Tunnel abnormity early warning and trend prediction method and system

Abstract

The invention relates to the technical field of tunnel engineering monitoring, in particular to a tunnel abnormity early warning and trend prediction method and system. Comprising the following steps: S1, data collection and arrangement; s2, tunnel health monitoring data anomaly recognition: monitoring a long-term change trend in a sequence by taking an LSTM method as a main method, and giving an early warning signal before an anomaly occurs; an ARIMA method is used as a supplement, a relatively strict threshold value judgment method is set, a sequence short-term change trend is monitored, and an early warning signal is timely identified and sent out, so that application and implementation of an overall data anomaly identification algorithm are realized; s3, decomposing tunnel health monitoring data elements, and constructing a prediction model; and S4, performing spatial-temporal change feature analysis on the decomposed elements. Compared with the prior art, the method is more suitable for a real-time anomaly cleaning algorithm of tunnel health monitoring data, a system constructed based on the whole method achieves dynamic graph drawing of measured values and model predicted values through input tunnel monitoring point historical numerical values and a model of a built-in algorithm, dynamic early warning of the tunnel health monitoring data is more visually reflected, and the method is more suitable for real-time anomaly cleaning of tunnel health monitoring data. And necessary services are provided for a tunnel intelligent operation and maintenance management platform.

Description

Technical Field

[0001] The invention relates to the technical field of tunnel engineering monitoring, in particular to a method and system for tunnel abnormality early warning and trend prediction. Background Art

[0002] The operating and service environment conditions of cross-river tunnels and cross-sea tunnels are harsh, especially for tunnels in soft soil environments. Under the repeated effects of external loads such as foundation settlement, traffic, tides, and siltation depth, various deformations and local damages will occur, affecting the normal use of the tunnel. If not handled properly, it may even cause serious safety accidents.

[0003] Health monitoring systems are an effective means of preventing catastrophic accidents caused by structural degradation. They have gradually replaced traditional manual detection methods and become a trend in the development of operation and maintenance management. In recent years, thanks to the popularization of the concept of the entire life cycle of engineering and the commissioning of a large number of underwater tunnels, coupled with the rapid development of sensor technology and big data technology, data mining and analysis in health monitoring systems have gradually gained attention and become a research hotspot and urgent need for tunnel structure operation and maintenance.

[0004] However, there are a lot of abnormal data in the raw data of the tunnel health monitoring system. Data anomalies may be structural anomalies that actually occur in the tunnel, or anomalies caused by sensor system failures. Related research such as raw data anomaly detection and trend prediction is still in its infancy, lacking in-depth analysis of anomaly recognition and early warning algorithms, system comparison, and application module development. Summary of the invention

[0005] In order to solve the problems raised in the above background technology, the purpose of the present invention is to provide a tunnel abnormality early warning and trend prediction method and system, comprising the following steps:

[0006] S1, data collection and compilation;

[0007] S2. Tunnel health monitoring data anomaly identification: mainly using the LSTM method to monitor the medium- and long-term trend of the sequence and give early warning signals before the anomaly occurs; supplemented by the ARIMA method, a more stringent threshold judgment method is set to monitor the short-term trend of the sequence, identify and issue early warning signals in a timely manner, and implement the application of the overall data anomaly identification algorithm;

[0008] S3. Decomposition of tunnel health monitoring data elements and construction of prediction model: Decomposition of health monitoring data into trend items, periodicity and random items to build a prediction model;

[0009] S4. Analyze the spatiotemporal variation characteristics of the decomposed elements: Based on the ANOVA variance analysis data development trend, a comparative analysis is conducted to obtain the spatiotemporal variation law of the tunnel deformation data.

[0010] The step S2 includes the following steps of the ARIMA-based model data anomaly identification algorithm:

[0011] S201, stationarity test, if the time series is stationary, proceed to step S22, if not, perform differential operation and then proceed to step S22;

[0012] S202, white noise test, if the time series is white noise, proceed to step S26, if not, proceed to step S23;

[0013] S203, model identification, using Akaike information criterion and Bayesian information criterion to set ARIMA model parameters;

[0014] S204, model fitting;

[0015] S205, model testing, testing the fit between the model and the data;

[0016] S206, model prediction, performing multi-step prediction based on single-step prediction;

[0017] S207, establishing a dynamic prediction model; fitting the model value and the actual value error, and updating the model parameters if the error average is greater than the dynamic threshold, otherwise the model parameters are not updated;

[0018] S208. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

[0019] The step S2 includes the following steps of the LSTM-based model data anomaly recognition algorithm:

[0020] S209, data training, divide the training set and the test set in a ratio of 5:1, and use the network to search for the optimal hyperparameters;

[0021] S210, construct an LSTM model network, the model structure is a two-layer LSTM network and a fully connected (Dense) layer;

[0022] S211, model training, using the adaptive moment estimation algorithm (Adam algorithm) for optimization calculation, using the mean square error (MSE) loss calculation and accuracy evaluation criteria for model training. Set the model training standard mean square error to when the test set error is greater than this value, the model will be retrained until the error requirement is met;

[0023] S212, setting the interval for model update to one hour, and setting a dynamic threshold every hour based on the standard deviation of the data of the previous hour;

[0024] S214. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

[0025] The step S3 includes the following steps of decomposing the time series elements based on STL:

[0026] S301, perform STL decomposition on deformation monitoring data to obtain trend terms, seasonal terms and residual terms.

[0027] S302, performing function fitting on the trend term, the seasonal term, and the residual term respectively.

[0028] S303, summing up the three fitted parts to obtain a combined model that can predict the subsequent development of tunnel convergence deformation.

[0029] The step S4 specifically includes the following steps:

[0030] S401. Set the number of different groups to k and the number of different samples to n. The subscript i represents the group, and each group has n i The calculation formula is as follows: The systematic difference formed by different test groups is called the between-group deviation (MSA). The other type is the random difference between individuals in the group, which is called the within-group deviation (MSE).

[0031] Group mean:

[0032] Overall mean:

[0033] S402, introduce F statistics to perform statistical hypothesis testing,

[0034] The calculation formula is as follows:

[0035]

[0036] S403. Compare MSA, MSE and F value. If MSA is greater than MSE, the F statistic is greater than 1. It is judged that the difference caused by the test group contributes greatly to the total difference, and it is considered that there are significant differences between different test groups. Otherwise, it is considered that there are no significant differences.

[0037] A tunnel anomaly warning and trend prediction system comprises an ARIMA-based anomaly warning module, an LSTM-based anomaly warning module and an STL-based convergence deformation prediction module.

[0038] The ARIMA-based abnormal warning module includes a data acquisition unit: used to import a table and obtain data from the table; a parameter setting unit: setting how many times the standard deviation exceeds the predicted data range, setting the prediction time range, and setting the parameters of the ARIMA (p, d, q) model; and a result generation unit: drawing a dynamic prediction graph and displaying the time and amount of abnormal data.

[0039] The LSTM-based abnormal warning module includes a data acquisition unit: used to import a table and obtain data from the table; a parameter setting unit: used to set how many times the standard deviation exceeds the predicted data range and set the prediction time range; and a result generation unit: used to draw a dynamic prediction graph and display the time when abnormal data occurs and the number of abnormal data.

[0040] The STL-based convergence deformation prediction module includes a data acquisition unit: used to import a table and obtain data from the table; a result generation unit: used to output ① trend item: draw the trend item curve and its fitting straight line, and display the slope of the fitting straight line. ② Season item: draw the seasonal item curve and display the seasonal index of each quarter. ③ Irregular item: draw the irregular item curve and its fitting flat line. ④ STL combined model: draw the original input curve and the model prediction curve, and predict the convergence value of the next two years.

[0041] Compared with the prior art, the present invention includes data collection and organization in the entire method, and is more suitable for real-time abnormal cleaning algorithms, data feature analysis, model construction, and dynamic model update mechanism design of tunnel health monitoring data. The system has built a relatively complete system for dynamic early warning of the health monitoring system. The system constructed based on the entire method realizes dynamic graph drawing of measured values ​​and model predicted values ​​through the input of historical values ​​of tunnel monitoring points and the model of built-in algorithms, which more intuitively reflects the dynamic early warning of tunnel health monitoring data and provides necessary services for the tunnel intelligent operation and maintenance management platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0043] Figure 2 This is the flow chart of the ARIMA model data anomaly identification algorithm;

[0044] Figure 3a is the ARIMA model result of the points in this embodiment;

[0045] Figure 3b This is the LSTM model result of the point in this embodiment;

[0046] Figure 4 This is the STL model result of this embodiment;

[0047] Figure 5This is the ARIMA abnormal warning module and example results of the embodiment of the present invention; Figure 6 This is the LSTM abnormal warning module and example results of an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0049] Example

[0050] like Figure 1 As shown, a tunnel abnormality early warning and trend prediction method includes the following steps:

[0051] S1, data collection and compilation;

[0052] S2. Tunnel health monitoring data anomaly identification: mainly using the LSTM method to monitor the medium- and long-term trend of the sequence and give early warning signals before the anomaly occurs; supplemented by the ARIMA method, a more stringent threshold judgment method is set to monitor the short-term trend of the sequence, identify and issue early warning signals in a timely manner, and implement the application of the overall data anomaly identification algorithm;

[0053] S3. Decomposition of tunnel health monitoring data elements and construction of prediction model: Decomposition of health monitoring data into trend items, periodicity and random items to build a prediction model;

[0054] S4. Analyze the spatiotemporal variation characteristics of the decomposed elements: Based on the ANOVA variance analysis data development trend, a comparative analysis is conducted to obtain the spatiotemporal variation law of the tunnel deformation data.

[0055] Step S2 includes the following steps of the ARIMA-based model data anomaly identification algorithm, such as Figure 2 As shown:

[0056] S201, stationarity test, if the time series is stationary, proceed to step S22, if not, perform differential operation and then proceed to step S22;

[0057] S202, white noise test, if the time series is white noise, proceed to step S26, if not, proceed to step S23;

[0058] S203, model identification, using Akaike information criterion and Bayesian information criterion to set ARIMA model parameters;

[0059] S204, model fitting;

[0060] S205, model testing, testing the fit between the model and the data;

[0061] S206, model prediction, performing multi-step prediction based on single-step prediction;

[0062] S207, establishing a dynamic prediction model; fitting the model value and the actual value error, and updating the model parameters if the error average is greater than the dynamic threshold, otherwise the model parameters are not updated;

[0063] S208. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

[0064] The ARIMA model calculation formula of step S203 includes the following:

[0065] in, is the difference operator, p is the autoregressive parameter, q is the moving average parameter, d is the difference order, B is the delay operator, X t is the time series, ε t is a zero-mean white noise sequence, Φ p (*) is the autoregressive coefficient polynomial, Θ q (*) is the sliding mean coefficient polynomial.

[0066] Step S205 model verification includes parameter significance test and model significance test. Parameter significance test is to test whether each introduced independent variable has a real impact on the corresponding variable. The probability that the parameter is not significant is close to zero, indicating that the autoregressive term coefficient of the model is significant. Model significance test: a model with high fitting degree can fully extract the information contained in the time series. The closer its residual sequence is to the white noise sequence, the better the model fitting degree is.

[0067] Step S208 specifically includes the following process: taking the standard deviation of the data in the previous hour as the fluctuation range benchmark of the monitoring data, taking 5.5 times, 7 times and 8 times the standard deviation as the threshold to judge the abnormal point. Calculate the error between the measured value and the predicted value. If the error is greater than a certain graded dynamic threshold, output the corresponding graded warning result, otherwise store the monitoring data and error.

[0068] This embodiment uses settlement monitoring data to identify anomalies in tunnel health monitoring data. Using the ARIMA model, firstly, the model is preliminarily fitted and predicted within a larger time range by using the set parameters (p, d, q). The confidence interval of the predicted value is calculated using the error between the predicted data and the actual monitoring data to determine whether there are abnormal points. When the monitoring value exceeds the predicted confidence interval, the program marks the moment as an abnormal point. The settlement value deviates greatly from the predicted value and is identified as an abnormal point, such as Figure 3a .

[0069] The step S2 includes the following steps of the LSTM-based model data anomaly recognition algorithm:

[0070] S209, data training, divide the training set and the test set in a ratio of 5:1, and use the network to search for the optimal hyperparameters;

[0071] S210, construct an LSTM model network, the model structure is a two-layer LSTM network and a fully connected (Dense) layer;

[0072] S211, model training, using the adaptive moment estimation algorithm (Adam algorithm) for optimization calculation, using the mean square error (MSE) loss calculation and accuracy evaluation criteria for model training. Set the model training standard mean square error to when the test set error is greater than this value, the model will be retrained until the error requirement is met;

[0073] S212, setting the interval for model update to one hour, and setting a dynamic threshold every hour based on the standard deviation of the data of the previous hour;

[0074] S214. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

[0075] Step S209 specifically includes the following steps:

[0076] S221, set up alternative sets for the number of neurons in each layer of the network and the input time window, and search for hyperparameters in the network;

[0077] S222, looping the model under each parameter combination in the candidate set ten times, and taking the ten averages of the mean absolute error on the test set as the measurement standard;

[0078] S223. Select the hyperparameter combination by comparing the measurement criteria, that is, the one with the smallest ten-fold mean of the mean absolute error of the test set is the optimal hyperparameter combination.

[0079] The LSTM network of step S210 includes a cell state that tracks the information content of the previous node, a forget gate, a memory gate, and an output gate that control the input and output of information. When the cell state progresses from time t-1 to time t, the input value at the current moment and the output of the model at the previous moment are added or removed to the cell state at the previous moment through the forget gate and the memory gate, completing an update of the cell state. The calculation formula in the LSTM network includes the following:

[0080] (1) Setting

[0081] (2) Forget gate: Γ f =σ(Wf [X t ,h t-1 ]+b f );

[0082] (3) Memory Gate:

[0083] (4) Output gate: Γ o =σ(W o [X t ,h t-1 ]+b o );

[0084] (5) Cell state: c t =tanh(W c [X t ,h t-1 ]+b c ), c t =Γ u *c t +Γ f *c t-1 ;

[0085] (6) Output value: h t =Γ o *c t .

[0086] Among them, tanh(*) represents the hyperbolic tangent function, σ(*) represents the sigmoid function, and c t is the memory cell state, h t is the hidden state at time t, h t-1 is the hidden state at the moment before time t, W f , W u , W o , W c is the correlation weight matrix, b f , b u , b o , b c is the relevant bias term.

[0087] Step S211 specifically includes the following steps:

[0088] S231, set random numbers, and perform iterative training on parameters through optimization calculation and loss calculation. The optimization calculation uses the adaptive moment estimation algorithm (Adam algorithm), and the loss calculation uses the mean squared error (MSE);

[0089] S232, the network is repeatedly trained 30 times, and the distribution of the mean square error on the test set is observed;

[0090] S233, determine the test set error, if the mean square error is greater than 2*10 -8 , then returns to step S231, otherwise the network training is completed.

[0091] Step S212 includes the following process: setting the interval of model update to one hour, retraining the LSTM model every hour based on the data of the previous hour, and checking whether it exceeds the mean square error standard of 2*10 -8 , if it exceeds, retrain the LSTM model.

[0092] Step S14 specifically includes the following process:

[0093] The standard deviation of the data from the previous hour is used as the benchmark for the fluctuation range of the monitoring data, and 5.5 times, 7 times and 8 times the standard deviation are taken as thresholds to determine abnormal points.

[0094] Calculate the error between the measured value and the predicted value. If the error is greater than a certain graded dynamic threshold, output the corresponding graded warning result. Otherwise, store the monitoring data and error.

[0095] This embodiment uses settlement monitoring data to identify anomalies in tunnel health monitoring data. Use the LSTM model to divide the training set and test set in a ratio of 5:1, and adjust the input parameters of the LSTM to keep it consistent with the average length of the ARIMA confidence interval. Set the model training standard MSE to, when the test set error is greater than this value, retrain the model until the error requirement is met. The settlement value deviates greatly from the predicted value, and the identified anomalies, such as Figure 3b .

[0096] At the modeling level, LSTM is less constrained by data features, but has higher requirements for sample capacity, and the model is more complex and difficult to explain. At the level of anomaly identification, ARIMA is very sensitive to changes in sequence values ​​in the short term, but cannot identify continuous anomalies in the sequence in the medium and long term, and can usually only be identified after the anomaly occurs, so the early warning is not sufficient. LSTM learns the potential characteristics of medium and long-term changes in the sequence, and is not as sensitive to short-term fluctuations, but has better control over the long-term trend of the sequence, and the early warning is more forward-looking.

[0097] Combining the two methods, more accurate anomaly recognition results are obtained. The LSTM method is used as the main method to monitor the medium- and long-term trend of the sequence; the ARIMA method is used as a supplement to monitor the anomalies that occur in the short- and medium-term of the sequence. It can identify and issue early warning signals in a timely manner, avoiding the possibility of missed detection and the inability to identify short-term sudden anomalies in scenarios with large fluctuations.

[0098] The step S3 includes the following steps of decomposing the time series elements based on STL:

[0099] S301, perform STL decomposition on deformation monitoring data to obtain trend terms, seasonal terms and residual terms.

[0100] S302, performing function fitting on the trend term, the seasonal term, and the residual term respectively.

[0101] S303, summing up the three fitted parts to obtain a combined model that can predict the subsequent development of tunnel convergence deformation.

[0102] Step S301 performs STL decomposition, including the following steps:

[0103] S310, setting the trend component to zero;

[0104] S311, removing the trend component from the original sequence to obtain a detrended sequence;

[0105] S312, using the LOESS (local weighted regression) smoothing method for the subsequence at each position to obtain a smoothed sequence;

[0106] S313, performing low-pass filtering on the smoothed sequence to eliminate noise;

[0107] S314, applying LOESS (locally weighted regression) smoothing method to the detrended sequence to obtain a deseasonalized sequence;

[0108] S315, calculating the seasonal series by subtracting the deseasonalized series from the original series;

[0109] S316. Obtain the trend series by using the LOESS (locally weighted regression) smoothing method on the seasonal series;

[0110] S317. Subtract the trend component and seasonal component from the original sequence to obtain the residual sequence.

[0111] S318, judging whether the residual term sequence is less than a given threshold, if so, then the process is completed, otherwise, proceeding to step S319;

[0112] S319. Introduce a weight coefficient in the loop to control the influence of outliers in the data, and repeat the loop until it is determined whether the residual term sequence is less than a given threshold.

[0113] This embodiment uses the STL model to model the time series of tunnel convergence, such as Figure 4 The STL model prediction results and actual measured values ​​are plotted, and the mean absolute error and standard deviation of the error are used to test the accuracy of the model prediction. The smaller the value, the better the model performance. The mean absolute error is 0.938 and the standard deviation of the error is 1.047. The STL model captures the seasonal variation and long-term expansion trend of the nonlinear convergence amount, and has a high prediction accuracy.

[0114] The step S4 specifically includes the following steps:

[0115] S401. Set the number of different groups to k and the number of different samples to n. The subscript i represents the group, and each group has n i The calculation formula is as follows: The systematic difference formed by different test groups is called the between-group deviation (MSA). The other type is the random difference between individuals in the group, which is called the within-group deviation (MSE).

[0116] Group mean:

[0117] Overall mean:

[0118] S402, introduce F statistics to perform statistical hypothesis testing,

[0119] The calculation formula is as follows:

[0120]

[0121] S403. Compare MSA, MSE and F value. If MSA is greater than MSE, the F statistic is greater than 1. It is judged that the difference caused by the test group contributes greatly to the total difference, and it is considered that there are significant differences between different test groups. Otherwise, it is considered that there are no significant differences.

[0122] A tunnel anomaly warning and trend prediction system comprises an ARIMA-based anomaly warning module, an LSTM-based anomaly warning module and an STL-based convergence deformation prediction module.

[0123] This embodiment uses Python's Tkinter module to implement the system's graphical interface, and uses the matplotlib module to achieve dynamic graph drawing of measured values ​​and model prediction values.

[0124] The ARIMA-based abnormal warning module includes a data acquisition unit: used to import tables and obtain data from the tables; a parameter setting unit: to set the number of times the standard deviation exceeds the predicted data range, set the prediction time range, and set the parameters of the ARIMA (p, d, q) model; and a result generation unit: to draw a dynamic prediction chart and display the time and number of abnormal data. Figure 5 shown.

[0125] The LSTM-based abnormal warning module includes a data acquisition unit: used to import tables and obtain data from the tables; a parameter setting unit: to set the number of times the standard deviation exceeds the predicted data range and the prediction time range; a result generation unit: to draw a dynamic prediction graph and display the time and amount of abnormal data. Figure 6 shown.

[0126] The STL-based convergence deformation prediction module includes a data acquisition unit: used to import tables and obtain data from the tables; a result generation unit: used to output ① Trend item: draw the trend item curve and its fitting straight line, and display the slope of the fitting straight line. ② Season item: draw the seasonal item curve and display the seasonal index of each quarter. ③ Irregular item: draw the irregular item curve and its fitting straight line. ④ STL combined model: draw the original input curve and the model prediction curve, and predict the convergence value in the next two years.

[0127] The present invention is a more intelligent and reliable anomaly detection method, which is helpful for improving structural reliability, ensuring its operational safety, reducing maintenance costs, conducting disaster forecasting, and improving service levels.

Claims

1. A tunnel abnormality early warning and trend prediction method, characterized by: The following steps are involved: S1, data collection and compilation; S2. Tunnel health monitoring data anomaly identification: mainly using the LSTM method to monitor the medium- and long-term trend of the sequence and give early warning signals before the anomaly occurs; supplemented by the ARIMA method, a more stringent threshold judgment method is set to monitor the short-term trend of the sequence, identify and issue early warning signals in a timely manner, and implement the application of the overall data anomaly identification algorithm; S3. Decomposition of tunnel health monitoring data elements and construction of prediction model: Decomposition of health monitoring data into trend items, periodicity and random items to build a prediction model; S4. Analyze the spatiotemporal variation characteristics of the decomposed elements: Based on the ANOVA variance analysis data development trend, a comparative analysis is conducted to obtain the spatiotemporal variation law of the tunnel deformation data.

2. A tunnel abnormality early warning and trend prediction method according to claim 1, characterized in that: The step S2 includes the following steps of the ARIMA-based model data anomaly identification algorithm: S201, stationarity test, if the time series is stationary, proceed to step S22, if not, perform differential operation and then proceed to step S22; S202, white noise test, if the time series is white noise, proceed to step S26, if not, proceed to step S23; S203, model identification, using Akaike information criterion and Bayesian information criterion to set ARIMA model parameters; S204, model fitting; S205, model testing, testing the fit between the model and the data; S206, model prediction, performing multi-step prediction based on single-step prediction; S207, establishing a dynamic prediction model; fitting the model value and the actual value error, and updating the model parameters if the error average is greater than the dynamic threshold, otherwise the model parameters are not updated; S208. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

3. A tunnel abnormality early warning and trend prediction method according to claim 1, characterized in that: The step S2 includes the following steps of the LSTM-based model data anomaly recognition algorithm: S209, data training, divide the training set and the test set in a ratio of 5:1, and use the network to search for the optimal hyperparameters; S210, construct an LSTM model network, the model structure is a two-layer LSTM network and a fully connected (Dense) layer; S211, model training, using the adaptive moment estimation algorithm (Adam algorithm) for optimization calculation, using the mean square error (MSE) loss calculation and accuracy evaluation criteria for model training. Set the model training standard mean square error to when the test set error is greater than this value, the model will be retrained until the error requirement is met; S212, setting the interval for model update to one hour, and setting a dynamic threshold every hour based on the standard deviation of the data of the previous hour; S214. Calculate the error of the difference between the predicted value and the measured value, set a dynamic threshold with reference to the standard deviation of the data in the previous hour, and output a graded warning after comparing the error with the dynamic threshold.

4. A tunnel abnormality early warning and trend prediction method according to claim 1, characterized in that: The step S3 includes the following steps of decomposing the time series elements based on STL: S301, perform STL decomposition on deformation monitoring data to obtain trend terms, seasonal terms and residual terms. S302, performing function fitting on the trend term, the seasonal term, and the residual term respectively. S303, summing up the three fitted parts to obtain a combined model that can predict the subsequent development of tunnel convergence deformation.

5. The method for tunnel abnormality early warning and trend prediction according to claim 1 is characterized by: The step S4 specifically includes the following steps: S401. Set the number of different groups to k and the number of different samples to n. The subscript i represents the group, and each group has n i The calculation formula is as follows: The systematic difference formed by different test groups is called the between-group deviation (MSA). The other type is the random difference between individuals in the group, which is called the within-group deviation (MSE). Group mean: Overall mean: S402, introduce F statistics to perform statistical hypothesis testing, The calculation formula is as follows: S403. Compare MSA, MSE and F value. If MSA is greater than MSE, the F statistic is greater than 1. It is judged that the difference caused by the test group contributes greatly to the total difference, and it is considered that there are significant differences between different test groups. Otherwise, it is considered that there are no significant differences.

6. A tunnel abnormality warning and trend prediction system, characterized by: The system comprises an abnormal warning module based on ARIMA, an abnormal warning module based on LSTM and an abnormal warning module based on STL convergence deformation prediction module.

7. A tunnel abnormality early warning and trend prediction system according to claim 6, characterized in that: The ARIMA-based abnormal warning module includes a data acquisition unit: used to import a table and obtain data from the table; a parameter setting unit: setting how many times the standard deviation exceeds the predicted data range, setting the prediction time range, and setting the parameters of the ARIMA (p, d, q) model; and a result generation unit: drawing a dynamic prediction graph and displaying the time and amount of abnormal data.

8. The tunnel abnormality early warning and trend prediction system according to claim 6 is characterized by: The LSTM-based abnormal warning module includes a data acquisition unit: used to import a table and obtain data from the table; a parameter setting unit: used to set how many times the standard deviation exceeds the predicted data range and set the prediction time range; and a result generation unit: used to draw a dynamic prediction graph and display the time when abnormal data occurs and the number of abnormal data.

9. The tunnel abnormality early warning and trend prediction system according to claim 6, characterized in that: The STL-based convergence deformation prediction module includes a data acquisition unit: used to import a table and obtain data from the table; a result generation unit: used to output ① trend item: draw the trend item curve and its fitting line, and display the slope of the fitting line. ②Seasonal term: draw the seasonal term curve and display the seasonal index of each quarter. ③Irregular term: draw the irregular term curve and its fitted straight line. ④STL combined model: draw the original input curve and the model forecast curve, and forecast the convergence value in the next two years.

Images