Ground penetrating radar data denoising method and system based on unsupervised learning
By creating a deep learning model based on unsupervised learning, the problem of insufficient generalization and adaptability of the denoising method of ground penetrating radar data in complex data is solved, and efficient denoising in complex ground penetrating radar data is achieved, maintaining signal details and shortening training time.
Patent Information
- Application Number
- CN202510392285.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-31
- Publication Date
- 2025-07-18
AI Technical Summary
The existing ground penetrating radar data denoising methods are insufficient in generalization and adaptability when facing complex data, and cannot effectively maintain signal details. Traditional methods such as mean filtering, median filtering and wavelet transformation have their own disadvantages and cannot meet the needs in actual applications.
Denoising autoencoder based on unsupervised learning is used to create a deep learning model. Through the data processing of the training set and the test set, corrupt data is introduced for training, and an improved denoising autoencoder is generated. The neural network composed of the encoder and the decoder is used for signal reconstruction, and noise characteristics are learned and denoising is performed.
It has achieved good generalization and adaptability in complex ground penetrating radar data, can maintain signal details, achieve better denoising effect, has denoising robustness, shorten training time and improve the generalization ability of the model.
Smart Images
Figure CN120337737A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for denoising ground penetrating radar data, and more particularly to a method for denoising ground penetrating radar data based on unsupervised learning, and further to a ground penetrating radar data denoising system employing the method for denoising ground penetrating radar data based on unsupervised learning. Background Art
[0002] As a non-destructive, fast and economical shallow subsurface geophysical exploration method, ground penetrating radar is widely used in fields such as infrastructure construction and engineering survey. In actual engineering detection, ground penetrating radar data is often contaminated by various random noises, and the noises in the ground penetrating radar data will reduce the signal-to-noise ratio and resolution of the data. Therefore, there are errors in the qualitative and quantitative analysis and interpretation of the ground penetrating radar noisy data. How to better achieve the denoising processing of the ground penetrating radar data has always been a technical problem to be solved in the industry.
[0003] Traditional methods for denoising ground penetrating radar data include mean filtering algorithm, median filtering algorithm, wavelet transform method, etc. These several traditional denoising methods all have drawbacks. Among them, the mean filtering algorithm suppresses noise by calculating the average value of pixel values within a window, and has a poor suppression effect on Gaussian noise, which will cause image blurring. The median filtering algorithm suppresses noise by calculating the median value of pixel values within a window, and there will be a problem of loss of image details when there is strong noise. The wavelet transform method analyzes the signal by converting it to the wavelet domain and selects an appropriate threshold to achieve noise suppression. However, the denoising effects of different wavelet bases are significantly different, and it needs to be selected specifically in practical applications, and its calculation amount is large and the operation speed is slow. Therefore, the above traditional methods for denoising ground penetrating radar data all have their own drawbacks, cannot meet the generalization and adaptability requirements in practical applications, and cannot well preserve the details of the signal. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method for denoising ground penetrating radar data based on unsupervised learning, aiming to improve the generalization and adaptability of the method for denoising ground penetrating radar data; and, in the face of complex ground penetrating radar data, it can have denoising robustness and can well preserve the details of the signal. On this basis, a ground penetrating radar data denoising system employing the method for denoising ground penetrating radar data based on unsupervised learning is further provided.
[0005] To this end, the present invention provides a method for denoising ground penetrating radar data based on unsupervised learning, including the following steps:
[0006] Step S1, collecting and preparing training data;
[0007] Step S2, training a deep learning model using the noisy data;
[0008] Step S3, verifying the performance of the deep learning model on the test set;
[0009] The step S2 comprises the following sub-steps:
[0010] Step S201, creating a deep learning model using a denoising autoencoder, wherein the denoising autoencoder includes an encoder and a decoder for generating reconstruction;
[0011] Step S202, introducing damaged data to train the denoising autoencoder to obtain an improved denoising autoencoder;
[0012] Step S203, training the improved denoising autoencoder using the training data in the training set.
[0013] A further improvement of the present invention is that step S1 includes the following sub-steps:
[0014] Step S101, using the finite difference time domain method to simulate and obtain original radar data, and then adding Gaussian white noise to the original radar data to obtain damaged data;
[0015] Step S102: divide the prepared original radar data and damaged data into equal parts, namely, a training set train and a test set test.
[0016] A further improvement of the present invention is that in step S101, the original radar data is X={x1,x2,x3,…x n}, the damaged data after adding Gaussian white noise is According to the original radar data X and the damage data The subscript of the original radar data X and the damaged data Establish a one-to-one correspondence; in the step S102, the training set train and the test set test respectively include equally divided original radar data, and respectively include damaged data corresponding to the subscripts of the original radar data.
[0017] A further improvement of the present invention is that the process of using a denoising autoencoder to create a deep learning model in step S201 is as follows: a denoising autoencoder is composed of two artificial neural networks, an encoder function and a decoder, damaged data is input for training, and the original undamaged data is predicted as output to serve as a deep learning model.
[0018] A further improvement of the present invention is that in step S202, the formula W=PQ -1 Calculate the encoding weight W of the denoising autoencoder loss function, and use the encoding weight W as the network parameter of the improved denoising autoencoder; wherein, P and Q are both intermediate parameters used to represent matrices; denotes the conjugate matrix of the original radar data X denotes the corrupted data transpose of denotes the corrupted data corresponding to the conjugate matrix
[0019] A further improvement of the present invention lies in that in the step S202, m different corrupted data are generated as inputs by adding different noises, and then m different corrupted data are introduced to train the denoising autoencoder respectively; m is the preset number of training times.
[0020] A further improvement of the present invention lies in that in the step S203, the improved denoising autoencoder is trained with the training data in the training set train.
[0021] A further improvement of the present invention lies in that in the step S203, the improved denoising autoencoder is composed of several single-layer denoising autoencoders, and the input signal is the corrupted data after noise addition processing.
[0022] A further improvement of the present invention lies in that in the step S3, the performance of the deep learning model is evaluated with the data in the test set test, and the difference between the output data and the original radar data is compared. When the difference is greater than the preset threshold, return to step S1 to regenerate new training data, and repeat the operation until step S203 for training the deep learning model until the difference is not greater than the preset threshold.
[0023] The present invention also provides a ground penetrating radar data denoising system based on unsupervised learning, which adopts the above-mentioned ground penetrating radar data denoising method based on unsupervised learning, and includes:
[0024] A training data generation module, which collects and prepares training data;
[0025] A model training module, which trains the deep learning model with noisy data;
[0026] A model verification module, which verifies the performance of the deep learning model on the test set.
[0027] Compared with the prior art, the beneficial effects of the present invention are as follows: First, collect and prepare training data, then use a denoising autoencoder to create a deep learning model, and introduce corrupted data to train the denoising autoencoder to obtain an improved denoising autoencoder. Finally, verify the performance of the deep learning model on the test set, and further complete the learning of noisy data and noise features by adding noise during the training process, generating an improved denoising autoencoder for training. Therefore, the present invention not only has good generalization and adaptability; moreover, when facing complex ground penetrating radar data, it also has denoising robustness, can well preserve the details of the signal, and achieve better denoising effect of ground penetrating radar data based on unsupervised learning. BRIEF DESCRIPTION OF THE DRAWINGS
[0028] Figure 1 is a schematic diagram of the work flow of an embodiment of the present invention;
[0029] Figure 2 is a schematic diagram of the model for creating a deep learning model using a denoising autoencoder in an embodiment of the present invention;
[0030] Figure 3 is a schematic diagram of the model principle of a denoising autoencoder in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0031] In the description of the present invention, if it involves orientation description, such as "up", "down", "front", "back", "left", "right", etc., the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation of the present invention. If a technical feature is referred to as "set", "fixed", "connected", "installed" on another technical feature, it can be directly set, fixed, connected, installed on another technical feature, or indirectly set, fixed, connected, installed on another technical feature.
[0032] In the description of the present invention, if it involves "several", its meaning is more than one; if it involves "multiple", its meaning is more than two; if it involves "greater than", "less than", "exceeding", it should be understood as not including the present number; if it involves "above", "below", "within", it should be understood as including the present number. If it involves "first", "second", etc., it should be understood that it is only used for the distinction of the same or similar technical feature names, and cannot be understood as implying / indicating the relative importance of the technical features, cannot be understood as implying / indicating the quantity of the technical features, nor can it be understood as implying / indicating the sequence relationship of the technical features.
[0033] The following further elaborates on the preferred embodiments of the present invention in conjunction with the accompanying drawings.
[0034] As Figures 1 to 3 shown, this embodiment provides a ground penetrating radar data denoising method based on unsupervised learning, including the following steps:
[0035] Step S1, collect and prepare training data;
[0036] Step S2, train a deep learning model using the noisy data;
[0037] And, step S3, verify the performance of the deep learning model on the test set;
[0038] The said step S2 includes the following sub-steps:
[0039] Step S201, create a deep learning model using a denoising autoencoder, and the denoising autoencoder includes an encoder and a decoder for generating reconstructions;
[0040] Step S202, introduce corrupted data to train the denoising autoencoder to obtain an improved denoising autoencoder;
[0041] Step S203, train the improved denoising autoencoder using the training data in the training set.
[0042] In this embodiment, step S1 is used to generate the original radar data to obtain the training set and the test set. The said step S1 preferably includes step S101 and step S102.
[0043] Step S101, simulate the original radar data using the finite-difference time-domain method, and then add Gaussian white noise to the original radar data to obtain corrupted data. Let the original radar data be X = {x1, x2, x3, … x n}, and the corrupted data after adding Gaussian white noise is According to the subscripts 1, 2, 3…, n of the original radar data X and the corrupted data , establish a one-to-one correspondence between the original radar data X and the corrupted data with the same subscript; n represents the total number of the original radar data.
[0044] Step S102, equally divide the prepared original radar data and the corrupted data into two parts: the training set train and the test set test. Among them, the training set train and the test set test respectively include the equally divided original radar data, and respectively include the corrupted data corresponding to the subscripts of the original radar data; that is, the training set train and the test set test each include half of the original radar data, and include the corrupted data corresponding one-to-one to the original radar data.
[0045] Step S2 in this embodiment is used to train a deep learning model with noisy data, preferably including steps S201 to S203.
[0046] Step S201 in this embodiment is used to create an improved deep learning model.
[0047] An autoencoder is a simple artificial neural network (ANN, i.e., Artificial Neural Network). After training, it can learn the encoded representation of input data. This unsupervised mechanism does not require labels. An autoencoder consists of two neural networks. The first half is called the encoder, and the second half is called the decoder. Both neural networks consist of single or multiple hidden layers with activation functions.
[0048] In step S201 of this embodiment, a denoising autoencoder is used to create a deep learning model. The implementation process is as follows: A denoising autoencoder is composed of two artificial neural networks, an encoder function and a decoder. The corrupted data is input for training, and the original uncorrupted data is predicted as the output to be used as the deep learning model. Its autoencoding network consists of two parts: an encoder function, denoted as h = f(x); and a decoder for generating the reconstruction, denoted as the function r = g(h); its structure is as Figure 2 shown. The denoising autoencoder refers to the Denoising Autoencoder, abbreviated as DAE.
[0049] During the training process of the deep learning model, the denoising autoencoder receives corrupted data as input. Through training, the decoder can accurately predict the original uncorrupted data as the output. In this way, the denoising autoencoder can learn the robust feature representation of the data, thereby improving the noise resistance of the deep learning model.
[0050] Step S202 in this embodiment trains the denoising autoencoder by introducing corrupted data to obtain an improved denoising autoencoder. The training process of the denoising autoencoder is as Figure 3 shown. This embodiment introduces a corruption process This conditional distribution represents the probability of generating a corrupted data sample given that the data sample is the original radar data X. of the sample.
[0051] The denoising autoencoder is trained to reconstruct clean data points from the corrupted samples i.e., the original radar data X. This can be achieved by minimizing the loss where, is the corrupted version of the sample of the original radar data X after passing through the corruption process . P decoderrepresents the distribution of a factor, and the distribution P of this factor decoder The average parameter of is given by the feedforward network g.
[0052] The autoencoder learns to reconstruct the distribution from the training data pairs according to the following process The process of is as follows: 1. Sample a training sample X from the training set. 2. Sample a corrupted sample from 3. Use as a training sample to estimate the reconstructed distribution of the autoencoder 3. Use as a training sample to estimate the reconstructed distribution of the autoencoder where h is the output of the encoder and P decoder is defined according to the decoding function g(h).
[0053] Generally, it is possible to simply perform approximate minimization of the negative log-likelihood -logP decoder (X|h) using a gradient-based method (such as mini-batch gradient descent). As long as the encoder is deterministic, the denoising autoencoder is a feedforward neural network and can be trained in exactly the same way as other feedforward neural networks.
[0054] Therefore, it can be considered that the denoising autoencoder performs stochastic gradient descent under the following expectation: where, P data (x) is the distribution of the training data.
[0055] The above denoising autoencoder belongs to the denoising autoencoder before improvement and still has two key defects: 1. High computational cost; 2. Lack of scalability for high-dimensional features.
[0056] Therefore, aiming at these two defects, this embodiment proposes an improved technical solution, that is, an improved denoising autoencoder. The improved denoising autoencoder in this embodiment is composed of several single-layer denoising autoencoders combined, so its input signal is the same as the input signal of the denoising autoencoder, both are corrupted signals after noise addition processing. That is, in step S203 described in this embodiment, the improved denoising autoencoder is composed of several single-layer denoising autoencoders combined, and the input signal is the corrupted data after noise addition processing.
[0057] If the squared reconstruction error is used as the loss function, its goal is to minimize the following formula: where, J DAE (W) represents the loss function, which is a metric for measuring the difference between the output of the autoencoder and the original input; W represents the encoding weight, which is used to adjust the influence of the input data; i represents the serial number of the ground penetrating radar data, i represents the serial number of the ground penetrating radar data when different corrupted data are input, and n represents the total number of the ground penetrating radar data.
[0058] To make the model more general, the experiment process will be repeated m times, where m is the preset number of training times, also known as the number of experiments; each time different noises are added to generate m different corrupted data as inputs. Thus, It can be rewritten as minimizing the overall squared reconstruction error: where is the original uncorrupted input x i is the j-th corrupted sample of. j represents the sample serial number in the corrupted data, J DAE′ (W) represents the loss function after the reconstruction error.
[0059] According to the relationship between the norm and the matrix, the formula can be further rewritten as: where X = [x1, x2, x3,... x n , represents the conjugate matrix of the original radar data X; has m elements, is the corrupted data corresponding to the conjugate matrix .
[0060] tr(A) represents the sum of all elements on the trace of matrix A, which is called the trace or trace number of matrix A, generally denoted as tr(A). The trace of matrix A represents the sum of all elements on the main diagonal of an n×n matrix A. Thus, the loss function of the improved denoising encoder can be converted to solve the following formula: W = PQ -1 . Where,
[0061] When the number of experiments m is large enough, P and Q in the formula W = PQ -1 will eventually converge to their expected values respectively. Let m → ∞, the formula W = PQ -1 can be converted to W = E[P]E[Q] -1 . E[P] represents the expected value of the intermediate parameter P; E[Q] represents the expected value of the intermediate parameter Q, and E[Q] -1 represents the reciprocal of the expected value E[Q].
[0062] Therefore, in step S202 of this embodiment, in step S202, the loss function W of the denoising autoencoder is calculated through the formula W = PQ -1 and this loss function W is used as the network parameter of the improved denoising autoencoder; where, Both P and Q are intermediate parameters used to represent matrices; represents the conjugate matrix of the original radar data X, represents the corrupted data transpose of, represents the conjugate matrix The corresponding damaged data.
[0063] Preferably, in step S202 of this embodiment, m different damaged data are generated as inputs by adding different noises, and then m different damaged data are introduced to train the denoising autoencoder respectively; m is the preset number of training times. By introducing different damaged data for training respectively in this embodiment, the required traversal times can be effectively reduced, and the overall training time can be shortened.
[0064] Therefore, the network parameter W can also be calculated by the formula W = E[P]E[Q]. -1 The network parameter W is calculated. From the above content, it can be known that no optimization algorithm is used in the process of solving the network parameter, and the values of E[P] and E[Q] can be obtained only by traversing the training data once. This is one of the advantages brought by using the improved denoising autoencoder in this embodiment, which can greatly shorten the training time.
[0065] Therefore, by adopting the improved denoising autoencoder in this embodiment, it has relatively strong learning ability and fast training speed; it can effectively reduce the number of intermediate parameters, realize faster model selection and convexity based on layer-by-layer training.
[0066] In step S203 of this embodiment, the improved denoising autoencoder is trained with the training data in the training set train.
[0067] Step S3 of this embodiment is used to verify the deep learning model using the improved denoising autoencoder. In step S3, the performance of the deep learning model is evaluated with the data in the test set test, and the difference between the output data and the original radar data is compared. When the difference is greater than the preset threshold, return to step S1 to regenerate new training data, and repeat the operation until step S203 for training the deep learning model until the difference is not greater than the preset threshold. The preset threshold refers to the difference threshold between the output data and the original radar data set in advance, which can be set and adjusted according to the actual situation.
[0068] This embodiment also provides a ground penetrating radar data denoising system based on unsupervised learning, which adopts the above-mentioned ground penetrating radar data denoising method based on unsupervised learning, and includes:
[0069] A training data generation module, which collects and prepares training data;
[0070] A model training module, which trains the deep learning model with noisy data;
[0071] And a model verification module, which verifies the performance of the deep learning model on the test set.
[0072] In summary, in this embodiment, training data is first collected and prepared to provide a data basis for signal denoising through feature extraction and training with a large amount of data, which has good generalization and adaptability. Then, a denoising autoencoder is used to create a deep learning model, and corrupted data is introduced to train the denoising autoencoder to obtain an improved denoising autoencoder, enabling it to complete the learning of the features of ground penetrating radar data and noise, and generating a deep learning model for prediction. The deep learning model uses the improved denoising autoencoder to separate the noise from the ground penetrating radar data containing noise, achieving the purpose of denoising. When facing complex ground penetrating radar data, such as ground penetrating radar signals with different features, it has denoising robustness, can well preserve the details of the signal, and achieves better denoising effect of ground penetrating radar data based on unsupervised learning.
[0073] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions or substitutions can still be made, and all should be regarded as belonging to the protection scope of the present invention.
Claims
1. A ground penetrating radar data denoising method based on unsupervised learning, characterized in that, It includes the following steps: Step S1, collect and prepare training data; Step S2, train a deep learning model using noisy data; Step S3, verify the performance of the deep learning model on the test set; Step S2 includes the following sub-steps: Step S201, create a deep learning model using a denoising autoencoder, where the denoising autoencoder includes an encoder and a decoder for generating reconstructions; Step S202, introduce corrupted data to train the denoising autoencoder to obtain an improved denoising autoencoder; Step S203, train the improved denoising autoencoder using the training data in the training set.
2. The ground penetrating radar data denoising method based on unsupervised learning according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101, simulate to obtain original radar data using the finite-difference time-domain method, and then add Gaussian white noise to the original radar data to obtain corrupted data; Step S102, equally divide the prepared original radar data and corrupted data into a training set train and a test set test.
3. The ground penetrating radar data denoising method based on unsupervised learning according to claim 2, characterized in that, In the step S101, let the original radar data be X = {x1, x2, x3, … x n}, and the damaged data after adding Gaussian white noise is According to the subscripts of the original radar data X and the damaged data of the same subscript, establish a one-to-one correspondence between the original radar data X and the damaged data ; in the step S102, the training set train and the test set test respectively include equally divided original radar data, and respectively include damaged data corresponding to the subscripts of the original radar data.
4. The ground penetrating radar data denoising method based on unsupervised learning according to any one of claims 1 to 3, characterized in that, The process of creating a deep learning model using a denoising autoencoder in Step S201 is as follows: compose a denoising autoencoder through two artificial neural networks, namely an encoder function and a decoder, input corrupted data for training, and predict the original uncorrupted data as the output to serve as the deep learning model.
5. The ground penetrating radar data denoising method based on unsupervised learning according to any one of claims 1 to 3, characterized in that In the step S202, the encoding weight W of the denoising autoencoder loss function is calculated through the formula W = PQ -1 and this encoding weight W is used as the network parameter of the improved denoising autoencoder; where both P and Q are intermediate parameters used to represent matrices; represents the conjugate matrix of the original radar data X, represents the corrupted data transpose of, represents the corrupted data corresponding to the conjugate matrix 6. The ground penetrating radar data denoising method based on unsupervised learning according to claim 5, characterized in that In Step S202, m different corrupted data are generated by adding different noises as inputs, and then m different corrupted data are introduced to train the denoising autoencoder respectively; m is a preset number of training times.
7. The ground penetrating radar data denoising method based on unsupervised learning according to claim 2 or 3, characterized in that, In Step S203, the improved denoising autoencoder is trained using the training data in the training set train.
8. The ground penetrating radar data denoising method based on unsupervised learning according to any one of claims 1 to 3, characterized in that, In Step S203, the improved denoising autoencoder is composed of several single-layer denoising autoencoders, and the input signal is the corrupted data after noise addition processing.
9. The ground penetrating radar data denoising method based on unsupervised learning according to claim 2 or 3, characterized in that In Step S3, the performance of the deep learning model is evaluated using the data in the test set test, and the difference between the output data and the original radar data is compared. When the difference is greater than the preset threshold, return to Step S1 to regenerate new training data and repeat the operation until Step S203 for training the deep learning model until the difference is not greater than the preset threshold.
10. A ground penetrating radar data denoising system based on unsupervised learning, characterized in that, The ground penetrating radar data denoising method based on unsupervised learning as described in any one of claims 1 to 9 is adopted, and it includes: A training data generation module, which collects and prepares training data; A model training module, which trains a deep learning model using noisy data; A model verification module, which verifies the performance of the deep learning model on the test set.