Deep foundation pit multi-source monitoring data generative adversarial network anomaly diagnosis method

Abstract

The invention discloses a deep foundation pit multi-source monitoring data generative adversarial network anomaly diagnosis method, and relates to the technical field of civil engineering, and the method comprises the following steps: collecting multi-source heterogeneous monitoring data from a deep foundation pit automatic monitoring system, and carrying out the normalization processing; based on the preprocessed normalized multivariate time sequence, constructing a condition vector containing static and dynamic context information for each sample, and forming an input form which can be directly processed by a network model; designing a generator and discriminator network for the constructed sample fragments and condition vectors; integrating a space-time constraint loss function into a total loss function of the generator, and physically constraining generated data; initializing network parameters and hyper-parameters, and alternately updating discriminator and generator parameters to obtain a generator model; and searching an optimal noise vector for a to-be-detected sample through an optimization process by utilizing the generator model for training convergence, calculating a comprehensive anomaly score, and judging anomaly according to a threshold value.

Description

Technical Field

[0001] This invention belongs to the field of civil engineering technology, and in particular relates to a method for diagnosing anomalies in deep foundation pit multi-source monitoring data using adversarial generative networks. Background Technology

[0002] As a key component of major infrastructure construction such as urban rail transit and high-rise buildings, deep foundation pit engineering relies on the real-time and accurate analysis of multi-source monitoring data to ensure the stability of the support structure and the safety of the surrounding environment during its construction. Deep foundation pit monitoring involves multi-source heterogeneous data such as horizontal displacement of the top of the support structure, axial force of the support, surface settlement and groundwater level. However, there are many technical bottlenecks in the current field of deep foundation pit monitoring data processing and anomaly diagnosis.

[0003] Traditional anomaly diagnosis methods fail to adequately utilize the spatiotemporal correlation of monitoring data. In the temporal dimension, they do not effectively constrain data continuity, leading to unreasonable jumps in data that interfere with diagnostic results. In the spatial dimension, they neglect the physical correlations between monitoring points, resulting in generated monitoring data samples that do not conform to the actual physical characteristics of the engineering project. Furthermore, existing generative model-based diagnostic techniques do not fully integrate the characteristics of deep foundation pit monitoring data into the generator and discriminator network design. The loss function does not incorporate spatiotemporal constraints, making it difficult to balance the realism and physical rationality of generated samples. This results in poor network training convergence, and the anomaly score calculation relies solely on a single error index, lacking dynamic threshold settings. Consequently, the accuracy of anomaly detection is low, with high false alarm and false negative rates, failing to meet the high-precision and real-time requirements of deep foundation pit engineering safety monitoring. Therefore, we propose an adversarial generative network anomaly diagnosis method based on multi-source monitoring data for deep foundation pits. Summary of the Invention

[0004] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0005] This invention provides a method for diagnosing anomalies in multi-source monitoring data of deep foundation pits using adversarial generative networks, comprising the following steps:

[0006] Step S1: Multi-source monitoring data acquisition and preprocessing: Collect multi-source heterogeneous monitoring data from the deep foundation pit automated monitoring system, perform time alignment, cleaning, filling in missing data and normalization processing to provide a data foundation for the subsequent construction of effective samples;

[0007] Step S2: Spatiotemporal sequence sample and condition vector construction: Based on the preprocessed normalized multivariate time series, the sliding window method is used to cut it into fixed-length sample segments, and a condition vector containing static and dynamic context information is constructed for each sample to form an input form that the network model can directly process.

[0008] Step S3: Generator and Discriminator Network Design: For the constructed sample segments and condition vectors, design generator and discriminator networks. The generator takes noise vectors and condition vectors as inputs and outputs the generated time series segments of multi-source monitoring data. The discriminator takes real / generated data sequences and condition vectors as inputs and outputs the probability that they are real data, providing the basic architecture for network adversarial training.

[0009] Step S4: Definition and integration of spatiotemporal constraint loss function: In order to enable the designed generator to generate samples that conform to the spatiotemporal patterns of deep foundation pit monitoring data, a spatiotemporal constraint loss function composed of time smoothing constraints and spatial physical constraints is defined and integrated into the generator's total loss function to impose reasonable constraints on the physical nature of the generated data.

[0010] Step S5: Network Training and Optimization: Based on the network structure and the defined loss function, initialize the network parameters and hyperparameters, alternately update the discriminator and generator parameters, and achieve network convergence through iterative optimization to obtain a generator model that can generate normal monitoring data patterns;

[0011] Step S6, Anomaly Diagnosis and Alarm: Using the trained and converged generator model, the optimal noise vector is found through the optimization process of the sample to be detected, the comprehensive anomaly score is calculated, and the anomaly is determined according to the threshold, triggering the alarm mechanism, thus completing the final output from data to anomaly diagnosis results.

[0012] Furthermore, step S1 includes the following steps:

[0013] Step S11: Collect multi-source heterogeneous monitoring data of horizontal displacement of the top of the support structure, axial force of the support, surface settlement and groundwater level from the deep foundation pit automated monitoring system, and align the data with a unified timestamp to form a multi-dimensional time series dataset with fixed time intervals.

[0014] Step S12: Clean the raw data, handle data loss caused by recording errors and communication interruptions, fill in short-term missing data using linear interpolation, and remove long-term missing data.

[0015] Step S13: Different monitoring data have different dimensions and ranges. To eliminate the influence of dimensions and bring the data into a unified range to facilitate subsequent network model processing, each type of monitoring data is normalized separately and mapped to the [-1,1] or [0,1] interval. The normalization formula is as follows:

[0016]

[0017] In the formula, X norm X represents the normalized monitoring data value, and X represents the original monitoring data value. min X is the minimum value among this type of monitoring data.max is the maximum value in this type of monitoring data, and 2-1 is a linear transformation that maps the data in the [0,1] interval to the [-1,1] interval.

[0018] Furthermore, step S2 includes the following steps:

[0019] Step S21: Use the sliding window method to segment the multivariate time series into sample segments of fixed length T, with a sliding step size of S. step Each sample is a matrix of dimension (S,T), where S is the total number of monitoring points;

[0020] Step S22: Construct a condition vector C for each sample segment. This vector contains the sample's pit partition number, excavation stage code, weather condition code, and adjacent construction activity flag information.

[0021] Furthermore, step S3 includes the following steps:

[0022] Step S31: The input of the generator G includes a noise vector Z and a conditional vector C sampled from a standard normal distribution. Z and C are concatenated and used as input. The network outputs a generated data matrix with dimension (S,T).

[0023] Step S32: The input of the discriminator D is a data sequence of dimension (s,t) and its condition vector C. Features are extracted through a series of one-dimensional convolutional layers, and the condition information is fused with the data features. Finally, a scalar probability value is output through a fully connected layer and a Sigmoid activation function.

[0024] Furthermore, step S4 includes the following steps:

[0025] Step S41: Deep foundation pit monitoring data is continuous in the time dimension, and the data changes between adjacent time points are usually relatively smooth. To constrain the temporal continuity of the generated data and avoid unreasonable jumps, a time smoothing constraint L is defined. T The formula is as follows:

[0026]

[0027] In the formula, L T The time-smoothing constraint loss value is given by S, where S is the total number of monitoring points and T is the time length of the sample segment. This is the averaging coefficient, used to normalize the loss value. To perform a summation operation on all monitoring points, To perform a summation operation over all adjacent time steps, G(Z,C) is the generated data matrix output by the generator. s,t+1 To generate the value of the s-th monitoring point at time t+1 in the data, G(Z,C)s,t To generate the value of the s-th monitoring point at time t in the data, It is the square of the L2 norm, used to measure the magnitude of the difference;

[0028] Step S42: There is a physical correlation between the monitoring points in the deep foundation pit (such as the mutual influence of deformation in different parts of the support structure). To ensure that the generated data conforms to the laws of space physics, a space physical constraint L is defined. S The formula is as follows:

[0029]

[0030] In the formula, L S This represents the spatial physical constraint loss value. This is the averaging coefficient, used to normalize the loss values ​​over the time dimension. To perform a summation operation over all time steps, M is a predefined spatial correlation matrix, G(Z,C):,t represents the values ​​of all monitoring points at time t in the generated data, and ":" indicates taking all rows;

[0031] Step S43: Taking into account the constraints of temporal continuity and spatial physical correlation on the generated data, the temporal smoothing constraint and spatial physical constraint are weighted and combined to obtain the spatiotemporal constraint loss L. ST The formula is as follows:

[0032] L ST =λ T L T +λ S L S ;

[0033] In the formula, L ST Let λ be the total loss value due to spatiotemporal constraints. T Let λ be the weight hyperparameter of the time smoothing constraint. S The weight hyperparameters for space physical constraints;

[0034] Step S44: The generator needs to simultaneously generate realistic samples (adversarial training target) and samples that conform to spatiotemporal patterns (constraint target). To balance these two objectives, the binary cross-entropy adversarial loss and the spatiotemporal constraint loss are combined to form the generator's total loss function, as shown in the following formula:

[0035] L G =L BCE (D(G(Z,C)),1)+λ×L ST ;

[0036] In the formula, L G L represents the total loss of the generator. BCEis the binary cross-entropy loss function, used to measure the difference between the classification probability and the true label, D(·) is the output of the discriminator, and λ is the weight hyperparameter of the spatiotemporal constraint loss.

[0037] Furthermore, step S5 includes the following steps:

[0038] Step S51: Initialize the generator and discriminator parameters, and set the number of training rounds, batch size, optimizer and its learning rate;

[0039] Step S52: For each batch of training data, sample the noise vector Z and the condition vector C, and the generator generates fake data G(Z,C);

[0040] Step S53: Update the discriminator: Calculate its loss with respect to the real data and the generated data, fix the generator parameters, and update the discriminator parameters through backpropagation;

[0041] Step S54: Update the generator: Fix the discriminator parameters, calculate the adversarial loss and spatiotemporal constraint loss of the generator, and update the generator parameters through backpropagation;

[0042] Step S55: Repeat the iteration until the network converges or reaches the maximum number of training rounds.

[0043] Furthermore, step S6 includes the following steps:

[0044] Step S61: Find the noise vector that enables the generator to reconstruct the data that is closest to the sample to be detected and conforms to the spatiotemporal pattern. Determine the optimal noise vector Z through the optimization process. * The formula is as follows:

[0045]

[0046] In the formula, Z * This is the optimal noise vector. To find the variable Z that minimizes the expression within the parentheses, MSE is the mean square error function, and X... test For the sample data to be detected, G(Z,C) test ) represents the generator under the input noise Z and the condition C of the sample to be detected. test Output at time, C test Let γ be the condition vector of the sample to be detected, and let γ be the weighting coefficient of the spatiotemporal constraint loss.

[0047] Step S62: Comprehensively evaluate the differences between the sample to be detected and the reconstructed sample, as well as the spatiotemporal rationality of the reconstructed sample, to more accurately determine whether the sample is abnormal, and calculate the reconstructed data X. recon =G(Z) * C test The formula for the comprehensive anomaly score AS is as follows:

[0048] AS = α × MSE(X) test ,X recon )+β×L ST (X recon );

[0049] In the formula, AS is the comprehensive anomaly score, α is the weighting coefficient of the mean square error, and X... recon Use the optimal noise Z for the generator * The generated data corresponding to the sample to be detected, where β is the weighting coefficient of the spatiotemporal constraint loss, and L ST (X recon () represents the spatiotemporal constraint loss value of the reconstructed data;

[0050] Step S63: Set a dynamic threshold. When AS exceeds the threshold, it is judged as abnormal and an alarm mechanism is triggered.

[0051] The present invention has the following beneficial effects:

[0052] 1. This invention designs a generator and discriminator network structure specifically for the characteristics of deep foundation pit monitoring data. The generator takes a noise vector and a condition vector containing information such as foundation pit partitioning and excavation stages as input. The discriminator extracts features and fuses the condition information through a one-dimensional convolutional layer. Simultaneously, it employs an alternating parameter update training method, combining adversarial loss and spatiotemporal constraint loss to optimize the network. This solves the problems of existing generative model network designs not fitting the characteristics of monitoring data and having poor training convergence. This design enables the network to stably converge to a state that generates normal monitoring data, improving the reliability of the model in generating normal data and laying a foundation for accurate anomaly diagnosis.

[0053] 2. This invention optimizes and determines the optimal noise vector during anomaly diagnosis to generate reconstructed data that closely matches the sample to be tested. It calculates anomaly scores by integrating the mean square error and the spatiotemporal constraint loss of the reconstructed data, and sets a dynamic threshold for anomaly determination. This solves the problems of traditional methods that rely on a single indicator for anomaly scores and lack a dynamic threshold, leading to high false alarm and false negative rates. This comprehensive evaluation method can fully reflect the anomaly situation of the samples, effectively improve the accuracy of anomaly determination, reduce false alarms and false negatives, meet the high precision and real-time requirements of deep foundation pit engineering safety monitoring, and provide reliable monitoring assurance for engineering construction safety.

[0054] 3. By integrating temporal smoothing constraints and spatial physical constraints into the generator's total loss function, the problem of insufficient utilization of the spatiotemporal correlation of monitoring data in traditional anomaly diagnosis methods is effectively solved. Temporal smoothing constraints can avoid unreasonable time jumps in the generated data, ensuring the continuity of data time. Spatial physical constraints can conform to the physical correlation patterns between monitoring points in deep foundation pits, making the generated samples both realistic and consistent with the actual characteristics of engineering projects. This compensates for the shortcomings of traditional methods in generating data that does not conform to the actual physical characteristics of engineering projects, providing high-quality data support for subsequent anomaly diagnosis and improving the adaptability of data samples to deep foundation pit monitoring scenarios.

[0055] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. 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.

[0057] Figure 1 This is a flowchart illustrating the anomaly diagnosis method of deep foundation pit multi-source monitoring data adversarial generative network according to the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1 As shown, this invention provides a method for diagnosing anomalies in deep foundation pit multi-source monitoring data using adversarial generative networks, comprising the following steps:

[0060] Step S1: Multi-source monitoring data acquisition and preprocessing: Collect multi-source heterogeneous monitoring data from the deep foundation pit automated monitoring system, perform time alignment, cleaning, filling in missing data and normalization processing to provide a data foundation for the subsequent construction of effective samples;

[0061] Step S2: Spatiotemporal sequence sample and condition vector construction: Based on the preprocessed normalized multivariate time series, the sliding window method is used to cut it into fixed-length sample segments, and a condition vector containing static and dynamic context information is constructed for each sample to form an input form that the network model can directly process.

[0062] Step S3: Generator and Discriminator Network Design: For the constructed sample segments and condition vectors, design generator and discriminator networks. The generator takes noise vectors and condition vectors as inputs and outputs the generated time series segments of multi-source monitoring data. The discriminator takes real / generated data sequences and condition vectors as inputs and outputs the probability that they are real data, providing the basic architecture for network adversarial training.

[0063] Step S4: Definition and integration of spatiotemporal constraint loss function: In order to enable the designed generator to generate samples that conform to the spatiotemporal patterns of deep foundation pit monitoring data, a spatiotemporal constraint loss function composed of time smoothing constraints and spatial physical constraints is defined and integrated into the generator's total loss function to impose reasonable constraints on the physical nature of the generated data.

[0064] Step S5: Network Training and Optimization: Based on the network structure and the defined loss function, initialize the network parameters and hyperparameters, alternately update the discriminator and generator parameters, and achieve network convergence through iterative optimization to obtain a generator model that can generate normal monitoring data patterns;

[0065] Step S6, Anomaly Diagnosis and Alarm: Using the trained and converged generator model, the optimal noise vector is found through the optimization process of the sample to be detected, the comprehensive anomaly score is calculated, and the anomaly is determined according to the threshold, triggering the alarm mechanism, thus completing the final output from data to anomaly diagnosis results.

[0066] Step S1 includes the following steps:

[0067] Step S11: Collect multi-source heterogeneous monitoring data of horizontal displacement of the top of the support structure, axial force of the support, surface settlement and groundwater level from the deep foundation pit automated monitoring system, and align the data with a unified timestamp to form a multi-dimensional time series dataset with fixed time intervals.

[0068] Step S12: Clean the raw data, handle data loss caused by recording errors and communication interruptions, fill in short-term missing data using linear interpolation, and remove long-term missing data.

[0069] Step S13: Different monitoring data have different dimensions and ranges. To eliminate the influence of dimensions and bring the data into a unified range to facilitate subsequent network model processing, each type of monitoring data is normalized separately and mapped to the [-1,1] or [0,1] interval. The normalization formula is as follows:

[0070]

[0071] In the formula, X norm X represents the normalized monitoring data value, and X represents the original monitoring data value. min X is the minimum value among this type of monitoring data. maxis the maximum value in this type of monitoring data, and 2-1 is a linear transformation that maps the data in the [0,1] interval to the [-1,1] interval.

[0072] Step S2 includes the following steps:

[0073] Step S21: Use the sliding window method to segment the multivariate time series into sample segments of fixed length T, with a sliding step size of S. step Each sample is a matrix of dimension (S,T), where S is the total number of monitoring points;

[0074] Step S22: Construct a condition vector C for each sample segment. This vector contains the sample's pit partition number, excavation stage code, weather condition code, and adjacent construction activity flag information.

[0075] Step S3 includes the following steps:

[0076] Step S31: The input to generator G includes a noise vector Z and a conditional vector C sampled from a standard normal distribution. Z and C are concatenated and used as input. The network outputs a generated data matrix of dimension (S,T).

[0077] Step S32: The input to the discriminator D is a data sequence of dimension (s,t) and its condition vector C. Features are extracted through a series of one-dimensional convolutional layers, and the condition information is fused with the data features. Finally, a scalar probability value is output through a fully connected layer and a Sigmoid activation function.

[0078] Step S4 includes the following steps:

[0079] Step S41: Deep foundation pit monitoring data is continuous in the time dimension, and the data changes between adjacent time points are usually relatively smooth. To constrain the temporal continuity of the generated data and avoid unreasonable jumps, a time smoothing constraint L is defined. T The formula is as follows:

[0080]

[0081] In the formula, L T The time-smoothing constraint loss value is given by S, where S is the total number of monitoring points and T is the time length of the sample segment. This is the averaging coefficient, used to normalize the loss value. To perform a summation operation on all monitoring points, To perform a summation operation over all adjacent time steps, G(Z,C) is the generated data matrix output by the generator. s,t+1 To generate the value of the s-th monitoring point at time t+1 in the data, G(Z,C) s,t To generate the value of the s-th monitoring point at time t in the data, It is the square of the L2 norm, used to measure the magnitude of the difference;

[0082] Step S42: There is a physical correlation between the monitoring points in the deep foundation pit (such as the mutual influence of deformation in different parts of the support structure). To ensure that the generated data conforms to the laws of space physics, a space physical constraint L is defined. S The formula is as follows:

[0083]

[0084] In the formula, L S This represents the spatial physical constraint loss value. This is the averaging coefficient, used to normalize the loss values ​​over the time dimension. To perform a summation operation over all time steps, M is a predefined spatial correlation matrix, G(Z,C):,t represents the values ​​of all monitoring points at time t in the generated data, and ":" indicates taking all rows;

[0085] Step S43: Taking into account the constraints of temporal continuity and spatial physical correlation on the generated data, the temporal smoothing constraint and spatial physical constraint are weighted and combined to obtain the spatiotemporal constraint loss L. ST The formula is as follows:

[0086] L ST =λ T L T +λ S L S ;

[0087] In the formula, L ST Let λ be the total loss value due to spatiotemporal constraints. T Let λ be the weight hyperparameter of the time smoothing constraint. S The weight hyperparameters for space physical constraints;

[0088] Step S44: The generator needs to simultaneously generate realistic samples (adversarial training target) and samples that conform to spatiotemporal patterns (constraint target). To balance these two objectives, the binary cross-entropy adversarial loss and the spatiotemporal constraint loss are combined to form the generator's total loss function, as shown in the following formula:

[0089] L G =L BCE (D(G(Z,C)),1)+λ×L ST ;

[0090] In the formula, L G L represents the total loss of the generator. BCE is the binary cross-entropy loss function, used to measure the difference between the classification probability and the true label, D(·) is the output of the discriminator, and λ is the weight hyperparameter of the spatiotemporal constraint loss.

[0091] Step S5 includes the following steps:

[0092] Step S51: Initialize the generator and discriminator parameters, and set the number of training rounds, batch size, optimizer and its learning rate;

[0093] Step S52: For each batch of training data, sample the noise vector Z and the condition vector C, and the generator generates fake data G(Z,C);

[0094] Step S53: Update the discriminator: Calculate its loss with respect to the real data and the generated data, fix the generator parameters, and update the discriminator parameters through backpropagation;

[0095] Step S54: Update the generator: Fix the discriminator parameters, calculate the adversarial loss and spatiotemporal constraint loss of the generator, and update the generator parameters through backpropagation;

[0096] Step S55: Repeat the iteration until the network converges or reaches the maximum number of training rounds.

[0097] Step S6 includes the following steps:

[0098] Step S61: Find the noise vector that enables the generator to reconstruct the data that is closest to the sample to be detected and conforms to the spatiotemporal pattern. Determine the optimal noise vector Z through the optimization process. * The formula is as follows:

[0099]

[0100] In the formula, Z * This is the optimal noise vector. To find the variable Z that minimizes the expression within the parentheses, MSE is the mean square error function, and X... test For the sample data to be detected, G(Z,C) test ) represents the generator under the input noise Z and the condition C of the sample to be detected. test Output at time, C test Let γ be the condition vector of the sample to be detected, and let γ be the weighting coefficient of the spatiotemporal constraint loss.

[0101] Step S62: Comprehensively evaluate the differences between the sample to be detected and the reconstructed sample, as well as the spatiotemporal rationality of the reconstructed sample, to more accurately determine whether the sample is abnormal, and calculate the reconstructed data X. recon =G(Z) * C test The formula for the comprehensive anomaly score AS is as follows:

[0102] AS = α × MSE(X) test ,X recon )+β×L ST (X recon );

[0103] In the formula, AS is the comprehensive anomaly score, α is the weighting coefficient of the mean square error, and X... recon Use the optimal noise Z for the generator * The generated data corresponding to the sample to be detected, where β is the weighting coefficient of the spatiotemporal constraint loss, and L ST (X recon () represents the spatiotemporal constraint loss value of the reconstructed data;

[0104] Step S63: Set a dynamic threshold. When AS exceeds the threshold, it is judged as abnormal and an alarm mechanism is triggered.

[0105] One specific application of this embodiment is:

[0106] Implementation scenario:

[0107] Taking the deep foundation pit project of Phase I of Metro Line 3 as an example, the foundation pit is 180m long, 25m wide, and 22m deep, and adopts a "diaphragm wall + internal bracing" support system with a construction period of 6 months. The project requires real-time monitoring of the stability of the support structure and the safety of the surrounding environment. Therefore, the method of this invention is used to conduct multi-source monitoring data anomaly diagnosis, with the specific monitoring range covering the entire foundation pit area and the surrounding 50m area.

[0108] Implementation steps:

[0109] Step 1:

[0110] 1. Data types of monitoring: Four core monitoring indicators were selected, with the following specific information: 30 monitoring points were set for the horizontal displacement of the top of the support structure, with a monitoring accuracy of ±0.1mm and a sampling interval of 1 hour; 20 monitoring points were set for the axial force of the support, with a monitoring accuracy of ±1kN and a sampling interval of 1 hour; 40 monitoring points were set for the surface settlement, with a monitoring accuracy of ±0.1mm and a sampling interval of 1 hour; and 15 monitoring points were set for the groundwater level, with a monitoring accuracy of ±1cm and a sampling interval of 1 hour.

[0111] Time alignment: Based on the GPS timestamp of the automated monitoring system, the time nodes of all monitoring data are unified to form a multi-dimensional time series dataset with an interval of 1 hour and a total duration of 180 days. The total amount of data is 180 days × 24 hours × (30 horizontal displacement monitoring points + 20 support axial force monitoring points + 40 surface settlement monitoring points + 15 groundwater level monitoring points) = 180 × 24 × 105 = 453,600 records.

[0112] 2. Error data processing: Abnormal values ​​caused by sensor malfunctions (such as sudden changes in horizontal displacement to 100mm or negative support axial force) were removed, and a total of 128 error data entries were processed;

[0113] Missing data processing: "Short-term missing" is defined as continuous missing data ≤ 3 hours, which is filled using linear interpolation (e.g., for missing groundwater level monitoring data, based on 12.5m at 7:00 and 12.7m at 11:00, interpolation yields 12.55m at 8:00, 12.60m at 9:00, and 12.65m at 10:00); "Long-term missing" is defined as continuous missing data > 6 hours, which is removed from the monitoring data corresponding to that time period, resulting in the removal of 3 long-term missing data sequences.

[0114] 3. Select the normalization interval as [-1, 1], and calculate using the formula described in the patent. Taking "horizontal displacement at the top of the support structure" as an example: In the original range of this type of data, the minimum value X min = -5.2mm (displacement to the outside of the pit is negative), maximum value X max = 8.6mm (displacement towards the inside of the pit is positive); If the original data X = 2.1mm, the normalization calculation process is as follows: First calculate the numerator 2.1 - (-5.2) = 7.3, the denominator 8.6 - (-5.2) = 13.8, divide the two to get 7.3 / 13.8 ≈ 0.529, multiply by 2 to get ≈ 1.06, and finally subtract 1 to get 0.06, which is the normalized value X of the data. norm =0.06; Similarly, normalize all four types of monitoring data to ensure consistent data units;

[0115] Step Two:

[0116] 1. Sliding window parameters: Set fixed length T = 24 (corresponding to 24 hours of continuous data, covering one complete monitoring cycle), sliding step size S step =12 (One sample is generated every 12 hours to balance the number of samples and overlap);

[0117] Sample dimension calculation: The total number of monitoring points S is 30 horizontal displacement monitoring points + 20 support axial force monitoring points + 40 surface settlement monitoring points + 15 groundwater level monitoring points = 105. Therefore, each sample is a matrix with dimension (105, 24).

[0118] Total sample size: The total monitoring time is 180 days × 24 hours = 4320 hours. The sample size is calculated as (total time - fixed sample length) / sliding step + 1, that is, (4320-24) / 12+1 = 359+1 = 360 samples, which meets the data requirements for network training.

[0119] 2. The condition vector C for each sample is a 4-dimensional vector. The encoding rules and examples for each dimension are as follows:

[0120] Pit zoning numbering: The pit is divided into 3 zones (zone 1, zone 2, and zone 3), which are coded as 1, 2, and 3 respectively. For example, the sample code corresponding to zone 2 is 2.

[0121] Excavation stage coding: The construction process is divided into 4 stages, namely the preparation stage (code 1), the excavation stage (code 2), the bottom sealing stage (code 3), and the backfilling stage (code 4). For example, the sample in the excavation stage is coded as 2.

[0122] Weather condition coding: Based on the actual weather, it is divided into sunny (code 1), cloudy (code 2), and rainy (code 3). For example, the sample code corresponding to rainy days is 3.

[0123] Adjacent construction activity markers: Based on the surrounding construction situation, they are divided into no construction (code 0), dewatering construction (code 1), and earthwork excavation (code 2). For example, the sample code for a dewatering construction period is 1.

[0124] Step 3:

[0125] 1. Input layer: The input contains a noise vector Z (dimension 100) and a conditional vector C (dimension 4) sampled from the standard normal distribution N(0,1). The two are concatenated to form an input vector with a dimension of 100 + 4 = 104.

[0126] Network structure: It consists of 3 fully connected layers and 2 one-dimensional deconvolutional layers, as shown in the following figure:

[0127] Fully connected layer 1: Input dimension 104, output dimension 105×6 (used to map to low-dimensional spatiotemporal features), activation function is ReLU;

[0128] Fully connected layer 2: Input dimension 105×6, output dimension 105×12, activation function is ReLU;

[0129] Fully connected layer 3: Input dimension 105×12, output dimension 105×24, activation function is ReLU;

[0130] One-dimensional deconvolutional layer 1: The kernel size is 3×1, the number of kernels is 64, and the output dimension is 105×24;

[0131] One-dimensional deconvolution layer 2: The kernel size is 3×1, the number of kernels is 32, and the output dimension is 105×24 (consistent with the sample dimension);

[0132] 2. Input layer: The input contains a data sequence (real data or generated data, with dimensions (105,24)) and a condition vector C (with dimensions 4). The two are concatenated to form an input vector with dimensions 105×24+4=2524.

[0133] Network structure: It consists of 3 one-dimensional convolutional layers and 2 fully connected layers, as shown in the following figure:

[0134] One-dimensional convolutional layer 1: The kernel size is 3×1, the number of kernels is 32, the stride is 1, and the activation function is LeakyReLU (where α=0.2);

[0135] One-dimensional convolutional layer 2: The kernel size is 3×1, the number of kernels is 64, the stride is 1, and the activation function is LeakyReLU (where α=0.2);

[0136] One-dimensional convolutional layer 3: The kernel size is 3×1, the number of kernels is 128, the stride is 1, and the activation function is LeakyReLU (where α=0.2);

[0137] Fully connected layer 1: Input dimension 128×24, output dimension 512, activation function is LeakyReLU (where α=0.2);

[0138] Fully connected layer 2: Input dimension 512, output dimension 1, activation function is Sigmoid;

[0139] Output: A scalar probability value in the range [0,1]. The closer the value is to 1, the more likely the input data is to be real and normal data; the closer the value is to 0, the more likely the input data is to be generated data or abnormal data.

[0140] Step Four:

[0141] 1. Given that the total number of monitoring points S = 105 and the sample time length T = 24, calculate the time smoothing constraint loss L according to the formula described in the patent. T First, calculate the averaging coefficients, 1 / S = 1 / 105 and 1 / (T-1) = 1 / 23, respectively. Then, perform a double summation on all monitoring points (from s = 1 to s = 105) and all adjacent time steps (from t = 1 to t = 23). The summation term is the squared L2 norm of the difference between the value at time t+1 and the value at time t of the s-th monitoring point in the generated data matrix G(Z,C). Finally, multiply the double summation result by the two averaging coefficients to obtain the time smoothing constraint loss L. T ;

[0142] For example, in a certain generated sample, the value of the 5th monitoring point (the horizontal displacement monitoring point at the top of the support structure) is 0.12 at time t=10 and 0.15 at time t=11. The squared L2 norm of the difference between the two is (0.15-0.12). 2 =0.0009; After traversing all 105 monitoring points and 23 adjacent time steps, the time smoothing constraint loss L for this sample is calculated. T =0.025;

[0143] 2. Predefined spatial correlation matrix M: Based on the mechanical characteristics of deep foundation pits, a correlation coefficient of 0.8 is set for adjacent monitoring points (such as axial force monitoring points of the same support and surface settlement monitoring points of adjacent areas), and a correlation coefficient of 0.2 is set for non-adjacent monitoring points. The matrix has a dimension of 105×105.

[0144] Given a sample time length T = 24, calculate the spatial physical constraint loss L according to the formula described in the patent. S First, calculate the averaging coefficient 1 / T = 1 / 24; then sum over all time steps (from t=1 to t=24). The summation term is the squared L2 norm of the product of the spatial correlation matrix M and the generated data matrix G(Z,C) at time t (i.e., the t-th column of G(Z,C), denoted as G(Z,C):,t); finally, multiply the summation result by the averaging coefficient to obtain the spatial physical constraint loss L. S ;

[0145] For example, at a certain time t=8, the data matrix G(Z,C) is generated, containing all monitoring point values ​​G(Z,C) at that time. G(Z,C) is a 105-dimensional vector. Multiplying the spatial correlation matrix M with this vector yields a spatial correlation correction vector, whose L2 norm squared is 0.36. After traversing 24 time steps, the spatial physical constraint loss L for this sample is calculated. S =0.032;

[0146] 3. Set the weight hyperparameter λ for the time smoothing constraint. T =0.3, the weight hyperparameter λ of the space physics constraints S =0.5, according to the formula described in the patent, the time smoothing constraint loss L T With space physical constraint loss L S After multiplying each value by its corresponding weight and summing them, we obtain the total spatiotemporal constraint loss L. ST The calculation process is as follows: L ST =0.3×0.025+0.5×0.032=0.0075+0.016=0.0235;

[0147] 4. Binary cross-entropy adversarial loss L BCE Assuming the discriminator outputs 0.3 for the generated data G(Z,C), according to the binary cross-entropy loss function formula, L BCE (D(G(Z,C)),1)=-ln(0.3)≈1.204 (where D(·) is the discriminator output)

[0148] The weight hyperparameter λ for the spatiotemporal constraint loss is set to 0.4. Based on the formula described in the patent, the binary cross-entropy adversarial loss L_BCE is compared with the total spatiotemporal constraint loss L... ST The sum of the results after multiplying by the weight λ yields the total generator loss L. GThe calculation process is as follows: L G =1.204 + 0.4 × 0.0235 ≈ 1.204 + 0.0094 = 1.2134;

[0149] Step 5:

[0150] 1. Network parameters: The weights of the generator and discriminator are initialized using the He normal method, and the bias parameters are initialized to 0;

[0151] Training hyperparameters: Set the number of training epochs (EPOCH) to 100 and the batch size (BATCH) to... SIZE =32, the optimizer used is the Adam optimizer (with parameters β1=0.5, β2=0.999), and the learning rate lr=0.0002.

[0152] 2. Generating fake data: In each batch of training, 32 noise vectors Z (each vector has a dimension of 100) are sampled from the standard normal distribution and matched with 32 corresponding condition vectors C. The noise vectors and condition vectors are input into the generator to obtain 32 fake data samples G(Z,C) (each sample has a dimension of (105,24)).

[0153] Input 32 real normal data samples and 32 fake data samples into the discriminator, and calculate the loss value according to the discriminator loss formula;

[0154] With the generator parameters fixed, the discriminator weight parameters are updated using the backpropagation algorithm. After multiple iterations, the discriminator loss converges from the initial 1.8 to the range of 0.6 ± 0.05.

[0155] 3. With the discriminator parameters fixed, calculate the generator's total loss L according to the generator's total loss formula. G (Including adversarial loss and spatiotemporal constraint loss);

[0156] By updating the generator's weight parameters using the backpropagation algorithm, after multiple iterations, the generator loss converged from the initial 3.2 to the range of 0.2 ± 0.05.

[0157] 4. When the generator loss and discriminator loss are both stable within the above convergence range for 10 consecutive training rounds, stop training to obtain a generator model G that can generate normal monitoring data patterns. * ;

[0158] Step Six:

[0159] 1. Sample to be tested: Monitoring data from 0:00 to 23:00 on a certain day is selected as sample X to be tested. test (Dimensions are (105, 24)), the conditional vector C corresponding to this sample test= (2,2,3,2), where 2 represents Zone 2 and the excavation stage, 3 represents rainy weather, and 2 represents earthwork excavation in the surrounding area;

[0160] The weighting coefficient γ for the spatiotemporal constraint loss is set to 0.3. According to the formula described in the patent, the optimal noise vector Z is the one that makes "the sample to be detected X..." test With the generated data G(Z,C) test The mean squared error (MSE) of the generated data G(Z,C) is added to 0.3 times the generated data. test The spatiotemporal constraint loss L) ST The noise vector that achieves the minimum value is Z, i.e., Z = argmin[MSE(X). test ,G(Z,C test ))+0.3×L ST (G(Z,C test ))];

[0161] The optimal noise vector Z is obtained after 50 iterations using the gradient descent method. * At this point, the mean square error (MSE) is 0.085, and the spatiotemporal constraint loss L... ST =0.042, the objective function value is minimized.

[0162] 2. Data Reconstruction: Reconstruct the optimal noise vector Z and the conditional vector C of the sample to be detected. test Input generator model G, obtain reconstructed data X recon =G(Z) * C test (Dimensions are (105, 24))

[0163] The weighting coefficient for the mean squared error is set to α = 0.6, and the weighting coefficient for the spatiotemporal constraint loss of the reconstructed data is set to β = 0.4. According to the formula described in the patent, the comprehensive anomaly score AS is calculated as follows: AS = α × MSE(X test ,X recon )+β×L ST (X recon Substituting the values, we get AS = 0.6 × 0.085 + 0.4 × 0.042 = 0.051 + 0.0168 = 0.0678;

[0164] Dynamic threshold setting: Based on the comprehensive abnormal score AS value of 360 normal samples in the training set, the 95th percentile is taken as the abnormal judgment threshold TH = 0.05, that is, the comprehensive abnormal score of 95% of normal samples is ≤0.05.

[0165] 3. Anomaly Detection: The overall anomaly score AS of the sample to be tested is 0.678, which is greater than the set dynamic threshold TH = 0.05. Therefore, the monitoring data corresponding to this sample is determined to be abnormal.

[0166] Anomaly localization: By comparing the sample X to be detected test With reconstructing data X recon The differences revealed that the mean square error (MSE) of the axial force monitoring points at supports 12-15 reached 0.15, far exceeding the MSE range of 0.02-0.05 for other monitoring points. Furthermore, the spatiotemporal constraint loss L of the reconstructed data in this region was also significant. ST =0.06, which is higher than the average spatiotemporal constraint loss of 0.03 for all monitoring points. Therefore, the source of the anomaly is determined to be the sudden change in axial force of supports 12-15.

[0167] Alarm Trigger: The system automatically pushes alarm information to the engineering management platform. The alarm information includes the specific location of the abnormal monitoring point (support axial force monitoring points 12-15), the abnormal time period, the abnormal index (the maximum deviation of the support axial force reaches 12kN), and targeted recommended measures (suspend earthwork excavation operations and immediately organize personnel to check the integrity of the support structure).

[0168] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0169] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for deep foundation pit multi-source monitoring data GAN (Generative Adversarial Network) anomaly diagnosis, characterized in that: The method comprises the following steps: Step S1: multi-source monitoring data collection and preprocessing: collecting multi-source heterogeneous monitoring data from the deep foundation automatic monitoring system, performing time alignment, cleaning, filling missing data and normalization processing, and providing a data basis for subsequent effective sample construction; Step S2: construction of time-space sequence samples and condition vectors: based on the preprocessed normalized multivariate time series, the sliding window method is used to cut it into fixed-length sample segments, and a condition vector containing static and dynamic context information is constructed for each sample, forming an input form that can be directly processed by the network model; Step S3: generator and discriminator network design: for the constructed sample segments and condition vectors, design the generator and discriminator network-the generator takes the noise vector and the condition vector as input, and outputs the generated multi-source monitoring data time series segment; the discriminator takes the real / generated data sequence and the condition vector as input, and outputs the probability that it is real data, providing a basic framework for network adversarial training; Step S4: definition and integration of spatio-temporal constraint loss function: in order to make the designed generator generate samples that conform to the spatio-temporal rules of deep foundation monitoring data, define a spatio-temporal constraint loss function composed of time smoothing constraint and spatial physical constraint weighting, and integrate it into the total loss function of the generator, to reasonably constrain the physics of the generated data; Step S5: network training and optimization: initialize network parameters and hyperparameters according to network structure and defined loss function, alternately update discriminator and generator parameters, and make the network converge through iterative optimization to obtain a generator model that can generate normal monitoring data patterns; Step S6, anomaly diagnosis and alarm: using the trained generator model, find the optimal noise vector for the sample to be detected through the optimization process, calculate the comprehensive anomaly score, and determine the anomaly according to the threshold to trigger the alarm mechanism, and complete the final output from data to anomaly diagnosis result.

2. The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method according to claim 1, characterized in that, The step S1 comprises the following steps: Step S11, collect multi-source heterogeneous monitoring data of support structure top horizontal displacement, support axial force, ground settlement and underground water level from the deep foundation automatic monitoring system, and perform time alignment based on the unified timestamp to form a multivariate time series data set with fixed time interval; Step S12, clean the original data, process data missing caused by recording errors and communication interruption, fill short missing data with linear interpolation method, and exclude long missing data; Step S13, the dimensions and ranges of different monitoring data are different, so eliminate the dimensional influence and make the data in the same interval to facilitate subsequent network model processing, normalize each kind of monitoring data to map it to the [-1, 1] or [0, 1] interval, and the normalization formula is as follows: wherein X norm is the normalized monitoring data value, X is the original monitoring data value, X min is the minimum value in the monitoring data of this type, X max is the maximum value in the monitoring data of this type, and 2-1 is a linear transformation mapping data in the interval [0, 1] to the interval [-1, 1]. 3.The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method of claim 1, wherein, The step S2 comprises the following steps: Step S21, using a sliding window method to cut the multivariate time series into sample segments with a fixed length T and a sliding step S step Each sample is a matrix with dimensions (S, T), where S is the total number of monitoring points; Step S22, construct a condition vector C for each sample segment, which contains the pit partition number, excavation stage code, weather condition code and adjacent construction activity flag information of the sample.

4. The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method according to claim 1, characterized in that, The step S3 comprises the following steps: Step S31, the input of the generator G includes a noise vector Z sampled from the standard normal distribution and a condition vector C, and the concatenated Z and C are taken as the input to output a generated data matrix with dimensions (S, T) through the network Step S32, the input of the discriminator D is a data sequence with dimensions (s, t) and its condition vector C, and the features are extracted through a series of one-dimensional convolutional layers, the condition information is fused with the data features, and finally the scalar probability value is output through the fully connected layer and the Sigmoid activation function.

5. The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method according to claim 1, characterized in that, The step S4 includes the following steps: Step S41, the deep foundation pit monitoring data has continuity in time dimension, the data change of adjacent time is usually relatively smooth, in order to constrain the time continuity of generated data and avoid unreasonable jump, define time smoothing constraint L T , the formula is as follows: In the formula, L T is a time-smoothed constraint loss value, S is the total number of monitoring points, T is the time length of the sample segment, is an averaging coefficient, used for normalizing the loss value, is a summation operation on all monitoring points, is a summation operation on all adjacent time steps, G(Z,C) is a generated data matrix output by the generator, G(Z,C) s,t+1 is the value of the s-th monitoring point at the t+1 time in the generated data, G(Z,C) s,t is the value of the s-th monitoring point at the t time in the generated data, is the square of the L2 norm, used to measure the size of the difference; Step S42, there is a physical correlation between each monitoring point of the deep foundation pit, in order to ensure that the generated data conforms to the spatial physical law, define the spatial physical constraint L S , the formula is as follows: wherein L S is a spatial physical constraint loss value, is an averaging coefficient, normalizing the loss value in the time dimension, is a summation operation over all time steps, M is a predefined spatial correlation matrix, G(Z,C):,t is the value of all monitoring points at the t-th time in the generated data, and " " is taking all rows. Step S43, considering the constraint effect of time continuity and space physical correlation on the generated data, the time smoothing constraint and the space physical constraint are combined by weighting to obtain a space-time constraint loss L ST The formula is as follows: L ST = λ T L T + λ S L S ; In the formula, L ST is the total loss value of space-time constraints, λ T is the weight hyperparameter of time smoothing constraints, λ S is the weight hyperparameter of space physical constraints; Step S44, the generator needs to generate realistic samples and samples conforming to the space-time rule at the same time, in order to balance the two goals, the binary cross-entropy adversarial loss is combined with the space-time constraint loss to form the total loss function of the generator, and the formula is as follows: L G = L BCE (D(G(Z,C)),1) + λ x L ST ; where L G is the total loss value of the generator, L BCE is the binary cross-entropy loss function for measuring the difference between the classification probability and the true label, D(·) is the output of the discriminator, and λ is the weight hyperparameter of the spatiotemporal constraint loss.

6. The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method according to claim 1, characterized in that, The step S5 includes the following steps: Step S51, initialize the parameters of the generator and the discriminator, set the training round number, the batch size, the optimizer and its learning rate; Step S52, for each batch of training data, sample the noise vector Z and the condition vector C, and the generator generates the false data G(Z, C); Step S53, update the discriminator: calculate its loss about the real data and the generated data, fix the generator parameters, and update the discriminator parameters through back propagation; Step S54, update the generator: fix the discriminator parameters, calculate the adversarial loss and the space-time constraint loss of the generator, and update the generator parameters through back propagation; Step S55, repeat the iteration until the network converges or the maximum training round number is reached.

7. The deep foundation pit multi-source monitoring data GAN anomaly diagnosis method according to claim 1, characterized in that, The step S6 includes the following steps: Step S61, find the noise vector that can make the generator reconstruct the data corresponding to the noise vector closest to the to-be-detected sample and conforming to the space-time rule, and determine the optimal noise vector Z through an optimization process * The formula is as follows: In the formula, Z * is the optimal noise vector, is the variable Z that makes the expression in the parentheses minimum, MSE is the mean square error function, X test is the sample data to be detected, G(Z, C test ) is the output of the generator under the input noise Z and the condition C test of the sample to be detected, C test is the condition vector of the sample to be detected, and γ is the weight coefficient of the space-time constraint loss. Step S62, the difference between the to-be-detected sample and the reconstructed sample is comprehensively evaluated, and the spatiotemporal rationality of the reconstructed sample is evaluated, so as to more accurately judge whether the sample is abnormal, and calculate the reconstructed data X recon = G (Z * , C test ) and the comprehensive anomaly score AS, the formula is as follows: AS = a x MSE(X test ,X recon ) + β x L ST (X recon ); In the formula, AS is a comprehensive abnormal score, a is a weight coefficient of mean square error, X recon is the optimal noise Z for the generator * The generated data corresponding to the sample to be detected, β is a weight coefficient of the spatiotemporal constraint loss, L ST (X recon ) is a spatiotemporal constraint loss value of the reconstructed data; Step S63: set a dynamic threshold, when the AS exceeds the threshold, it is determined to be abnormal and the alarm mechanism is triggered.

Images