DIC and LSTM-based overhanging rock instability early warning method

Abstract

The invention belongs to the technical field of geological disaster monitoring, and particularly discloses a DIC and LSTM-based overhanging rock instability early warning method. The method comprises the following steps: regularly acquiring surface images of the overhanging rock; carrying out geometric correction on the image and removing noise; analyzing the image by using a DIC technology to obtain a strain field mean value mu and a strain field standard deviation sigma; establishing an LSTM model comprising an input layer, an LSTM layer, a full connection layer and an output layer; training an LSTM model by adopting the training set and the verification set, and optimizing the performance of the LSTM model by adjusting hyper-parameters of the LSTM model; performing normalization processing on the strain field mean value mu and the strain field standard deviation sigma, dividing the strain field mean value mu and the strain field standard deviation sigma into a plurality of sliding windows, inputting data of each window into an LSTM model, outputting a prediction result of each window, judging the prediction result to be normal or abnormal, and triggering early warning when the windows are continuously judged to be abnormal. The method has the characteristics of high timeliness, less human intervention and accurate judgment.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster monitoring, and in particular to a DIC and LSTM-based overhang rock instability early warning method with high timeliness, less human intervention and accurate judgment. Background Art

[0002] Overhang rock is a rock mass with potential collapse risk formed on slopes or cliffs due to external forces such as weathering, rainfall, and earthquakes. The instability and collapse of overhang rock not only poses a threat to engineering structures, but may also cause major safety accidents. Therefore, accurately predicting the stability of overhang rock and conducting effective early warning have become important issues in the current prevention and control of geological disasters.

[0003] At present, traditional monitoring of overhanging rock stability mostly relies on manual inspection, survey, radar, sensor monitoring and other means, but manual inspection and survey have problems such as poor timeliness and reliance on human experience, while radar and sensor monitoring have the problem of limited data, which makes the accuracy of early warning low. In recent years, digital image correlation technology (DIC) is a non-contact, non-interference full-field deformation measurement method. It compares the digital images of the object before and after deformation, and uses the feature points or feature areas in the image to match the images before and after deformation, so as to calculate the displacement and deformation of the object surface. DIC is widely used in rock deformation monitoring due to its advantages such as extremely loose experimental environment requirements, strong anti-interference ability, and high measurement accuracy.

[0004] In the prior art, DIC is applied to overhang rock monitoring, which can accurately capture the slight changes in rock deformation by collecting displacement field or strain field images on the surface of the rock mass. However, the captured rock deformation information is generally displayed in the form of three-dimensional images, curve graphs, charts, etc., and then researchers intuitively analyze the deformation of the rock mass. However, the analysis based on images and charts by researchers is not only slow and labor-intensive, resulting in low timeliness, but also limited by the knowledge level and sense of responsibility of the researchers, making the accuracy and reliability of the analysis conclusions insufficient. Therefore, how to efficiently analyze and automatically judge the instability of overhang rocks is still a technical problem. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention provides an overhanging rock instability early warning method based on DIC and LSTM, which has high timeliness, less human intervention and accurate judgment.

[0006] The present invention is implemented in this way: including data collection, image preprocessing, deformation field calculation, early warning model construction, early warning model training, monitoring and early warning steps, the specific contents are: A. Data collection: Regularly collect surface images of key structural parts of overhanging rocks under external forces; B. Image preprocessing: geometric correction and noise removal of the acquired images; C. Deformation field calculation: The pre-processed images are analyzed using DIC technology to obtain the displacement field and strain field of all image sequences, which are then exported as a strain field data matrix. The strain field data matrix is ​​statistically analyzed to obtain the strain field mean μ and the strain field standard deviation. s : , , Where: m and n are the total number of rows and columns of the strain field data matrix respectively; ε ij is the first i Line j The corresponding strain values ​​of the columns; D. Early warning model construction: Establish an LSTM model including input layer, LSTM layer, fully connected layer and output layer, where the input layer receives the strain field mean μ and the strain field standard deviation s The sliding window data is divided by time points. The LSTM layer extracts the features of each sliding window and captures the changing trend and potential mutation of the strain field. The fully connected layer outputs a single value, and the output layer outputs the window prediction result as a label of "normal" or "abnormal"; E. Early warning model training: Divide the sliding window data set of existing overhang rock instability into a training set and a validation set, and ensure that the validation set contains some abnormal points. Then input the training set and the validation set into the LSTM model, use the labels to supervise the training process of the LSTM model, monitor the loss value and accuracy of the LSTM model during the training process, and optimize the performance of the LSTM model by adjusting the hyperparameters of the LSTM model; F. Monitoring and early warning: The mean value of the strain field and the standard deviation σ of the strain field obtained in step C are received and normalized. Then, the normalized data is divided into multiple sliding windows according to the time point. Then, the data of each sliding window is input into the trained LSTM model. The LSTM model captures the changing trend and potential mutation of the strain field, and then outputs the prediction result of each sliding window and determines whether the sliding window is "normal" or "abnormal". When the LSTM model continuously determines that the window is "abnormal", an early warning is triggered immediately.

[0007] Furthermore, the specific process of step A is as follows: A1. Select monitoring area: Select the key structural parts of overhanging rock including dangerous rock mass, main controlling joint surface and bedrock as monitoring area; A2. Monitoring equipment installation: Install high-resolution camera equipment and set up light sources around the monitoring area of ​​the overhanging rock; A3. Image acquisition: Through high-resolution camera equipment, based on the stability of the monitored overhanging rock mass, surface images of the overhanging rock in the monitoring area under the action of external forces are regularly collected.

[0008] Furthermore, in the step B, perspective distortion and / or lens distortion correction is performed on the captured image, and then image noise is removed by Gaussian filtering or median filtering.

[0009] Furthermore, the specific process of step C is as follows: C1. Image input: input the preprocessed image into Ncorr software; C2. Data matrix: Ncorr software uses the first image or the designated image in the preprocessed image as the reference image, and the remaining images as the deformed images, and compares them with the reference image one by one. Ncorr software calculates the displacement of each pixel by identifying the changes in the surface pattern of the overhanging rock in the image, and then generates a two-dimensional displacement field reflecting the movement of the overhanging rock surface at different time points; then Ncorr software analyzes the gradient change of the displacement in the displacement field, obtains the strain field composed of the strain values ​​of each pixel point, and then saves the strain field data as a strain field matrix at different time points; C3. Matrix calculation: Statistical analysis is performed on the strain field data matrix to obtain the strain field mean μ and the strain field standard deviation s : , , Where: m and n are the total number of rows and columns of the strain field data matrix respectively; ε ij is the first i Line j The corresponding strain values ​​of the column.

[0010] Furthermore, in the step D, the LSTM model includes 2 to 3 LSTM layers, and each LSTM layer contains 100 to 150 neurons; the output layer uses a Sigmoid activation function to map the output to a probability between 0 and 1, and then uses a binary classification method to output the window prediction result as a label of "normal" or "abnormal"; wherein the Sigmoid activation function expression is: , Where: x is the linear output of the LSTM model; y It is the probability value calculated by the Sigmoid function, ranging from 0 to 1; e is the base of natural logarithms; The binary classification method is to convert the probability value output by the Sigmoid functiony ≥With preset value k Compare, when y ≥ k When the prediction result is “abnormal”, the output prediction result is “normal”.

[0011] Furthermore, in the step E, the sliding window data set of existing overhang rock instability is divided into k subsets, k-1 subsets are used as training sets, and the remaining 1 subset is used as a validation set, and it is ensured that the validation set contains some abnormal points. The LSTM model is trained and validated alternately using the aforementioned training set and validation set, and finally the loss value and accuracy are calculated to evaluate the performance of the LSTM model, and then the performance of the LSTM model is optimized by adjusting the hyperparameters of the LSTM model including the number of LSTM layers, the number of neurons, the learning rate and / or regularization according to the performance.

[0012] Furthermore, the loss value is obtained using a binary cross entropy loss function: , Where: L is the average loss value; N is the total number of samples; y i For the i The true label of each sample (0 or 1); y i ' It is i The probability of a sample being predicted by the LSTM model is between 0 and 1. The greater the deviation between the prediction result of the LSTM model and the true label, the greater the loss value, thereby guiding the LSTM model to adjust parameters to gradually reduce the error. The adjustment of the number of LSTM layers and the number of neurons is to optimize the number of LSTM layers and the number of neurons in each layer through grid search or random search to find the optimal complexity of the LSTM model; the learning rate adjustment is to dynamically adjust the learning rate using a learning rate decay or an adaptive learning rate optimization algorithm to avoid gradient explosion or disappearance during training; the regularization introduces a regularization method to randomly discard a part of neurons to prevent the LSTM model from overfitting.

[0013] Furthermore, in the F step, the normalization process is performed using Min-Max normalization, which converts the strain field mean μ and the strain field standard deviation s All are scaled to between 0 and 1 to avoid the impact of too large or too small values ​​on the model; , Where: x i is the mean or standard deviation in the received data; xmin and x max are the minimum and maximum values ​​of the group of data respectively; x΄ is the normalized data, with a range of [0,1]; Dividing the entire time series data into multiple sliding windows is to take a predetermined number of consecutive time points starting from the first time point. p The corresponding data is used as the first window; then the window is slid backward from the first time point q At a certain point in time, q The time point is taken as the first time point, and the consecutive time points are taken p The corresponding data forms the second window; repeat the above process until the entire time series is divided into multiple sliding windows; among them, q < p .

[0014] Furthermore, in the F step, the LSTM model extracts the strain field mean μ and strain field standard deviation corresponding to each sliding window from the input data of each sliding window. s Two features ， Form a window length p × two-dimensional feature matrix of two features; then set the strain field mean threshold and strain field standard deviation threshold, use the LSTM model to compare the values ​​of the two-dimensional feature matrix with the threshold one by one, and capture the change trend and potential mutation of the strain field; finally, output the prediction result of each sliding window and judge the sliding window as "normal" or "abnormal", and trigger an early warning immediately when the continuous judgment window is "abnormal"; The threshold is set according to the data characteristics, and a strain field mean threshold and a strain field standard deviation threshold are set respectively. Then the LSTM model compares the prediction results of each sliding window with the corresponding thresholds. When the prediction results exceed their respective thresholds, the sliding window is output as an "abnormal" label, otherwise the sliding window is output as a "normal" label.

[0015] LSTM (Long Short-Term Memory) is a special recurrent neural network (RNN) that uses a memory gating mechanism (input gate, forget gate, output gate) to decide which information to remember and which to forget, thereby effectively solving the long-range dependency and gradient vanishing problems in RNN, and can capture long-term and short-term dependencies in time series.

[0016] The beneficial effects of the LSTM invention are: 1. The present invention combines digital image correlation technology (DIC) and long short-term memory network (LSTM) model, which can not only give full play to the advantages of DIC such as extremely loose requirements on experimental environment, strong anti-interference ability and high measurement accuracy, but also automatically analyze the strain field statistical curve generated by DIC through the LSTM model, so as to predict the stability of overhanging rock in real time and generate multi-level early warning signals, providing decision-making basis for geological disaster prevention, and effectively avoiding the problems of low timeliness, accuracy and reliability of existing manual analysis of DIC charts.

[0017] 2. The present invention evaluates the performance of the LSTM model by calculating the loss value and accuracy during the early warning model training, and then adjusts the hyperparameters of the LSTM model such as the number of LSTM layers, the number of neurons, the learning rate and the regularization according to the performance, thereby optimizing the performance of the LSTM model. It can not only effectively match the optimal complexity of the LSTM model to improve the computational efficiency, but also avoid the gradient explosion or disappearance during the LSTM model training process, and avoid overfitting of the LSTM model.

[0018] 3. The present invention has achieved innovative applications in the intersection of optical monitoring, machine learning and statistics. The full-field deformation data of the surface of the overhanging rock mass is obtained through DIC technology. Combined with the time series data analysis capability of the LSTM model, it is possible to capture potential instability characteristics and make accurate time series predictions of deformation trends. At the same time, statistical methods are used to normalize the data, process outliers and set thresholds to ensure the reliability and robustness of the model in complex environments. This multidisciplinary integration enables the present invention to quickly respond to the early warning needs of geological disasters, provide efficient and reliable decision-making support for relevant departments, and greatly improve the level of intelligence in geological disaster prevention.

[0019] 4. The present invention is based on the trainability of machine learning models and optimizes model parameters to cope with different monitoring scenarios, thereby forming a set of universal and adaptable intelligent monitoring and early warning methods that can meet the needs of different geological conditions. Therefore, it can not only be used for monitoring the instability of overhanging rocks, but also for monitoring other types of geological disasters such as landslides and slope collapses. This universal design provides a wider application prospect for geological disaster monitoring.

[0020] In summary, the present invention has the characteristics of high timeliness, less human intervention and accurate judgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flow chart of the present invention; Figure 2 It is a flow chart of the monitoring and early warning steps in the present invention; Figure 3 A schematic diagram of a typical overhanging rock of the present invention; Figure 4 The strain field mean value-time curve and the strain field standard deviation-time curve in the embodiment; Figure 5 This is an example diagram of sliding window division in the embodiment; In the figure: 1-bedrock, 2-dangerous rock mass, 3-free surface, 4-main controlling joint surface. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0023] like Figure 1 As shown, the present invention includes data collection, image preprocessing, deformation field calculation, early warning model construction, early warning model training, monitoring and early warning steps, and the specific contents are: A. Data collection: Regularly collect surface images of key structural parts of overhanging rocks under external forces; B. Image preprocessing: geometric correction and noise removal of the acquired images; C. Deformation field calculation: The pre-processed images are analyzed using DIC technology to obtain the displacement field and strain field of all image sequences, which are then exported as a strain field data matrix. The strain field data matrix is ​​statistically analyzed to obtain the strain field mean μ and the strain field standard deviation. s : , , Where: m and n are the total number of rows and columns of the strain field data matrix respectively; ε ij is the first i Line j The corresponding strain values ​​of the columns; D. Early warning model construction: Establish an LSTM model including input layer, LSTM layer, fully connected layer and output layer, where the input layer receives the strain field mean μ and the strain field standard deviation s The sliding window data is divided by time points. The LSTM layer extracts the features of each sliding window and captures the changing trend and potential mutation of the strain field. The fully connected layer outputs a single value, and the output layer outputs the window prediction result as a label of "normal" or "abnormal"; E. Early warning model training: Divide the sliding window data set of existing overhang rock instability into a training set and a validation set, and ensure that the validation set contains some abnormal points. Then input the training set and the validation set into the LSTM model, use the labels to supervise the training process of the LSTM model, monitor the loss value and accuracy of the LSTM model during the training process, and optimize the performance of the LSTM model by adjusting the hyperparameters of the LSTM model; F. Monitoring and early warning: The mean value of the strain field and the standard deviation σ of the strain field obtained in step C are received and normalized. Then, the normalized data is divided into multiple sliding windows according to the time point. Then, the data of each sliding window is input into the trained LSTM model. The LSTM model captures the changing trend and potential mutation of the strain field, and then outputs the prediction result of each sliding window and determines whether the sliding window is "normal" or "abnormal". When the LSTM model continuously determines that the window is "abnormal", an early warning is triggered immediately.

[0024] The specific process of step A is as follows: A1. Select monitoring area: Select the key structural parts of overhanging rock including dangerous rock mass, main controlling joint surface and bedrock as monitoring area; A2. Monitoring equipment installation: Install high-resolution camera equipment and set up light sources around the monitoring area of ​​the overhanging rock; A3. Image acquisition: Through high-resolution camera equipment, based on the stability of the monitored overhanging rock mass, surface images of the overhanging rock in the monitoring area under the action of external forces are regularly collected.

[0025] In step B, perspective distortion and / or lens distortion correction is performed on the captured image, and then image noise is removed by Gaussian filtering or median filtering.

[0026] The specific process of step C is as follows: C1. Image input: input the preprocessed image into Ncorr software; C2. Data matrix: Ncorr software uses the first image or the designated image in the preprocessed image as the reference image, and the remaining images as the deformed images, and compares them with the reference image one by one. Ncorr software calculates the displacement of each pixel by identifying the changes in the surface pattern of the overhanging rock in the image, and then generates a two-dimensional displacement field reflecting the movement of the overhanging rock surface at different time points; then Ncorr software analyzes the gradient change of the displacement in the displacement field, obtains the strain field composed of the strain values ​​of each pixel point, and then saves the strain field data as a strain field matrix at different time points; C3. Matrix calculation: Statistical analysis is performed on the strain field data matrix to obtain the strain field mean μ and the strain field standard deviation s : , , Where: m and n are the total number of rows and columns of the strain field data matrix respectively; ε ij is the first i Line j The corresponding strain values ​​of the column.

[0027] The step C also includes obtaining curves: plotting the strain field mean and strain field standard deviation σ of the aforementioned images in time sequence to obtain the strain field mean-time curve and strain field standard deviation-time curve. The curve gives a schematic diagram of the corresponding parameter evolution trend, and the trend of mutation can be seen; at the same time, it can also intuitively display the division of windows.

[0028] In the step D, the LSTM model includes 2 to 3 LSTM layers, and each LSTM layer contains 100 to 150 neurons; the output layer uses the Sigmoid activation function to map the output to a probability between 0 and 1, and then uses a binary classification method to output the window prediction result as a label of "normal" or "abnormal"; wherein the Sigmoid activation function expression is: , Where: x is the linear output of the LSTM model (x is the numerical representation of the overall state of the input sequence by the LSTM network. Before being processed by the activation function, it can be any real number, which contains the model's prediction results for the input data, but has not yet been normalized. In the instability warning model, the larger the x value, the higher the possibility of "instability" in the model; conversely, the higher the possibility of "normality"); y It is the probability value calculated by the Sigmoid function, ranging from 0 to 1; e is the base of natural logarithms; The binary classification method is to convert the probability value output by the Sigmoid function y With the default value k Compare, when y ≥ k When the prediction result is “abnormal”, the output prediction result is “normal”.

[0029] In the step E, the sliding window data set of existing overhang rock instability is divided into k subsets, k-1 subsets are used as training sets, and the remaining 1 subset is used as a validation set, and it is ensured that the validation set contains some abnormal points. The LSTM model is trained and validated alternately using the aforementioned training set and validation set, and finally the loss value and accuracy are calculated to evaluate the performance of the LSTM model. Subsequently, the performance of the LSTM model is optimized by adjusting the hyperparameters of the LSTM model including the number of LSTM layers, the number of neurons, the learning rate and / or regularization according to the performance.

[0030] The loss value is obtained using the binary cross entropy loss function: , Where: L is the average loss value; N is the total number of samples; y i For the i The true label of each sample (0 or 1); y i ' It is i The probability of a sample being predicted by the LSTM model is between 0 and 1. The greater the deviation between the prediction result of the LSTM model and the true label, the greater the loss value, thereby guiding the LSTM model to adjust parameters to gradually reduce the error. The adjustment of the number of LSTM layers and the number of neurons is to optimize the number of LSTM layers and the number of neurons in each layer through grid search or random search to find the optimal complexity of the LSTM model; the learning rate adjustment is to dynamically adjust the learning rate using a learning rate decay or an adaptive learning rate optimization algorithm to avoid gradient explosion or disappearance during training; the regularization introduces a regularization method to randomly discard a part of neurons to prevent the LSTM model from overfitting.

[0031] In the F step, the normalization process is performed using Min-Max normalization, which converts the strain field mean μ and the strain field standard deviation s All are scaled to between 0 and 1 to avoid the impact of too large or too small values ​​on the model; , Where: x i is the mean or standard deviation in the received data; x min and x max are the minimum and maximum values ​​of the group of data respectively; x΄ is the normalized data, with a range of [0,1]; Dividing the entire time series data into multiple sliding windows is to take a predetermined number of consecutive time points starting from the first time point. p The corresponding data is used as the first window; then the window is slid backward from the first time point q At a certain point in time, q The time point is taken as the first time point, and the consecutive time points are taken p The corresponding data forms the second window; repeat the above process until the entire time series is divided into multiple sliding windows; among them, q < p .

[0032] In the F step, the LSTM model extracts the strain field mean μ and strain field standard deviation corresponding to each sliding window from the input data of each sliding window. s Two features ， Form a window length p × two-dimensional feature matrix of two features; then set the strain field mean threshold and strain field standard deviation threshold, use the LSTM model to compare the values ​​of the two-dimensional feature matrix with the threshold one by one, and capture the change trend and potential mutation of the strain field; finally, output the prediction result of each sliding window and judge the sliding window as "normal" or "abnormal", and trigger an early warning immediately when the continuous judgment window is "abnormal"; The threshold is set according to the data characteristics, and a strain field mean threshold and a strain field standard deviation threshold are set respectively. Then the LSTM model compares the prediction results of each sliding window with the corresponding thresholds. When the prediction results exceed their respective thresholds, the sliding window is output as an "abnormal" label, otherwise the sliding window is output as a "normal" label.

[0033] Example 1: Overhanging rock instability early warning method based on DIC and LSTM of the present invention S100: Regularly collect surface images of key structural parts of overhanging rocks under external forces. The specific process is as follows: S110: Select the key structural parts of the overhanging rock including the dangerous rock mass, the main controlling joint surface and the bedrock as the monitoring area; ensure that the monitoring area covers the key structural parts so as to accurately capture the deformation information of the rock mass.

[0034] S120: Install high-resolution cameras with at least 50 megapixels around the monitoring area of ​​the overhanging rock, and set up light sources to ensure stable lighting conditions. The high-resolution cameras are installed on a stable bracket to avoid vibration interference.

[0035] S130: Using high-resolution camera equipment, based on the stability of the monitored overhanging rock mass, regularly collect surface images of the overhanging rock in the monitoring area under the action of external forces. The collection frequency can be adjusted based on the stability of the rock mass.

[0036] S200: performing geometric corrections such as perspective distortion and / or lens distortion correction on the acquired image, and then removing image noise by Gaussian filtering or median filtering to ensure image quality.

[0037] S300: The pre-processed images are analyzed using DIC technology to obtain the displacement field and strain field of all image sequences, which are then exported as a strain field data matrix. The strain field data matrix is ​​statistically analyzed to obtain the strain field mean μ and the strain field standard deviation. s The specific process is as follows: S310: Input the pre-processed image into the DIC analysis software to facilitate the subsequent deformation field calculation.

[0038] S320: Analyze the input pre-processed image using DIC technology to obtain the displacement field and strain field of all image sequences, and then derive the strain field data matrix and save it; S330: Perform statistical analysis on the strain field data matrix to obtain the strain field mean μ and the strain field standard deviation s : , , Where: m and n are the total number of rows and columns of the strain field data matrix respectively; ε ij is the first i Line j The corresponding strain values ​​of the column.

[0039] S340: Figure 4 As shown, the strain field mean μ and strain field standard deviation of the above images are s By plotting in time sequence, we can obtain the strain field mean value-time curve and the strain field standard deviation-time curve.

[0040] S400: Establish an LSTM model including an input layer, an LSTM layer, a fully connected layer and an output layer, wherein the input layer receives the sliding window data of the strain field mean and the strain field standard deviation σ divided by time points, the LSTM layer extracts the features of each sliding window and captures the changing trend and potential mutation of the strain field, the fully connected layer (Dense Layer) outputs a single value to indicate whether the sliding window is abnormal, and the output layer outputs the window prediction result as a label of "normal" or "abnormal".

[0041] In this embodiment, the LSTM model includes 2 to 3 LSTM layers, and each LSTM layer contains 100 to 150 neurons. The LSTM layer extracts information from the input time series to capture the changing trend and potential mutation of the strain field; the output layer uses the Sigmoid activation function to map the output to a probability between 0 and 1, and then uses a binary classification method to output the window prediction result as a label of "normal" or "abnormal"; wherein, the Sigmoid activation function expression is: , Where: x is the linear output of the LSTM model; y It is the probability value calculated by the Sigmoid function, ranging from 0 to 1; e is the base of natural logarithms.

[0042] The binary classification method is to convert the probability value output by the Sigmoid function y With the default value k Compare, when y ≥ k If the prediction result is "abnormal", the prediction result is output as "normal". k Take 0.5, then y When ≥0.5, the prediction result is output as “abnormal”, otherwise the prediction result is output as “normal”.

[0043] S500: In order to ensure the stability and generalization ability of the LSTM model, cross-validation is introduced to verify the LSTM model. Using k-fold cross-validation, the sliding window dataset of the existing overhang rock instability is divided into k subsets, with k-1 subsets as training sets and the remaining 1 subset as validation sets (usually 80% of the data is used as training sets and 20% of the data is used as validation sets), and it is ensured that the validation set contains some abnormal points; then the training set and validation set are used to train and validate the LSTM model in turn, and the labels ("normal" and "abnormal") are used to supervise the training process of the LSTM model. During the training process, the loss value and accuracy are calculated to evaluate the performance of the LSTM model, and then the performance of the LSTM model is optimized by adjusting the hyperparameters of the LSTM model including the number of LSTM layers, the number of neurons, the learning rate and / or regularization according to the performance.

[0044] The loss value is obtained using the binary cross entropy loss function: , Where: L is the average loss value; N is the total number of samples; y i For the i The true label of each sample (0 or 1);y i ' It is i The probability of a sample being predicted by the LSTM model is between 0 and 1. The greater the deviation between the prediction result of the LSTM model and the true label, the greater the loss value, thereby guiding the LSTM model to adjust parameters to gradually reduce the error.

[0045] Among them, the adjustment of the number of LSTM layers and the number of neurons is to optimize the number of LSTM layers and the number of neurons in each layer through grid search or random search to find the optimal complexity of the LSTM model; the learning rate adjustment is to use learning rate decay or adaptive learning rate optimization algorithm (such as Adam) to dynamically adjust the learning rate to avoid gradient explosion or disappearance during training; regularization is to introduce regularization methods such as Dropout to randomly discard a part of neurons to prevent overfitting of the LSTM model.

[0046] S600: During the actual monitoring process, receive the strain field mean value and strain field standard deviation obtained in step S300 s And normalize it; Figure 5 As shown in the figure, the normalized data is divided into multiple sliding windows starting from the first time point, and the data corresponding to 50 consecutive time points are taken as the first window, which means that each sliding window will contain data from 50 consecutive time points; in order to better capture mutations and changes in time series, the sliding windows should overlap (usually 50% window overlap is selected), so the window is then slid 25 time points backward from the first time point, and then starting from the 25th time point as the first time point, the data corresponding to 50 consecutive time points are taken to form the second window; repeat the above process until the entire time series is divided into multiple sliding windows. In this way, each time the sliding window slides, the new sliding window will share half of the data with the previous sliding window.

[0047] Then, the strain field mean threshold and strain field standard deviation threshold are set; then, the LSTM model extracts the strain field mean μ and strain field standard deviation corresponding to each sliding window from the input data of each sliding window. s Two features ， Form a window length p× A two-dimensional feature matrix of two features; it can also include derived features: (1) Calculate the rate of change of the mean and standard deviation in each window. For example, calculate the difference or rate of change of adjacent time points so that the model can capture the mutation characteristics of the time series; (2) Calculate the sliding average of the mean and standard deviation in the window, which can further smooth the data and help the model capture trends. Then, the LSTM model compares the values ​​of the feature matrix with the pre-set threshold one by one, and captures the changing trend and potential mutation of the strain field; when the strain field mean μ or strain field standard deviation of several consecutive sliding windows is monitored s When the threshold is exceeded, or the LSTM model continuously outputs "abnormal" windows (prediction probability is greater than 0.5), it can be considered that the overhang rock has cracked, and an early warning is immediately triggered to remind relevant departments or personnel to conduct further overhang rock monitoring or reinforcement.

[0048] Among them, the normalization process is to use Min-Max normalization, which converts the strain field mean μ and the strain field standard deviation s They are all scaled to between 0 and 1 to avoid the impact of too large or too small values ​​on the model.

[0049] , Where: x i is the mean or standard deviation in the received data; x min and x max are the minimum and maximum values ​​of the group of data respectively; x΄ is the normalized data, and its range is [0,1].

[0050] The setting of mean and standard deviation thresholds can be based on statistical analysis of historical data and experimental data. Specifically, a large amount of strain field data of overhanging rocks can be obtained through long-term monitoring, and the distribution of their mean and standard deviation can be statistically analyzed. Based on abnormal statistical methods (such as the 3σ rule), the mean or standard deviation exceeding 99.7% of the data distribution is defined as an abnormal point. Combined with field monitoring and experimental data, the thresholds are further verified and adjusted to ensure the rationality and effectiveness of the threshold setting. For example, empirical threshold standards for different types of rock masses can be established through stability experiments on different rock masses.

[0051] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for early warning of overhanging rock instability based on DIC and LSTM, characterized by: It includes data collection, image preprocessing, deformation field calculation, early warning model construction, early warning model training, monitoring and early warning steps. The specific contents are as follows: A. Data collection: Regularly collect surface images of key structural parts of overhanging rocks under external forces; B. Image preprocessing: geometric correction and noise removal of the acquired images; C. Deformation field calculation: The preprocessed images are analyzed using DIC technology to obtain the displacement field and strain field of all image sequences, which are then exported as a strain field data matrix. The strain field data matrix is ​​statistically analyzed to obtain the strain field mean μ and strain field standard deviation σ: ， ， Where: m and n are the total number of rows and columns of the strain field data matrix, respectively; ε ij is the first i Line j The corresponding strain values ​​of the columns; D. Early warning model construction: Establish an LSTM model including input layer, LSTM layer, fully connected layer and output layer, where the input layer receives the strain field mean μ and the strain field standard deviation σ The sliding window data is divided by time points. The LSTM layer extracts the features of each sliding window and captures the changing trend and potential mutation of the strain field. The fully connected layer outputs a single value, and the output layer outputs the window prediction result as a label of "normal" or "abnormal"; E. Early warning model training: Divide the sliding window data set of existing overhang rock instability into a training set and a validation set, and ensure that the validation set contains some abnormal points. Then input the training set and the validation set into the LSTM model, use the labels to supervise the training process of the LSTM model, monitor the loss value and accuracy of the LSTM model during the training process, and optimize the performance of the LSTM model by adjusting the hyperparameters of the LSTM model; F. Monitoring and early warning: Receive the strain field mean μ and strain field standard deviation obtained in step C σ The normalized data is then divided into multiple sliding windows according to the time points. The data of each sliding window is then input into the trained LSTM model. The LSTM model captures the changing trend and potential mutation of the strain field, and then outputs the prediction results of each sliding window and determines whether the sliding window is "normal" or "abnormal". When the LSTM model continuously determines that the window is "abnormal", an early warning is triggered immediately.

2. The overhang rock instability early warning method based on DIC and LSTM according to claim 1 is characterized by: The specific process of step A is as follows: A1. Select monitoring area: Select the key structural parts of overhanging rock including dangerous rock mass, main controlling joint surface and bedrock as monitoring area; A2. Monitoring equipment installation: Install high-resolution camera equipment and set up light sources around the monitoring area of ​​the overhanging rock; A3. Image acquisition: Through high-resolution camera equipment, based on the stability of the monitored overhanging rock mass, surface images of the overhanging rock in the monitoring area under the action of external forces are regularly collected.

3. The overhang rock instability early warning method based on DIC and LSTM according to claim 1 is characterized by: In step B, perspective distortion and / or lens distortion correction is performed on the captured image, and then image noise is removed by Gaussian filtering or median filtering.

4. The overhanging rock instability early warning method based on DIC and LSTM according to claim 1 is characterized by: The specific process of step C is as follows: C1. Image input: input the preprocessed image into Ncorr software; C2. Data matrix: Ncorr software uses the first image or the designated image in the preprocessed image as the reference image, and the remaining images as the deformed images, and compares them with the reference image one by one. Ncorr software calculates the displacement of each pixel by identifying the changes in the surface pattern of the overhanging rock in the image, and then generates a two-dimensional displacement field reflecting the movement of the overhanging rock surface at different time points; then Ncorr software analyzes the gradient change of the displacement in the displacement field, obtains the strain field composed of the strain values ​​of each pixel point, and then saves the strain field data as a strain field matrix at different time points; C3. Matrix calculation: Statistical analysis is performed on the strain field data matrix to obtain the strain field mean μ and the strain field standard deviation σ : ， ， Where: m and n are the total number of rows and columns of the strain field data matrix, respectively; ε ij is the first i Line j The corresponding strain values ​​of the column.

5. The overhang rock instability early warning method based on DIC and LSTM according to claim 1 is characterized by: In the step D, the LSTM model includes 2 to 3 LSTM layers, and each LSTM layer contains 100 to 150 neurons; the output layer uses the Sigmoid activation function to map the output to a probability between 0 and 1, and then uses a binary classification method to output the window prediction result as a label of "normal" or "abnormal"; wherein the Sigmoid activation function expression is: ， Where: x is the linear output of the LSTM model; y It is the probability value calculated by the Sigmoid function, ranging from 0 to 1; e is the base of natural logarithms; The binary classification method is to convert the probability value output by the Sigmoid function y With the default value k Compare, when y ≥ k When the prediction result is "abnormal", the output prediction result is "normal".

6. The overhang rock instability early warning method based on DIC and LSTM according to claim 1 is characterized by: In the step E, the sliding window data set of existing overhang rock instability is divided into k subsets, k-1 subsets are used as training sets, and the remaining 1 subset is used as a validation set, and it is ensured that the validation set contains some abnormal points. The LSTM model is trained and validated alternately using the aforementioned training set and validation set, and finally the loss value and accuracy are calculated to evaluate the performance of the LSTM model. Subsequently, the performance of the LSTM model is optimized by adjusting the hyperparameters of the LSTM model including the number of LSTM layers, the number of neurons, the learning rate and / or regularization according to the performance.

7. The overhanging rock instability early warning method based on DIC and LSTM according to claim 6 is characterized by: The loss value is obtained using the binary cross entropy loss function: ， Where: L is the average loss value; N is the total number of samples; y i For the i The true label of each sample (0 or 1); y i ' It is i The probability of a sample being predicted by the LSTM model is between 0 and 1. The greater the deviation between the prediction result of the LSTM model and the true label, the greater the loss value, thereby guiding the LSTM model to adjust parameters to gradually reduce the error. The adjustment of the number of LSTM layers and the number of neurons is to optimize the number of LSTM layers and the number of neurons in each layer through grid search or random search to find the optimal complexity of the LSTM model; the learning rate adjustment is to dynamically adjust the learning rate using a learning rate decay or an adaptive learning rate optimization algorithm to avoid gradient explosion or disappearance during training; the regularization introduces a regularization method to randomly discard a part of neurons to prevent the LSTM model from overfitting.

8. The overhang rock instability early warning method based on DIC and LSTM according to any one of claims 1 to 7, characterized in that: In the F step, the normalization process is performed using Min-Max normalization, which converts the strain field mean μ and the strain field standard deviation σ All are scaled to between 0 and 1 to avoid the impact of too large or too small values ​​on the model; ， Where: x i is the mean or standard deviation in the received data; x min and x max are the minimum and maximum values ​​of the group of data respectively; x ΄ is the normalized data, with a range of [0,1]; Dividing the entire time series data into multiple sliding windows is to take a predetermined number of consecutive time points starting from the first time point. p The corresponding data is used as the first window; then the window is slid backward from the first time point q At a certain point in time, q The time point is taken as the first time point, and the consecutive time points are taken p The corresponding data forms the second window; repeat the above process until the entire time series is divided into multiple sliding windows; among them, q < p .

9. The overhanging rock instability early warning method based on DIC and LSTM according to claim 8 is characterized by: In the F step, the LSTM model extracts the strain field mean μ and strain field standard deviation corresponding to each sliding window from the input data of each sliding window. σ Two features ， Form a window length p × two-dimensional feature matrix of two features; then set the strain field mean threshold and strain field standard deviation threshold, use the LSTM model to compare the values ​​of the two-dimensional feature matrix with the threshold one by one, and capture the change trend and potential mutation of the strain field; finally output the prediction result of each sliding window and judge the sliding window as "normal" or "abnormal", and trigger an early warning immediately when the continuous judgment window is "abnormal"; The threshold is set according to the data characteristics, and a strain field mean threshold and a strain field standard deviation threshold are set respectively. Then the LSTM model compares the prediction results of each sliding window with the corresponding thresholds. When the prediction results exceed their respective thresholds, the sliding window is output as an "abnormal" label, otherwise the sliding window is output as a "normal" label.

Images