CN112149502A - Unfavorable geology positioning forecasting method based on convolutional neural network - Google Patents

Abstract

The present invention provides a method for predicting poor geological location based on convolutional neural network, comprising: firstly establishing a geophysical detection image data set; then building a target prediction and positioning neural network model based on feature extraction, using the image data set, The target prediction and positioning neural network model is trained to obtain a trained target prediction and positioning neural network model; finally, the image result data obtained by a certain physical detection method is input into the trained target prediction and positioning neural network model, and the actual bad geology is carried out. Positioning prediction. The beneficial effects of the invention are: the location scale and nature state of the unfavorable geological bodies within the range passed in the construction process of underground engineering such as tunnels can be accurately predicted, the decision basis for engineering design and construction management is provided, and the existing geological prediction geophysical prospecting is reduced. It can improve the safety of engineering construction by solving the problems of low interpretability of the method, relying on expert experience, and low prediction accuracy.

Description

A prediction method of bad geological location based on convolutional neural network

technical field

The invention relates to the technical field of bad geological positioning prediction, in particular to a bad geological positioning forecasting method based on a convolutional neural network.

Background technique

In some actual engineering projects, we found that the poor geological conditions of the water-rich broken zone bring major risks to the tunnel construction. Currently, the geophysical prospecting method is used to carry out advanced geological forecasting of tunnels. The waveform data image detected by the sensor needs to be interpreted by experienced experts. In the prediction process, there are problems such as low interpretability of the measurement result map obtained by geophysical sensing, relying on expert experience, and low prediction accuracy. It is very important to accurately predict and locate a certain unfavorable geology (such as the water-rich fracture zone) in the advance forecast image.

With the rapid development of Deep Learning in recent years, Convolutional Neural Networks (CNN, Convolutional Neural Networks), through its features of feature sampling, weight sharing, and operational dimensionality reduction, are used in the fields of image classification, target detection, and image understanding. has been widely used. This method is based on the deep convolutional neural network to automatically identify and probabilize the detection images of geophysical methods (geo-radar, TSP, transient electromagnetic, etc.) Interpretability and accuracy of the unfavorable geology of the water-rich fracture zone.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention provides a method for predicting poor geological location based on a convolutional neural network, which mainly includes the following steps:

This application uses an example of an unfavorable geological advance forecast of the water-rich fracture zone, and this method can be applied to geophysical image data to achieve accurate prediction of the water-rich fracture zone or other bad geology.

A method for predicting poor geological location based on convolutional neural network, comprising the following steps:

S101: establish a geophysical detection image data set;

S102: construct a target prediction and positioning neural network model based on feature extraction;

S103: Using the image data set, train the target prediction and positioning neural network model to obtain a trained target prediction and positioning neural network model;

S104: Input the image result data obtained by a geophysical method into the trained neural network model for target prediction and positioning, and perform actual bad geological positioning prediction.

Further, in step S101, a geophysical detection image data set is established; specifically, it includes:

S201: Establish a preliminary image data set of geophysical detection results through advanced geological prediction case collection and actual on-site project data collection; the preliminary image data set includes a plurality of result images obtained from advanced geological prediction based on the geophysical method;

S202: Predict the abnormal geology on each image in the preliminary image data set according to the prediction mechanism of the fractured zone of the water-rich zone and its image characteristics;

S203: Combine the prediction conclusion, and use the expert experience method to re-determine the position of the water-rich crushing zone on each image in the preliminary image data set, and use the location area of the water-rich crushing zone as the label content to perform data analysis on each image in the preliminary image data set Annotate, get the initial image dataset after annotation;

S204: Using a variety of image data augmentation methods, expand and augment the marked preliminary image data set, and couple with some geophysical methods to detect negative sample data of the water-rich broken zone to obtain a final image data set.

Further, in step S201, the geophysical method includes TSP, geological radar, transient electromagnetic, etc.; the sample size in the preliminary image data set is greater than 500 images.

Further, in step S102, the target prediction and positioning neural network model based on feature extraction includes a feature extraction basic network and a positioning prediction network that are connected in sequence;

Wherein, the feature extraction basic network includes sequentially connected: CBR block, maximum pooling layer, BaseRN1 layer, 2 BaseRN0 layers, BaseRN1 layer, 3 BaseRN0 layers, BaseRN1 layer, 5 BaseRN0 layers;

The localization prediction network includes a max pooling layer, a fully connected layer and a sigmoid layer that are connected in sequence.

Further, the CBR block includes a convolutional base layer, a regularization layer and an activation layer connected in sequence; wherein, C represents the convolution layer, B represents the regularization layer Batch Norm, and R represents the activation layer, using the Leaky-ReLU activation function.

Further, both the BaseRN1 layer and the BaseRN0 layer are designed based on the residual network idea, and the residual term is introduced to facilitate the construction of a deep network model; among them, the backbones in the BaseRN0 layer and the BaseRN1 layer are CBR-CBR-CB modules, and the BaseRN0 layer and BaseRN1 The parameters of the backbone CBR-CBR-CB modules in the layer are the same, specifically: the filter size of the first CBR block is 1×1, 64 channels; the filter size of the second CBR block is 3×3, 64 channels. ; The filter size of the third CB block is 1 × 1, 256 channels; the filter size of the branch CB block in the BaseRN0 layer is consistent with the filter size of the backbone CB block, and the filter size of the branch CBR block in BaseRN1 is the same as that of the backbone. The filter size of the first CBR block is the same; the CB block includes a convolutional base layer and a regularization layer connected in sequence.

Further, in step S103, in the feature extraction basic network, the output feature map of the feature extraction layer is down-sampled by a factor of 16 in four levels, and a compromise is made between the spatial resolution and the extracted feature strength, and the activation layer Leaky - The final output of ReLU is 14×14×1024; the feature extraction network uses four stages to downsample the input image by 2 times to achieve 16 times downsampling: the details are as follows:

The first stage: input a three-channel RGB image with a size of W×H×C, and enter a CBR block in the first stage; where W=256 is the width of the image, H=256 is the height of the image, and C=3 Represents three channels of the image; the filter size in this CBR block is 1×1, 64 channels, after the CBR block is downsampled once, the output image size is 128×128×64;

The second stage: first pass a max pooling layer with a filter of 3 × 3 and stride 2, and then pass through a BaseRN1 layer and 2 BaseRN0 layers; the size of the output image after passing through the second stage is 64 × 64 × 256;

The third stage: Input the output of the second stage to 1 BaseRN1 layer and 3 BaseRN0 layers connected in sequence, and obtain a deeper network model structure through the combination of the BaseRN0 layer and the BaseRN1 layer, and the output graph size is 32×32× 512;

The fourth stage: Input the output of the third stage to 1 BaseRN1 layer and 5 BaseRN0 layers connected in sequence, and further obtain a deeper network model structure through the combination of BaseRN0 and BaseRN1, and the output image size is 16×16×1024 .

Further, in the localization prediction network, the output of the feature extraction network is input into the localization prediction network, the maximum pooling layer is used to effectively reduce the offset effect of the feature estimation mean caused by the model parameter error, and finally the fully connected layer and the sigmoid layer are connected. Perform regression positioning, and realize simultaneous detection of position and classification through the design of multi-task loss function.

Further, in step S103, during training, in order to achieve accurate positioning prediction for a certain unfavorable geology, a multi-task coupling loss function is designed, and the sigmoid regression method is finally used in the positioning prediction network to realize the probability confidence judgment and positioning of the unfavorable geology. Area data output; specific:

The design of the loss function is aimed at the two tasks of judging the confidence of bad geological target prediction and target positioning, and correspondingly constructing the target loss function, among which:

Target prediction confidence: the target prediction confidence loss function L _conf (o,p), which represents the probability that the predicted target area is the bad geology to be monitored, and is represented by binary cross-entropy loss:

表示第i个预测框中是否存在目标的概率，通过对目标预测定位神经网络模型输出的结果p_i进行sigmoid求值得到，

属于(0,1)；i＝1,2,…,I，I为预测的结果框个数；L_conf(o,p)的值受预测框个数和值的影响，范围不定，L_conf(o,p)越小，损失越小说明预测结果越准确；Among them, o _i ∈ {0,1} indicates whether there is a target in the i-th prediction box, 0 indicates no existence, and 1 indicates existence;

Indicates the probability of whether there is a target in the i-th prediction box, which is obtained by sigmoid evaluation of the result p _i output by the target prediction and positioning neural network model,

Belongs to (0,1); i=1,2,...,I, I is the number of predicted result boxes; the value of L _conf (o,p) is affected by the number and value of predicted boxes, and the range is indeterminate, L _conf The smaller the (o, p), the smaller the loss, the more accurate the prediction result;

Target positioning: target positioning loss function L _loc (μ,σ), in order to improve the accuracy of the result positioning, the Gaussian distribution modeling is introduced into the predicted bounding box results (x, y, w, h) of the positioning box, and the results of each parameter are obtained. Mean and variance; Boundingbox is a rectangular box, (x, y) represents the position of its center point, and (w, h) represents its size;

The target localization loss function is represented by a negative log likelihood loss (NLL, negative log likelihood loss):

in,

Among them, W×H is the width and height of the input image, and the input image is divided into grids of H rows and W columns; I is the number of prediction frames, and the i, j, and k indices get the i-th row of the k-th row and the j-th column. Prediction frame and its base coordinates; γ _t is a hyperparameter that affects the weight of (x, y, w, h) prediction results, and is a hyperparameter for model training. As the training gradually approaches a stable value, the training results can be effectively guaranteed accuracy; G _ijk represents the true value {x, y, w, h} _truth of the result of the i-th prediction frame at the pixel position of the k-th row and the j-th column in the H×W coordinate system, satisfying the mean value μ , the normal distribution G～N(μ,σ ² ) with variance σ; μ _t ,σ _t , t∈{x,y,w,h} is the model prediction result, which is transformed to the (0,1) range by the sigmoid function , represents the center coordinates in the prediction frame of the current t=i, j, k index, μ _w , μ _h represent the length and width of the rectangular frame of the prediction area with μ _x and μ _y as the center points; at the same time, the variance σ is used _t , t∈{x,y,w,h} represents the reliability of the prediction result, 0 means reliable, 1 means unreliable;

Synthesize L _conf (o,p) and L _loc (μ,σ) to obtain a multi-task loss function, and couple the two loss functions with weight factors λ ₁ , λ ₂ to obtain a comprehensive bad geological location prediction loss function L(o,p, μ,σ):

L(o,p,μ,σ)＝λ ₁ L _conf (o,p)+λ ₂ L _loc (μ,σ)

Among them, the two weighting factors λ ₁ and λ ₂ are set according to specific task requirements.

Further, the image data set constructed in S101 belongs to the small sample data set; in step S103 , the target prediction and positioning neural network model is trained by the transfer learning method to improve the training speed.

The beneficial effects brought about by the technical solution provided by the present invention are: the technical solution proposed in the present application can accurately predict the location scale and nature state of the unfavorable geological bodies within the range passed during the construction of underground projects such as tunnels, which is beneficial for engineering design and construction. The construction management part provides the basis for decision-making, reduces the problems of low interpretability of existing geophysical prediction methods, relies on expert experience, and low prediction accuracy, and improves the safety of engineering construction.

Description of drawings

The present invention will be further described below in conjunction with the accompanying drawings and embodiments, in which:

1 is a flowchart of a method for predicting poor geological location based on a convolutional neural network in an embodiment of the present invention;

FIG. 2 is a structural diagram of a target prediction and positioning neural network in an embodiment of the present invention.

Detailed ways

In order to have a clearer understanding of the technical features, objects and effects of the present invention, the specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

Embodiments of the present invention provide a method for predicting poor geological location based on a convolutional neural network.

Please refer to FIG. 1. FIG. 1 is a flowchart of a method for predicting poor geological location based on a convolutional neural network in an embodiment of the present invention, which specifically includes the following steps:

S101: establish a geophysical detection image data set;

S102: construct a target prediction and positioning neural network model based on feature extraction;

S103: Using the image data set, train the target prediction and positioning neural network model to obtain a trained target prediction and positioning neural network model;

S104: Input the image result data obtained by a geophysical method into the trained neural network model for target prediction and positioning, and perform actual bad geological positioning prediction.

In order to solve the problems of over-fitting of training and low prediction accuracy caused by the small sample size of the detection data in the advance geological prediction of the geophysical method, the sample expansion of the data set is realized based on the image data augmentation method.

In step S101, a geophysical detection image data set is established; specifically, it includes:

S201: Establish a preliminary image data set of geophysical detection results through advanced geological prediction case collection and actual field project data collection; the preliminary image data set includes advanced geological prediction based on geophysical methods (TSP, geological radar, transient electromagnetic, etc.). The obtained multiple result images are collected and arranged to obtain a sample set, which is the original data to be processed; the sample capacity in the preliminary image data set is greater than 500 images;

S202: Predict the abnormal geology on each image in the preliminary image data set according to the prediction mechanism of the water-rich fracture zone and its image characteristics; and guide the network structure design of the feature extraction model;

S203: Combine the prediction conclusion, and use the expert experience method to re-determine the position of the water-rich crushing zone on each image in the preliminary image data set, and use the location area of the water-rich crushing zone as the label content to perform data analysis on each image in the preliminary image data set Annotate, get the initial image dataset after annotation;

S204: Using a variety of image data augmentation methods to expand and expand the labeled preliminary image data set, and couple with some geophysical methods to detect the negative sample data of the water-rich fractured zone (that is, the detection result indicates that the water-rich fractured zone is not included) data) to obtain the final image data set; the image data augmentation method includes: image mirroring, translation, scaling, rotation, cropping, Gaussian noise, etc.;

Among them, the image data in the final image data set is more than 1000; according to the ratio of 8:2, it is divided into a training data set and a test data set, and the construction of the image data set is completed.

Aiming at the difficulty of identifying and classifying unfavorable geological types in the complex geological environment of tunnel engineering, based on the feature extraction method of deep learning, a neural network model with the function of target prediction and positioning is designed to determine whether there is a water-rich fracture zone in the image.

According to the graphic imagery mechanism of various bad geology on the detection data, the appropriate target prediction and positioning network model structure is selected and designed. Based on a general principle, deeper networks can learn more complex functions and input representations, so that the model performance can be improved. Better, this method chooses to introduce the ResNet residual network concept to design the basic module, and design the super-deep convolutional network model structure.

Please refer to FIG. 2. FIG. 2 is a structural diagram of a target prediction and positioning neural network in an embodiment of the present invention; in step S102, the feature extraction-based target prediction and positioning neural network model includes a feature extraction basic network and a positioning prediction network that are sequentially connected. network;

Wherein, the feature extraction basic network includes sequentially connected: CBR block, maximum pooling layer, BaseRN1 layer, 2 BaseRN0 layers, BaseRN1 layer, 3 BaseRN0 layers, BaseRN1 layer, 5 BaseRN0 layers;

Wherein, the CBR block includes a convolutional base layer, a regularization layer and an activation layer connected in sequence; wherein, C represents a convolutional layer, B represents a regularization layer Batch Norm, and R represents an activation layer; in the embodiment of the present invention, Leaky -ReLU activation function; avoid the problem of silent neurons in the negative interval of the traditional ReLU activation function;

Both the BaseRN1 layer and the BaseRN0 layer are designed based on the residual network idea, and the residual term is introduced to facilitate the construction of a deep network model; among them, the backbone in the BaseRN0 layer and the BaseRN1 layer are CBR-CBR-CB modules, and the backbone in the BaseRN0 layer and the BaseRN1 layer The CBR-CBR-CB module parameters are the same, specifically: the filter size of the first CBR block is 1×1, 64 channels; the filter size of the second CBR block is 3×3, 64 channels; the third The filter size of each CB block is 1×1, 256 channels; the filter size of the branch CB block in the BaseRN0 layer is the same as the filter size of the backbone CB block, and the filter size of the branch CBR block in BaseRN1 is the same as the first one of the backbone. The filter size of each CBR block is the same; the CB block includes a volume base layer and a regularization layer connected in sequence;

The localization prediction network includes a max pooling layer, a fully connected layer and a sigmoid layer that are connected in sequence.

In step S103, in the feature extraction basic network, the output feature map of the feature extraction layer is down-sampled in four levels by a factor of 16, and a compromise is made between the spatial resolution and the extracted feature strength, and the activation layer Leaky-ReLU The final output is 14×14×1024; the feature extraction network uses four stages to downsample the input image by 2 times to achieve 16 times downsampling: the details are as follows:

The first stage: input a three-channel RGB image with a size of W×H×C, and enter a CBR block in the first stage, as shown in Figure 2; where W=256 is the width of the image, and H=256 is the image The height of C=3 represents three channels of the image; the filter size in this CBR block is 1×1, 64 channels, after the CBR block is downsampled once, the output image size is 128×128×64;

The second stage: first pass a max pooling layer with a filter of 3 × 3 and stride 2, and then pass through a BaseRN1 layer and 2 BaseRN0 layers; the size of the output image after passing through the second stage is 64 × 64 × 256;

The BaseRN0 layer and the BaseRN1 layer are both designed based on the idea of the residual network, which is an improved and upgraded version of the ResNet residual module; the BaseRN0 layer and the BaseRN1 layer are connected to the leaky-ReLU activation function, which is a general processing method;

The third stage: Input the output of the second stage to 1 BaseRN1 layer and 3 BaseRN0 layers connected in sequence, and obtain a deeper network model structure through the combination of the BaseRN0 layer and the BaseRN1 layer, and the output graph size is 32×32× 512;

The fourth stage: Input the output of the third stage to 1 BaseRN1 layer and 5 BaseRN0 layers connected in sequence, and further obtain a deeper network model structure through different combinations of BaseRN0 and BaseRN1, and the output image size is 16×16× 1024;

Among them, in the third and fourth stages, the model depth can also be improved by adding or reducing the BaseRNO layer and the BaseRN1 layer according to the needs.

In the localization prediction network, the output of the feature extraction network is input into the localization prediction network, and the maximum pooling layer is used to effectively reduce the offset of the feature estimation mean caused by the model parameter error. Finally, the fully connected layer and the sigmoid layer are connected for regression positioning. , the simultaneous detection of prediction and positioning is realized through the design of multi-task loss function; prediction is to judge whether there is bad geology to be detected; positioning is to locate the area of bad geology if there is such bad geology;

In step S103, during training, in order to realize the accurate positioning prediction of a certain unfavorable geology, a multi-task coupling loss function is designed, and the sigmoid regression method is finally used in the positioning prediction network to realize the probability confidence judgment of the unfavorable geology and the data output of the location area. ;specific:

The design of the loss function is aimed at the two tasks of judging the confidence of bad geological target prediction and target positioning, and correspondingly constructing the target loss function, among which:

Target prediction confidence: the target prediction confidence loss function L _conf (o,p), which represents the probability that the predicted target area is the bad geology to be monitored, and is represented by binary cross-entropy loss:

表示第i个预测框中是否存在目标的概率，通过对目标预测定位神经网络模型输出的结果p_i进行sigmoid求值得到，

属于(0,1)；i＝1,2,…,I，I为预测框个数；L_conf(o,p)的值受预测的结果框个数和值的影响，范围不定，L_conf(o,p)越小，损失越小说明预测结果越准确；Among them, o _i ∈ {0,1} indicates whether there is a target in the i-th prediction box, 0 indicates no existence, and 1 indicates existence;

Indicates the probability of whether there is a target in the i-th prediction box, which is obtained by sigmoid evaluation of the result p _i output by the target prediction and positioning neural network model,

Belongs to (0,1); i=1,2,...,I, I is the number of prediction boxes; the value of L _conf (o,p) is affected by the number and value of the predicted result boxes, the range is indeterminate, L _conf The smaller the (o, p), the smaller the loss, the more accurate the prediction result;

与方差

通过对均值、方差都训练得到损失最小的结果实现预测精度的提升，作为输出特征张量中定位区域的表征，这种方式不仅有区域定位，同时对定位的不确定性进行了评价，得到的结果是定位最精确的结果；Boundingbox为一个长方形框，(x,y)表示其中心点位置，(w,h)表示其大小；Target positioning: target positioning loss function L _loc (μ,σ), in order to improve the accuracy of the result positioning, the Gaussian distribution modeling is introduced into the predicted bounding box results (x, y, w, h) of the positioning box, and the results of each parameter are obtained. mean

with variance

The prediction accuracy is improved by training the mean and variance to obtain the result with the smallest loss, which is used as the representation of the positioning area in the output feature tensor. This method not only has regional positioning, but also evaluates the uncertainty of positioning. The result is The most accurate result of positioning; Boundingbox is a rectangular box, (x, y) represents the position of its center point, and (w, h) represents its size;

The target localization loss function is represented by a negative log likelihood loss (NLL, negative log likelihood loss):

in,

Among them, W×H is the width and height of the input image, and the input image is divided into grids of H rows and W columns; I is the number of prediction frames,

Since the result of each prediction is multiple target prediction frames, j∈[1,W] represents the column, k∈[1,H] represents the row index to the corresponding grid, and i∈[1,I] represents the grid For the i-th prediction frame area of grid prediction, the i-th prediction frame and its base coordinates of the k-th row and the j-th column are obtained through the i, j, and k indices. There are multiple scales of prediction frames on each grid to make the result. approach;

γ _t is a hyperparameter that affects the weight of (x, y, w, h) prediction results, and is a hyperparameter for model training. As the training gradually approaches a stable value, the accuracy of the training results can be effectively guaranteed; G _ijk represents In the H×W coordinate system, the true value {x,y,w,h} _truth of the result of the i-th prediction frame at the pixel point position of the k-th row and the j-th column satisfies a normal distribution with a mean value of μ and a variance of σ G～N(μ,σ ² ); μ _t , σ _t , t∈{x,y,w,h} are the model prediction results, μ _x , μ _y are transformed into the range of (0,1) by the sigmoid function, Represents the center coordinates in the prediction frame of the current t=kji index, μ _w , μ _h represent the length and width of the prediction frame with μ _x and μ _y as the center points;

At the same time, the variance σ _t , t∈{x,y,w,h} is used to represent the reliability of the prediction results, 0 means reliable, 1 means unreliable, and the solution process is as follows:

t₁∈{x,y}

t ₁ ∈{x,y}

t₂∈{w,h}

t ₂ ∈{w,h}

t₃∈{x,y,w,h}

t ₃ ∈{x,y,w,h}

The multi-task loss function of target positioning and target prediction confidence probability is comprehensively obtained, and the combined two loss functions are coupled by weight factors λ ₁ , λ ₂ to obtain a comprehensive bad geological positioning prediction loss function L(o,p,μ,σ):

L(o,p,μ,σ)＝λ ₁ L _conf (o,p)+λ ₂ L _loc (μ,σ)

Among them, the two weight factors λ ₁ , λ ₂ are set according to the specific task requirements. In some scenarios, it is necessary to prioritize whether the target is abnormal geological conditions and its probability accuracy, then the weight of λ ₁ is increased; some task scenarios focus on the location of abnormal geological conditions Prediction accuracy, the λ ₂ weight is improved.

The image data set constructed in S101 belongs to the small sample data set; in step S103, the training speed is improved by the transfer learning method; the specific steps are:

S301: first pre-train the original target prediction and positioning neural network model based on the large sample data set A; obtain the target prediction and positioning neural network model A after preliminary training; the large sample data set can use ImageNet, etc.;

S302: Freeze the parameters of the first t layers of the model A, and use the training data set in the image data set as the small sample data set B to perform migration training on the model A; obtain the trained target prediction and positioning neural network model B; wherein, t belongs to (1, L), which are preset values based on experience;

S303: Use the test data set in the image data set to test the trained target prediction and positioning neural network model; and determine whether the test meets the requirements; if so, use the current target prediction and positioning neural network model as the trained neural network model Target prediction and positioning neural network model; otherwise, use the training data set to retrain the parameters of the last t to L layers of the model B, and feed back the training according to the training data set label to obtain all the parameters of the model B; at the same time, the model can be fine-tuned. parameters to optimize the results until the trained target prediction and positioning neural network model B can meet the task requirements.

Fine-tuning Fine-tune actually trains all parameters of the original target prediction and positioning neural network model on the large-sample ImageNet data set, and then freezes the parameters of the first t layer of A. The parameters of the first t layer do not need to be retrained; then use The small-sample training data set constructed is used to train the parameters of the t-L layer of A to obtain the parameters of the t-L layer; according to the results of the test evaluation (the evaluation threshold can be specifically set according to the accuracy of the task requirements), it is judged whether the parameters are well trained, Whether the model can meet the task requirements, if not, repeat the above steps for training, and adjust parameters such as the number of frozen layers to optimize the results.

The training process requires a training data set and a test data set. The training process is to repeat N epochs to continuously update the parameters until the result is optimal. In each training process, the model parameters are obtained by training the training data set first, and then pass the test. The data set evaluates the set of parameters, and then performs the training of the next epoch until the set N epoch training is completed or the evaluation results meet the preset requirements

In practical applications, input the image result data obtained by a geophysical method into the trained neural network model for target prediction and positioning, and automatically output whether there is a bad geological probability confidence level to be predicted and its location, reducing the need for professionals and professionals. Rely on expert experience to improve work efficiency.

The beneficial effects of the present invention include:

1) Constructed a geological advance forecast data set, which is the basis for deep learning and network model training;

2) Based on the residual network idea, an ultra-high-depth deep feature extraction network model is designed, which can detect the recessive and abstract information in the geological advance forecast image through the deep activation of the model, and complete the extraction of the corresponding bad geological features;

3) After the feature extraction is completed, a multi-task loss function is designed to realize the probability confidence judgment and regional positioning of a certain unfavorable geological type;

4) The method can be continuously upgraded and evolved with the continuous collection and expansion of image data sets, improving performance, detection accuracy, range, and so on.

The technical solution proposed in this application can accurately predict the location, scale and nature of unfavorable geological bodies within the scope of the construction of tunnels and other underground projects, provide decision-making basis for engineering design and construction management, and reduce existing geological prediction and geophysical exploration. It can improve the safety of engineering construction by solving the problems of low interpretability of the method, relying on expert experience, and low prediction accuracy.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (10)

1. A bad geological positioning forecasting method based on a convolutional neural network is characterized by comprising the following steps: the method comprises the following steps:

s101: establishing a geophysical prospecting image data set;

s102: constructing a target prediction positioning neural network model based on feature extraction;

s103: training the target prediction positioning neural network model by adopting the image data set to obtain a trained target prediction positioning neural network model;

s104: and inputting image result data obtained by a geophysical prospecting method into the trained target prediction positioning neural network model to perform actual unfavorable geological positioning prediction.

2. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 1, wherein: in the step S101, a geophysical prospecting image data set is established; the method specifically comprises the following steps:

s201: establishing a geophysical prospecting method detection result preliminary image data set through advanced geological forecast case collection or actual field project data acquisition; the preliminary image data set comprises a plurality of result images obtained by performing advanced geological forecast based on a geophysical prospecting method;

s202: forecasting abnormal geology on each image in the preliminary image data set according to a water-rich fractured zone forecasting mechanism and the imaging characteristics of the water-rich fractured zone forecasting mechanism;

s203: determining the position of a water-rich broken band on each image in the preliminary image data set again by adopting an expert experience method in combination with a forecast conclusion, and performing data annotation on each image in the preliminary image data set by taking a water-rich broken band positioning area as label content to obtain an annotated preliminary image data set;

s204: and adopting an image data augmentation method to perform capacity expansion augmentation on the marked preliminary image data set, and coupling a partial geophysical prospecting method to detect the sample data of the water-rich broken belt to obtain a final image data set.

3. The method for forecasting the unfavorable geologic localization of a convolutional neural network as claimed in claim 2, wherein: in step S201, the geophysical prospecting method includes TSP, geological radar, and transient electromagnetism; the preliminary image dataset has a sample capacity greater than 500 images.

4. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 1, wherein: in step S102, the target prediction positioning neural network model based on feature extraction includes a feature extraction basic network and a positioning prediction network that are connected in sequence;

wherein, the characteristic extraction basic network comprises the following connected in sequence: CBR block, max pooling layer, BaseRN1 layer, 2 BaseRN0 layers, BaseRN1 layers, 3 BaseRN0 layers, BaseRN1 layers, 5 BaseRN0 layers;

the positioning prediction network comprises a maximum pooling layer, a full connection layer and a Sigmoid layer which are sequentially connected.

5. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 4, wherein: the CBR block comprises a volume base layer, a regularization layer and an activation layer which are sequentially connected; wherein C represents a convolution layer, B represents a regularization layer Batch Norm, R represents an activation layer, and a Leaky-ReLU activation function is adopted.

6. The method for forecasting unfavorable geologic positioning based on a convolutional neural network as claimed in claim 5, wherein: the BaseRN1 layer and the BaseRN0 layer are designed based on a residual error network concept, and a residual error item is introduced, so that a depth network model is conveniently constructed; wherein, the trunks in the BaseRN0 layer and the BaseRN1 layer are both CBR-CBR-CB modules, and the parameters of the trunks CBR-CBR-CB modules in the BaseRN0 layer and the BaseRN1 layer are consistent, and the concrete steps are as follows: the filter size of the first CBR block is 1 × 1, 64 channels; the filter size of the second CBR block is 3 × 3, 64 channels; the filter size of the third CB block is 1 × 1, 256 channels; the filter size of the branched CB block in the BaseRN0 layer is consistent with the CB block filter size of the trunk, and the filter size of the branched CBR block in the BaseRN1 is consistent with the filter size of the first CBR block of the trunk; the CB block comprises a volume base layer and a regularization layer which are connected in sequence.

7. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 4, wherein: in step S103, in the feature extraction basic network, the output feature mapping of the feature extraction layer is subjected to four-level down-sampling by a factor of 16 times, a tradeoff is made between spatial resolution and extracted feature strength, and the final output of the active layer leak-ReLU is 14 × 14 × 1024; the feature extraction network performs 2-time down-sampling on the input image in four stages to realize 16-time down-sampling: the method comprises the following specific steps:

the first stage is as follows: inputting a three-channel RGB image with the size of W multiplied by H multiplied by C, and inputting the three-channel RGB image into a CBR block at a first stage; w-256 is the width of the image, H-256 is the height of the image, and C-3 represents three channels of the image; the filter size in the CBR block is 1 multiplied by 1, 64 channels are sampled by the CBR block for one time, and the size of the output image is 128 multiplied by 64;

and a second stage: first pass through a maximum pooling layer of 3 × 3 and stride 2 with a filter, then pass through a BaseRN1 layer and 2 BaseRN0 layers; the size of the image output after the second stage is 64 × 64 × 256;

and a third stage: inputting the output of the second stage to 1 BaseRN1 layer and 3 BaseRN0 layers which are connected in sequence, obtaining a deeper network model structure through the combination of the BaseRN0 layer and the BaseRN1 layer, and enabling the size of an output graph to be 32 multiplied by 512;

a fourth stage: the output of the third stage is input to 1 BaseRN1 layer and 5 BaseRN0 layers connected in sequence, a further deeper network model structure is obtained by the combination of BaseRN0 and BaseRN1, and the size of the output image is 16 × 16 × 1024.

8. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 4, wherein: in the positioning prediction network, the output result of the feature extraction network is input into the positioning prediction network, the offset influence of the feature estimation mean value caused by the parameter error of the model is effectively reduced through the maximum pooling layer, finally, the full-connection layer and the Sigmoid layer are connected for regression positioning, and the simultaneous detection of prediction and positioning is realized through the design of a multi-task loss function.

9. The method for forecasting unfavorable geologic positioning based on a convolutional neural network as claimed in claim 8, wherein: in the step S103, during training, in order to realize accurate positioning forecast of a certain unfavorable geology, a multitask coupling loss function is designed, and finally probability confidence judgment and positioning area data output of the unfavorable geology are realized through a sigmoid regression method in a positioning prediction network; specifically, the method comprises the following steps:

the design of the loss function aims at two tasks of judging the prediction confidence coefficient of the unfavorable geological target and positioning the target, and the target loss function is respectively and correspondingly constructed, wherein:

target prediction confidence: target prediction confidence loss function L_conf(o, p) representing the probability that the predicted target region is the unfavorable geology to be monitored, and adopting binary cross entropy loss characterization:

wherein o is_iE {0,1} represents whether a target exists in the ith prediction box, 0 represents absence and 1 represents existence;

representing the probability of whether the target exists in the ith prediction box, and locating the result p output by the neural network model through predicting the target_iThe sigmoid is evaluated to obtain the result,

belongs to (0, 1); i is 1,2, …, and I is the number of prediction frames; l is_confThe value of (o, p) is influenced by the number and value of the prediction frames, the range is not fixed, L_confThe smaller the (o, p), the smaller the loss, the more accurate the prediction result;

target positioning: target location loss function L_loc(mu, sigma), in order to improve the result positioning accuracy, Gaussian distribution modeling is introduced to the prediction bounding box result (x, y, w, h) of the positioning box to obtain the mean value and the variance of each parameter; the bounding box is a rectangular box, (x, y) represents the center position of the box, and (w, h) represents the size of the box;

the target positioning loss function is characterized by adopting negative log likelihood loss:

wherein,

w multiplied by H is the width and the height of an input image, and the input image is divided into H rows and W columns of grids; i is the number of the prediction frames, and the ith prediction frame and the base coordinate of the jth row and the jth column of the kth line are obtained by indexing I, j and k; gamma ray_tIs a hyper-parameter affecting the (x, y, w, h) prediction result weight; g_ijkRepresenting the real value { x, y, W, H } of the result of the ith prediction box at the position of the jth row pixel point on the kth line under the H multiplied by W coordinate system_truthA positive distribution G to N (mu, sigma) satisfying a mean value of mu and a variance of sigma²)；μ_t,σ_tAnd t ∈ { x, y, w, h } is the model prediction result, passing sigmThe oid function translates to a (0,1) range, representing the center coordinate, μ, within the current t ═ i, j, k indexed prediction box_w,μ_hIs expressed in μ_x,μ_yThe length and width of a rectangular frame of the prediction region which is a central point; using the square error amount sigma simultaneously_tThe t belongs to { x, y, w, h } representation prediction result reliability, 0 represents reliable, and 1 represents unreliable;

synthesis L_conf(o, p) and L_loc(mu, sigma) to obtain a multitask loss function by a weighting factor lambda₁,λ₂And coupling the two loss functions to obtain a comprehensive bad geology positioning forecast loss function L (o, p, mu, sigma):

L(o,p,μ,σ)＝λ₁L_conf(o,p)+λ₂L_loc(μ,σ)

wherein λ is₁,λ₂These two weighting factors are set according to the specific task requirements.

10. The method for forecasting unfavorable geologic positioning based on convolutional neural network as defined in claim 1, wherein: the image dataset constructed in S101 belongs to a small sample dataset; in step S103, the target prediction positioning neural network model is trained by a transfer learning method to increase the training speed.

Images