Rock-soil characteristic parameter prediction method based on non-uniform normalization and spectrum enhancement deep fusion network

Abstract

The invention provides a rock-soil characteristic parameter prediction method based on a non-uniform normalization and spectrum strengthening depth fusion network, which comprises the following steps: step 1, performing data processing on rock-soil characteristic parameters by using a non-uniform normalization NUN method, the rock-soil characteristic parameters comprising a foundation bearing capacity characteristic value, a compression modulus and a rock-soil body state; changing the geotechnical characteristic parameters from a text type to an image type to obtain image data of the geotechnical characteristic parameters; step 2, training a spectrum enhanced deep fusion network SEDFN by using the image data obtained in the step 1; and step 3, predicting an actual geological profile map of a certain project by using the SEDFN trained in the step 2, and evaluating a prediction result by using a similarity index SI. According to the invention, the accuracy, reliability and precision of investigation work can be improved.

Description

Technical Field

[0001] The present invention relates to the field of geotechnical engineering, and in particular to a method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network. Background Art

[0002] Engineering geological survey is an indispensable and important part of modern engineering construction. It provides a scientific basis for the design and construction of the project by conducting detailed investigations and studies on geological conditions, topography, and geomorphology, so as to ensure the safety and stability of the project. In actual operation, geological survey work faces many challenges, one of which is how to conduct effective surveys in the presence of existing buildings. In this case, traditional methods often find it difficult to obtain accurate and complete geotechnical survey data in the target area, which brings uncertainty and risks to subsequent design and construction. When there are buildings above the target survey area, traditional direct survey methods such as drilling and sampling are often restricted, making it difficult to obtain the true situation of the underground structure in the area. At this time, geotechnical engineers usually make inferences based on their own experience and refer to the survey data of nearby areas. However, this method is highly subjective, and its results may deviate greatly from the actual situation, which in turn affects the safety and stability of the entire project.

[0003] In order to overcome this limitation, some new technologies and methods have emerged in recent years, such as non-invasive survey technology, remote sensing technology, and machine learning technology, which can make up for the shortcomings of traditional methods to a certain extent and improve the accuracy and reliability of survey work. However, in terms of improving the accuracy of engineering geological surveys, especially when facing target areas covered by buildings, efficient and reliable methods are still rare. This means that when such situations occur in actual projects, engineers have to rely on experience and judgment, and combine data from surrounding areas to make predictions. However, this method that relies on experience and intuitive judgment is very subjective and may lead to deviations in the assessment of geological conditions, which in turn affects the safety and stability of the project. Summary of the invention

[0004] In view of this, the present invention provides a method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network, which can improve the accuracy, reliability and precision of survey work.

[0005] A method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network includes the following steps:

[0006] Step 1: Use the non-uniform normalized NUN method to process the geotechnical characteristic parameters, wherein the geotechnical characteristic parameters include the characteristic value of foundation bearing capacity, compression modulus and geotechnical body state; change the geotechnical characteristic parameters from text type to image type to obtain image data of the geotechnical characteristic parameters;

[0007] Step 2: Use the image data obtained in step 1 to train the spectrum enhancement deep fusion network SEDFN;

[0008] Step 3: Use the SEDFN network trained in step 2 to predict the actual geological profile of a project, and use the similarity index SI to evaluate the prediction results.

[0009] Furthermore, step 1 specifically includes:

[0010] Step 1.1, collect the original data of geotechnical engineering characteristic parameters, and select the corresponding data of three geotechnical characteristic parameters from the original data according to the engineering needs;

[0011] Step 1.2: Use the NUN-based data processing method to process the selected raw data:

[0012]

[0013] Among them, max L represents the peak pixel value of the RGB image, which is set to 255; N(c) represents the processed data, c represents the channels represented by different feature parameters, and f r represents the original geotechnical characteristic parameter data, r is the category parameter, which is used to distinguish different geotechnical characteristic parameters; if the target characteristic parameter is described in text type, the peak pixel value of the RGB image is divided equally according to the number of types, and the divided RGB value is assigned to the feature;

[0014] Step 1.3: Convert the data processed in step 1.2 into image data for visual analysis. Further, step 1.2 specifically includes:

[0015] Step 1.2.1, determine the category parameter r: adjust the r value through trial and error algorithm to find the best nonlinear transformation parameter to optimize data distribution; substitute the parameter value in the original data into the NUN formula, preset different r values ​​and calculate the parameter value after processing, compare the processing results under different r values, and select the optimal value of r according to the following requirements: the mean should be close to 122.5; the standard deviation should be as large as possible, the skewness should be as small as possible, and the kurtosis should be as negative as possible;

[0016] Step 1.2.2: Process the original data using the category parameter r value selected in step 1.2.1.

[0017] Furthermore, step 1.3 specifically includes:

[0018] Step 1.3.1, repeat the operation in step 1.2 three times to process the three geotechnical characteristic parameters;

[0019] Step 1.3.2: Correspond the three parameters of each rock and soil layer to the red, green and blue channels of the RGB pixel value in turn, and then generate the RGB pixel value corresponding to each rock and soil layer, and finally generate image data.

[0020] Furthermore, step 2 specifically includes:

[0021] Step 2.1, input image preprocessing: Convert the input image to Y'CbCr color space from RGB, and only retain the brightness channel Y':

[0022]

[0023] Among them, R, G, and B correspond to the three components of the pixel value of the RGB three-channel image;

[0024] Step 2.2: Use the modified SRCNN network mSRCNN to extract low-frequency features and train the network for low-frequency processing; use a two-layer mSRCNN convolutional network to perform feature extraction and nonlinear mapping on the input image as a training set, and then perform image reconstruction:

[0025]

[0026] The first convolution kernel is W1, the bias term is B1, the last convolution kernel is W2, the bias term is B2, and X is the input block of the previous layer. is an intermediate process function used to represent the content after feature extraction. Leaky ReLU is a nonlinear mapping activation function used to control the output of negative input. is the output result of the first layer, The final output contains low-frequency features.

[0027] Step 2.3: Use frequency domain high-preserving fast Fourier convolution (FHP-FFC) to extract high-frequency features and train the network to process high-frequency features;

[0028] Step 2.4, parallel fusion of high and low frequency features: Integrate low frequency and high frequency features, and fuse them through a 1×1 convolution kernel, and finally perform an inverse transformation from Y' to RGB to obtain the final RGB image;

[0029] Step 2.5: Compare the result obtained in step 2.4 with the actual situation, calculate the difference between the network training result and the actual situation (i.e., calculate the loss function), and use the back propagation algorithm to calculate the gradient of the loss function with respect to each weight; repeat steps 2.1 to 2.4 until the preset number of training iterations is reached.

[0030] Furthermore, step 2.3 specifically includes:

[0031] Step 2.3.1. Perform fast Fourier convolution on the brightness channel ' obtained in step 2.1:

[0032]

[0033] where F(k,l) is the complex value at position (k,l) in the frequency domain after transformation, including the frequency and The amplitude and phase of the frequency component of , f(m,n) is the pixel value of the image at position (m,n) in the spatial domain, M and N are the width and height of the image respectively, and i is the imaginary unit;

[0034] Step 2.3.2: Separate the real part and the imaginary part in step 2.3.1, and separate the real part and the imaginary part of the frequency domain information into two channels, so that the information of the two channels exists in the real number domain:

[0035]

[0036] in, represents the real part after FFT processing, represents the processed imaginary part;

[0037] Step 2.3.3: Retain high-frequency information by weighted means: First, determine the high- and low-frequency allocation thresholds; then, assign a high-frequency weight coefficient greater than or equal to 0.75 to the high-frequency information, so as to ensure that this step can mainly process the high-frequency information; at the same time, assign a low-frequency information weight coefficient greater than or equal to 0.1 and less than or equal to 0.25 to the low-frequency information:

[0038] T c =Q a％ (F abs,c )

[0039] Among them, T c Indicates the threshold for distinguishing high and low frequencies on the c channel, Q a％ represents the operation of calculating the percentile of a, F abs,c is the absolute value set of all frequency amplitudes on the cth channel, according to the determined threshold T c To assign corresponding weights to each frequency component, if the absolute value of a frequency component is greater than T c, then the frequency component is considered to be a high-frequency component, otherwise it is a low-frequency component:

[0040]

[0041] Among them, w ij is the element of the output weight tensor, k and l represent its position coordinates, w l is the corresponding weight of the low-frequency component, w h is the weight corresponding to the high frequency component, F abs,kl is the absolute value of the frequency at the corresponding position, T c is the threshold of the cth channel; the weight determined above is applied to the processed spectrum to enhance the high-frequency features in the frequency domain information:

[0042]

[0043] in, is the weighted frequency domain tensor, k and l represent its position information, represents the frequency domain features that separate the real and imaginary parts, and W(k,l) is the weight determined in the previous part of the formula;

[0044] Step 2.3.4, high-frequency feature learning and inverse fast Fourier transform: Based on the weighted frequency domain information in step 2.3.3, the learning of high-frequency features in the frequency domain is completed by performing frequency domain convolution:

[0045]

[0046] Among them, G(k,l) is the value of the output feature map corresponding to the position (k,k) in the frequency domain, is the weighted input frequency domain feature map, W freq is the weight matrix, i.e. the convolution kernel; B is the bias term; then Leaky ReLU is used to perform nonlinear activation processing on the output feature map:

[0047] G a (k,l)=LeakyReLU(G(k,l))=max(G(k,l),α·G(k,l))

[0048] Among them, G a (k, l) is the feature map after nonlinear activation, α is the hyperparameter of the slope in the negative area, and α is a positive value much less than 1; then, the processed two-channel feature tensor is converted into a single-channel complex form:

[0049]

[0050] in, is the frequency domain feature tensor in complex form, is a two-channel three-dimensional real tensor before processing, k and l represent the location of the feature, and c is the number of channels; then, the processed complex form feature map is inverse Fourier transformed to restore it to the spatial domain:

[0051]

[0052] Among them, g(m,n) is the spatial domain feature restored by iFFT, (m,n) represents its position, It is a frequency domain feature tensor in complex form.

[0053] Furthermore, step 2.4 specifically includes:

[0054] Step 2.4.1: Integrate the low-frequency features in step 2.2 and the high-frequency features in step 2.3 into a two-channel three-dimensional tensor:

[0055]

[0056] Among them, h and w are the indexes of height and width, and c is the channel index;

[0057] Step 2.4.2: Apply a set of 1×1 convolution kernels to each position of each channel on the integrated feature tensor C:

[0058]

[0059] Among them, O f [h][w] is the value of the f-th channel at position (h,w) in the output feature, h and w are the indices of the height and width, c[h][w][c] is the value of the c-th input channel of the input tensor C at position (h,w), is the weight of the 1×1 convolution kernel of the cth input channel corresponding to the fth output channel. The initial value of the weight is a random number that conforms to the Gaussian distribution. f is the bias term of the f-th output channel. The initial value of the bias is also a random number that conforms to the Gaussian distribution. Since the final output is a single-channel feature, f should be 0 during the calculation process.

[0060] Step 2.4.3, convert the result in step 2.4.2 from Y' to RGB:

[0061] R=Y-16+1.402×(Cr-128)

[0062] G=Y-16-0.34414×(Cb-128)-0.71414×(Cr-128)

[0063] B=Y-16+1.772×(Cb-128)

[0064] Among them, R, G, B are the red, green, and blue pixel values ​​corresponding to the RGB image, Y is the brightness component, Cb and Cr represent the color difference components, and Cb and Cr are 0 in the above formula.

[0065] Furthermore, step 2.5 specifically includes:

[0066] Step 2.5.1. Calculate the loss function: The loss function measures the gap between the predicted result and the actual result. The mean square error (MSE) is used as the loss function:

[0067]

[0068] Step 2.5.2, back propagation and weight update: The back propagation algorithm calculates the gradient of the loss function with respect to each parameter, starting from the output layer, back propagating the gradient to each layer, and using the chain rule to calculate the gradient of each layer; the gradient is usually obtained by taking the derivative of the loss function, and then the weight and bias terms in the network are updated using the gradient descent method to minimize the loss function; back propagation and parameter update are performed alternately, and this process is repeated until the specified number of iterations is reached:

[0069]

[0070] Where l is the number of layers, i is the number of iterations, W is the weight, W is also called the convolution kernel in the convolution layer, and the initial value of W is randomly generated, and its distribution satisfies the Gaussian distribution with a mean of 0 and a variance of 0.0001; is the derivative of the loss function, which represents the gradient of the gradient descent method during the back propagation process, L is the loss function MSE, and η is the learning rate.

[0071] Furthermore, step 3 specifically includes:

[0072] Use SI to evaluate the prediction results more intuitively: the larger the SI, the more accurate the prediction effect; conversely, if the SI value is small, it indicates that the prediction accuracy needs to be improved:

[0073]

[0074] Among them, w and h represent the width and height of the image, Respectively represent the true values ​​of the red, green, and blue pixel values ​​at position (i, j), They represent the predicted values ​​of the red, green, and blue pixel values ​​at position (i, j), t is the allowable error, and I(A) is the indicator function, which means that when condition A is met, the indicator function I outputs 1, otherwise the indicator function I outputs 0; usually, the allowable error t r ,t g ,t bThe value of is the same; According to the Technical Specification for Building Foundation Testing (JGJ 340-2015) (hereinafter referred to as the "Specification"), if the difference between the extreme value and the average value does not exceed 15%, the data prediction is considered accurate;

[0075]

[0076] Among them, N(r) min ,N(g) min ,N(b) min They represent the minimum values ​​of the characteristic data after processing by the NUN data characterization method, r, g, and b represent the red, green, and blue channels respectively; the true value after processing is regarded as the average value, and according to the allowable error limit in the "Specification", 15% of the minimum characteristic value is rounded down and finally used as the allowable error.

[0077] Compared with the prior art, the present invention has the following beneficial effects:

[0078] 1. The geotechnical characteristic parameter characterization and prediction method of the present invention combines the super-resolution reconstruction concept, which can basically meet the needs of obtaining accurate geotechnical characteristic parameters in actual engineering. The data visualization characterization part of the present invention solves the visualization requirement of the original data in the prediction process using the super-resolution reconstruction concept, and ensures the integrity of the data after visualization characterization; the prediction part realizes the prediction of three different geotechnical characteristic parameters at the same time, greatly shortening the prediction time and improving the prediction efficiency.

[0079] 2. The characterization and prediction method of the present invention can accurately predict the characteristic parameters of geotechnical materials. A large number of experiments have verified that the prediction accuracy of the present invention exceeds 85%. Compared with the linear interpolation method commonly used in engineering practice, the parameter acquisition accuracy of this method is improved by 10% under complex geological conditions, which can meet the needs of actual engineering.

[0080] 3. The geotechnical characteristic parameter prediction and evaluation method of the present invention is easy to operate, has a clear calculation method, high accuracy and strong practicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 It is a flow chart of a method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network (NUN-SEDFN) of the present invention;

[0082] Figure 2 It is a schematic diagram of SEDFN network structure;

[0083] Figure 3 Schematic diagram of the mSRCNN network structure of an embodiment of the present invention.

[0084] Figure 4It is a schematic diagram of the FHP-FFC network structure of the present invention.

[0085] Figure 5 The original soil profile image, where the red box is the known drilling site, the blue box is the verification site, and from left to right are a, b, c, d, e drilling sites;

[0086] Figure 6 Predicted stratigraphic profile image, the red box is the known drilling site, the blue box is the verification site, from left to right are a, b, c, d, e drilling sites;

[0087] Figure 7 : This is a comparison between the traditional interpolation method and the method of the present invention. From left to right, they are a, b, c, d, and e drilling sites. In each group of comparison pictures, the left side is the actual situation, the middle is the NUN-SEDFN prediction situation, and the right side is the linear interpolation situation. DETAILED DESCRIPTION

[0088] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0089] See also Figure 1 The embodiment of the present invention provides a method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network, comprising the following steps:

[0090] 1. Use the non-uniform normalization (NUN) method to process the geotechnical characteristic parameters. The processed data is changed from text type to image type, providing original image data for subsequent predictions.

[0091] 1.1 Taking the geological survey data of an actual project in Wuhan as the original data source (as shown in Table 1), according to the needs of the project, the corresponding data of three geotechnical characteristic parameters are selected from the original data. In this example, the three parameters of foundation bearing capacity characteristic value, compression modulus and geotechnical state are selected for processing.

[0092] Table 1 Original data of geotechnical engineering characteristic parameters

[0093]

[0094]

[0095] 1.2 Use the NUN-based data processing method to process the selected raw data.

[0096]

[0097] Among them, max L represents the peak pixel value of the RGB image, which is set to 255. N(c) represents the processed data, c represents the channels represented by different feature parameters, and f r represents the original geotechnical characteristic parameter data, which can be replaced by the foundation bearing capacity characteristic value f ak , compression modulus E s , r is a category parameter used to distinguish different geotechnical characteristic parameters. If the description of the geotechnical state S is text type, the peak pixel value (255) of the RGB image is divided equally according to the number of types, and the RGB value after the division is assigned to the feature. Taking the geotechnical state S as an example, its description is text type. In actual engineering, there are usually 5 different geotechnical states, so it is divided at intervals of 50. The processed data is shown in Table 2.

[0098] Table 2 Corresponding values ​​of soil and rock states

[0099]

[0100] 1.2.1 Determine the category parameter r. Adjust the r value through trial and error to find the best nonlinear transformation parameter to optimize the data distribution. Substitute the parameter value in the original data into the NUN formula. Preset different r values ​​and calculate the parameter value after processing. Compare the processing results under different r values ​​and select the optimal value of r according to the following requirements: the mean should be close to 122.5 (that is, half of the extreme value of the RGB pixel value); the standard deviation should be as large as possible; the skewness should be as small as possible; and the kurtosis should be as negative as possible. The category parameter r corresponding to the characteristic value of the foundation bearing capacity is r For example, r r The processed data and statistical indicators with values ​​of 0.01, 0.005, 0.0025 and 0.001 are shown in Tables 3 and 4.

[0101] Table 3. After processing based on NUN geotechnical characteristic parameter characterization method, f ak value

[0102]

[0103]

[0104] Table 4. After processing based on NUN geotechnical characteristic parameter characterization method, f ak Value statistics

[0105]

[0106] It can be found from Table 4 above that when the category parameter is 0.0025, the mean of the processed data is closer to the median of RGB pixel extreme values ​​122.5, and the degree of dispersion is better than other groups of data. In terms of skewness, the best value is obtained when the category parameter is 0.005, but when the absolute value of the skewness is less than 0.5, the data can be considered to be symmetrically distributed. Therefore, from the perspective of skewness, the category parameter can be 0.0025 or 0.005. In terms of kurtosis, when the category parameters are 0.0025 and 0.005, the kurtosis values ​​are both less than 0. The data after processing are relatively flat, and there are no many extreme values. Based on the above standards, the category parameter of the characteristic value of foundation bearing capacity is determined to be 0.0025. The category parameter r of the compression modulus is g The method for determining the value of is the same as above, and is determined to be 0.05.

[0107] 1.2.2 Use the formula to calculate the processed parameter value. The original data is processed by the category parameter r value selected in 1.2.1.

[0108] 1.3 Data imaging processing: Convert the processed data into image data for easy visual analysis.

[0109] 1.3.1 Repeat the operation in 1.2 to process the three geotechnical characteristic parameters.

[0110] 1.3.2 The three parameters of each rock and soil layer are processed and corresponded to the red, green and blue channels in the RGB pixel value in turn, and then the RGB pixel value corresponding to each rock and soil layer is generated (as shown in Table 5), and finally the image data is generated.

[0111] Table 5 Corresponding values ​​of geotechnical engineering characteristic parameters after NUN processing

[0112]

[0113]

[0114] 2. Use the image data obtained in step 1 to train the spectrum enhanced deep fusion network SEDFN.

[0115] 2.1 Input image preprocessing: Convert the input image to Y'CbCr color space from RGB, and only keep the brightness channel Y'.

[0116]

[0117] Among them, R, G, and B correspond to the three components of the pixel value of the RGB three-channel image respectively.

[0118] 2.2 Extracting low-frequency features using the modified SRCNN network (mSRCNN) (see Figure 2 and Figure 3 ). A two-layer mSRCNN convolutional network is used to extract features and perform nonlinear mapping on the input image, and then reconstruct the image.

[0119]

[0120] The first convolution kernel is W1, the bias term is B1, the last convolution kernel is W2, the bias term is B2, and X is the input block of the previous layer. is an intermediate process function used to represent the content after feature extraction. Leaky ReLU is a nonlinear mapping activation function used to control the output of negative input. is the output result of the first layer, The final output contains low-frequency features.

[0121] 2.3 Extract high-frequency features using frequency-domain high-preserving fast Fourier convolution (FHP-FFC) (see Figure 4 ).

[0122] 2.3.1 Perform fast Fourier convolution on the brightness channel ′ obtained in step 2.1.

[0123]

[0124] where F(k,l) is the complex value at position (k,l) in the frequency domain after transformation, including the frequency and The amplitude and phase of the frequency component of , f(m,n) is the pixel value of the image at position (m,n) in the spatial domain, M and N are the width and height of the image respectively, and i is the imaginary unit.

[0125] 2.3.2 Separate the real part and the imaginary part in 2.3.1. Separate the real part and the imaginary part of the frequency domain information into two channels so that the information of the two channels exists in the real number domain.

[0126]

[0127] in, represents the real part after FFT processing, Represents the processed imaginary part.

[0128] Step 2.3.3, retain high-frequency information by weighted means: first, determine the high- and low-frequency allocation thresholds; then, assign a high-frequency weight coefficient greater than or equal to 0.75 to the high-frequency information, so as to ensure that this step can mainly process the high-frequency information; at the same time, assign a low-frequency information weight coefficient greater than or equal to 0.1 and less than or equal to 0.25 to the low-frequency information.

[0129] T c =Q a％ (Fabs,c )

[0130] Among them, T c Indicates the threshold for distinguishing high and low frequencies on the c channel, Q a％ represents the operation of calculating the percentile of a, F abs,c is the absolute value set of all frequency amplitudes on the cth channel. According to the determined threshold T c To assign corresponding weights to each frequency component. If the absolute value of a frequency component is greater than T c , then the frequency component is considered to be a high-frequency component, otherwise it is a low-frequency component.

[0131]

[0132] Among them, w ij is the element of the output weight tensor, k and l represent its position coordinates, w l is the corresponding weight of the low-frequency component, w h is the weight corresponding to the high frequency component, F abs,kl is the absolute value of the frequency at the corresponding position, T c is the threshold of the cth channel. The weight determined above is applied to the processed spectrum to enhance the high-frequency features in the frequency domain information.

[0133]

[0134] in, is the weighted frequency domain tensor, k and l represent its position information, represents the frequency domain features that separate the real and imaginary parts, and W(k,l) is the weight determined in the previous part of the formula.

[0135] 2.3.4 High-frequency feature learning and inverse fast Fourier transform Based on the weighted frequency domain information in 2.3.3, the learning of high-frequency features in the frequency domain is completed by performing frequency domain convolution.

[0136]

[0137] Among them, G(k,l) is the value of the output feature map corresponding to the position (k,k) in the frequency domain, is the weighted input frequency domain feature map, W freq is the weight matrix, i.e., the convolution kernel. B is the bias term. Leaky ReLU is then used to perform nonlinear activation processing on the output feature map.

[0138] G a (k,k)=LeakyReLU(G(k,l))=max(G(k,l),α·G(k,l))

[0139] Among them, G a(k, l) is the feature map after nonlinear activation, α is the hyperparameter of the slope in the negative region, and α is usually a positive value much less than 1. Subsequently, the processed two-channel feature tensor is converted into a single-channel complex form.

[0140]

[0141] in, is the frequency domain feature tensor in complex form, is a two-channel three-dimensional real tensor before processing, k and l represent the location of the feature, and c is the number of channels. Subsequently, the processed complex form feature map is inverse Fourier transformed to restore it to the spatial domain.

[0142]

[0143] Among them, g(m,n) is the spatial domain feature restored by iFFT, (m,n) represents its position, It is a frequency domain feature tensor in complex form.

[0144] 2.4 Parallel fusion of high and low frequency features. Integrate low frequency and high frequency features and fuse them through 1×1 convolution kernel, and finally perform inverse transformation from Y' to RGB to obtain the final RGB image.

[0145] 2.4.1 Integrate the low-frequency features in 2.2 and the high-frequency features in 2.3 into a two-channel three-dimensional tensor.

[0146]

[0147] Where h and w are the height and width indices, and c is the channel index.

[0148] 2.4.2 Apply a set of 1×1 convolution kernels to each position of each channel on the integrated feature tensor C.

[0149]

[0150] Among them, O f [h][w] is the value of the fth channel at position (h,w) in the output feature, h and w are the indices of the height and width, C[h][w][c] is the value of the cth input channel of the input tensor C at position (h,w), W c f is the weight of the 1×1 convolution kernel of the cth input channel corresponding to the fth output channel. The initial value of the weight is a random number that conforms to the Gaussian distribution. f is the bias term of the fth output channel, and the initial value of the bias is also a random number that conforms to the Gaussian distribution. Since the final output is a single-channel feature, f should be 0 during the calculation process.

[0151] 2.4.3 Convert the result in 2.4.2 from Y' to RGB.

[0152] R=Y-16+1.402×(Cr-128)

[0153] G=Y-16-0.34414×(Cb-128)-0.71414×(Cr-128)

[0154] B=Y-16+1.772×(Cb-128)

[0155] Among them, R, G, B are the red, green, and blue pixel values ​​corresponding to the RGB image, Y is the brightness component, Cb and Cr represent the color difference components, and Cb and Cr are 0 in the above formula.

[0156] 2.5 Use the back propagation algorithm to optimize the SEDFN network parameters and repeat steps 2.1 to 2.4 until the preset number of training times is reached. The number of training times is determined according to actual needs.

[0157] 3. Use the SEDFN network trained in step 2 to predict the actual geological profile of a project in Wuhan, and use the similarity index (SI) to evaluate the prediction results. Select an address profile from the project for prediction. The original data and predicted data are as follows: Figure 5 and Figure 6 shown.

[0158] Use SI to evaluate the prediction results more intuitively. The larger the SI, the more accurate the prediction. On the contrary, if the SI value is small, it indicates that the prediction accuracy needs to be improved.

[0159]

[0160] Among them, w and h represent the width and height of the image, Respectively represent the true values ​​of the red, green, and blue pixel values ​​at position (i, j), They represent the predicted values ​​of the red, green, and blue pixel values ​​at position (i, j), t is the allowable error, and I(A) is the indicator function, which means that when condition A is met, the indicator function I outputs 1, otherwise the indicator function I outputs 0. Usually, the allowable error t r ,t g ,t b According to the Technical Specification for Building Foundation Testing (JGJ 340-2015) (hereinafter referred to as the "Specification"), if the difference between the extreme value and the average value does not exceed 15%, the data prediction is considered accurate.

[0161]

[0162] Among them, N(r) min ,N(g) min ,N(b) min Respectively represent the minimum value of the characteristic data after processing by the NUN data characterization method, r, g, b represent the red, green and blue channels respectively. The true value after processing is regarded as the average value. According to the allowable error limit in the "Code", 15% of the minimum characteristic value is rounded down and finally used as the allowable error. When characterizing the state of the rock and soil body, the pixel value interval corresponding to each state is 50, so the parameter V corresponding to the state of the rock and soil body is b , the allowable error is t b Set to 25.

[0163] Combining the data in Table 5-5 with the above formula, the allowable error t of the other two channels is obtained. r and t g The value of is 5, and substituting it into the previous formula gives the actual calculation formula of SI:

[0164]

[0165] The prediction results are evaluated according to the above formula. The SI values ​​of the five predicted pile positions in this example are as follows: Figure 7 As shown in Table 6.

[0166] Table 6 Comparison of prediction results of NUN-SEDFN and linear interpolation (Note: the drilling depth of each point is 63m)

[0167]

[0168] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network, characterized in that: The steps include: Step 1: Use the non-uniform normalized NUN method to process the geotechnical characteristic parameters, wherein the geotechnical characteristic parameters include the characteristic value of foundation bearing capacity, compression modulus and geotechnical body state; change the geotechnical characteristic parameters from text type to image type to obtain image data of the geotechnical characteristic parameters; Step 2: Use the image data obtained in step 1 to train the spectrum enhancement deep fusion network SEDFN; Step 3: Use the SEDFN network trained in step 2 to predict the actual geological profile of a project, and use the similarity index SI to evaluate the prediction results.

2. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 1 is characterized in that: Step 1 specifically includes: Step 1.1, collect the original data of geotechnical engineering characteristic parameters, and select the corresponding data of three geotechnical characteristic parameters from the original data according to the engineering needs; Step 1.2: Use the NUN-based data processing method to process the selected raw data: Among them, max L represents the peak pixel value of the RGB image, which is set to 255; N(c) represents the processed data, c represents the channels represented by different feature parameters, and f r represents the original geotechnical characteristic parameter data, r is the category parameter, which is used to distinguish different geotechnical characteristic parameters; if the target characteristic parameter is described in text type, the peak pixel value of the RGB image is divided equally according to the number of types, and the divided RGB value is assigned to the feature; Step 1.3: Convert the data processed in step 1.2 into image data for easy visual analysis.

3. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 2 is characterized by: Step 1.2 specifically includes: Step 1.2.1, determine the category parameter r: adjust the r value through trial and error algorithm to find the best nonlinear transformation parameter to optimize data distribution; substitute the parameter value in the original data into the NUN formula, preset different r values ​​and calculate the parameter value after processing, compare the processing results under different r values, and select the optimal value of r according to the following requirements: the mean should be close to 122.5; the standard deviation should be as large as possible, the skewness should be as small as possible, and the kurtosis should be as negative as possible; Step 1.2.2: Process the original data using the category parameter r value selected in step 1.2.

1.

4. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 2 is characterized in that: Step 1.3 specifically includes: Step 1.3.1, repeat the operation in step 1.2 three times to process the three geotechnical characteristic parameters; Step 1.3.2: Correspond the three parameters of each rock and soil layer to the red, green and blue channels of the RGB pixel value in turn, and then generate the RGB pixel value corresponding to each rock and soil layer, and finally generate image data.

5. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 1, characterized in that: Step 2 specifically includes: Step 2.1, input image preprocessing: Convert the input image to a color space from RGB to Y'CbCr color space, and only retain the brightness channel Y': Among them, R, G, and B correspond to the three components of the pixel value of the RGB three-channel image; Step 2.2: Use the modified SRCNN network mSRCNN to extract low-frequency features and train the network for low-frequency processing; use a two-layer mSRCNN convolutional network to perform feature extraction and nonlinear mapping on the input image as a training set, and then perform image reconstruction: The first convolution kernel is W1, the bias term is B1, the last convolution kernel is W2, the bias term is B2, and X is the input block of the previous layer. is an intermediate process function used to represent the content after feature extraction. Leaky ReLU is a nonlinear mapping activation function used to control the output of negative input. is the output result of the first layer, The final output contains low-frequency features. Step 2.3: Use frequency domain high-preserving fast Fourier convolution (FHP-FFC) to extract high-frequency features and train the network to process high-frequency features; Step 2.4: Parallel fusion of high- and low-frequency features: Integrate low-frequency and high-frequency features and fuse them through a 1×1 convolution kernel, and finally perform an inverse transformation from Y′ to RGB to obtain the final RGB image; Step 2.5: Compare the result obtained in step 2.4 with the actual situation, calculate the difference between the network training result and the actual situation, and use the back propagation algorithm to calculate the gradient of the loss function with respect to each weight; repeat steps 2.1 to 2.4 until the preset number of training iterations is reached.

6. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 5 is characterized in that: Step 2.3 specifically includes: Step 2.3.

1. Perform fast Fourier convolution on the brightness channel ' obtained in step 2.1: Where F(k, l) is the complex value at position (k, l) in the frequency domain after transformation, including the frequency and The amplitude and phase of the frequency component of , f(m, n) is the pixel value of the image at position (m, n) in the spatial domain, M and N are the width and height of the image respectively, and i is the imaginary unit; Step 2.3.2: Separate the real part and the imaginary part in step 2.3.1, and separate the real part and the imaginary part of the frequency domain information into two channels, so that the information of the two channels exists in the real number domain: in, represents the real part after FFT processing, represents the processed imaginary part; Step 2.3.3: Retain high-frequency information by weighted means: First, determine the high- and low-frequency allocation thresholds; then, assign a high-frequency weight coefficient greater than or equal to 0.75 to the high-frequency information, so as to ensure that this step can mainly process the high-frequency information; at the same time, assign a low-frequency information weight coefficient greater than or equal to 0.1 and less than or equal to 0.25 to the low-frequency information: T c =Q a％ (F abs,c ) Among them, T c Indicates the threshold for distinguishing high and low frequencies on the c channel, Q a％ represents the operation of calculating the percentile of a, F abs,c is the absolute value set of all frequency amplitudes on the cth channel, according to the determined threshold T c To assign corresponding weights to each frequency component, if the absolute value of a frequency component is greater than T c , then the frequency component is considered to be a high-frequency component, otherwise it is a low-frequency component: Among them, w ij is the element of the output weight tensor, k and l represent its position coordinates, w l is the corresponding weight of the low-frequency component, w h is the weight corresponding to the high frequency component, F abs,kl is the absolute value of the frequency at the corresponding position, T c is the threshold of the cth channel; the weight determined above is applied to the processed spectrum to enhance the high-frequency features in the frequency domain information: in, is the weighted frequency domain tensor, k and l represent its position information, represents the frequency domain features that separate the real and imaginary parts, and W(k, l) is the weight determined in the previous part of the formula; Step 2.3.4, high-frequency feature learning and inverse fast Fourier transform: Based on the weighted frequency domain information in step 2.3.3, the learning of high-frequency features in the frequency domain is completed by performing frequency domain convolution: Among them, G(k, l) is the value of the output feature map corresponding to the position (k, l) in the frequency domain, is the weighted input frequency domain feature map, W freq is the weight matrix, i.e. the convolution kernel; s is the bias term; then Leaky ReLU is used to perform nonlinear activation processing on the output feature map: G a (k,l)=LeakyReLU(G(k,l))=max(G(k,l),α G(k,l)) Among them, G a (k, l) is the feature map after nonlinear activation, α is the hyperparameter of the slope in the negative area, and α is a positive value much less than 1; then, the processed two-channel feature tensor is converted into a single-channel complex form: in, is the frequency domain feature tensor in complex form, is a two-channel three-dimensional real tensor before processing, k and l represent the location of the feature, and c is the number of channels; then, the processed complex form feature map is inverse Fourier transformed to restore it to the spatial domain: Among them, g(m,n) is the spatial domain feature restored by iFFT, (m,n) represents its position, It is a frequency domain feature tensor in complex form.

7. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 5, characterized in that: Step 2.4 specifically includes: Step 2.4.1: Integrate the low-frequency features in step 2.2 and the high-frequency features in step 2.3 into a two-channel three-dimensional tensor: Among them, h and w are the indexes of height and width, and c is the channel index; Step 2.4.2: Apply a set of 1×1 convolution kernels to each position of each channel on the integrated feature tensor C: Among them, O f [h][w] is the value of the f-th channel at position (h, w) in the output feature, h and w are the indices of the height and width, C[h][w][c] is the value of the c-th input channel of the input tensor C at position (h, w), is the weight of the 1×1 convolution kernel of the cth input channel corresponding to the fth output channel. The initial value of the weight is a random number that conforms to the Gaussian distribution. f is the bias term of the f-th output channel. The initial value of the bias is also a random number that conforms to the Gaussian distribution. Since the final output is a single-channel feature, f should be 0 during the calculation process. Step 2.4.3, convert the result in step 2.4.2 from Y′ to RGB: R=Y-16+1.402×(Cr-128) G=Y-16-0.34414×(Cb-128)-0.71414×(Cr-128) B=Y-16+1.772×(Cb-128) Among them, R, G, B are the red, green, and blue pixel values ​​corresponding to the RGB image, Y is the brightness component, Cb and Cr represent the color difference components, and Cb and Cr are 0 in the above formula.

8. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 5, characterized in that: Step 2.5 specifically includes: Step 2.5.

1. Calculate the loss function: The loss function measures the gap between the predicted result and the actual result. The mean square error (MSE) is used as the loss function: Step 2.5.2, back propagation and weight update: The back propagation algorithm calculates the gradient of the loss function with respect to each parameter, starting from the output layer, back propagating the gradient to each layer, and using the chain rule to calculate the gradient of each layer; the gradient is usually obtained by taking the derivative of the loss function, and then the weight and bias terms in the network are updated using the gradient descent method to minimize the loss function; back propagation and parameter update are performed alternately, and this process is repeated until the specified number of iterations is reached: Where l is the number of layers, i is the number of iterations, W is the weight, W is also called the convolution kernel in the convolution layer, and the initial value of W is randomly generated, and its distribution satisfies the Gaussian distribution with a mean of 0 and a variance of 0.0001; is the derivative of the loss function, which represents the gradient of the gradient descent method during the back propagation process, L is the loss function MSE, and η is the learning rate.

9. The method for predicting geotechnical characteristic parameters based on non-uniform normalization and spectrum enhancement deep fusion network as claimed in claim 5, characterized in that: Step 3 specifically includes: Use SI to evaluate the prediction results more intuitively: the larger the SI, the more accurate the prediction effect; conversely, if the SI value is small, it indicates that the prediction accuracy needs to be improved: Among them, w and h represent the width and height of the image, Respectively represent the true values ​​of the red, green, and blue pixel values ​​at position (i, j), They represent the predicted values ​​of the red, green, and blue pixel values ​​at position (i, j), t is the allowable error, and I(A) is the indicator function, which means that when condition A is met, the indicator function outputs 1, otherwise the indicator function outputs 0; Among them, N(r) min ,N(g) min ,N(b) min They represent the minimum values ​​of the characteristic data after processing by the NUN data characterization method, r, g, and b represent the red, green, and blue channels, respectively. The true value after processing is regarded as the average value. According to the allowable error limit in the Technical Code for Building Foundation Testing (JGJ 340-2015), 15% of the minimum characteristic value is rounded down and finally used as the allowable error.

Images