sampling frequency, communication delay, occasional missing measurements, and noise disturbances. This results in the displacement sequence and the water level sequence not being completely synchronized at the same time. Moreover, the response of groundwater level changes to the displacement of the retaining structure has a significant time lag and changes with the stage. Traditional methods using a single displacement threshold or fixed time lag correlation analysis are prone to false alarms and missed alarms. Especially in the case of short-term precipitation disturbances, local rebounds, and sensor jumps, alarms lack consistency verification and anomaly attribution. On-site handling often relies on experience verification, and the operability of early warning is insufficient.

[0034] In this scenario, the monitoring system deploys multiple horizontal displacement monitoring points and groundwater level monitoring points around the foundation pit. Data acquisition terminals record sampled values ​​and timestamps according to a preset sampling period and aggregate them to form an original collaborative monitoring sequence. In actual operation, it is observed that the displacement sampling period is relatively long, while the groundwater level sampling is more frequent, and timestamp offsets and missing measurements occur during certain periods due to network jitter. To make the two types of data comparable and computable, this invention first constructs a unified aligned time axis based on timestamp information, maps the original sampling records to unified aligned time points, and generates an aligned collaborative monitoring sequence. During the alignment process, duplicate records are merged, and missing measurement points are filled using adjacent sampling interpolation or boundary value filling to ensure sequence continuity. Subsequently, a candidate time delay set search is performed on the aligned and coordinated monitoring sequence within a sliding window. By constructing a pairing relationship between the water level lag sequence and the displacement alignment value sequence under each candidate time delay, the coupling consistency index is calculated to select the optimal time delay. Continuity constraints are applied to the optimal time delay of adjacent windows to ensure that the current time delay is updated smoothly over time, avoiding time delay jitter caused by instantaneous correlation alone. Time delay offset is performed on the water level alignment value sequence based on the current time delay to form a coordinated pairing sequence after time delay alignment. On this basis, a coupling feature flow is constructed to describe the coordinated change pattern of the displacement alignment value, water level lag value, and their adjacent increments. At the same time, according to the preset coupling response mapping relationship, the water level lag value is converted into a displacement estimate and forms an alignment residual flow with the displacement alignment value to describe the degree of mismatch in the coordinated relationship.

[0035] The two inputs are respectively fed into the ChangeFinder online change point detection module to obtain two change point scoring sequences, which are then fused to form a collaborative scoring sequence. The scoring is then corrected using collaborative consistency gating. The gating input consists of the water level-side driving strength, displacement-side response strength, and residual consistency index within the gating window. The gating weights distinguish and verify situations such as strong water level disturbances but weak displacement response, sudden displacement changes but stable water level, and significant deterioration of residual consistency, making the corrected scoring sequence more consistent with the collaborative mechanism.

[0036] The change point determination uses a scoring threshold and a number of consecutive determination steps to avoid single-point spikes. After identifying the change point event and recording the change point time, the change point window is extracted to form change point window data. Within the window, the cooperative pattern discrimination quantity is obtained by accumulating the absolute values ​​of adjacent differences. The cooperative pattern category is determined according to rules and used as the anomaly attribution result. At the same time, the warning threshold parameter and warning determination strategy matching the cooperative pattern category are switched. The internal state of ChangeFinder is updated hierarchically. The short-term state is reset or updated quickly to adapt to the local statistical characteristics after the mutation. The long-term state is updated slowly under the constraint of limited migration to maintain baseline stability. Finally, the corrected scoring sequence in the backward interval of the change point window is subjected to hierarchical warning determination and outputs hierarchical warning results, thereby providing the field with an executable output of anomaly level, attribution category and time location.

[0037] To demonstrate its beneficial effects, monitoring data during continuous operation were statistically analyzed without specifying specific times and locations. The data covered multiple construction phases and included various typical events. Event labels were generated from on-site verification records and multi-source monitoring comparisons. Three commonly used engineering methods were selected as baselines for comparison: a single-variable early warning method based solely on displacement thresholds, a correlation discrimination method with fixed time delay plus single-path change point detection, and a dual-path change point detection fusion method without collaborative consistency gating. The method of this invention was compared and evaluated under the same data and alarm response requirements. Specific comparative experiments are shown in Table 1. Table 1. Performance Comparison of Different Methods in Coordinated Monitoring of Displacement and Water Level in Deep Foundation Pit

[0038] Table 1 shows that the method of the present invention achieves an event detection rate of 0.93, which is higher than 0.78 for single-variable displacement threshold, 0.83 for fixed time delay correlation and single-path change point, and 0.88 for dual-path change point fusion without gate. Simultaneously, it reduces both the false alarm rate and the missed alarm rate to 0.07, significantly lower than 0.21 and 0.22, 0.16 and 0.17, and 0.13 and 0.12. Regarding the stability of change point positioning, the present invention reduces the median positioning deviation to 4 alignment steps and the 90th percentile to 10 alignment steps, both superior to 9 and 24, 7 and 19, and 6 and 16. Furthermore, it increases the median effective early warning lead time to 7 alignment steps, higher than 3, 4, and 5. This indicates that the present invention has comprehensive advantages in terms of the reliability of collaborative anomaly identification, suppression of false alarms and missed alarms, and early warning capability. The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition, characterized in that, Includes the following steps: Collect monitoring data of deep foundation pits to obtain the original collaborative monitoring sequence and timestamp information; The original collaborative monitoring sequence is time-aligned based on the timestamp information to obtain an aligned collaborative monitoring sequence. Within a sliding window, a candidate time delay set search is performed on the aligned collaborative monitoring sequence, and a continuity constraint is imposed on the search results to obtain the current time delay updated over time. Based on the current time delay, the aligned collaborative monitoring sequence is aligned with the time delay to generate a coupled feature flow and an alignment residual flow; ChangeFinder online change point detection is performed on the coupled feature stream and the aligned residual stream respectively to obtain two change point score sequences. The two change point score sequences are then fused to form a collaborative score sequence. Perform collaborative consistency gating on the collaborative scoring sequence to obtain a corrected scoring sequence, and determine the change point based on the corrected scoring sequence, identify the change point event, and record the change point time; The system captures the change point window around the change point moment and determines the collaborative mode category. It then switches the warning threshold parameters and warning judgment strategy corresponding to the collaborative mode category, performs hierarchical updates on the internal state of ChangeFinder online change point detection, and outputs hierarchical warning results after completing the anomaly attribution judgment.

2. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The acquisition of the original collaborative monitoring sequence and timestamp information specifically includes: Horizontal displacement monitoring points and groundwater level monitoring points are set up in the deep foundation pit monitoring area, and data acquisition terminals are configured for each monitoring point to form a data acquisition channel; The data acquisition terminal samples the horizontal displacement monitoring point and the groundwater level monitoring point according to the preset sampling period, and generates the original sequence of horizontal displacement and the original sequence of groundwater level respectively. The sampling time under a unified clock reference is written into the sampling record to generate the horizontal displacement timestamp information and the groundwater level timestamp information. The original sequence of horizontal displacement, the original sequence of groundwater level, and the timestamp information of horizontal displacement and groundwater level are collected to form the original collaborative monitoring sequence and timestamp information.

3. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The specific steps to obtain the alignment and collaborative monitoring sequence include: Read the original collaborative monitoring sequence and timestamp information, perform unified clock reference verification, duplicate record merging, and time sequence arrangement on the timestamp information to generate the original sampling record set; A unified alignment timeline is constructed based on the original sampled record set. The start and end times of the unified alignment are determined, and the unified time step is set as the preset alignment period. Alignment time points are generated sequentially from the start time to the end time according to the unified time step. The original sampling record set is mapped to a unified alignment time axis. For each alignment time point, the adjacent previous and next sampling times in the corresponding subset are located. The alignment value of the current alignment time point is calculated based on the previous and next sampling values. If the alignment time point exceeds the sampling range of the corresponding subset, the boundary sampling value is used as the alignment value. Each alignment time point and alignment value are collected to form an alignment collaborative monitoring sequence.

4. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The current time delay is obtained specifically in the following ways: Read the alignment and coordination monitoring sequence, set the sliding window length, sliding step size, number of candidate time delays and upper limit of candidate time delays, and extract window data window by window along the alignment and coordination monitoring sequence with the sliding step size. For each sliding window, obtain the displacement alignment value sequence, water level alignment value sequence and corresponding alignment time point sequence. Within each sliding window, a candidate time delay set is constructed. The candidate time delay set consists of different time delay candidate values. Each time delay candidate value is generated by a unified time step and candidate number. Within the sliding window, the time delay offset operation of the water level sequence is performed one by one according to the candidate time delay set. The water level alignment value sequence is moved forward or backward by the corresponding time delay step to form the water level lag sequence under the current candidate time delay. The water level lag sequence is then paired one by one with the unoffset displacement alignment value sequence according to the same alignment time point to form the paired sequence under the candidate time delay. For each candidate time delay, a coupling consistency index is calculated for the paired sequences. The adjacent sampling point difference is performed on the water level lag sequence and the displacement alignment value sequence to obtain the water level increment sequence and the displacement increment sequence. The cooperative change intensity is obtained by summing the products of the water level increment sequence and the displacement increment sequence at the same position in the window. The energy quantization value is calculated for the water level increment sequence and the displacement increment sequence respectively, and the cooperative change intensity is normalized to obtain the coupling consistency index of the candidate time delay. After traversing the candidate time delay set to obtain the coupling consistency index of each candidate time delay, the candidate time delay that makes the coupling consistency index reach the maximum value is selected as the search result of the current sliding window and recorded as the optimal time delay. Apply a continuity constraint to the optimal time delay of adjacent sliding windows, set the maximum allowable range of time delay variation, and read the time delay value of the previous sliding window. When the difference between the optimal time delay of the current sliding window and the time delay value of the previous sliding window does not exceed the maximum range, output the current time delay of the current sliding window updated over time. Otherwise, recursively update the time delay value of the previous sliding window with the maximum range according to the direction of the difference, and at the same time obtain the current time delay of the current sliding window updated over time.

5. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The generation of the coupled feature stream and the aligned residual stream specifically includes: Read the current time delay and convert it into the number of time delay steps based on a uniform time step. A time-delay offset operation is performed on the water level alignment value sequence based on the time delay step number. The water level alignment value sequence is backtracked forward according to the time delay step number to obtain the value. This makes each water level lag value in the water level lag sequence form a driving and response pairing relationship with the displacement alignment value at the same alignment time point. When the backtracking index exceeds the starting position of the water level alignment value sequence, the boundary value of the water level alignment value sequence is used to fill it, resulting in a time-delay aligned cooperative pairing sequence. Based on the time-delay aligned collaborative pairing sequence, the displacement increment and water level lag increment are calculated between adjacent sampling points according to the alignment time point order. The displacement alignment value, water level lag value, displacement increment, and water level lag increment are combined to form a coupled feature vector sequence that is updated with the alignment time point, and output as a coupled feature stream. Based on the time-delayed aligned cooperative pairing sequence, a pre-defined coupling response mapping relationship is established and the water level lag value is converted into a displacement estimate. The displacement estimate is then compared with the displacement alignment value at the same alignment time point to obtain the alignment residual. The alignment residuals at each alignment time point are then collected in chronological order to form an alignment residual sequence, which is then output as an alignment residual stream.

6. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The formation of the collaborative scoring sequence specifically includes: The coupled feature stream and the aligned residual stream are arranged into the first input sequence and the second input sequence respectively according to the alignment time point, and the parameters are set to update the forgetting coefficient and the fusion coefficient of the two-stage score. ChangeFinder online change point detection is performed on the first input sequence and the second input sequence respectively. At each alignment time point, a prediction update is performed on the input sequence to obtain the prediction residual and generate the first stage score value. The first stage score value is input into the second stage prediction update to obtain the second stage score value. During the parameter update process, a dual-rate mechanism of short-term update and long-term update is adopted for the first stage and the second stage respectively to form a short-term state and a long-term state. The short-term state performs fast update on the samples in the latest window, and the long-term state performs slow update on the samples in the historical window. Based on the parameter update forgetting coefficient, the parameters of the short-term state and the long-term state are updated exponentially. At the same time, a restricted migration constraint is applied to the parameter change amplitude. At each alignment time point, the residual scale parameter is updated robustly according to the magnitude of the prediction residual and the residual exceeding the preset magnitude upper limit is truncated. The second stage score value is used as the change point score value of the current alignment time point and is collected in chronological order to form two change point score sequences. The two variable point scoring sequences are fused. Within a preset fusion window, the two variable point scoring sequences are subjected to scale unification processing to obtain two normalized scoring sequences. A first fusion weight and a second fusion weight are set. Based on the fusion coefficient of the two-stage scoring, the two normalized scores are weighted and summed according to the corresponding fusion weight to obtain the collaborative scoring value. The collaborative scoring values ​​at each aligned time point are then collected in chronological order to form a collaborative scoring sequence.

7. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The specific methods for obtaining the changing point event and the changing point time include: Read the collaborative scoring sequence, and set collaborative consistency gating parameters and change point determination parameters. The change point determination parameters include a scoring threshold and a number of consecutive determination steps. Establish a storage structure for the corrected scoring sequence and a recording structure for change point events. Based on the alignment of the collaborative monitoring sequence and the alignment of the residual flow, the gated input is calculated. The gated input includes the water level-side driving intensity, the displacement-side response intensity, and the residual consistency index. The driving intensity value, response intensity value, and consistency index value are subjected to scale unification processing to obtain normalized values. At the same time, the normalized values ​​are weighted and combined to generate the gated weight. Based on the gating weight, the collaborative score sequence is subjected to collaborative consistency gating. For each alignment time point, the collaborative score value at the alignment time point is weighted with the gating weight at the alignment time point to obtain the corrected score value. The corrected score values ​​at each alignment time point are then aggregated in chronological order to form the corrected score sequence. The change point determination is performed based on the modified scoring sequence. The modified score value is compared with the scoring threshold point by point in the order of the alignment time point. The alignment time points that meet the condition of not being less than the scoring threshold are continuously counted. When the continuous count reaches the number of continuous determination steps, the change point event is determined, and the earliest alignment time point that meets the continuous counting condition corresponding to the change point event is recorded as the change point time.

8. The method for coordinated monitoring of horizontal displacement and groundwater level in deep foundation pits based on pattern recognition according to claim 1, characterized in that, The output of the tiered early warning results specifically includes: Determine the change point time, set the forward and backward lengths of the change point window, determine the start and end alignment time points of the change point window on the unified alignment time axis, extract the values ​​of the alignment collaborative monitoring sequence, alignment residual sequence, and correction scoring sequence within the change point window in the order of alignment time points, and collect them to form the change point window data; Based on the variable point window data, adjacent difference is performed, and the absolute value of the difference result is taken and accumulated within the variable point window to obtain the cooperative mode discrimination quantity. Set up a set of collaborative mode categories and collaborative mode determination rules. Perform category determination on the collaborative mode discriminant according to the collaborative mode determination rules to obtain the collaborative mode category. Record the collaborative mode category as the anomaly attribution result. Establish a correspondence table between collaborative mode categories and early warning threshold parameters and early warning judgment strategies. Locate the early warning threshold parameters and early warning judgment strategies that match the collaborative mode categories in the correspondence table, and switch to the matching early warning threshold parameters and early warning judgment strategies. Based on the switched warning threshold parameters and warning judgment strategy, the internal state of ChangeFinder online change point detection is updated hierarchically. Based on the warning threshold parameters and warning judgment strategy, the corrected score sequence in the backward interval of the change point window is judged point by point in a hierarchical warning manner to obtain the hierarchical warning result. The hierarchical warning result is then associated with the anomaly attribution result and recorded.

Images