Method and system for processing ground surface settlement monitoring data

Abstract

The invention discloses a method and system for processing ground surface settlement monitoring data, and relates to the technical field of tunnel monitoring data processing, and the method comprises the steps: collecting the ground surface settlement monitoring data of each measuring point, and carrying out the outlier recognition and processing; extracting a feature image of the monitoring data of each measuring point by adopting a Mel-frequency cepstrum coefficient feature extraction method, and converting the ground surface settlement monitoring data into an MFCC feature map; the MFCC feature map is input into a pre-trained MFCC-Transform model for data classification, and measuring point data affected by tunnel excavation construction are reserved; an improved ELM time sequence prediction algorithm is adopted to fill data missing values; and decomposing the monitoring data into a plurality of intrinsic mode functions (IMF) by adopting a VMD time sequence decomposition algorithm, and selecting the IMF with the highest similarity as the monitoring data after noise reduction in combination with a similarity criterion. According to the invention, the full-automatic processing of the ground surface settlement monitoring data is realized, and the reliability of data processing is ensured while the data processing speed is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tunnel monitoring data processing, and particularly relates to a ground surface settlement monitoring data processing method and system. BACKGROUND

[0002] It is more and more common to build a subway or high-speed railway tunnel in an environmentally sensitive area such as an airport, which puts forward very high requirements for settlement control. Therefore, in order to prevent the settlement from exceeding the limit value, it is necessary to monitor and warn the ground surface settlement. Obtaining high-quality real-time monitoring data has become a prerequisite for settlement prediction and risk analysis. How to improve the accuracy and real-time performance of monitoring within the budget limit has become a key problem, which has also led to the emergence of automatic monitoring technology. In order to further improve the accuracy of automatic monitoring of ground surface settlement, some researchers have introduced InSAR (Interferometric Synthetic Aperture Radar) and laser technology (such as TLS near-ground laser scanner and total station), and developed a series of automatic monitoring systems. Among them, the InSAR technology has a large coverage capability and can realize continuous monitoring all day long, and is not affected by light and clouds. However, its data needs to be combined with multi-source data inversion and is easily affected by atmospheric interference. This technology is suitable for large-scale survey and dynamic tracking, but it is difficult to meet the strict requirements of high-precision continuous monitoring. The TLS technology obtains high-density point cloud data through laser pulses to realize high-precision three-dimensional modeling. This technology has strong anti-interference ability and can be combined with unmanned aerial vehicles or mobile platforms to realize real-time dynamic monitoring. However, the TLS device has a high cost, is easily affected by environmental factors, has a limited scanning range, and the data processing is complex.

[0003] The efficiency and accuracy of data anomaly processing have become a difficult problem that must be solved for automatic monitoring systems to improve real-time performance. In particular, in the case of long-distance multi-target, the existing automatic monitoring and data processing system performance is difficult to balance the monitoring accuracy and real-time performance requirements. Unlike indoor monitoring, settlement monitoring in natural conditions is easily affected by weather, light, visibility and other measurement conditions. The data noise, measurement uncertainty and diversity of abnormal types generated in the monitoring process make it more difficult to process abnormal data. In order to process monitoring data anomalies, a series of algorithms such as time series anomaly detection, outlier processing, noise reduction, missing value filling and the like are required for abnormal data processing.

[0004] However, data anomaly processing is highly dependent on empirical design, such as Kalman filtering commonly used in noise reduction, and each algorithm needs to be adjusted in parameters when processing different samples. In addition, due to the lack of data validity discrimination criteria and data quality evaluation indicators, the quality of the data set is often related to the data processing method. At the same time, with the continuous growth of data volume and data acquisition speed, in order to improve the efficiency of data processing and maintain the real-time of monitoring data, minimize manual intervention in the data processing process, it is necessary to use fully automatic, standardized and efficient data processing methods, and optimize the design according to the inherent characteristics of civil engineering data. SUMMARY

[0005] To solve the above technical problems, the present application provides a kind of ground surface settlement monitoring data processing method and system, to realize the full-automatic intelligent processing of ground surface settlement monitoring data.

[0006] According to one aspect of the present application, a ground surface settlement monitoring data processing method is provided, characterized by comprising: collecting ground surface settlement monitoring data of each measuring point and performing outlier identification and processing; using a Mel frequency cepstral coefficient feature extraction method to extract feature images of monitoring data of each measuring point, and converting ground surface settlement monitoring data into MFCC feature images; inputting the MFCC feature images into a pre-trained MFCC-Transformer model for classification, the MFCC-Transformer model uses a Transformer architecture to receive MFCC feature image input, and divides the measuring points into three categories: affected by tunnel excavation construction, not affected by tunnel excavation construction and abnormal measuring points; removing the monitoring data of measuring points not affected by tunnel excavation construction and abnormal measuring points, and retaining only the monitoring data of measuring points affected by tunnel excavation construction; using an improved ELM time series prediction algorithm to fill in the missing values of the retained monitoring data, wherein the improved ELM time series prediction algorithm includes using a dynamic monitoring time interval and a recursive prediction strategy to automatically identify missing segments and fill them in real time; using a VMD time series decomposition algorithm to decompose the monitoring data into multiple intrinsic mode functions (IMF), and selecting the IMF with the highest similarity as the denoised monitoring data according to the similarity criterion.

[0007] Optionally, collecting ground surface settlement monitoring data of each measuring point and performing outlier identification and processing comprises: arranging multiple monitoring devices at different monitoring points to monitor ground surface settlement data in real time, and constructing an original monitoring data set; using the resample function of MATLAB to obtain effective resampling points from the original monitoring data set, and ensuring the consistency of the time interval to obtain a time-aligned resampled data set; identifying outliers in the resampled data set and marking the outliers by using a local outlier factor and an isolation forest algorithm; and replacing the marked outliers using linear interpolation combined with a shift median moving window method.

[0008] Optionally, the feature image of the monitoring data of each measuring point is extracted by using a mel-frequency cepstral coefficient feature extraction method, and the ground surface settlement monitoring data is converted into an MFCC feature image, including: representing the ground surface settlement signal through convolution operation of an acoustic pulse and a filter; sending the ground surface settlement signal to a high-pass filter for pre-emphasis to improve the clarity of the signal; after high-pass filtering, the signal is divided into N equal-length subsegments with a length of 512 through frame processing, the overlapping length of each frame is 20, and Hamming window is used as a window function for windowing processing; the fast Fourier transform is applied to each windowed frame signal to obtain the frequency spectrum of each frame of the ground surface settlement time sequence; the frequency spectrum is sent to a filter bank composed of M overlapping triangular band-pass filters based on the Mel frequency scale to obtain the log energy spectrum output by each filter; the obtained log energy spectrum is subjected to discrete cosine transform to obtain the MFCC coefficient; the MFCC coefficients of each measuring point are arranged in time sequence to generate a two-dimensional feature matrix, wherein the rows of the matrix correspond to the MFCC coefficients, and the columns correspond to the time; the two-dimensional feature matrix is visualized as an MFCC feature heat map, and the space-time distribution of the settlement features is represented by color mapping.

[0009] Optionally, the MFCC-Transformer model is composed of a feature embedding layer, an encoder layer and a multilayer perceptron, wherein the feature embedding layer divides the input MFCC feature image into fixed-size non-overlapping blocks by using an image blocking strategy, then performs local feature coding on each block through linear projection to generate a high-dimensional feature vector sequence with spatial position information; the encoder layer adopts a standard Transformer architecture, performs layer normalization and self-attention mechanism calculation in sequence in the feature processing process, establishes global feature dependency relationship through a multi-head attention module, and respectively introduces a random inactivation mechanism in the attention weight calculation and feedforward network output stages, thereby preventing overfitting by suppressing neuron co-adaptation and improving model training efficiency; the multilayer perceptron adopts a double-layer feedforward network containing a bottleneck structure, the first layer full connection layer expands the input dimension to twice the original dimension, and the GELU activation function is used to enhance the non-linear expression ability; the second layer full connection performs dimension reduction operation, and cooperates with the phased Dropout regularization strategy to compress the parameter space while retaining effective features.

[0010] Optionally, the improved ELM time series prediction algorithm is used to fill in the missing values of the monitoring data, including: using a random distillation method to randomly and automatically insert the missing values into the equidistant points of the monitoring time, and at the same time, the coordinates, the elevation and the mileage number remain consistent with the data before and after; using the position information of the measuring point as an influence parameter, the position information of the measuring point including longitude, latitude, elevation and mileage number; using the influence parameter, the time stamp and the corresponding settlement value as input parameters, and the settlement value corresponding to the next time stamp as output, training an ELM time series prediction model; and using the trained ELM time series prediction model to fill in the missing values of the monitoring data.

[0011] Optionally, the IMF with the highest similarity is selected as the denoised monitoring data in combination with the similarity criterion, including: calculating the Pearson correlation coefficient of each IMF component and the original settlement monitoring signal : for a given original settlement monitoring signal and the kth IMF component , the similarity Pearson correlation coefficient is defined as: ; wherein, is the similarity of the kth IMF component and the original signal; is the value of the original signal at the ith sampling point, is the value of the kth IMF component at the ith sampling point, is the mean value of the original signal, is the mean value of the kth IMF component, is the total number of sampling points of the signal; the IMF component with the maximum is selected as the denoised signal.

[0012] Optionally, after selecting the IMF with the highest similarity as the denoised monitoring data, the method includes: integrating the processing method of the ground surface settlement monitoring data into an automatic processing system of the ground surface settlement monitoring data, encapsulating the algorithms and models in the function functions, and embedding the data input, output and visualization functions, so as to realize the function of automatically outputting the original data and the settlement time curve of the processed data after inputting the monitoring data.

[0013] According to another aspect of the present application, there is also provided a ground surface settlement monitoring data processing system based on the ground surface settlement monitoring data processing method as described above, comprising a data collection module, an outlier processing module, a feature map conversion module, a data classification module, a data elimination module, a missing value filling module and a data noise reduction module, wherein the data collection module is configured to collect ground surface settlement monitoring data of each measuring point; the outlier processing module is configured to perform outlier identification and processing; the feature map conversion module is configured to extract a feature image of the monitoring data of each measuring point by using a mel-frequency cepstral coefficient feature extraction method, and convert the ground surface settlement monitoring data into an MFCC feature map; the data classification module is configured to input the MFCC feature map into a pre-trained MFCC-Transformer model for classification, wherein the MFCC-Transformer model adopts a Transformer architecture and is configured to receive the MFCC feature map input and divide the measuring points into three categories, i.e., measuring points affected by tunnel excavation construction, measuring points not affected by tunnel excavation construction and abnormal measuring points; the data elimination module is configured to eliminate the monitoring data of the measuring points not affected by tunnel excavation construction and the abnormal measuring points, and only keep the monitoring data of the measuring points affected by tunnel excavation construction; the missing value filling module is configured to fill the missing values of the kept monitoring data by using an improved ELM time series prediction algorithm, wherein the improved ELM time series prediction algorithm improves the dynamic monitoring time interval and the recursive prediction strategy to automatically identify missing sections and fill them online; and the data noise reduction module is configured to decompose the monitoring data into a plurality of intrinsic mode functions (IMFs) by using a VMD time series decomposition algorithm, and select the IMF with the highest similarity as the noise-reduced monitoring data in combination with a similarity criterion.

[0014] According to another aspect of the present application, there is also provided a computer device comprising a memory and a processor, wherein the memory stores a computer program, and wherein the processor implements the steps of the ground surface settlement monitoring data processing method as described above when executing the computer program.

[0015] According to another aspect of the present application, there is also provided a computer readable storage medium having a computer program stored thereon, wherein the computer program implements the steps of the ground surface settlement monitoring data processing method as described above when executed by a processor.

[0016] The present application has the following advantages: The application realizes the full-process automation from data acquisition to analysis output by data monitoring, outlier cleaning, data classification, missing filling, noise reduction processing and other processing procedures, and analyzes the data effectiveness classification and abnormal data characteristics according to the ground subsidence characteristics; and on this basis, a plurality of algorithms without parameter adjustment are adopted to realize the full-automatic intelligent processing of the ground subsidence monitoring data; the manual intervention is reduced, the automation level of data processing is improved, the real-time processing demand of large-scale monitoring data can be met, and the effectiveness of improving the tunnel monitoring accuracy and efficiency is verified through the Kunming data set.

[0017] The application introduces a classification method based on MFCC-Transformer on the basis of the previous algorithm, solves the problem of manual screening of effective data, improves the ELM time series prediction model, realizes the filling of missing values, and the VMD algorithm can effectively reduce data noise, reduce sample volatility, and eliminate the dependence on professional knowledge for noise suppression parameter design or setting; the system framework provided by the application improves the data processing speed while ensuring the reliability of data processing, and proves to have good application value. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings described herein are used to provide further understanding of the application, and form a part of the application, the illustrative embodiments of the application and the description thereof are used to explain the application, and do not constitute improper limitation on the application. In the drawings: Figure 1 It is a flow chart of a ground subsidence monitoring data processing method in embodiment 1 of the application; Figure 2 It is a flow chart of a ground subsidence monitoring data processing method in embodiment 2 of the application; Figure 3 It is a subsidence sample classification diagram; Figure 4 It is a typical cross-sectional view of a tunnel; Figure 5 It is a ground subsidence data monitoring area division schematic diagram; Figure 6 It is a subsidence-affected sample feature diagram; Figure 7 It is a data cleaning result schematic diagram; Figure 8 It is an MFCC-Transformer flow chart; Figure 9 It is an MFCC-Transformer evaluation index diagram; Figure 10 It is a fitting condition schematic diagram of ELM prediction value and true value; Figure 11A matching degree diagram of different missing time lengths and the true value; Figure 12 A typical IMF feature and its corresponding frequency feature diagram; Figure 13 A change diagram of entropy before and after noise reduction; Figure 14 An effect diagram of the full-automatic data processing system. DETAILED DESCRIPTION

[0019] In order to enable personnel in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work should fall within the scope of protection of the present application. It should be noted that, in the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0020] Secondly, the "one embodiment" or "embodiment" referred to herein can include specific features, structures or characteristics contained in at least one implementation of the present application. The "in one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or selective embodiment that excludes other embodiments.

[0021] In the specification and claims of the present application, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, product or device including a series of steps or units need not be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0022] Embodiment 1: Reference Figure 1 , Figure 1 is a flowchart of a ground subsidence monitoring data processing method in an embodiment of the present application, as shown in the figure, the method comprises: Figure 1 S1, collecting ground subsidence monitoring data of each measuring point, and performing outlier identification and processing; ​By arranging multiple monitoring points in the monitoring area, monitoring is carried out at each monitoring point, and a group of settlement data over time is obtained for each monitoring point. In an example, a plurality of Leica Nova MS60 3D and TM60 intelligent monitoring robots are arranged to monitor ground settlement data in real time at different monitoring points, and an original data set is constructed therefrom. Then, the MATLAB resample function is used to obtain effective resampling points from the original data set, and the consistency of the time interval is ensured, the problem of inconsistent collection frequencies is eliminated, and a time-aligned resampled data set S0_resampled is obtained; if the time interval of the original data is not uniform, the resample function of MATLAB can be converted to uniform sampling through interpolation, and the time interval consistency of the output signal is automatically maintained.

[0023] Local outlier factor, isolation forest and moving window algorithms are used to identify and process abnormal data in the monitoring data. Specifically, by using local outlier factor (LOF) and isolation forest (Isolation Forest, IF), the outliers in the resampled data (S0_resampled) are identified, and the two algorithms are run respectively, and the union (as long as one algorithm considers that the point is an outlier) or intersection (both algorithms consider that the point is an outlier) is marked to mark the outliers; after the outlier identification is completed, the linear interpolation is combined with the filloutliers function to replace the outliers. Among them, the linear interpolation is a linear interpolation using the two normal values before and after. Considering that outliers may appear continuously, linear interpolation may not be applicable (because the points before and after may also be outliers), so a moving window can be used to calculate the median, and then the median is used to replace the outliers. If the points before and after the outliers are normal values, linear interpolation (linear interpolation of the normal values before and after) is used, and if the points before and after the outliers are missing (such as at the boundary) or the points before and after the outliers are also outliers, the median of the moving window (excluding outliers in the window) is used to replace. By identifying and processing outliers, the accuracy of subsequent analysis and processing can be improved.

[0024] S2, a mel frequency cepstral coefficient (MFCC) feature extraction method is used to extract feature images of monitoring data of each measuring point, and ground settlement monitoring data is converted into MFCC feature images; In feature engineering-based dataset optimization, data quality assessment can select more relevant datasets. Similarly, feature construction and data augmentation can provide additional data to help algorithms train and understand datasets more effectively. Since there is currently no uniform standard or rule to clearly distinguish between valid and invalid samples, the embodiment of the present application introduces the MFCC-Transformer method to solve this problem. By converting the monitoring data into an MFCC feature image and then using the Transformer algorithm for image classification, this process provides a new solution for data classification, enabling it to be analyzed through a unified model.

[0025] In the embodiment of the present application, S2 comprises: S21, the surface subsidence signal is represented by the convolution operation of the acoustic pulse and the filter, and the formula is: ; wherein, represents the output continuous signal; represents the acoustic pulse; represents the filter; represents the convolution operation; The formula of the above surface subsidence signal can be explained as the subsidence signal in the process of stress re-equilibrium through a filter , which is equivalent to the physical response of the data acquisition system, and is equivalent to the influence of rock and soil disturbance, formation loss and monitoring error. The output continuous signal is the subsidence signal filtered by the data acquisition system.

[0026] S22, the surface subsidence signal is sent to a high-pass filter for pre-emphasis to improve the clarity of the signal, and the transfer function of the high-pass filter is defined as: ; wherein, is the high-pass filter function, is the weighting coefficient, represents a unit delay operator, that is, the operation of delaying the signal by one sampling period; this process is essentially a first-order difference operation, which improves the distinguishability of the transient characteristics of the signal: ; represents the output result of the filter; represents the input signal at the current time; represents the input signal at the previous time; S23, after high-pass filtering, the complete monitoring signal is divided into N equal-length sub-segments of 512 in length through frame segmentation, with an overlap length of 20 for each frame, and a Hamming window is used as the window function for windowing. Framing is used to divide a continuous signal into short, stationary segments. Windowing is to multiply each frame of data by a window function (a Hamming window is used in this invention) to reduce spectral leakage. Directly truncating the signal introduces discontinuities at the frame boundaries. Windowing allows the two ends of the frame to transition smoothly to zero, thereby reducing spectral leakage during Fourier transform.

[0027] The signal after windowing is The formula is expressed as follows: ; in, The signal before windowing processing; It is called hamming window.

[0028] S24, Apply Fast Fourier Transform to each frame of the windowed signal to obtain the spectrum of each frame of the land subsidence time series; The Fast Fourier Transform (FFT) is a mathematical operation used to obtain the spectrum of land subsidence time series. The formula is as follows: , ; in, The spectrum for each frame; For the number of frames; This represents the number of sampling points for the Fourier transform. For frequency domain variables; S25, the spectrum is fed into a filter bank consisting of M overlapping triangular bandpass filters constructed based on the Mel frequency scale, to obtain the logarithmic energy spectrum of each filter output: ; The frequency of the triangular filter is: ; In the Mel frequency cepstral coefficient (MFCC) extraction process of speech recognition and audio processing, M triangular filters are a set of bandpass filters constructed on the Mel frequency scale to simulate the human ear's perception of different frequencies. The filter bank constructed based on the Mel frequency scale better matches the human ear's frequency perception characteristics, enabling better extraction of features related to land subsidence and improving the effectiveness of feature extraction. In the above formula, Indicates the first The frame signal is in the first Logarithmic energy output on a Mel filter; For the first Logarithmic energy spectrum over each filter; For the first A Mel filter at frequency point The magnitude response weight at the location; For filter index; For the first The center frequency of each filter; For the first The left boundary frequency of the filter; For the first The right boundary frequency of each filter; S26. Perform Discrete Cosine Transform (DCT) on the obtained logarithmic energy spectrum (Mel spectrum) to obtain the MFCC coefficients: ; in, Indicates the first Mel-frequency cepstral coefficients of the frame signal; Index for Mel filter; For signal frame index; S27. Arrange the MFCC coefficients of each measurement point in chronological order to generate a two-dimensional feature matrix, where the rows of the matrix correspond to the MFCC coefficients and the columns correspond to the time. S28, further visualize the two-dimensional feature matrix as an MFCC feature heatmap, and characterize the spatiotemporal distribution pattern of settlement features through color mapping.

[0029] Through the above steps, the settlement time series of each measuring point is converted into an MFCC feature map.

[0030] S3, input the MFCC feature map into the pre-trained MFCC-Transformer model for classification. The MFCC-Transformer model adopts the Transformer architecture to receive the MFCC feature map input and classifies the measurement points into three categories: those affected by tunnel excavation construction, those not affected by tunnel excavation construction, and abnormal measurement points. To fully utilize MFCC feature images, this invention constructs an image classification neural network based on the Transformer architecture. Its core structure consists of three modules: a feature embedding layer, an encoder layer, and a multilayer perceptron (MLP). In the feature embedding layer, the image patching strategy is adopted to reconstruct the spatial dimension of the input MFCC image: first, the obtained MFCC image is divided into fixed-size non-overlapping patches, and then each patch is locally encoded by linear projection to generate a high-dimensional feature vector sequence with spatial position information, effectively capturing the fine-grained visual features of the MFCC image.

[0031] The encoder layer adopts a standard Transformer architecture, which performs layer normalization and self-attention mechanism calculation in sequence during feature processing, and establishes global feature dependency through the multi-head attention module. To enhance the generalization ability of the model, the layer introduces the Dropout mechanism in the attention weight calculation and the output stage of the feedforward network, which prevents overfitting by suppressing neuron co-adaptation and improves the training efficiency of the model.

[0032] The multi-layer perceptron module adopts a double-layer feedforward network containing a bottleneck structure: the first fully connected layer expands the input dimension to twice the original dimension, and the GELU (Gaussian Error Linear Unit) activation function is used to enhance the non-linear representation ability; the second fully connected layer performs dimension reduction operation, and cooperates with the Dropout regularization strategy to compress the parameter space while preserving effective features. This "expansion-contraction" dimension transformation design significantly improves the model's representation and classification ability for MFCC feature images.

[0033] By training the constructed Transformer model, the MFCC-Transformer algorithm model for surface subsidence data classification is obtained. During the training process, the accuracy, precision, recall, and F1-score are used to verify the performance of the model.

[0034] The MFCC feature image is input into the pre-trained MFCC-Transformer model, and the model outputs the classification of the MFCC feature image. The monitoring points of surface subsidence are divided into three categories: points affected by tunnel excavation construction, points not affected by tunnel excavation construction, and abnormal points. Among them, when the cumulative settlement is not zero and shows a large settlement trend (exceeding the preset threshold), the sample is classified as a construction-affected sample. Conversely, if the settlement value is stable at 0 mm and there is no obvious trend, it is determined as a non-construction-affected sample. In addition, data deviating from the identified settlement trend (which means that a data sample is inconsistent with the overall or local settlement trend, showing abnormal or unreasonable fluctuations) or whose settlement trend is masked by noise due to noise interference are classified as abnormal samples.

[0035] S4, the monitoring data of the measurement points not affected by the tunnel excavation construction and the abnormal measurement points are removed, and only the monitoring data of the measurement points affected by the tunnel excavation construction are reserved; If the measurement point is attributed to the measurement point not affected by the tunnel excavation or the abnormal measurement point, the measurement point does not enter the subsequent data processing flow. By classifying the samples, the data can be processed in a targeted manner, the amount of data to be processed and the amount of calculation are reduced, and the data set quality and the subsequent prediction accuracy are improved.

[0036] S5, an improved ELM time series prediction algorithm is used to fill in the missing values of the reserved monitoring data; ELM (Extreme Learning Machine) is widely used in time series prediction problems. Due to its high computational efficiency, it is particularly suitable for processing large-scale data sets. In addition, the time series prediction algorithm can use historical data to predict and infer missing values. The present application uses an improved ELM time series prediction model to fill in the missing values of the monitoring data, and the pseudo code is as follows: Pseudocode 1

[0037] The original ELM lacks the ability to directly process the structure of time series, cannot automatically detect and locate missing values, is difficult to dynamically adapt to feature changes in different time periods, and usually needs batch training, making it difficult to directly support real-time streaming prediction. The missing value filling model based on the improved ELM proposed in the present application automatically identifies the missing segment (the time interval exceeds 5 hours) in the time series through dynamic monitoring of the time interval (pseudo code lines 2-3), and generates a time point sequence with filling automatically; adopts a recursive prediction strategy (pseudo code lines 7-9), uses the prediction result of the previous step as the input of the next step, forms a time-dependent chain, so that the model can capture the continuity of time series data, and realizes an online filling real-time prediction model without retraining.

[0038] In order to realize the filling of missing values, a suitable data set needs to be constructed. In the data set, the position information of the measurement point is regarded as an influence parameter, including longitude (X), latitude (Y), elevation (H), and construction mileage (Nm). In addition to the above four position information parameters, the input parameters also include the time stamp (Ts) and the corresponding settlement value (S0). At the same time, the output of the data set is the settlement value (S0') corresponding to the next time stamp.

[0039] In addition, a pretreatment step is required before filling the missing values. In the ELM model of the application, the missing timestamps in the data set are automatically assigned to the equally spaced monitoring time points, and the missing values are randomly and automatically inserted into the equally spaced points of the monitoring time points by using the random distillation method. At the same time, the coordinates, elevations and mileages are kept consistent with the data before and after. Thus, it can be approximately considered that the output value of the prediction model is the missing value of the settlement change. Finally, the generated prediction value is regarded as the monitoring value. Random distillation refers to the concept of generating random missing data, inputting a complete historical monitoring data set (containing timestamps, positions and settlement values), outputting a data set with random distillation missing marks (coordinates / elevations / mileages remain unchanged), obtaining a data set for training an ELM missing value filling model, and using the data set to complete model training, and using the trained ELM time series prediction model to predict and fill the real missing values.

[0040] Based on this, the missing values in the ground settlement monitoring data can be filled.

[0041] S6, using a VMD time series decomposition algorithm, decomposing the monitoring data into a plurality of intrinsic mode functions (IMF), and selecting the IMF with the highest similarity as the denoised monitoring data according to a similarity criterion; VMD (variational mode decomposition) has the ability to overcome mode mixing and end effect. Therefore, the application uses VMD to decompose the ground settlement monitoring data with extremely strong nonlinearity and instability into a plurality of intrinsic mode functions (IMF). In the application, the monitoring data can be regarded as being composed of an actual settlement curve and noise, wherein the intrinsic mode function (IMF) with the highest similarity is used as the actual settlement curve for approximate processing, and the remaining IMF functions jointly constitute the noise.

[0042] The basic premise of VMD is to construct and solve a finite variational problem. In the construction of the settlement signal, and denote the modal components and the center frequency: ; wherein, denotes the original input signal; denotes the constraint condition (subject to); denotes the modal component index; denotes the total number of preset modes; denotes the first modal component; denotes the second modal component; denotes the corresponding center frequency; denotes the corresponding center frequency; The constrained variational expression is as follows (the core of VMD is to solve the following optimization problem): ; wherein, denotes a time gradient operator; is a Dirac function; denotes an imaginary unit.

[0043] By introducing a quadratic penalty term and a Lagrange multiplier , the extreme value constraint problem is converted into an unconstrained problem, as follows: ; is an original input signal; is an introduced Lagrange multiplier; is the th modal component at the th moment; denotes a time variable; The algorithm iteration steps include: First, initialize , , and , , wherein is the number of iterations, and in each round of iteration, the modal component , the center frequency and the Lagrange multiplier are updated; Subsequently, update , according to the VMD algorithm formula; Finally, update the Lagrange multiplier until the iteration constraint condition is met.

[0044] Based on this, the ground subsidence monitoring data can be decomposed into a plurality of (K) intrinsic mode functions IMFs ( ).

[0045] In the VMD-based subsidence monitoring data denoising method, the similarity criterion refers to an evaluation system for selecting the optimal denoising component by quantifying the matching degree of each IMF component and the original signal. According to the similarity criterion, the low-frequency intrinsic mode function is approximated as an intrinsic curve, i.e., an ideal ground subsidence monitoring curve, and the remaining IMFs collectively constitute noise, thereby realizing the denoising processing of the ground subsidence monitoring data. Specifically, it includes: Calculate the Pearson correlation coefficient of each IMF component and the original subsidence monitoring signal: For a given original settlement monitoring signal and the kth IMF component , the similarity Pearson correlation coefficient is defined as: ; wherein, is the similarity of the kth IMF component to the original signal; is the value of the original signal at the ith sampling point, is the value of the kth IMF component at the ith sampling point, is the mean value of the original signal, is the mean value of the kth IMF component, is the total number of sampling points of the signal.

[0046] The Pearson correlation coefficient as the core index of the similarity criterion can effectively measure the degree of retention of the main features of each IMF component to the original signal. Generally, the IMF component with high similarity is retained to reconstruct the denoised signal, and the environmental noise component with low correlation to the original signal is discarded. In the embodiment of the present application, the largest IMF component is taken as the denoised signal, and the remaining K−1 IMF components are regarded as noise.

[0047] Further, after step S6, the algorithm and model in steps S1 to S6 are packaged as a function in the embodiment of the present application, and the processing flow of sequentially calling the algorithm and model in steps S1 to S6 is integrated into an automatic processing system of surface settlement monitoring data, and the data input, output and visualization functions are embedded, so as to realize the function of automatically outputting the original data and settlement time series data (the curve of the settlement amount changing with time) after inputting the monitoring data.

[0048] In summary, the present application arranges multiple monitoring points in the monitoring area, carries out monitoring for each monitoring point, obtains a group of settlement data (drawn into a graph called a settlement time history curve) for each monitoring point over time, collects the settlement data obtained by monitoring all monitoring points, performs outlier processing on these data, then performs measurement point screening to eliminate the measurement points irrelevant to the influence of tunnel construction and only retain the settlement data corresponding to the relevant measurement points, on this basis, fills in the missing values or interrupted values of the data of these relevant measurement points, thereby forming complete data, further performs data decomposition to identify the data having a guiding effect on the project, this process is called denoising, and finally visualizes and outputs. Through these processes, intelligent and automatic processing of surface settlement monitoring data is realized.

[0049] Embodiment 2: refer to Figure 2 , Figure 2A flow chart showing the processing method of ground settlement monitoring data in the embodiment 2 of the present application is shown. In the embodiment of the present application, after the ground settlement data is automatically collected, data resampling and cleaning are first performed to establish a corresponding data set. Then, the data set is classified by the MFCC-Transformer algorithm. Subsequently, the ELM time series prediction model is used to identify and fill in the missing values of the samples affected by settlement. Then, the VMD is used to denoise the data set after filling.

[0050] Automatic monitoring involves a large amount of computing work, and if there is no pre-screening process, a large amount of computing resources will be used to process invalid samples, thereby reducing the processing efficiency. It is worth noting that in ground settlement monitoring, the main concern is the area affected by excavation, so the monitoring of the area without impact is of limited significance. In order to improve the processing efficiency, the present application proposes three classifications of settlement monitoring data, as shown in Figure 3 , which are: affected by tunnel excavation (affected by excavation), not affected by tunnel excavation (not affected), and abnormal measuring points (invalid).

[0051] As shown in Figure 3 , when the settlement value of a sample fluctuates at 0 mm and no significant settlement trend is observed, it is considered as a non-impact sample. On the contrary, when the cumulative settlement of a sample is not zero and a significant settlement trend is observed, it is considered as a sample affected by excavation. Similarly, data inconsistent with the determined settlement trend, or data in which these trends are obscured by noise, are classified as “invalid samples”.

[0052] In analyzing the length of missing measurements, the present application assumes that the missing condition occurs in a time period other than the detection interval being located at 0 to 2 hours or equal to 2 hours and 6 hours, and the relevant results are shown in Table 1. As mentioned earlier, the average collection interval is 4.58 hours, which indicates that a day of monitoring missing will result in an average of about 4.8 missing values.

[0053] Table 1 Maximum missing time statistics for each sample Interval 0~1 1~2 2~3 3~4 4~5 5~6 6~7 7> Sample 1 2 6 504 58 274 103 442 Proportion 0.07% 0.14% 0.43% 36.26% 4.17% 19.71% 7.41% 31.80% In order to achieve the filling of missing values, it is necessary to create a suitable data set. In this data set, the measurement point position information is considered as an impact parameter, including longitude (X), latitude (Y), altitude (H) and construction mileage (Nm).

[0054] In addition to the above four position information parameters, the input parameters also include a timestamp (Ts) and the corresponding settlement value (S0). Meanwhile, the output of this data set is the settlement value (S0') corresponding to the next timestamp. The missing timestamp in the data set is automatically assigned to an equal monitoring time interval. At the same time, the coordinates, elevations, and distances are consistent before and after. Thus, the approximate prediction model output value is the missing value of the settlement accumulation. Finally, the generated prediction value is designated as the monitoring value.

[0055] VMD has the ability to overcome modal aliasing and end effects. Therefore, the present application uses VMD to decompose the ground settlement sequence with extremely strong nonlinearity and instability into multiple intrinsic mode functions, selects the intrinsic sequence according to the similarity criterion, and finally removes the noise sequence to obtain high-quality and high-precision data after full-automatic intelligent processing.

[0056] Embodiment 3: The present scheme will be described in detail below with specific cases: A certain railway tunnel is 2700 meters long, with a depth of between 39 and 60 meters. The tunnel passes through the flight management area (FMA), the southern work area, the internal road, and the external urban road of the Kunming airport from north to south. A 267-meter-long section of the tunnel passing through the airport is selected to predict ground settlement, which contains 159 monitoring points. As shown in FIG. 1, the tunnel passes through the FMA, the southern work area, the internal road, and the external urban road of the Kunming airport from north to south. Figure 4 , Figure 4 A typical cross section of the tunnel is depicted, containing three layers of rock and soil, with a depth of between 44 and 47 meters. As shown in FIG. 2, a typical cross section of the tunnel is depicted, containing three layers of rock and soil, with a depth of between 44 and 47 meters. Figure 5 To ensure measurement accuracy, seven Leica Nova MS60 3D scanning robots are arranged inside the FMA, and five TM60 precision monitoring automated measurement robots are deployed outside the FMA to monitor ground settlement in real time.

[0057] In this embodiment, 287,297 samples were collected from 159 monitoring points, with an average sampling interval of 3.85 hours, and a continuous monitoring of 290 days, generating an average of 1,806 data points for each monitoring point. To verify the reliability of the algorithm, the data set is divided into training set, validation set, and test set according to the region.

[0058] For samples affected by settlement, the average stable settlement is calculated to be 10.57 mm, and the average settlement rate is about 0.1 mm / day. During the monitoring period, the maximum settlement was 39 mm, and the settlement rate was 0.36 mm / day. Figure 6Features of samples affected by settlement obtained in the construction of a high-speed railway tunnel in Kunming (hereinafter referred to as the Kunming dataset) are shown. Numerical anomalies can be defined as data points where the value of a single or consecutive single sample point deviates significantly from the previous and subsequent measurements. Fluctuation anomalies can also be defined as abnormal noise, which has different frequency domain characteristics or amplitudes from the monitoring sequence. Fault anomalies can be defined as an increase in value after a specific monitoring time. Missing value anomalies are caused by missing values due to abnormal monitoring collection intervals.

[0059] Among the 703 samples included in the training and test sets, the classification process obtained a total of 50 normal samples, 51 unaffected samples, 158 invalid samples, and 186 abnormal samples (of which 29 samples required missing value filling). Then, the ELM missing value filling model was trained using 208 samples (without 29 samples) in the training set. Finally, after filling in the missing values, VMD denoising verification was performed on these 237 samples.

[0060] Due to the presence of missing values in the original data and the inclusion of a large number of abnormal values in some samples, the data was re-collected at a frequency of every 2 hours to reduce the impact of these problems on the MFCC features. Subsequently, the local outlier factor (LOF) was used to identify and remove abnormal values. The results of data preprocessing are shown in Figure 7 This process effectively removes numerical anomalies and reduces the impact of data quality on MFCC features.

[0061] The calculation results of the MFCC feature values show that their feature values are all located in the interval [-12.5, 12.5]. Therefore, the color mapping interval is specified as [-12.5, 12.5] to establish a color mapping relationship, and the MFCC feature heat map (i.e., a feature map obtained by mapping MFCC feature values to a color interval) is obtained. Based on the obtained MFCC feature heat map, image classification is realized based on the Transformer algorithm, and the flowchart is shown in Figure 8 .

[0062] To verify the performance of MFCC-Transformer classification, the accuracy (accuracy), precision (precision), recall (recall), and F1-score are used for evaluation. These four indicators are key indicators for evaluating the performance of classification models, especially for binary classification problems. In this invention, the confusion matrix of the two algorithms is reorganized into the form shown in Table 2.

[0063] Table 2 Confusion matrix MFCC-Transformer Ture False Positive TP FP Negative FN TN Wherein, the calculation formulas of accuracy, precision, recall and F1-score are respectively: ; Positive indicates that the model considers it to be the target class, and in the present application, Positive refers to being affected by tunnel excavation; Negative indicates that the model considers it to be a non-target class, and in the present application, it refers to not being affected by excavation and abnormal measuring points; TP indicates that the model predicts a positive class, and the actual one is also a positive class, that is, the model correctly identifies the samples affected by excavation; FP indicates that the model predicts a positive class, and the actual one is a negative class, that is, the model incorrectly judges the unaffected samples as affected; FN indicates that the model predicts a negative class, and the actual one is a positive class, that is, the model misses the actual affected samples; TN indicates that the model predicts a negative class, and the actual one is also a negative class, that is, the model correctly identifies the samples not affected by excavation; The MFCC maps of the monitoring samples in each region are merged into a classification data set, and the training set is used to complete the training of the MFCC-Transformer classification model. The original MFCC map is trained, and the overall accuracy on the validation set is 91.76%. The evaluation indicators of each category are as follows Figure 9 (a), the classification model shows satisfactory overall performance in each category, and there is no error in the classification of the no-impact category. However, the invalid category is slightly lower, indicating that the model shows false classification when identifying this category. The detailed classification results of the validation set are shown in Figure 9 (b). Only two "affected" samples are not identified, and the omission rate is only 4.5%. This indicates that although the MFCC-Transformer classification may have certain deficiencies in accurately identifying the "invalid" category, it can still ensure that more than 95% of the valid data is not missed, while filtering out a sufficient number of invalid data.

[0064] The missing value processing model can be obtained by ELM model training. Then, the coefficient of determination (R²) and the mean absolute error (MAE) are used to evaluate the accuracy of filling and processing. The R² value ranges from 0 to 1, and the closer to 1 indicates the better the model fitting. MAE calculates the average of the absolute values of the differences between the predicted values and the actual values, and the smaller the value is, the better.

[0065] As Figure 10As shown, the R² and MAE of the predicted values of ELM on ground surface settlement on the validation set are 0.99167 and 0.44489, respectively, and the R² and MAE on the training set are 0.99141 and 0.44667, respectively. This shows that the ELM time series prediction model is reliable and there is no overfitting problem. In addition, most of the data points in the figure are concentrated on the straight line of the function y=x, and only a small number of points deviate significantly from the main distribution. This phenomenon shows that ELM has sufficient prediction accuracy and can effectively complete the missing value filling process.

[0066] In order to evaluate the effectiveness of the ELM missing value filling algorithm, the present application constructs a validation set with different lengths of continuous missing values, as shown in the first row of Table 3. Each different length of continuous missing value corresponds to 24 samples for statistical analysis, and the corresponding accuracy evaluation indicators are listed in Table 3. Figure 11 The matching degree of representative samples with true values under different missing time lengths is shown.

[0067] Table 3 Filling accuracy under each missing time length Missing time length 1 10 20 30 40 50 60 70 80 90 [R 2 ]] 0.98 0.98 0.99 1.00 0.94 0.98 0.99 0.97 0.89 0.91 MAPE 32.41% 37.66% 19.82% 48.07% 14.04% 25.34% 26.64% 27.00% 119.97% 72.04% The algorithm shows strong tracking ability of the original data fluctuation trend when the missing length of the filled data ranges from 10 to 90. However, as the missing length increases to 70 (12 days), the filled data shows a certain lag in trend change and the detail restoration ability is weakened, but the overall trend is still accurate. It is worth noting that the continuous missing length exceeding 12 days accounts for about 25% of the samples, and the algorithm can accurately and reliably fill it. Therefore, the algorithm shows sufficient effectiveness in meeting the established requirements.

[0068] Subsequently, the filled data is imported into the VMD denoising process to obtain the IMF signals and their spectral characteristics of the samples, as shown in Figure 12 During the VMD decomposition process, a total of 10 IMF components are obtained. It can be clearly seen that the principal component frequency gradually increases, while the high intensity part synchronously decreases. Except for IMF1, the remaining signals all fluctuate around 0 with a maximum fluctuation amplitude of 1 mm, which is consistent with the established measurement accuracy. Therefore, it can be concluded that IMF2 to IMF10 are all noise signals.

[0069] As shown in Figure 13 The information entropy values of the 208 samples before and after VMD denoising are measured. The results show that the average information entropy of the noise is 2.89, while the average information entropy before and after denoising is about 2.00. In addition, the data curves before and after denoising are basically coincident, which shows that the entropy values of most original samples are higher than those of the samples after denoising. This result shows that most samples are not significantly affected by noise, and the VMD method can effectively reduce the noise in the samples.

[0070] All the above algorithms are integrated into the fully automatic data processing system of the present application. The system processes 1,390 samples using the Kunming dataset, collects all information and results of the fully automatic data processing system of the present application, and part of the results are shown in Table 4. Figure 14 The results show that the present application effectively processes the numerical anomalies, fluctuation anomalies, breakthrough anomalies, and mainly missing value anomalies in the data. As shown in Table 4, the data processing operation takes about 4.7 hours, and most of the time (more than 95%) is allocated to the VMD denoising step. The MFCC-Transformer classification and the improved ELM missing value filling step are the most efficient, accounting for only 0.25% of the total processing time. This result highlights the fast response capability of the present fully automatic data processing system.

[0071] Table 4 Data processing time consumption Processing step Resample & Outlier MFCC extraction Transformer classification Missing value filling VMD denoising Time-consuming 234.6s 489.7s 35.5s 7.4s 4.5h In the table, Resample & Outlier represents the data resampling and outlier removal process; MFCC extraction represents the MFCC feature extraction process; Transformer classification represents the Transformer classification process; Missing value filling represents the ELM missing value filling process; and VMD denoising represents the VMD denoising process.

[0072] In the present embodiment, a ground surface settlement monitoring data processing system is also provided for implementing the above embodiments and preferred embodiments, which have been described and will not be repeated. As used below, the term "module" "unit" can be a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware implementation is also possible and contemplated.

[0073] The surface subsidence monitoring data processing system in the embodiment comprises a data acquisition module, an outlier processing module, a feature map conversion module, a data classification module, a data elimination module, a missing value filling module and a data noise reduction module, wherein the data acquisition module is configured to collect surface subsidence monitoring data of each measuring point; the outlier processing module is configured to perform outlier identification and processing; the feature map conversion module is configured to extract feature images of the monitoring data of each measuring point by using a mel-frequency cepstral coefficient feature extraction method, and convert the surface subsidence monitoring data into MFCC feature maps; the data classification module is configured to input the MFCC feature maps into a pre-trained MFCC-Transformer model for classification, the MFCC-Transformer model adopts a Transformer architecture and is configured to receive MFCC feature map input, and divide the measuring points into three categories, i.e., measuring points affected by tunnel excavation construction, measuring points not affected by tunnel excavation construction and abnormal measuring points; the data elimination module is configured to eliminate the monitoring data of the measuring points not affected by tunnel excavation construction and the abnormal measuring points, and only keep the monitoring data of the measuring points affected by tunnel excavation construction; the missing value filling module is configured to fill the missing values of the kept monitoring data by using an improved ELM time series prediction algorithm, wherein the improved ELM time series prediction algorithm includes the following improved contents: a dynamic monitoring time interval and a recursive prediction strategy are used to automatically identify missing sections and fill them online; and the data noise reduction module is configured to decompose the monitoring data into a plurality of intrinsic mode functions (IMFs) by using a VMD time series decomposition algorithm, and select the IMF with the highest similarity as the noise-reduced monitoring data according to a similarity criterion.

[0074] The embodiment of the present application further provides a computer device, comprising a memory and a processor, the memory stores a computer program, and the processor is configured to execute the computer program to perform the steps in any of the above method embodiments.

[0075] The embodiment of the present application further provides a storage medium, which stores a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running. Optionally, in the embodiment, the storage medium can include but is not limited to a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk and various storage media that can store computer programs.

[0076] The above serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0077] In the above-mentioned embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments. The above-mentioned is only the preferred embodiment of the present application, and it should be pointed out that, for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method of processing ground settlement monitoring data, characterized by, The method comprises the following steps: Collecting ground surface settlement monitoring data of each measuring point and performing outlier identification and processing; Using a mel-frequency cepstral coefficient feature extraction method to extract feature images of the monitoring data of each measuring point, and converting the ground surface settlement monitoring data into MFCC feature maps; Inputting the MFCC feature maps into a pre-trained MFCC-Transformer model for classification, wherein the MFCC-Transformer model adopts a Transformer architecture and is used to receive MFCC feature map input, and the measuring points are divided into three categories: affected by tunnel excavation construction, not affected by tunnel excavation construction and abnormal measuring points; Removing the monitoring data of the measuring points not affected by tunnel excavation construction and the abnormal measuring points, and only retaining the monitoring data of the measuring points affected by tunnel excavation construction; Using an improved ELM time series prediction algorithm to fill in the missing values of the retained monitoring data, wherein the improved ELM time series prediction algorithm improves the dynamic monitoring time interval and the recursive prediction strategy to automatically identify missing segments and fill in the missing segments online; Using a VMD time series decomposition algorithm to decompose the monitoring data into multiple intrinsic mode functions (IMFs), and selecting the IMF with the highest similarity as the denoised monitoring data according to a similarity criterion.

2. The method of processing ground settlement monitoring data according to claim 1, wherein, Collecting ground surface settlement monitoring data of each measuring point and performing outlier identification and processing comprises: Arranging multiple monitoring devices at different monitoring points to monitor ground surface settlement data in real time, and constructing an original monitoring data set; Using the resample function of MATLAB to obtain effective resampling points from the original monitoring data set, ensuring the consistency of the time interval, and obtaining a time-aligned resampling data set; Identifying outliers in the resampling data set by using the local outlier factor and the isolation forest algorithm, and marking the outliers; Using linear interpolation combined with the offset median moving window method to replace the marked outliers.

3. The method of processing ground settlement monitoring data according to claim 1, wherein, Using a mel-frequency cepstral coefficient feature extraction method to extract feature images of the monitoring data of each measuring point, and converting the ground surface settlement monitoring data into MFCC feature maps comprises: Expressing the ground surface settlement signal through convolution operation of the sound pulse and the filter; Sending the ground surface settlement signal to a high-pass filter for pre-emphasis to improve the clarity of the signal; After high-pass filtering, dividing the signal into N equal-length subsegments with a length of 512 through frame processing, with a frame overlap length of 20, and using a Hamming window as a window function for windowing processing; Applying fast Fourier transform to each windowed frame signal to obtain the frequency spectrum of each frame of the ground surface settlement time series; Sending the frequency spectrum to a filter bank composed of M overlapping triangular bandpass filters based on the Mel frequency scale to obtain the log energy spectrum output by each filter; Performing discrete cosine transform on the obtained log energy spectrum to obtain MFCC coefficients; Arranging the MFCC coefficients of each measuring point in time sequence to generate a two-dimensional feature matrix, wherein the rows of the matrix correspond to the MFCC coefficients, and the columns correspond to the time; Visualizing the two-dimensional feature matrix as an MFCC feature heat map to represent the temporal and spatial distribution of the settlement features through a color mapping table.

4. The method of processing ground settlement monitoring data according to claim 1, wherein, The MFCC-Transformer model is composed of a feature embedding layer, an encoder layer and a multilayer perceptron, Among them, The feature embedding layer adopts an image block strategy to divide the input MFCC feature map into fixed-size non-overlapping blocks, and then performs local feature coding on each block through linear projection to generate a high-dimensional feature vector sequence with spatial position information; The encoder layer adopts a standard Transformer architecture, and sequentially performs layer normalization and self-attention mechanism calculation in the feature processing process, establishes global feature dependency relationship through the multi-head attention module, and respectively introduces the random inactivation mechanism in the attention weight calculation and feedforward network output stage, to prevent overfitting by suppressing neuron co-adaptation, while improving the model training efficiency; The multilayer perceptron adopts a double-layer feedforward network containing a bottleneck structure, the first full connection layer expands the input dimension to twice the original dimension, and the GELU activation function is used to enhance the non-linear expression ability; The second full connection performs dimension reduction operation, and cooperates with the stage Dropout regularization strategy to compress the parameter space while retaining effective features.

5. The method of processing ground settlement monitoring data according to claim 1, wherein, An improved ELM time series prediction algorithm is used to fill in missing values of monitoring data, including: Using a random distillation method, the missing values are randomly and automatically inserted into the equally spaced points at the monitoring time, and the coordinates, elevations and mileages are kept consistent with the data before and after; Using the location information of the measuring point as the influence parameter, the location information of the measuring point includes longitude, latitude, elevation and mileage; The influence parameter, timestamp and corresponding settlement value are used as input parameters, and the settlement value corresponding to the next timestamp is used as output to train the ELM time series prediction model; The trained ELM time series prediction model is used to fill in the missing values of the monitoring data.

6. The method of processing ground settlement monitoring data according to claim 1, wherein, Combined with the similarity criterion, the IMF with the highest similarity is selected as the denoised monitoring data, including: calculating a pearson correlation coefficient of each IMF component with the original settlement monitoring signal : For a given original settlement monitoring signal and the kth IMF component , the similarity Pearson correlation coefficient is defined as: ； wherein, is the similarity of the k-th IMF component to the original signal; is the value of the original signal at the i-th sampling point, is the value of the k-th IMF component at the i-th sampling point, is the mean value of the original signal, is the mean value of the k-th IMF component, is the total number of sampling points of the signal; Selecting The largest IMF component is taken as the denoised signal.

7. The method of processing ground settlement monitoring data according to claim 1, wherein, After selecting the IMF with the highest similarity as the denoised monitoring data, the method includes: The processing method of the ground subsidence monitoring data is integrated into an automatic processing system for ground subsidence monitoring data, the algorithms and models in the system are packaged as function functions, and the data input, output and visualization functions are embedded to realize the function of automatically outputting the original data and settlement time curve of the processed data after inputting the monitoring data.

8. A processing system of ground subsidence monitoring data based on the processing method of ground subsidence monitoring data according to any one of claims 1 to 7, characterized in that, Including: Data acquisition module, outlier processing module, feature map conversion module, data classification module, data rejection module, missing value filling module and data denoising module, wherein, The data acquisition module is used to collect the ground subsidence monitoring data of each measuring point; The outlier processing module is used for outlier identification and processing; The feature map conversion module is used to extract the feature image of each measuring point monitoring data by using the Mel frequency cepstral coefficient feature extraction method, and converts the ground subsidence monitoring data into an MFCC feature map; The data classification module is configured to input the MFCC feature map into a pre-trained MFCC-Transformer model for classification, the MFCC-Transformer model adopts a Transformer architecture and is configured to receive the MFCC feature map input, and divide the measuring points into three categories, i.e., measuring points affected by the tunnel excavation construction, measuring points not affected by the tunnel excavation construction, and abnormal measuring points; The data elimination module is configured to eliminate the monitoring data of the measuring points not affected by the tunnel excavation construction and the abnormal measuring points, and only keep the monitoring data of the measuring points affected by the tunnel excavation construction; The missing value filling module is configured to fill the missing values of the kept monitoring data by using an improved ELM time series prediction algorithm, wherein the improved ELM time series prediction algorithm includes the following improved contents: a dynamic monitoring time interval and a recursive prediction strategy are used to automatically identify missing segments and fill them online. The data denoising module is configured to decompose the monitoring data into a plurality of intrinsic mode functions (IMFs) by using a VMD time series decomposition algorithm, and select the IMF with the highest similarity as the denoised monitoring data according to a similarity criterion. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the ground settlement monitoring data processing method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the ground settlement monitoring data processing method of any one of claims 1 to 7.

Images