CN113112447A - Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network - Google Patents
Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network Download PDFInfo
- Publication number
- CN113112447A CN113112447A CN202010149085.5A CN202010149085A CN113112447A CN 113112447 A CN113112447 A CN 113112447A CN 202010149085 A CN202010149085 A CN 202010149085A CN 113112447 A CN113112447 A CN 113112447A
- Authority
- CN
- China
- Prior art keywords
- image
- surrounding rock
- neural network
- training
- layer
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 239000011435 rock Substances 0.000 title claims abstract description 61
- 238000000034 method Methods 0.000 title claims abstract description 45
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 57
- 238000013528 artificial neural network Methods 0.000 claims abstract description 22
- 238000012360 testing method Methods 0.000 claims abstract description 15
- 238000011156 evaluation Methods 0.000 claims abstract description 11
- 230000006870 function Effects 0.000 claims description 54
- 238000012545 processing Methods 0.000 claims description 35
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 12
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000010606 normalization Methods 0.000 claims description 10
- 238000007781 pre-processing Methods 0.000 claims description 10
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 8
- 230000004913 activation Effects 0.000 claims description 8
- 238000013434 data augmentation Methods 0.000 claims description 8
- 238000003491 array Methods 0.000 claims description 7
- 238000005457 optimization Methods 0.000 claims description 7
- 238000007477 logistic regression Methods 0.000 claims description 6
- 238000013519 translation Methods 0.000 claims description 6
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000012937 correction Methods 0.000 claims description 5
- 239000011159 matrix material Substances 0.000 claims description 5
- 230000003213 activating effect Effects 0.000 claims description 3
- 238000010801 machine learning Methods 0.000 claims description 3
- 230000002194 synthesizing effect Effects 0.000 claims description 3
- 239000002689 soil Substances 0.000 abstract description 5
- 238000002474 experimental method Methods 0.000 abstract description 3
- 206010063385 Intellectualisation Diseases 0.000 abstract 1
- 238000011176 pooling Methods 0.000 description 12
- 238000010276 construction Methods 0.000 description 5
- 238000013461 design Methods 0.000 description 5
- 239000000779 smoke Substances 0.000 description 4
- 238000010586 diagram Methods 0.000 description 3
- 210000002569 neuron Anatomy 0.000 description 3
- 238000012795 verification Methods 0.000 description 3
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 3
- 230000003321 amplification Effects 0.000 description 2
- 238000005422 blasting Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000011835 investigation Methods 0.000 description 2
- 238000003199 nucleic acid amplification method Methods 0.000 description 2
- 239000000047 product Substances 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000008094 contradictory effect Effects 0.000 description 1
- 239000000428 dust Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 239000013589 supplement Substances 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000009423 ventilation Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/40—Image enhancement or restoration using histogram techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/73—Deblurring; Sharpening
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The tunnel surrounding rock level intelligent judgment method based on the VGG convolutional neural network comprises the following steps: acquiring high-resolution image data of a tunnel face and constructing a sample library; randomly selecting samples from the sample library obtained in the step one as a training set, and taking the rest samples as a testing set; constructing a VGG deep convolution neural network, reducing loss function values through training, and updating network weight parameters to obtain learned network weight parameters; and step four, selecting the image needing to be subjected to surrounding rock grading as the input of the convolutional neural network, and returning a surrounding rock grading evaluation result corresponding to the tunnel face according to the obtained surrounding rock grade prediction. The method is based on the convolutional neural network to intelligently evaluate the grade of the surrounding rock, can realize the grade judgment of the surrounding rock without carrying out on-site or indoor rock-soil body mechanics experiments, can avoid the danger caused by contact acquisition of surrounding rock data information, and realizes the automation and the intellectualization of the grading work of the surrounding rock of the tunnel and the underground engineering.
Description
Technical Field
The invention relates to the field of investigation design of tunnels and underground engineering, is suitable for the investigation design of tunnels and underground engineering of various traffic, water conservancy and hydropower and the like, and particularly relates to an intelligent judgment method for tunnel surrounding rock level based on a VGG convolutional neural network.
Background
The tunnel is a cavern structure system built in various surrounding rock media with certain stress history and structural signs, the tunnel surrounding rock is rock-soil mass which influences the stability of the tunnel in a certain range around the tunnel, and the stratum around the tunnel, whether rock mass or soil mass, is called the tunnel surrounding rock. In order to meet the requirements of tunnel engineering construction, the stability degree of the surrounding rock is divided into a plurality of levels by using comprehensive indexes or single indexes, and the classification is called surrounding rock classification. The correct evaluation of the surrounding rock level of the tunnel not only relates to the design and construction scheme of the tunnel and the engineering cost of the tunnel, but also relates to the safety and stability of the tunnel during construction and operation. Therefore, the surrounding rock grading scheme for objectively evaluating rock mass is an important basis for tunnel design and construction stages.
Although relevant descriptions of surrounding rocks of all levels are given according to specifications, an accurate relation is not established between each index of the surrounding rock grading and the surrounding rock grade at present, and no clear limit exists between partial surrounding rock grading indexes, so that the tunnel surrounding rock grading is greatly influenced by subjective factors. The classification of the tunnel surrounding rock is a field with strong experience, and comprises various uncertain factors, and the classification of the tunnel surrounding rock is determined by experts with abundant engineering experience in many cases.
At present, tunnel surrounding rock grading mainly depends on-site collection of quantitative and qualitative indexes and is selected according to a standard requirement mode. The field information acquisition also has the defects of high acquisition difficulty, large workload, large error, low efficiency and poor safety in the acquisition process. Therefore, it is significant to research a method for extracting tunnel surrounding rock features based on images (non-contact type) and automatically grading the surrounding rocks, so that the accuracy and the safety of surrounding rock grading can be ensured to a great extent.
Disclosure of Invention
In order to overcome the technical defects in the prior art, the invention discloses an intelligent tunnel surrounding rock level judgment method based on a VGG convolutional neural network.
The invention discloses an intelligent tunnel surrounding rock grade judging method based on a VGG convolutional neural network, which comprises the following steps of:
acquiring high-resolution image data of a tunnel face and constructing a sample library;
randomly selecting samples with the quantity of more than 70% in proportion from the sample library obtained in the step one as a training set, taking the rest samples as a test set, and expanding the sample set by adopting a data augmentation method to ensure that the sample set meets the quantity of the samples required by the subsequent neural network training;
constructing a VGG deep convolution neural network for learning training, continuously reducing loss function values and updating network weight parameters through training, and obtaining the neural network which updates the network weight parameters after learning after training for multiple times;
selecting an image needing surrounding rock grading as the input of the convolutional neural network, iterating the convolutional neural network obtained in the step three, applying network weight parameters obtained by training and learning of a corresponding surrounding rock grading training set, and outputting prediction of corresponding grades of the surrounding rocks; and returning the surrounding rock grading evaluation result corresponding to the tunnel face of the tunnel according to the obtained surrounding rock grade prediction.
Preferably, the first step further comprises a step of preprocessing the image; the specific operation steps are as follows:
decomposing R, G, B three pixel channels of the image into single channels, respectively carrying out histogram equalization on the three channels, and synthesizing the results after equalization processing to obtain the equalized image.
Preferably, the data augmentation in step 2 may introduce an imputilis image processing packet to process the image, and the specific steps are as follows:
calling a path _ images () function of the packet to search all images of the sample library obtained in the first step, capturing image paths, listing all files, saving the paths into variable image paths, capturing the image paths, loading each image into a memory, initializing data and labels arrays in an image processing packet, circularly traversing the image paths, and adjusting the original pixel intensity to a range [0,1] to complete image data preprocessing;
reading in image data by using an imread () interface of an image processing pack cv2 module, modifying the image size to 224 x 224 by using a resize () interface, converting the image into an array by using an img _ to _ array () function, and storing the converted array into a data array;
extracting class labels from the image path, updating a label list to finish analysis of the multiple classes of labels, adding label names into label arrays of an image processing package, importing the label names into a machine learning library scinit-lean library, and finishing label binaryzation of the label arrays by using a LabelBinarizer () function;
importing a train _ test _ split () data packet from a sklean.model _ selection function, and transmitting a data set data and a label set labels into the train _ test _ split () data packet as parameters;
and dividing the image data into a training set and a testing set and performing data augmentation.
Preferably, in the third step, the convolution neural network performs linear operation on each layer of the result
Wherein wk [l]K-th convolution kernel, d, representing the l-th layer1Number of convolution kernels for layer l, b[l]Corresponding layer deviation of the l-th layer; a is[l-1]Output data representing the upper layer, i.e., (l-1) th layer;
activating the linear operation result by applying an activation function g to obtain an input layer a of the next layer[l]Output characteristics of the l-th layer a[l]Can be expressed as
a[l]=g(zl)
Preferably, in step three, the loss function equation used in the logistic regression is:
loss function:
the corresponding cost function equation is:
cost function
Wherein Y isjIs (the jth predictor),representing the output value of the jth training sample, m is the number of input values, and w and b represent different convolution kernels and deviations;
the logistic regression gradient descent calculation formula is as follows:
y is a predicted value and is a predicted value,representing the output value of the training sample.
Preferably, in the third step, the method for keeping the same distribution of the inputs of each layer of neural network by using the intermediate value of a hidden layer in the neural network specifically includes the following steps:
Z(i)for each of the input values, the value of the input value,
m is the number of samples in each run (batch);
μ is the calculated mean;
σ2is the calculated variance;
gamma is a scale factor used for adjusting the numerical value;
β is a translation factor for increasing the offset.
Preferably, in the training process of the third step, the loss function value is reduced and the network weight parameter is updated by an Adam algorithm, which specifically includes:
initialization parameters of the Adam optimization algorithm:
Vdw=0,Sdw=0,Vdb=0,Sdb=0;
at the t iteration:
Vdw=β1Vdw+(1-β1)dw,Vdb=β1Vdb+(1-β1)db
Sdw=β2Sdw+(1-β2)(dw)2,Sdb=β2Sdb+(1-β2)(db)2;
t number of iterations; an alpha learning rate;
β1: first momentAn exponential decay rate of the array estimate; t represents the attenuation rate under different iteration times t;
β2: the exponential decay rate of the second order matrix estimation; t represents the attenuation rate under different iteration times t;
e is as follows: to prevent a stable parameter with a denominator of 0;
w: a weight;
b: an offset amount;
dw: a weight derivative;
db: a derivative of the offset;
(dw)2the representation is the square of the gradient of the weight w, also called the square of the differential;
Vdm、Vdb、Sdb、Sdwis an intermediate variable used to hold an exponentially weighted average (moving average).
Preferably, in the training process of the third step, the deviation correction is considered to obtain a corrected parameter:
Vdm、Vdb、Sdb、Sdwthe addition of the superscript CORRECTED represents the CORRECTED parameter.
According to the method, the intelligent evaluation of the grade of the tunnel surrounding rock is realized based on the digital image and the convolutional neural network, the grade judgment of the surrounding rock can be realized without performing on-site or indoor rock-soil body mechanics experiments, and the danger caused by the contact type acquisition of surrounding rock data information can be avoided; the method can be well combined with various optimization problems, has strong universality and has certain adaptability to uncertain information in the image. The method is economical, practical, simple to operate and suitable for grade judgment and exploration design work of surrounding rocks of tunnel engineering of various traffic, water conservancy and hydropower and the like.
Drawings
Fig. 1 is a schematic flow chart of an embodiment of the determination method according to the present invention.
FIG. 2 is a schematic diagram of image comparison before image processing of an exemplary tunnel face and after image processing using the present invention;
fig. 3 is a schematic diagram of image comparison before image processing of another exemplary tunnel face and after image processing by the present invention.
FIG. 4 is a schematic diagram of an embodiment of the VGG16 convolutional neural network according to the present invention;
in fig. 4, CONV denotes a convolutional layer, POOL denotes that this layer is also a pooling layer, FC denotes a fully connected layer, 3X3 for each layer denotes the convolution kernel size, and the numbers 64,128,256,512 of each convolutional layer denote the number of corresponding convolutional layer convolution kernels, respectively; SIZE represents the number of input image data of several layers below, and the number in the fully connected layer represents the number of neurons in the fully connected layer.
Detailed Description
The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.
The invention discloses an intelligent tunnel surrounding rock grade judging method based on a VGG convolutional neural network, which comprises the following steps of:
step one, acquiring high-resolution image data of a tunnel face and constructing a sample library.
The image acquisition in the first step may be acquired according to the following method.
The method comprises the steps of constructing data sets according to different shooting contents, shooting time, shooting environments and shooting modes, and subdividing data according to different categories of the above elements.
The shooting content is divided by the distance from the camera to the tunnel face and the display area during shooting: the image data which can completely shoot the full-face image of the face under the condition of no amplification is taken as the full-face shooting type data of the face, and the image data which can embody partial characteristics of the face such as different lithological transition parts, fault fracture zones, joint cracks, water outlet parts and the like under the condition of no amplification is taken as the partial shooting type data of the face.
The shooting time is classified according to the time after normal ventilation after blasting, and each shooting is carried out at a certain interval of time, such as 10 min; the shooting environment is classified into factors such as external environment brightness and illumination direction during shooting, and the shooting mode is classified into an operating mode of the shooting equipment, such as image data under the condition that a flash lamp is not turned on in a normal mode and under the condition of different iso values.
And secondly, carrying out convolutional neural network training on different data sets constructed according to the step I, selecting a data set combination with the optimal training accuracy according to a training result, and further determining an image acquisition standard. The tunnel face full-view shooting type data and one or more combinations of partial local shooting type data can be used for subsequent surrounding rock classification.
For example, when it is found that the training accuracy of the data set of the fault fracture zone in the ISO ═ 200 mode is optimal after training 20 minutes after the day blasting, the shooting element of the data set is used as the image acquisition standard.
In the first step, image preprocessing can be carried out on the image by combining field data information; the image smoke and dust removing processing can be carried out on the image by using a defogging algorithm based on dark channel processing.
The image preprocessing technology can also perform color image histogram equalization on the image based on an image enhancement method, and the specific operation steps are as follows:
decomposing R, G, B three pixel channels of the image into single channels, respectively carrying out histogram equalization on the three channels, and synthesizing the results after equalization processing to obtain the equalized image. The purpose of doing so is to eliminate the shooting error, improve sample accuracy.
And carrying out image smoke removing treatment on the image smoke by using a defogging algorithm based on dark channel treatment.
Step two, randomly selecting 80% from the sample library obtained in the step one as a training set, selecting the rest 20% as a test set, and expanding the sample set by adopting a data augmentation method to meet the sample quantity required by the subsequent neural network training; the data augmentation method includes but is not limited to image processing modes such as random rotation, offset, magnification and flipping.
And b, introducing an imitils image processing package to process the image, calling a path _ images () function of the package to search all images of the sample library obtained in the step one, capturing an image path, listing all files, storing the path into a variable imagagepath, capturing the image path, loading each image into a memory, initializing data and labels arrays in the image processing package, circularly traversing the imagagepaths, and adjusting the original pixel intensity to a range [0,1] to complete image data preprocessing.
Then reading in image data by using an imread () interface of the image processing package cv2 module, and modifying the image size to be the same by using a resize () interface224*224Utilizing img _ to _ array ()Function(s)And converting the picture into an array, and storing the converted array into the data array, so that the next calculation is facilitated.
Extracting class labels from the image path, updating a label list to finish analysis of the multiple classes of labels, adding label names into a label array of an image processing package, importing the label names into a machine learning library scinit-lean library, and finishing label binarization of the label array by using a LabelBinarizer () function. A "model _ selection" function is imported from a sklean.model _ selection function, where the "train _ test _ split" function is a function that randomly partitions a training set provided in a processing packet, and a data set data and a tag set labels are passed as parameters into the "train _ test _ split" function.
The data is divided into training and test sets, and the parameter test _ size is usually set to 0.2, i.e. 80% of the data is used for training and the remaining 20% is used for testing. Image generator imagedata generator () is imported from a keras frame image file (the file is the image file in the keras frame data preprocessing and data supplement module preprocessing), and the generator is the imagedata generator of the keras frame image generator, and is used for enhancing data in the batch, expanding the size of a data set, and enhancing the generalization capability of a model, such as rotation, deformation, normalization and the like.
An image generator for data expansion is constructed, and specific parameter settings are as follows: the angle rotation _ range of the picture random flip 25, the width of the picture random horizontal shift width _ shift _ range 0.1, the width of the picture random vertical shift height _ shift _ range 0.1, the shear strength shear _ range 0.2, the random zoom-in zoom _ range 0.1, the random horizontal flip value horizontal _ flip True, and the parameter file _ mode 'nearest' may be set, in the case of an imputis image processing packet, the file _ mode may be set to one of 'constant', 'nearest', 'refresh' or 'nearest', and points beyond the boundary when the transform is performed will be processed according to the method given by the present parameters.
After all the parameters are set, the operation function expands the data of the sample base, so that the quantity of the sample base meets the requirement of neural network training.
Constructing a VGG (visual Geometry group) neural network for learning and training, continuously reducing loss function values and updating network weight parameters through training, and obtaining the neural network which updates the network weight parameters after learning after training for multiple times; the convolutional neural network is alternately implemented by a plurality of convolutional layers and pooling layers.
The two most important parts of the Convolutional neural network are a Convolutional layer and a pooling layer, and the core of the Convolutional layer (CONV layer) is a convolution operation. For all or most of the pixels in the image, the pixels in the local window with the pixel as the center and the convolution kernel are subjected to inner product operation, and the calculation result is used as a new value of the pixel. Traversing all or most of pixels in the image, performing the inner product operation, adding deviation, applying an activation function to activate, completing one-time filtering, and obtaining an output feature map with the same size as the original image or smaller than the original image.
The specific calculation process is as follows:
the current I layer input data is output data a of the (l-1) layer of the previous layer[l-1]Performing a convolution operation on the input data, wk [l]The kth convolution kernel representing the l-th layer, which has a total of d1A convolution kernel, and taking into account the corresponding layer deviation b of the l-th layer[l]The linear operation result of the layer can be expressed as:
activating the linear operation result by applying an activation function g to obtain an input layer a of the next layer[l]Output characteristics of the l-th layer a[l]Can be expressed as
a[l]=g(zl)
A coiled Linear Units layer (ReLU layer) is used as an activation function used after each layer of Linear operation, so that the nonlinear characteristic of the convolutional neural network is enhanced while the convolutional layer is unchanged, and the formula is as follows:
f (x) max (0, x); the formula carries out nonlinear processing on input data, and negative values in the input data are all replaced by 0.
Reducing the block of the input feature map by using a Pooling Layer (Pooling Layer), and combining the feature activation; the calculation speed is improved, the robustness of the extracted features is improved, and the noise influence is reduced.
A VGG16 neural network was built, the VGG16 neural network comprising 13 convolutional layers, 3 fully-connected layers, and 5 pooling layers, as shown in fig. 4.
The construction method of the neural network provides that an RGB image with the image input size of 224 x 224 is input, 3x3 convolution kernels are used in convolution layers, the step size of each convolution layer is set to be 1 pixel, and the filling of the 3x3 convolution layers is set to be zero padding (same); the pooling layer adopts a maximum pool (max pooling), the pooling layers have 5 layers, the convolutional layers are inserted into the pooling layers at intervals, each pooling layer is arranged behind a plurality of convolutional layers, and the window of the max-pooling is 3x 3; the convolutional layer is followed by three fully-connected layers (FC layers). 4096 neurons are arranged in the first two full connection layers, 1024 neurons are arranged in the third full connection layer, and the full connection layers of all the networks are configured identically; classification was performed after full connectivity using logistic regression (Softmax) classification using the Loss function equation (Loss function) of:
loss function:
the corresponding Cost function equation (Cost function) is:
cost function
In the above formula, m is the number of input values; w, b represent different weights and biases.
The logistic regression gradient descent calculation formula is as follows:
all hidden layers (located in the middle of each convolutional layer (conv layer)) of the VGG16 network use a ReLU function as an activation function, so that the nonlinearity of a decision function is improved; after activation, normalization processing is performed using the Batchnormalization () function to solve the sample Point Change (Covariate Shift) problem.
The hidden layer is other layers except the input layer and the output layer; the method for realizing that the input of each layer of neural network keeps the same distribution (namely BatchNorm) in the middle value of a certain hidden layer in the neural network is concretely as follows:
Z(i)For each of the input values, the value of the input value,
m in the above-mentioned mean and variance calculation formula is batch size, i.e. the number of samples in each batch (batch);
μ is the calculated mean;
σ2is the calculated variance;
for normalization of the results of the processing, the purpose of normalization is: and (4) the given data is subjected to division and unitization, and the data of the same type is divided by the sum of the data to obtain the proportion of the data in the sum.
E is to prevent variance from being 0 and ensure the stability of the value.
Gamma is a scale factor used for adjusting the numerical value;
β is a translation factor for increasing the offset.
Because the normalized input is basically limited under normal distribution, the expression capacity of the network is reduced, and in a multi-index evaluation system, because the properties of each evaluation index are different, the evaluation indexes generally have different dimensions and orders of magnitude. When the levels of the indexes are greatly different, if the original index values are directly used for analysis, the function of the indexes with higher numerical values in the comprehensive analysis is highlighted, and the function of the indexes with lower numerical levels is relatively weakened.
The normalization process changes the number into a decimal number between (0, 1), makes each feature contribute to the same result, and improves the convergence rate and the accuracy of the model by using the normalization. However, after normalization processing, the range of the original index with a higher numerical value is the same as that of the original index with a lower numerical value, and the range characteristic that each index is different from other indexes is not highlighted. Therefore, in order to solve the problem of insufficient expression capacity, a scale factor gamma and a translation factor beta are introduced, each index is reduced to the original level, and the expression capacity is recovered, wherein the scale factor gamma and the translation factor beta are parameters which can be continuously updated and learned in training.
And a regularization processing layer (dropout layer) is adopted in the middle of the full-connection layer, and the forgetting rate can be 0.25, so that the occurrence of overfitting is prevented.
Inputting the data set image calibration classes prepared in the second step into a designed convolutional neural network for training, and reducing a loss function value and updating a network weight parameter through an Adam efficient optimization algorithm, wherein the Adam optimization algorithm is formed by combining a Momentum gradient descent algorithm and an RMSprop gradient descent algorithm.
Initialization parameters of the Adam optimization algorithm:
Vdw=0,Sdw=0,Vdb=0,Sdb=0。
at the t iteration:
Vdw=β1Vdw+(1-β1)dw,Vdb=β1Vdb+(1-β1)db
Sdw=β2Sdw+(1-β2)(dw)2,Sdb=β2Sdb+(1-β2)(db)2
after the deviation correction is considered, finally obtaining the corrected parameters
The deviation of the Adam algorithm corrects each latest data value and depends on previous data results, but at the moment, a plurality of calculation results in the initial stage have larger deviation with a real average value, namely, the calculation results in the initial stage of iteration have larger deviation with the real value, so that the deviation correction can be carried out by using the formula, and as the iteration number t increases, beta t is closer to 0, and the correction effect on the early stage is better.
In the formula:
t being the time step, i.e. the number of iterations
α: the learning rate, also referred to as a step factor, controls the update rate of the weights. A larger value of the learning rate (e.g., 0.3) results in faster initial learning before the learning rate is updated, while a smaller value (e.g., 1.0E-5) results in the training converging on better performance.
β1: exponential decay rate of the first order matrix estimate; t represents the attenuation rate under different iteration times t;
β2: the exponential decay rate of the second order matrix estimation; t represents the attenuation rate under different iteration times t;
e is as follows: to prevent a stable parameter with a denominator of 0;
w: a weight;
b: an offset amount;
dw: a derivative of the weight;
db: a derivative of the offset;
(dw)2expressed is the square of the gradient of the weight w, also called the square of the differential;
Vdm、Vdb、Sdb、Sdwis a medium variable for storing an exponentially weighted average (moving average), and the superscript correct represents the CORRECTED parameter, the medium parameter being set for updating the weight w and the offset b.
In this embodiment, the following parameter settings may be used:
α: taking 0.015;
β1: taking a default value of 0.9;
β2: value 0.9999;
epsilon: get 10-8;
One specific embodiment is as follows: setting 100 rounds (epoch) of neural network training, setting an initial learning rate alpha to be 0.015, monitoring a value of val _ loss by using a callback function, setting a learning rate attenuation factor to be 0.5, setting a value of a learning rate reduction trigger to be 3, triggering learning rate attenuation when the performance of a model is not improved after 3 rounds of passing, setting a lower limit of the learning rate to be 0.00001, setting a Batch size to be 32, setting an image size to be 224 × 3, and adopting a binary cross entropy loss function of binary _ cross entropy as a loss function. And after multiple times of training, obtaining the learned network weight parameters.
As can be seen from the pre-and post-processing images obtained in fig. 2 and 3; if the influence factors in the image acquisition process before processing are excessive, such as smoke, insufficient exposure or overexposure of a flash lamp, the shot image can not accurately reflect the tunnel face characteristics, and the image processing method can remove the influence of the external environment on the image quality as much as possible after the image processing, restore and reproduce the real image of the tunnel face and improve the sample quality.
Step four: and selecting the image needing surrounding rock grading as the input of the convolutional neural network, iterating the convolutional neural network obtained in the step three, applying the network weight parameters obtained by training and learning of the corresponding surrounding rock grading training set, and outputting corresponding grading evaluation of the surrounding rock. And returning the surrounding rock grading evaluation result corresponding to the tunnel face of the tunnel according to the obtained surrounding rock grading evaluation.
The invention has the beneficial effects that:
(1) the surrounding rock classification is carried out by adopting non-contact image data acquisition, so that the working efficiency is improved, and the operation danger caused by contact acquisition of surrounding rock data information is avoided;
(2) the method does not need on-site or indoor rock-soil body mechanical experiments, is economical and practical, simple to operate, less in parameter setting, high in convergence rate, capable of being well combined with various optimization problems, high in universality and capable of having certain adaptability to uncertain information in the problems.
Taking a picture of a tunnel face of a double-hole tunnel of a certain expressway as an example, image processing, model training and surrounding rock grade prediction are carried out according to the flow of the figure 1. A total of 360 images were acquired, wherein 125 images of the v-level surrounding rock, 156 images of the iv-level surrounding rock, 79 images of the iii-level surrounding rock, the total image size was 423.36MB, 80% of the images were used for model training, 20% of the images were used for model verification, the read image size was 224 × 224 and contained three channels, the training was 100 rounds (epoch), the initial learning rate was set to 0.015, and the Batch size (Batch size) was set to 4. The training results are shown in the following table:
the training results are shown in the following table:
number of rounds | Loss function value of training set | Training set accuracy | Validation set loss function values | Verification set accuracy |
Epoch 1/100 | loss:1.4601 | acc:0.5891 | val_loss:4.6479 | val_acc:0.6157 |
|
loss:1.1364 | acc:0.5914 | val_loss:2.3311 | val_acc:0.6389 |
Epoch 3/100 | loss:0.9668 | acc:0.6331 | val_loss:1.4895 | val_acc:0.5880 |
Epoch 4/100 | loss:0.9634 | acc:0.6065 | val_loss:1.6413 | val_acc:0.6528 |
Epoch 5/100 | loss:1.0030 | acc:0.5903 | val_loss:0.6723 | val_acc:0.7546 |
… | … | … | … | … |
Epoch 96/100 | loss:0.6224 | acc:0.6678 | val_loss:0.5332 | val_acc:0.6898 |
Epoch 97/100 | loss:0.6077 | acc:0.6690 | val_loss:0.5125 | val_acc:0.7037 |
Epoch 98/100 | loss:0.6866 | acc:0.6412 | val_loss:0.5077 | val_acc:0.6991 |
Epoch 99/100 | loss:0.6720 | acc:0.6574 | val_loss:0.5092 | val_acc:0.6898 |
Epoch100/100 | loss:0.6911 | acc:0.6389 | val_loss:0.5093 | val_acc:0.7037 |
In this embodiment, the accuracy of the training set reaches 63.89%, and the accuracy of the verification set reaches 70.37%.
The foregoing is a description of preferred embodiments of the present invention, and the preferred embodiments in each of the preferred embodiments are used in any combination, if not explicitly contradictory or prerequisite for a particular preferred embodiment, and can be stacked, the examples and specific parameters in the examples are merely for the purpose of clearly illustrating the inventor's process of verifying the invention, the invention is not limited to the protection scope of the patent, the English in the parentheses of the present invention is the English name of Chinese in the computer language or function before the parentheses, and the English in the parentheses is the name of the mathematical function or database without strict Chinese translation.
Claims (8)
1. The tunnel surrounding rock grade intelligent judgment method based on the VGG convolutional neural network is characterized by comprising the following steps of:
acquiring high-resolution image data of a tunnel face and constructing a sample library;
randomly selecting samples with the quantity of more than 70% in proportion from the sample library obtained in the step one as a training set, taking the rest samples as a test set, and expanding the sample set by adopting a data augmentation method to ensure that the sample set meets the quantity of the samples required by the subsequent neural network training;
constructing a VGG deep convolution neural network for learning training, continuously reducing loss function values and updating network weight parameters through training, and obtaining the neural network which updates the network weight parameters after learning after training for multiple times;
selecting an image needing surrounding rock grading as the input of the convolutional neural network, iterating the convolutional neural network obtained in the step three, applying network weight parameters obtained by training and learning of a corresponding surrounding rock grading training set, and outputting prediction of corresponding grades of the surrounding rocks; and returning the surrounding rock grading evaluation result corresponding to the tunnel face of the tunnel according to the obtained surrounding rock grade prediction.
2. The intelligent decision method according to claim 1, wherein the first step further comprises the step of preprocessing the image; the specific operation steps are as follows:
decomposing R, G, B three pixel channels of the image into single channels, respectively carrying out histogram equalization on the three channels, and synthesizing the results after equalization processing to obtain the equalized image.
3. The intelligent judgment method according to claim 1, wherein the data augmentation in the step 2 can introduce imputilis image processing packets to process the image, and the specific steps are as follows:
calling a path _ images () function of the packet to search all images of the sample library obtained in the first step, capturing image paths, listing all files, saving the paths into variable image paths, capturing the image paths, loading each image into a memory, initializing data and labels arrays in an image processing packet, circularly traversing the image paths, and adjusting the original pixel intensity to a range [0,1] to complete image data preprocessing;
reading in image data by using an imread () interface of an image processing pack cv2 module, modifying the image size to 224 x 224 by using a resize () interface, converting the image into an array by using an img _ to _ array () function, and storing the converted array into a data array;
extracting class labels from the image path, updating a label list to finish analysis of the multiple classes of labels, adding label names into label arrays of an image processing package, importing the label names into a machine learning library scinit-lean library, and finishing label binaryzation of the label arrays by using a LabelBinarizer () function;
importing a train _ test _ split () data packet from a sklean.model _ selection function, and transmitting a data set data and a label set labels into the train _ test _ split () data packet as parameters;
and dividing the image data into a training set and a testing set and performing data augmentation.
4. The intelligent decision method of claim 1, wherein in step three, the convolutional neural network performs linear operation on each layer
Wherein wk [l]K-th convolution kernel, d, representing the l-th layer1Number of convolution kernels for layer l, b[l]Corresponding layer deviation of the l-th layer; a is[l-1]Output data representing the upper layer, i.e., (l-1) th layer;
activating the linear operation result by applying an activation function g to obtain an input layer a of the next layer[l]Output characteristics of the l-th layer a[l]Can be expressed as
a[l]=g(zl)。
5. The intelligent decision method according to claim 4, wherein in step three, the logistic regression uses a loss function equation of:
loss function:
the corresponding cost function equation is:
cost function
Wherein Y isjIs (the jth predictor),representing the output value of the jth training sample, m is the number of input values, and w and b represent different convolution kernels and deviations;
the logistic regression gradient descent calculation formula is as follows:
6. The intelligent decision method according to claim 4, wherein in step three, the method for keeping the same distribution of the inputs of each layer of neural network is implemented by using the intermediate values of a hidden layer in the neural network, specifically as follows:
Z(i)for each of the input values, the value of the input value,
m is the number of samples in each run (batch);
μ is the calculated mean;
σ2is the calculated variance;
gamma is a scale factor used for adjusting the numerical value;
β is a translation factor for increasing the offset.
7. The intelligent judgment method according to claim 1, wherein in the training process of the third step, the loss function value is reduced and the network weight parameter is updated by an Adam algorithm, specifically:
initialization parameters of the Adam optimization algorithm:
Vdw=0,Sdw=0,Vdb=0,Sdb=0;
at the t iteration:
Vdw=β1Vdw+(1-β1)dw,Vdb=β1Vdb+(1-β1)db
Sdw=β2Sdw+(1-β2)(dw)2,Sdb=β2Sdb+(1-β2)(db)2;
t number of iterations; an alpha learning rate;
β1: exponential decay rate of the first order matrix estimate; t represents the attenuation rate under different iteration times t;
β2: the exponential decay rate of the second order matrix estimation; t represents the attenuation rate under different iteration times t;
e is as follows: to prevent a stable parameter with a denominator of 0;
w: a weight;
b: an offset amount;
dw: a weight derivative;
db: a derivative of the offset;
(dw)2the representation is the square of the gradient of the weight w, also called the square of the differential;
Vdm、Vdb、Sdb、Sdwis an intermediate variable used to hold an exponentially weighted average (moving average).
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010149085.5A CN113112447A (en) | 2020-03-05 | 2020-03-05 | Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010149085.5A CN113112447A (en) | 2020-03-05 | 2020-03-05 | Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113112447A true CN113112447A (en) | 2021-07-13 |
Family
ID=76709144
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010149085.5A Pending CN113112447A (en) | 2020-03-05 | 2020-03-05 | Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113112447A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114140448A (en) * | 2021-12-06 | 2022-03-04 | 广西路信云信息技术有限公司 | Tunnel face surrounding rock intelligent grading method and device based on deep learning |
CN114547721A (en) * | 2021-12-21 | 2022-05-27 | 沈阳工业大学 | Blasting control method for differential directional fracture of local area of deep-buried tunnel |
CN116452511A (en) * | 2023-03-17 | 2023-07-18 | 西南交通大学 | Intelligent identifying method, device and medium for surrounding rock level of tunnel face of drilling and blasting method |
CN117077027A (en) * | 2023-07-24 | 2023-11-17 | 西南交通大学 | Surrounding rock sub-level grading method and device based on intelligent grading model grading probability |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635618A (en) * | 2018-08-07 | 2019-04-16 | 南京航空航天大学 | Visible images vein developing method based on convolutional neural networks |
CN110082283A (en) * | 2019-05-23 | 2019-08-02 | 山东科技大学 | A kind of Atmospheric particulates SEM image recognition methods and system |
-
2020
- 2020-03-05 CN CN202010149085.5A patent/CN113112447A/en active Pending
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109635618A (en) * | 2018-08-07 | 2019-04-16 | 南京航空航天大学 | Visible images vein developing method based on convolutional neural networks |
CN110082283A (en) * | 2019-05-23 | 2019-08-02 | 山东科技大学 | A kind of Atmospheric particulates SEM image recognition methods and system |
Non-Patent Citations (2)
Title |
---|
ZENRRAN: "简单理解Momentum,RMSprop,Adam优化算法", 《CSDN》 * |
李汪石: "公路隧道围岩云分级系统的开发与应用", 《中国优秀硕士学位论文全文数据库 (工程科技Ⅱ辑)》 * |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114140448A (en) * | 2021-12-06 | 2022-03-04 | 广西路信云信息技术有限公司 | Tunnel face surrounding rock intelligent grading method and device based on deep learning |
CN114547721A (en) * | 2021-12-21 | 2022-05-27 | 沈阳工业大学 | Blasting control method for differential directional fracture of local area of deep-buried tunnel |
CN114547721B (en) * | 2021-12-21 | 2022-10-25 | 沈阳工业大学 | Blasting control method for differential directional fracture of local area of deep-buried tunnel |
CN116452511A (en) * | 2023-03-17 | 2023-07-18 | 西南交通大学 | Intelligent identifying method, device and medium for surrounding rock level of tunnel face of drilling and blasting method |
CN116452511B (en) * | 2023-03-17 | 2024-05-03 | 西南交通大学 | Intelligent identifying method, device and medium for surrounding rock level of tunnel face of drilling and blasting method |
CN117077027A (en) * | 2023-07-24 | 2023-11-17 | 西南交通大学 | Surrounding rock sub-level grading method and device based on intelligent grading model grading probability |
CN117077027B (en) * | 2023-07-24 | 2024-03-15 | 西南交通大学 | Surrounding rock sub-level grading method and device based on intelligent grading model grading probability |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113112446A (en) | Tunnel surrounding rock level intelligent judgment method based on residual convolutional neural network | |
CN113112447A (en) | Tunnel surrounding rock grade intelligent determination method based on VGG convolutional neural network | |
CN110728654B (en) | Automatic pipeline detection and classification method based on deep residual error neural network | |
CN107169956B (en) | Color woven fabric defect detection method based on convolutional neural network | |
CN110097554B (en) | Retina blood vessel segmentation method based on dense convolution and depth separable convolution | |
CN111507884A (en) | Self-adaptive image steganalysis method and system based on deep convolutional neural network | |
CN112036513B (en) | Image anomaly detection method based on memory-enhanced potential spatial autoregression | |
CN108846413B (en) | Zero sample learning method based on global semantic consensus network | |
CN111275660B (en) | Flat panel display defect detection method and device | |
CN110991511A (en) | Sunflower crop seed sorting method based on deep convolutional neural network | |
CN110751644B (en) | Road surface crack detection method | |
CN112381763A (en) | Surface defect detection method | |
CN110930378B (en) | Emphysema image processing method and system based on low data demand | |
CN113971764B (en) | Remote sensing image small target detection method based on improvement YOLOv3 | |
CN113297929A (en) | Convolutional neural network microseismic monitoring waveform identification method based on whole-process visualization | |
CN113989257A (en) | Electric power comprehensive pipe gallery settlement crack identification method based on artificial intelligence technology | |
CN113362277A (en) | Workpiece surface defect detection and segmentation method based on deep learning | |
CN111145145A (en) | Image surface defect detection method based on MobileNet | |
CN110991247B (en) | Electronic component identification method based on deep learning and NCA fusion | |
CN114511521B (en) | Tire flaw detection method based on multiple representation and multiple sub-field self-adaption | |
CN114647760A (en) | Intelligent video image retrieval method based on neural network self-temperature cause and knowledge conduction mechanism | |
CN116206227B (en) | Picture examination system and method for 5G rich media information, electronic equipment and medium | |
CN117911437A (en) | Buckwheat grain adhesion segmentation method for improving YOLOv x | |
CN112541916A (en) | Waste plastic image segmentation method based on dense connection | |
CN115346091B (en) | Method and device for generating Mura defect image data set |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210713 |
|
RJ01 | Rejection of invention patent application after publication |